More path collapse for metaprogram.

not sure what the standalone repo is going to be yet, but it will be relatively permissive since this is prototype/educational setup

.gitignore

@@ -17,3 +17,7 @@ toolchain/PSn00bSDK
 .vscode/settings.json
 toolchain/lfs
 toolchain/lpeg
+
+scratch
+toolchain/libpsn00b
+scripts/pcsx_debug_helper.zip

.vscode/cozy-and-windy/README.md

@@ -0,0 +1,26 @@
+# Cozy and Windy
+
+Editor theme ported from the Rider scheme of the same name.
+
+It colors the editor surface, C/C++ syntax, and tape-atom DSL keywords
+emitted by `local.tape-atom-syntax`. It does not change workbench chrome.
+
+## Install
+
+```powershell
+cd C:\projects\Pikuma\ps1\.vscode\cozy-and-windy
+npm run package
+code --install-extension .\cozy-and-windy-0.1.0.vsix --force
+```
+
+Reload the window. Select **Cozy and Windy** as the color theme, or set
+`workbench.colorTheme` to `Cozy and Windy` in the PS1 workspace settings.
+
+Keep `local.tape-atom-syntax` installed. This theme colors those token
+types; it does not classify them.
+
+## Inspect
+
+Open `hello_camera.atom.c` and run **Developer: Inspect Editor Tokens and Scopes**
+on `MipsAtom_`, an atom name, `atom_info`, `R_PrimCursor`, a `gte_*` call,
+and a `mac_*` call.

.vscode/cozy-and-windy/package.json

@@ -0,0 +1,25 @@
+{
+	"name": "cozy-and-windy",
+	"displayName": "Cozy and Windy",
+	"description": "Editor theme ported from the Rider Cozy and Windy scheme. Colors C/C++ and tape-atom DSL keywords.",
+	"publisher": "local",
+	"version": "0.1.0",
+	"engines": {
+		"vscode": "^1.80.0"
+	},
+	"categories": [
+		"Themes"
+	],
+	"scripts": {
+		"package": "npx --yes @vscode/vsce@3.6.1 package --allow-missing-repository --skip-license --out cozy-and-windy-0.1.0.vsix"
+	},
+	"contributes": {
+		"themes": [
+			{
+				"label": "Cozy and Windy",
+				"uiTheme": "vs-dark",
+				"path": "./themes/cozy-and-windy-color-theme.json"
+			}
+		]
+	}
+}

.vscode/cozy-and-windy/themes/cozy-and-windy-color-theme.json

@@ -0,0 +1,132 @@
+{
+	"name": "Cozy and Windy",
+	"type": "dark",
+	"semanticHighlighting": true,
+	"colors": {
+		// 121212
+		// 111212
+		// 211f1e
+		// 191817
+		"editor.background": "#191817",
+		"editor.foreground": "#dfc6ba",
+		"editor.lineHighlightBackground": "#1c1c1c",
+		"editor.selectionBackground": "#164371",
+		"editor.selectionForeground": "#c8c8c8",
+		"editorLineNumber.foreground": "#43c3c3",
+		"editorLineNumber.activeForeground": "#00fff4",
+		"editorIndentGuide.background1": "#181818",
+		"editorIndentGuide.activeBackground1": "#202020",
+		"editorRuler.foreground": "#505050",
+		"editorGutter.background": "#211f1e",
+		"editorBracketMatch.background": "#3b514d",
+		"editor.foldBackground": "#0c0c0c6a",
+		"editor.wordHighlightBackground": "#211f1e4d",
+		"editor.wordHighlightStrongBackground": "#303030",
+		"editorCursor.foreground": "#00fff4",
+		"editorWhitespace.foreground": "#181818",
+		// "editorLineHighlightBorder": "#1c1c1c",
+		"editorWidget.background": "#211f1e",
+		"editorSuggestWidget.background": "#2c334b",
+		"editorHoverWidget.background": "#2c334b"
+	},
+	"semanticTokenColors": {
+		"comment": { "foreground": "#868686", "fontStyle": "italic" },
+		"keyword": { "foreground": "#d8bd5b" },
+		"string": { "foreground": "#d46a54" },
+		"number": { "foreground": "#b5cea8" },
+		"operator": { "foreground": "#be8e78" },
+		"class": { "foreground": "#54a4d6" },
+		"struct": { "foreground": "#54a4d6" },
+		"enum": { "foreground": "#54a4d6" },
+		"type": { "foreground": "#54a4d6" },
+		"interface": { "foreground": "#7984ab" },
+		"function": { "foreground": "#cccab5" },
+		// "function": { "foreground": "#6090a9" },
+		"method": { "foreground": "#6090a9" },
+		"variable": { "foreground": "#bc966c" },
+		"parameter": { "foreground": "#ce8365" },
+		"property": { "foreground": "#acb8c8" },
+		"*.static": { "foreground": "#9e95c6" },
+		"macro": { "foreground": "#5ea852" },
+		"namespace": { "foreground": "#8e8e8e" },
+		"typeParameter": { "foreground": "#b8d7a3" },
+		"enumMember": { "foreground": "#a373b0" },
+		"label": { "foreground": "#c8c8c8", "fontStyle": "bold" },
+		"tapeAtomKeyword": { "foreground": "#d8bd5b", "fontStyle": "bold" },
+		"tapeAtomName": { "foreground": "#b1b7d6", "fontStyle": "bold" },
+		"tapeComponentKeyword": { "foreground": "#d68a36", "fontStyle": "bold" },
+		"tapeComponentName": { "foreground": "#b1b7d6", "fontStyle": "bold" },
+		"tapeAnnotation": { "foreground": "#d8bd5b" },
+		"tapeBindType": { "foreground": "#54a4d6" },
+		"tapePhase": { "foreground": "#b8d7a3", "fontStyle": "italic" },
+		"tapeLabel": { "foreground": "#959595", "fontStyle": "bold" },
+		// "tapeCpuInstruction": { "foreground": "#6d9aa0" },
+		// "tapeCpuInstruction": { "foreground": "#cf7539" },
+		// "tapeCpuInstruction": { "foreground": "#d16b3a" },
+		"tapeCpuInstruction": { "foreground": "#d5895a" },
+		"tapeGteInstruction": { "foreground": "#988bcb" },
+		"tapeGpuInstruction": { "foreground": "#bf7dac" },
+		"tapeComponentInstruction": { "foreground": "#8baa5d" },
+		// "tapeGprRegister": { "foreground": "#92d4d9" },
+		"tapeGprRegister": { "foreground": "#a2bfa8" },
+		"tapeCop2Register": { "foreground": "#945cd9" },
+		"tapeDuffleType": { "foreground": "#54a4d6" },
+		"tapeAttribute": { "foreground": "#73a07c" },
+		"tapeGprRegister.tapeRead": { "foreground": "#5bb8b0", "fontStyle": "italic" },
+		"tapeGprRegister.tapeWrite": { "foreground": "#2d8f8c", "fontStyle": "bold" },
+		"tapeCop2Register.tapeRead": { "foreground": "#b08ae0", "fontStyle": "italic" },
+		"tapeCop2Register.tapeWrite": { "foreground": "#7b3ec4", "fontStyle": "bold" },
+		// "*.tapeAuto": { },
+		"tapeControlFlow": { "foreground": "#63d169", "fontStyle": "bold" },
+		"tapeDelaySlot": { "foreground": "#ff5647" }
+	},
+	"tokenColors": [
+		{ "scope": ["comment", "comment.block", "comment.line", "comment.block.documentation"], "settings": { "foreground": "#868686", "fontStyle": "italic" } },
+		{ "scope": ["keyword", "keyword.control", "keyword.other"], "settings": { "foreground": "#d8bd5b" } },
+		{ "scope": ["string", "string.quoted"], "settings": { "foreground": "#d46a54" } },
+		{ "scope": ["string.quoted.other"], "settings": { "foreground": "#d69d85" } },
+		{ "scope": ["constant.numeric"], "settings": { "foreground": "#b5cea8" } },
+		{ "scope": ["punctuation", "keyword.operator"], "settings": { "foreground": "#be8e78" } },
+		{ "scope": ["keyword.operator.overload"], "settings": { "foreground": "#b87e76" } },
+		{ "scope": ["entity.name.type", "entity.name.type.class", "entity.name.type.struct", "entity.name.type.enum"], "settings": { "foreground": "#54a4d6" } },
+		{ "scope": ["entity.name.type.interface"], "settings": { "foreground": "#7984ab" } },
+		{ "scope": ["entity.name.function"], "settings": { "foreground": "#cccab5" } },
+		{ "scope": ["entity.name.function.member"], "settings": { "foreground": "#6090a9" } },
+		{ "scope": ["variable.other.local"], "settings": { "foreground": "#bc966c" } },
+		{ "scope": ["variable.parameter"], "settings": { "foreground": "#ce8365" } },
+		{ "scope": ["variable.other.property"], "settings": { "foreground": "#acb8c8" } },
+		{ "scope": ["variable.other.constant"], "settings": { "foreground": "#9e95c6" } },
+		{ "scope": ["variable.other.global", "variable.other.defaultLibrary"], "settings": { "foreground": "#bf7dac" } },
+		{ "scope": ["support.type", "support.function"], "settings": { "foreground": "#8baa5d" } },
+		{ "scope": ["meta.preprocessor"], "settings": { "foreground": "#5ea852" } },
+		// { "scope": ["variable.parameter.preprocessor"], "settings": { "foreground": "#636363" } },
+		{ "scope": ["variable.parameter.preprocessor"], "settings": { "foreground": "#bc966c" } },
+		{ "scope": ["keyword.control.directive"], "settings": { "foreground": "#d68a36" } },
+		{ "scope": ["entity.name.namespace"], "settings": { "foreground": "#8e8e8e" } },
+		{ "scope": ["entity.name.type.parameter"], "settings": { "foreground": "#b8d7a3" } },
+		{ "scope": ["variable.other.enummember"], "settings": { "foreground": "#a373b0" } },
+		{ "scope": ["entity.name.type.concept"], "settings": { "foreground": "#76ff7d" } },
+		{ "scope": ["entity.name.type.dependent"], "settings": { "foreground": "#448b5a", "fontStyle": "bold" } },
+		{ "scope": ["entity.name.label"], "settings": { "foreground": "#c8c8c8", "fontStyle": "bold" } },
+		{ "scope": ["invalid"], "settings": { "foreground": "#ff5647" } },
+		{ "scope": ["keyword.codetag.todo"], "settings": { "foreground": "#c10000", "fontStyle": "bold italic" } },
+		{ "scope": ["keyword.control.duffle.atom"], "settings": { "foreground": "#d8bd5b", "fontStyle": "bold" } },
+		{ "scope": ["entity.name.function.duffle.atom"], "settings": { "foreground": "#cccab5", "fontStyle": "bold" } },
+		{ "scope": ["keyword.control.duffle.component"], "settings": { "foreground": "#d68a36", "fontStyle": "bold" } },
+		{ "scope": ["entity.name.function.duffle.component"], "settings": { "foreground": "#6090a9", "fontStyle": "bold" } },
+		{ "scope": ["support.function.duffle.annotation"], "settings": { "foreground": "#d8bd5b" } },
+		{ "scope": ["entity.name.type.duffle.bind"], "settings": { "foreground": "#54a4d6" } },
+		{ "scope": ["entity.name.tag.duffle.phase"], "settings": { "foreground": "#b8d7a3", "fontStyle": "italic" } },
+		{ "scope": ["entity.name.label.duffle.atom"], "settings": { "foreground": "#c8c8c8", "fontStyle": "bold" } },
+		{ "scope": ["support.function.duffle.cpu"], "settings": { "foreground": "#6d9aa0" } },
+		{ "scope": ["support.function.duffle.gte"], "settings": { "foreground": "#988bcb" } },
+		{ "scope": ["support.function.duffle.gpu"], "settings": { "foreground": "#bf7dac" } },
+		{ "scope": ["support.function.duffle.component"], "settings": { "foreground": "#8baa5d" } },
+		{ "scope": ["keyword.control.duffle.branch"], "settings": { "foreground": "#76ff7d", "fontStyle": "bold" } },
+		{ "scope": ["keyword.operator.duffle.delayslot"], "settings": { "foreground": "#ff5647" } },
+		{ "scope": ["variable.other.constant.duffle.gpr"], "settings": { "foreground": "#3fa8a6" } },
+		{ "scope": ["variable.other.constant.duffle.cop2"], "settings": { "foreground": "#945cd9" } },
+		{ "scope": ["storage.type.duffle.type"], "settings": { "foreground": "#54a4d6" } },
+		{ "scope": ["storage.modifier.duffle.attr"], "settings": { "foreground": "#73a07c" } }
+	]
+}

.vscode/install_extensions.ps1

@@ -0,0 +1,43 @@
+# Package and install the local VS Code Insiders extensions under .vscode/.
+# Usage:
+#   .\install_extensions.ps1
+#   .\install_extensions.ps1 -SkipPackage
+
+param([switch] $SkipPackage)
+
+$path_vscode   = $PSScriptRoot
+$code_insiders = "C:\apps\Microsoft VS Code Insiders\bin\code-insiders.cmd"
+if (-not (test-path -literalpath $code_insiders)) {
+	$found = get-command code-insiders -erroraction silentlycontinue
+	if ($found) { $code_insiders = $found.source }
+}
+
+if (-not (test-path -literalpath $code_insiders)) { throw "code-insiders not found. Install VS Code Insiders or add it to PATH." }
+
+$extensions = @(
+	(join-path $path_vscode "tape-atom-syntax"),
+	(join-path $path_vscode "cozy-and-windy")
+)
+
+foreach ($extension in $extensions) {
+	$package_json = join-path $extension "package.json"
+	if (-not (test-path -literalpath $package_json)) { throw "missing $package_json" }
+
+	$manifest = get-content -literalpath $package_json -raw | convertfrom-json
+	$vsix     = join-path $extension ("{0}-{1}.vsix" -f $manifest.name, $manifest.version)
+
+	if (-not $SkipPackage) {
+		if (-not $manifest.scripts.package) { throw "$package_json has no scripts.package" }
+		write-host "packaging $($manifest.displayName) ($($manifest.name)@$($manifest.version))"
+		& npm --prefix $extension run package
+		if ($LASTEXITCODE -ne 0) { throw "npm run package failed for $extension" }
+	}
+
+	if (-not (test-path -literalpath $vsix)) { throw "missing $vsix" }
+
+	write-host "installing $vsix"
+	& $code_insiders --install-extension $vsix --force
+	if ($LASTEXITCODE -ne 0) { throw "install failed for $vsix" }
+}
+
+write-host "done. reload the Insiders window (Developer: Reload Window)."

.vscode/launch.json

@@ -74,41 +74,7 @@
 			]
 		},
 		{
-			"name": "Debug: Hello GTE Psy-Q!",
-			"type": "gdb",
-			"request": "attach",
-			"target": "localhost:3333",
-			"remote": true,
-			"cwd": "${workspaceRoot}/build",
-			"valuesFormatting": "parseText",
-			"registerLimit": "1-32",
-			"frameFilters": false,
-			"showDevDebugOutput": false,
-			"printCalls": false,
-			"stopAtConnect": true,
-			"gdbpath": "gdb-multiarch",
-			"windows": {
-				"gdbpath": "gdb-multiarch.exe"
-			},
-			"osx": {
-				"gdbpath": "gdb"
-			},
-			"executable": "${workspaceRoot}/build/hello_gte.elf",
-			"setupCommands": [
-				{ "text": "set mi-async off" },
-				{ "text": "set remotetimeout 0" },
-				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
-				{ "text": "set logging redirect on" }
-			],
-			"autorun": [
-				"monitor reset shellhalt",
-				"load hello_gte.elf",
-				"tbreak main",
-				"continue"
-			]
-		},
-		{
-			"name": "Debug: Hello GTE Psy-Q! (atoms debug — DWARF-injected)",
+			"name": "Debug: Hello GTE!",
 			"type": "gdb",
 			"request": "attach",
 			"target": "localhost:3333",
@@ -138,7 +104,76 @@
 				"monitor reset shellhalt",
 				"load build/hello_gte.dwarf-injected.elf",
 				"source scripts/gdb/gdb_tape_atoms.gdb",
-				"source build/gen/hello_gte.gdbinit",
+				"tbreak main",
+				"continue"
+			]
+		},
+		{
+			"name": "Debug: Hello Joypad!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${workspaceRoot}",
+			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
+			"stopAtConnect": true,
+			"gdbpath": "gdb-multiarch",
+			"windows": {
+				"gdbpath": "gdb-multiarch.exe"
+			},
+			"osx": {
+				"gdbpath": "gdb"
+			},
+			"executable": "${workspaceRoot}/build/hello_joypad.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_joypad.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_joypad.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"tbreak main",
+				"continue"
+			]
+		},
+		{
+			"name": "Debug: Hello Camera!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${workspaceRoot}",
+			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
+			"stopAtConnect": true,
+			"gdbpath": "gdb-multiarch",
+			"windows": {
+				"gdbpath": "gdb-multiarch.exe"
+			},
+			"osx": {
+				"gdbpath": "gdb"
+			},
+			"executable": "${workspaceRoot}/build/hello_camera.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_camera.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_camera.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
 				"tbreak main",
 				"continue"
 			]

.vscode/tape-atom-syntax/classifier.js

@@ -0,0 +1,222 @@
+"use strict";
+
+const { nearestCall } = require("./lexer");
+const { mergeIndexes, scanSource } = require("./source-index");
+
+const TOKEN_TYPES = [
+	"tapeAtomKeyword",
+	"tapeAtomName",
+	"tapeComponentKeyword",
+	"tapeComponentName",
+	"tapeAnnotation",
+	"tapeBindType",
+	"tapePhase",
+	"tapeLabel",
+	"tapeCpuInstruction",
+	"tapeControlFlow",
+	"tapeGteInstruction",
+	"tapeGpuInstruction",
+	"tapeComponentInstruction",
+	"tapeDelaySlot",
+	"tapeGprRegister",
+	"tapeCop2Register",
+	"tapeDuffleType",
+	"tapeAttribute",
+	"keyword",
+	"macro",
+];
+
+const TOKEN_MODIFIERS = ["declaration", "tapeRead", "tapeWrite", "tapeAuto"];
+const TOKEN_TYPE_INDEX = new Map(TOKEN_TYPES.map((name, index) => [name, index]));
+const TOKEN_MODIFIER_INDEX = new Map(TOKEN_MODIFIERS.map((name, index) => [name, index]));
+
+const ATOM_KEYWORDS = new Set(["MipsAtom_", "MipsAtom_Proc_"]);
+const COMPONENT_KEYWORDS = new Set(["MipsAtomComp_", "MipsAtomComp_Proc_"]);
+const ANNOTATIONS = new Set([
+	"atom_info", "atom_bind", "atom_reads", "atom_writes", "atom_label",
+	"atom_offset", "atom_reg", "atom_type", "atom_ctx", "atom_phase",
+	"atom_auto_reg", "phase_auto_reg", "atom_dbg_skip",
+]);
+
+const DSL_KEYWORDS = new Set([
+	"FI_", "I_", "NI_", "Relative_", "Struct_", "Enum_", "Union_", "Array_",
+	"Slice_", "TypeR_", "TypeV_", "align_", "internal", "local_persist", "global",
+	"RO_", "LP_", "gknown", "expect_", "cexpr_",
+	"asm", "asm_words", "asm_rpins", "asm_clobber",
+	"O_", "S_", "C_", "T_", "tmpl", "glue", "r_", "v_", "tr_", "tv_",
+	"rgcc", "r_use", "r_set", "r_mod", "r_imm", "r_mem",
+	"u1_", "u2_", "u4_", "u8_", "s1_", "s2_", "s4_", "s8_",
+	"u1_r", "u2_r", "u4_r", "u8_r", "u1_v", "u2_v", "u4_v", "u8_v",
+]);
+
+const DELAY_SLOT_KEYWORDS = new Set(["LdSlot_", "BdSlot_", "DmaSlot_", "GteDelay_"]);
+
+const CONTROL_FLOW_PREFIXES = /^(?:branch_|jump_|call_)/;
+
+const ROLE_TO_TYPE = {
+	atomName: "tapeAtomName",
+	componentName: "tapeComponentName",
+	bindType: "tapeBindType",
+	duffleType: "tapeDuffleType",
+	gprRegister: "tapeGprRegister",
+	cop2Register: "tapeCop2Register",
+};
+
+function registerType(name, index) {
+	const kind = index.registers.get(name);
+	if (kind === "gpr" || /^R_[A-Za-z0-9_]+$/.test(name)) return "tapeGprRegister";
+	if (kind === "cop2" || /^(?:C2_|gte_cr_)[A-Za-z0-9_]+$/.test(name)) return "tapeCop2Register";
+	return null;
+}
+
+function instructionType(name, index) {
+	const domain = index.macros.get(name);
+	if (domain === "control") return "tapeControlFlow";
+	if (domain === "cpu") return "tapeCpuInstruction";
+	if (domain === "gte") return "tapeGteInstruction";
+	if (domain === "gpu") return "tapeGpuInstruction";
+	if (domain === "component") {
+		if (/^mac_gte_/.test(name)) return "tapeGteInstruction";
+		if (/^mac_gp/.test(name)) return "tapeGpuInstruction";
+		if (/^mac_/.test(name)) return "tapeComponentInstruction";
+		return "macro";
+	}
+	if (domain === "utility") return "macro";
+	if (/^gte_(?!cr_)/.test(name)) return "tapeGteInstruction";
+	if (/^gp[01]_/.test(name)) return "tapeGpuInstruction";
+	if (/^mac_gte_/.test(name)) return "tapeGteInstruction";
+	if (/^mac_gp/.test(name)) return "tapeGpuInstruction";
+	if (/^mac_/.test(name)) return "tapeComponentInstruction";
+	return null;
+}
+
+function modifierMask(modifiers) {
+	let mask = 0;
+	for (const modifier of modifiers) {
+		const index = TOKEN_MODIFIER_INDEX.get(modifier);
+		if (index !== undefined) mask |= (1 << index);
+	}
+	return mask;
+}
+
+function isRegUseAccess(tokens, tokenIndex) {
+	const prev = tokens[tokenIndex - 1];
+	if (!prev || prev.text !== ".") return false;
+	const prevPrev = tokens[tokenIndex - 2];
+	if (!prevPrev || prevPrev.kind !== "identifier") return false;
+	const next = tokens[tokenIndex + 1];
+	if (next && next.text === ".") return false;
+	if (prevPrev.text === "r") return true;
+	const prev3 = tokens[tokenIndex - 3];
+	const prev4 = tokens[tokenIndex - 4];
+	if (prev3 && prev3.text === "." && prev4 && prev4.kind === "identifier" && prev4.text === "r") return true;
+	return false;
+}
+
+function classifyDocument(source, filePath, workspaceIndex, shouldCancel = () => false) {
+	const scanned = scanSource(source, filePath);
+	const index = mergeIndexes(workspaceIndex, scanned.index);
+	const spans = [];
+
+	for (let tokenIndex = 0; tokenIndex < scanned.tokens.length; tokenIndex += 1) {
+		if (shouldCancel()) break;
+		const token = scanned.tokens[tokenIndex];
+		if (token.kind !== "identifier") continue;
+
+		let type = null;
+		let modifiers = [];
+		const declaration = scanned.declarations.get(token.start);
+		const context = nearestCall(scanned.contexts, tokenIndex);
+
+		if (declaration) {
+			type = ROLE_TO_TYPE[declaration.role] || null;
+			modifiers = declaration.modifiers.slice();
+		} else if (ATOM_KEYWORDS.has(token.text)) {
+			type = "tapeAtomKeyword";
+		} else if (COMPONENT_KEYWORDS.has(token.text)) {
+			type = "keyword";
+		} else if (ANNOTATIONS.has(token.text)) {
+			type = "tapeAnnotation";
+		} else if (context && context.callee === "atom_bind" && context.argIndex === 0) {
+			type = "tapeBindType";
+		} else if (context && context.callee === "atom_phase" && context.argIndex === 0) {
+			type = "tapePhase";
+			modifiers = ["declaration"];
+		} else if (context && context.callee === "atom_ctx" && context.argIndex === 0) {
+			type = "tapeAtomName";
+		} else if (context && context.callee === "atom_label" && context.argIndex === 0) {
+			type = "tapeLabel";
+			modifiers = ["declaration"];
+		} else if (context && context.callee === "atom_offset" && context.argIndex <= 1) {
+			type = "tapeLabel";
+		} else if (context && context.callee === "atom_reads") {
+			type = registerType(token.text, index);
+			if (type) modifiers = ["tapeRead"];
+		} else if (context && context.callee === "atom_writes") {
+			type = registerType(token.text, index);
+			if (type) modifiers = ["tapeWrite"];
+		} else if (context && context.callee === "atom_auto_reg") {
+			if (context.argIndex === 0) type = "tapeAtomName";
+			if (context.argIndex === 1) {
+				type = "tapeGprRegister";
+				modifiers = ["declaration", "tapeAuto"];
+			}
+		} else if (context && context.callee === "phase_auto_reg") {
+			if (context.argIndex === 0) type = "tapePhase";
+			if (context.argIndex === 1) {
+				type = "tapeGprRegister";
+				modifiers = ["declaration", "tapeAuto"];
+			}
+		}
+
+		if (!type && index.bindTypes.has(token.text)) type = "tapeBindType";
+		if (!type && DSL_KEYWORDS.has(token.text)) type = "keyword";
+		if (!type && index.types.has(token.text)) type = "tapeDuffleType";
+		if (!type && index.attributes.has(token.text)) type = "tapeAttribute";
+		if (!type) type = registerType(token.text, index);
+		if (!type && DELAY_SLOT_KEYWORDS.has(token.text)) type = "tapeDelaySlot";
+		if (!type) {
+			const domain = index.macros.get(token.text);
+			if (domain === "control" || (domain && CONTROL_FLOW_PREFIXES.test(token.text))) {
+				type = "tapeControlFlow";
+			}
+		}
+		if (!type && isRegUseAccess(scanned.tokens, tokenIndex)) type = "tapeGprRegister";
+		if (!type) type = instructionType(token.text, index);
+		if (!type && /^(?:Slice_|A[0-9]+_)/.test(token.text)) type = "tapeDuffleType";
+		if (!type && /_[RV]$/.test(token.text)) type = "tapeDuffleType";
+		if (!type && index.atoms.has(token.text)) type = "tapeAtomName";
+		if (!type && index.components.has(token.text)) type = "tapeComponentName";
+		if (!type && index.phases.has(token.text)) type = "tapePhase";
+		if (!type && index.labels.has(token.text)) type = "tapeLabel";
+		if (!type) continue;
+
+		spans.push({
+			text: token.text,
+			type,
+			typeIndex: TOKEN_TYPE_INDEX.get(type),
+			modifiers,
+			modifierMask: modifierMask(modifiers),
+			start: token.start,
+			length: token.end - token.start,
+			line: token.line,
+			character: token.character,
+		});
+	}
+
+	spans.sort((left, right) => left.start - right.start || left.length - right.length);
+	const nonOverlapping = [];
+	for (const span of spans) {
+		const previous = nonOverlapping[nonOverlapping.length - 1];
+		if (!previous || previous.start + previous.length <= span.start) nonOverlapping.push(span);
+	}
+
+	return { spans: nonOverlapping, errors: scanned.errors };
+}
+
+module.exports = {
+	TOKEN_MODIFIERS,
+	TOKEN_TYPES,
+	classifyDocument,
+	modifierMask,
+};

.vscode/tape-atom-syntax/extension.js

@@ -0,0 +1,111 @@
+"use strict";
+
+const vscode = require("vscode");
+const { TOKEN_MODIFIERS, TOKEN_TYPES, classifyDocument } = require("./classifier");
+const { createIndex, mergeIndexes, scanSource } = require("./source-index");
+
+const SOURCE_GLOB = "**/*.{c,h,cc,cpp,cxx,hh,hpp,hxx}";
+const EXCLUDE_GLOB = "**/{gen,build,.slop_cache,toolchain,node_modules}/**";
+const EXCLUDED_SEGMENTS = new Set(["gen", "build", ".slop_cache", "toolchain", "node_modules"]);
+
+function isExcluded(uri) {
+	const segments = uri.fsPath.replaceAll("\\", "/").split("/");
+	return segments.some((segment) => EXCLUDED_SEGMENTS.has(segment));
+}
+
+function formatError(filePath, error) {
+	return `${filePath}:${error.offset}: ${error.kind}`;
+}
+
+async function activate(context) {
+	const output = vscode.window.createOutputChannel("Tape Atom DSL");
+	const emitter = new vscode.EventEmitter();
+	const legend = new vscode.SemanticTokensLegend(TOKEN_TYPES, TOKEN_MODIFIERS);
+	let workspaceIndex = createIndex();
+	let rebuildGeneration = 0;
+	let debounceHandle = null;
+
+	async function rebuildIndex() {
+		const generation = ++rebuildGeneration;
+		const files = await vscode.workspace.findFiles(SOURCE_GLOB, EXCLUDE_GLOB);
+		let nextIndex = createIndex();
+
+		for (const uri of files) {
+			if (generation !== rebuildGeneration) return;
+			if (isExcluded(uri)) continue;
+			try {
+				const bytes = await vscode.workspace.fs.readFile(uri);
+				const source = Buffer.from(bytes).toString("utf8");
+				const result = scanSource(source, uri.fsPath);
+				nextIndex = mergeIndexes(nextIndex, result.index);
+				for (const error of result.errors) output.appendLine(formatError(uri.fsPath, error));
+			} catch (error) {
+				output.appendLine(`${uri.fsPath}: ${error.stack || error.message || error}`);
+			}
+		}
+
+		if (generation !== rebuildGeneration) return;
+		workspaceIndex = nextIndex;
+		emitter.fire();
+	}
+
+	function scheduleRebuild(uri) {
+		if (uri && isExcluded(uri)) return;
+		if (debounceHandle !== null) clearTimeout(debounceHandle);
+		debounceHandle = setTimeout(() => {
+			debounceHandle = null;
+			rebuildIndex().catch((error) => output.appendLine(error.stack || String(error)));
+		}, 100);
+	}
+
+	const provider = {
+		onDidChangeSemanticTokens: emitter.event,
+		provideDocumentSemanticTokens(document, cancellationToken) {
+			try {
+				const result = classifyDocument(
+					document.getText(),
+					document.uri.fsPath,
+					workspaceIndex,
+					() => cancellationToken.isCancellationRequested
+				);
+				const builder = new vscode.SemanticTokensBuilder(legend);
+				for (const span of result.spans) {
+					if (cancellationToken.isCancellationRequested) break;
+					builder.push(span.line, span.character, span.length, span.typeIndex, span.modifierMask);
+				}
+				for (const error of result.errors) {
+					output.appendLine(formatError(document.uri.fsPath || document.uri.toString(), error));
+				}
+				return builder.build();
+			} catch (error) {
+				output.appendLine(`${document.uri}: ${error.stack || error.message || error}`);
+				return new vscode.SemanticTokensBuilder(legend).build();
+			}
+		},
+	};
+
+	const selector = [
+		{ language: "c", scheme: "file" },
+		{ language: "c", scheme: "untitled" },
+		{ language: "cpp", scheme: "file" },
+		{ language: "cpp", scheme: "untitled" },
+	];
+	const watcher = vscode.workspace.createFileSystemWatcher(SOURCE_GLOB);
+
+	context.subscriptions.push(
+		output,
+		emitter,
+		watcher,
+		watcher.onDidCreate(scheduleRebuild),
+		watcher.onDidChange(scheduleRebuild),
+		watcher.onDidDelete(scheduleRebuild),
+		vscode.languages.registerDocumentSemanticTokensProvider(selector, provider, legend),
+		{ dispose() { if (debounceHandle !== null) clearTimeout(debounceHandle); } }
+	);
+
+	await rebuildIndex();
+}
+
+function deactivate() {}
+
+module.exports = { activate, deactivate };

.vscode/tape-atom-syntax/lexer.js

@@ -0,0 +1,186 @@
+"use strict";
+
+function isIdentifierStart(code) {
+	return code === 95 ||
+		(code >= 65 && code <= 90) ||
+		(code >= 97 && code <= 122);
+}
+
+function isIdentifierContinue(code) {
+	return isIdentifierStart(code) || (code >= 48 && code <= 57);
+}
+
+function lex(source) {
+	if (typeof source !== "string") throw new TypeError("source must be a string");
+
+	const tokens = [];
+	const errors = [];
+	let offset = 0;
+	let line = 0;
+	let character = 0;
+
+	function advance() {
+		if (source[offset] === "\r" && source[offset + 1] === "\n") {
+			offset += 2;
+			line += 1;
+			character = 0;
+			return;
+		}
+		if (source[offset] === "\n") {
+			offset += 1;
+			line += 1;
+			character = 0;
+			return;
+		}
+		offset += 1;
+		character += 1;
+	}
+
+	function pushToken(kind, start, startLine, startCharacter) {
+		tokens.push({
+			kind,
+			text: source.slice(start, offset),
+			start,
+			end: offset,
+			line: startLine,
+			character: startCharacter,
+		});
+	}
+
+	while (offset < source.length) {
+		const ch = source[offset];
+
+		if (/\s/.test(ch)) {
+			advance();
+			continue;
+		}
+
+		if (ch === "/" && source[offset + 1] === "/") {
+			while (offset < source.length && source[offset] !== "\r" && source[offset] !== "\n") advance();
+			continue;
+		}
+
+		if (ch === "/" && source[offset + 1] === "*") {
+			const start = offset;
+			advance();
+			advance();
+			let closed = false;
+			while (offset < source.length) {
+				if (source[offset] === "*" && source[offset + 1] === "/") {
+					advance();
+					advance();
+					closed = true;
+					break;
+				}
+				advance();
+			}
+			if (!closed) errors.push({ kind: "unterminated-block-comment", offset: start });
+			continue;
+		}
+
+		if (ch === "\"" || ch === "'") {
+			const quote = ch;
+			const start = offset;
+			advance();
+			let closed = false;
+			while (offset < source.length) {
+				if (source[offset] === "\\") {
+					advance();
+					if (offset < source.length) advance();
+					continue;
+				}
+				if (source[offset] === quote) {
+					advance();
+					closed = true;
+					break;
+				}
+				if (source[offset] === "\n" || source[offset] === "\r") break;
+				advance();
+			}
+			if (!closed) errors.push({ kind: "unterminated-literal", offset: start });
+			continue;
+		}
+
+		const code = source.charCodeAt(offset);
+		if (isIdentifierStart(code)) {
+			const start = offset;
+			const startLine = line;
+			const startCharacter = character;
+			advance();
+			while (offset < source.length && isIdentifierContinue(source.charCodeAt(offset))) advance();
+			pushToken("identifier", start, startLine, startCharacter);
+			continue;
+		}
+
+		const start = offset;
+		const startLine = line;
+		const startCharacter = character;
+		advance();
+		pushToken("punctuation", start, startLine, startCharacter);
+	}
+
+	return { tokens, errors };
+}
+
+function buildCallContexts(tokens) {
+	const contexts = Array.from({ length: tokens.length }, () => []);
+	const calls = [];
+	const errors = [];
+	const stack = [];
+
+	for (let tokenIndex = 0; tokenIndex < tokens.length; tokenIndex += 1) {
+		const token = tokens[tokenIndex];
+
+		if (token.text === ")") {
+			if (stack.length === 0) {
+				errors.push({ kind: "unmatched-close-paren", offset: token.start });
+			} else {
+				const frame = stack.pop();
+				if (frame.callee !== null) calls.push({ ...frame, closeTokenIndex: tokenIndex });
+			}
+		}
+
+		contexts[tokenIndex] = stack
+			.filter((frame) => frame.callee !== null)
+			.map((frame) => ({
+				callee: frame.callee,
+				calleeTokenIndex: frame.calleeTokenIndex,
+				openTokenIndex: frame.openTokenIndex,
+				argIndex: frame.argIndex,
+			}));
+
+		if (token.text === "(") {
+			const previous = tokens[tokenIndex - 1];
+			const hasCallee = previous && previous.kind === "identifier";
+			stack.push({
+				callee: hasCallee ? previous.text : null,
+				calleeTokenIndex: hasCallee ? tokenIndex - 1 : -1,
+				openTokenIndex: tokenIndex,
+				argIndex: 0,
+			});
+			continue;
+		}
+
+		if (token.text === "," && stack.length > 0) {
+			const frame = stack[stack.length - 1];
+			if (frame.callee !== null) frame.argIndex += 1;
+		}
+	}
+
+	for (const frame of stack) {
+		errors.push({ kind: "unmatched-open-paren", offset: tokens[frame.openTokenIndex].start });
+	}
+
+	return { contexts, calls, errors };
+}
+
+function nearestCall(contexts, tokenIndex, callee) {
+	const entries = contexts[tokenIndex] || [];
+	for (let contextIndex = entries.length - 1; contextIndex >= 0; contextIndex -= 1) {
+		const entry = entries[contextIndex];
+		if (callee === undefined || entry.callee === callee) return entry;
+	}
+	return null;
+}
+
+module.exports = { buildCallContexts, lex, nearestCall };

.vscode/tape-atom-syntax/package.json

@@ -0,0 +1,85 @@
+{
+	"name": "atomasm-psx",
+	"displayName": "AtomAsm-PSX",
+	"description": "Semantic highlighting for the PS1 Tape/Atom MIPS macro DSL",
+	"publisher": "local",
+	"version": "0.3.0",
+	"engines": { "vscode": "^1.80.0" },
+	"categories": ["Programming Languages"],
+	"activationEvents": ["onLanguage:c", "onLanguage:cpp"],
+	"main": "./extension.js",
+	"files": [
+		"classifier.js",
+		"extension.js",
+		"lexer.js",
+		"source-index.js",
+		"syntaxes/tape_atom.tmLanguage.json"
+	],
+	"scripts": {
+		"test": "node --test test/*.test.js",
+		"package": "npx --yes @vscode/vsce@3.6.1 package --allow-missing-repository --skip-license --out atomasm-psx-0.3.0.vsix"
+	},
+	"contributes": {
+		"semanticTokenTypes": [
+			{ "id": "tapeAtomKeyword", "superType": "keyword", "description": "Tape atom declaration keyword" },
+			{ "id": "tapeAtomName", "superType": "function", "description": "Tape atom name" },
+			{ "id": "tapeComponentKeyword", "superType": "keyword", "description": "Tape atom component declaration keyword" },
+			{ "id": "tapeComponentName", "superType": "function", "description": "Tape atom component name" },
+			{ "id": "tapeAnnotation", "superType": "macro", "description": "Tape atom annotation" },
+			{ "id": "tapeBindType", "superType": "type", "description": "Tape bind structure type" },
+			{ "id": "tapePhase", "superType": "label", "description": "Tape atom phase" },
+			{ "id": "tapeLabel", "superType": "label", "description": "Tape atom branch label" },
+			{ "id": "tapeCpuInstruction", "superType": "macro", "description": "MIPS CPU instruction emitter" },
+			{ "id": "tapeControlFlow", "superType": "keyword", "description": "MIPS branch or jump instruction" },
+			{ "id": "tapeGteInstruction", "superType": "macro", "description": "GTE instruction emitter" },
+			{ "id": "tapeGpuInstruction", "superType": "macro", "description": "GPU command emitter" },
+			{ "id": "tapeComponentInstruction", "superType": "macro", "description": "Tape atom component invocation" },
+			{ "id": "tapeDelaySlot", "superType": "keyword", "description": "Load or branch delay slot annotation" },
+			{ "id": "tapeGprRegister", "superType": "variable", "description": "MIPS GPR alias" },
+			{ "id": "tapeCop2Register", "superType": "variable", "description": "COP2 data or control register alias" },
+			{ "id": "tapeDuffleType", "superType": "type", "description": "Duffle type or type constructor" },
+			{ "id": "tapeAttribute", "superType": "keyword", "description": "Duffle linkage or storage attribute" },
+			{ "id": "keyword", "description": "Standard keyword (DSL built-in macros)" },
+			{ "id": "macro", "description": "Standard macro (utility #define with no instruction domain)" }
+		],
+		"semanticTokenModifiers": [
+			{ "id": "tapeRead", "description": "Register declared in atom_reads" },
+			{ "id": "tapeWrite", "description": "Register declared in atom_writes" },
+			{ "id": "tapeAuto", "description": "Auto-allocated register" }
+		],
+		"semanticTokenScopes": [
+			{
+				"language": "c",
+				"scopes": {
+					"tapeAtomKeyword": ["keyword.control.duffle.atom"],
+					"tapeAtomName": ["entity.name.function.duffle.atom"],
+					"tapeComponentKeyword": ["keyword.control.duffle.component"],
+					"tapeComponentName": ["entity.name.function.duffle.component"],
+					"tapeAnnotation": ["support.function.duffle.annotation"],
+					"tapeBindType": ["entity.name.type.duffle.bind"],
+					"tapePhase": ["entity.name.tag.duffle.phase"],
+					"tapeLabel": ["entity.name.label.duffle.atom"],
+					"tapeCpuInstruction": ["support.function.duffle.cpu"],
+					"tapeControlFlow": ["keyword.control.duffle.branch"],
+					"tapeGteInstruction": ["support.function.duffle.gte"],
+					"tapeGpuInstruction": ["support.function.duffle.gpu"],
+					"tapeComponentInstruction": ["support.function.duffle.component"],
+					"tapeDelaySlot": ["keyword.operator.duffle.delayslot"],
+					"tapeGprRegister": ["variable.other.constant.duffle.gpr"],
+					"tapeCop2Register": ["variable.other.constant.duffle.cop2"],
+					"tapeDuffleType": ["storage.type.duffle.type"],
+					"tapeAttribute": ["storage.modifier.duffle.attr"],
+				"keyword": ["keyword"],
+				"macro": ["entity.name.function.preprocessor"]
+				}
+			}
+		],
+		"grammars": [
+			{
+				"scopeName": "tape_atom.injection",
+				"path": "./syntaxes/tape_atom.tmLanguage.json",
+				"injectTo": ["source.c", "source.cpp"]
+			}
+		]
+	}
+}

.vscode/tape-atom-syntax/source-index.js

@@ -0,0 +1,341 @@
+"use strict";
+
+const path = require("node:path");
+const { buildCallContexts, lex, nearestCall } = require("./lexer");
+
+const BASE_TYPES = [
+	"B1", "B2", "B4", "B8", "F4", "F8", "S1", "S2", "S4", "S8",
+	"U1", "U2", "U4", "U8", "MipsAtom", "MipsCode", "Reg",
+];
+
+const C_BUILTINS = new Set([
+	"void", "type", "char", "short", "int", "long", "float", "double",
+	"unsigned", "signed", "bool", "size_t", "uint8_t", "uint16_t", "uint32_t",
+	"int8_t", "int16_t", "int32_t",
+]);
+
+const BASE_ATTRIBUTES = [
+	"FI_", "I_", "NI_", "Relative_", "Struct_", "Enum_", "Union_", "Array_",
+	"Slice_", "TypeR_", "TypeV_", "align_", "internal", "local_persist", "global",
+	"RO_", "LP_", "gknown", "expect_", "cexpr_",
+	"asm", "asm_words", "asm_rpins", "asm_clobber",
+	"O_", "S_", "C_", "T_", "tmpl", "glue", "r_", "v_", "tr_", "tv_",
+	"rgcc", "r_use", "r_set", "r_mod", "r_imm", "r_mem",
+	"u1_", "u2_", "u4_", "u8_", "s1_", "s2_", "s4_", "s8_",
+	"u1_r", "u2_r", "u4_r", "u8_r", "u1_v", "u2_v", "u4_v", "u8_v",
+];
+
+function createIndex() {
+	return {
+		atoms: new Set(),
+		components: new Set(),
+		componentAliases: new Set(),
+		macros: new Map(),
+		registers: new Map(),
+		bindTypes: new Set(),
+		types: new Set(BASE_TYPES),
+		phases: new Set(),
+		labels: new Set(),
+		attributes: new Set(BASE_ATTRIBUTES),
+		componentCallees: new Map(),
+	};
+}
+
+function cloneIndex(source) {
+	const result = createIndex();
+	for (const key of ["atoms", "components", "componentAliases", "bindTypes", "types", "phases", "labels", "attributes"]) {
+		for (const value of source[key]) result[key].add(value);
+	}
+	for (const [name, domain] of source.macros) result.macros.set(name, domain);
+	for (const [name, domain] of source.registers) result.registers.set(name, domain);
+	for (const [name, callees] of source.componentCallees) result.componentCallees.set(name, callees.slice());
+	return result;
+}
+
+function mergeIndexes(...sources) {
+	const result = createIndex();
+	for (const source of sources) {
+		if (!source) continue;
+		for (const key of ["atoms", "components", "componentAliases", "bindTypes", "types", "phases", "labels", "attributes"]) {
+			for (const value of source[key]) result[key].add(value);
+		}
+		for (const [name, domain] of source.macros) {
+			const existing = result.macros.get(name);
+			if (!existing || domainRank(domain) >= domainRank(existing)) result.macros.set(name, domain);
+		}
+		for (const [name, domain] of source.registers) result.registers.set(name, domain);
+		for (const [name, callees] of source.componentCallees) {
+			const existing = result.componentCallees.get(name) || [];
+			result.componentCallees.set(name, existing.concat(callees));
+		}
+	}
+	return resolveComponentDomains(result);
+}
+
+function domainFromPath(filePath) {
+	const base = path.basename(filePath.replaceAll("\\", "/")).toLowerCase();
+	if (base === "mips.h") return "cpu";
+	if (base === "gte.h") return "gte";
+	if (base === "gp.h") return "gpu";
+	return null;
+}
+
+function prefixDomain(name) {
+	if (/^(?:branch_|jump_|call_)/.test(name)) return "control";
+	if (/^gte_(?!cr_)/.test(name) || name.startsWith("mac_gte_") || name.startsWith("ac_gte_")) return "gte";
+	if (/^gp[01]_/.test(name) || name.startsWith("mac_gp_") || name.startsWith("ac_gp_")) return "gpu";
+	return null;
+}
+
+function collectBraceIdentifiers(tokens, openBraceIndex) {
+	const names = [];
+	let depth = 0;
+	for (let tokenIndex = openBraceIndex; tokenIndex < tokens.length; tokenIndex += 1) {
+		if (tokens[tokenIndex].text === "{") depth += 1;
+		if (tokens[tokenIndex].text === "}") {
+			depth -= 1;
+			if (depth === 0) break;
+		}
+		if (tokens[tokenIndex].kind === "identifier") names.push(tokens[tokenIndex].text);
+	}
+	return names;
+}
+
+function resolveComponentDomains(index) {
+	const hardwareRank = { cpu: 1, gpu: 2, gte: 3, control: 4 };
+	let changed = true;
+	while (changed) {
+		changed = false;
+		for (const [alias, callees] of index.componentCallees) {
+			let best = index.macros.get(alias) || "component";
+			let bestRank = hardwareRank[best] || 0;
+			for (const callee of callees) {
+				const domain = prefixDomain(callee) || index.macros.get(callee);
+				const rank = hardwareRank[domain] || 0;
+				if (rank > bestRank) {
+					best = domain;
+					bestRank = rank;
+				}
+			}
+			if (bestRank > 0 && index.macros.get(alias) !== best) {
+				index.macros.set(alias, best);
+				changed = true;
+			}
+		}
+	}
+	return index;
+}
+
+function domainRank(domain) {
+	if (domain === "control") return 4;
+	if (domain === "cpu" || domain === "gte" || domain === "gpu") return 3;
+	if (domain === "component") return 2;
+	return 1;
+}
+
+function registerKind(name) {
+	if (/^R_[A-Za-z0-9_]+$/.test(name)) return "gpr";
+	if (/^(?:C2_|gte_cr_)[A-Za-z0-9_]+$/.test(name)) return "cop2";
+	return null;
+}
+
+function componentAlias(name) {
+	return name.startsWith("ac_") ? `mac_${name.slice(3)}` : null;
+}
+
+function findFunctionNameBefore(tokens, calleeTokenIndex) {
+	let closeIndex = calleeTokenIndex - 1;
+	while (closeIndex >= 0 && tokens[closeIndex].kind === "identifier" && tokens[closeIndex].text === "atom_dbg_skip") {
+		closeIndex -= 1;
+	}
+	if (!tokens[closeIndex] || tokens[closeIndex].text !== ")") return null;
+
+	let depth = 1;
+	for (let tokenIndex = closeIndex - 1; tokenIndex >= 0; tokenIndex -= 1) {
+		if (tokens[tokenIndex].text === ")") depth += 1;
+		if (tokens[tokenIndex].text === "(") depth -= 1;
+		if (depth !== 0) continue;
+		const name = tokens[tokenIndex - 1];
+		return name && name.kind === "identifier" ? name : null;
+	}
+	return null;
+}
+
+function scanSource(source, filePath) {
+	const lexical = lex(source);
+	const balanced = buildCallContexts(lexical.tokens);
+	const tokens = lexical.tokens;
+	const contexts = balanced.contexts;
+	const index = createIndex();
+	const declarations = new Map();
+	const domain = domainFromPath(filePath);
+
+	function mark(token, role, modifiers = ["declaration"]) {
+		declarations.set(token.start, { role, modifiers });
+	}
+
+	function addComponent(token) {
+		index.components.add(token.text);
+		mark(token, "componentName");
+		const alias = componentAlias(token.text);
+		if (alias) {
+			index.componentAliases.add(alias);
+			index.macros.set(alias, prefixDomain(alias) || prefixDomain(token.text) || "component");
+		}
+	}
+
+	function bindComponentCallees(alias, callees) {
+		if (!alias) return;
+		index.componentAliases.add(alias);
+		index.componentCallees.set(alias, callees);
+		if (!index.macros.has(alias)) index.macros.set(alias, "component");
+	}
+
+	for (let tokenIndex = 0; tokenIndex < tokens.length; tokenIndex += 1) {
+		const token = tokens[tokenIndex];
+		if (token.kind !== "identifier") continue;
+
+		const kind = registerKind(token.text);
+		if (kind) {
+			index.registers.set(token.text, kind);
+			if (tokens[tokenIndex + 1] && tokens[tokenIndex + 1].text === "=") {
+				mark(token, kind === "gpr" ? "gprRegister" : "cop2Register");
+			}
+		}
+
+		const context = nearestCall(contexts, tokenIndex);
+		if (context && context.argIndex === 0) {
+			if (context.callee === "MipsAtom_") {
+				index.atoms.add(token.text);
+				mark(token, "atomName");
+			}
+			if (context.callee === "MipsAtomComp_") addComponent(token);
+			if (context.callee === "atom_bind") index.bindTypes.add(token.text);
+			if (context.callee === "atom_phase" || context.callee === "phase_auto_reg") index.phases.add(token.text);
+			if (context.callee === "atom_label" || context.callee === "atom_offset") index.labels.add(token.text);
+		}
+
+		const isWrappedType = context && (
+			((context.callee === "Struct_" || context.callee === "Union_") && context.argIndex === 0) ||
+			(context.callee === "Enum_" && context.argIndex === 1)
+		);
+		if (isWrappedType) {
+			index.types.add(token.text);
+			mark(token, token.text.startsWith("Binds_") ? "bindType" : "duffleType");
+			if (token.text.startsWith("Binds_")) index.bindTypes.add(token.text);
+		}
+
+		if (context && context.callee === "atom_offset" && context.argIndex === 1) index.labels.add(token.text);
+
+		if (context && context.callee === "atom_auto_reg") {
+			if (context.argIndex === 0) index.atoms.add(token.text);
+			if (context.argIndex === 1) {
+				index.registers.set(token.text, "gpr");
+				mark(token, "gprRegister", ["declaration", "tapeAuto"]);
+			}
+		}
+
+		if (context && context.callee === "phase_auto_reg" && context.argIndex === 1) {
+			index.registers.set(token.text, "gpr");
+			mark(token, "gprRegister", ["declaration", "tapeAuto"]);
+		}
+
+		if (token.text === "define" && tokens[tokenIndex - 1] && tokens[tokenIndex - 1].text === "#") {
+			const name = tokens[tokenIndex + 1];
+			if (name && name.kind === "identifier" && name.line === token.line) {
+				if (/^(?:RegUse_|Struct_|Enum_|Union_|TypeR_|TypeV_|Relative_|Binds_)/.test(name.text)) {
+					index.types.add(name.text);
+				} else if (/^(?:ac_|mac_)/.test(name.text)) {
+					const alias = name.text.startsWith("ac_") ? componentAlias(name.text) : name.text;
+					const rest = [];
+					for (let restIndex = tokenIndex + 2; restIndex < tokens.length && tokens[restIndex].line === name.line; restIndex += 1) {
+						if (tokens[restIndex].kind === "identifier") rest.push(tokens[restIndex].text);
+					}
+					if (alias) {
+						index.componentAliases.add(alias);
+						index.macros.set(alias, prefixDomain(alias) || "component");
+						if (rest.length) index.componentCallees.set(alias, rest);
+					}
+				} else {
+					index.macros.set(name.text, domain || "utility");
+				}
+			}
+		}
+
+		if (token.text === "typedef") {
+			let endIndex = tokenIndex + 1;
+			let hasBrace = false;
+			let lastIdentifier = null;
+			while (endIndex < tokens.length && tokens[endIndex].text !== ";") {
+				if (tokens[endIndex].text === "{") hasBrace = true;
+				if (tokens[endIndex].kind === "identifier" && !C_BUILTINS.has(tokens[endIndex].text)) lastIdentifier = tokens[endIndex];
+				endIndex += 1;
+			}
+			if (!hasBrace && lastIdentifier && !C_BUILTINS.has(lastIdentifier.text)) {
+				index.types.add(lastIdentifier.text);
+				mark(lastIdentifier, "duffleType");
+			}
+		}
+
+		if (token.text === "MipsAtom_Proc_") {
+			const functionName = findFunctionNameBefore(tokens, tokenIndex);
+			if (functionName) {
+				const atomName = functionName.text.endsWith("_proc")
+					? functionName.text.slice(0, -5)
+					: functionName.text;
+				index.atoms.add(atomName);
+				index.atoms.add(functionName.text);
+				mark(functionName, "atomName");
+			}
+		}
+
+		if (token.text === "MipsAtomComp_Proc_") {
+			const functionName = findFunctionNameBefore(tokens, tokenIndex);
+			if (functionName) addComponent(functionName);
+		}
+	}
+
+	for (const call of balanced.calls) {
+		if (call.callee === "MipsAtomComp_") {
+			const name = tokens[call.openTokenIndex + 1];
+			const brace = tokens[call.closeTokenIndex + 1];
+			if (name && name.kind === "identifier" && brace && brace.text === "{") {
+				bindComponentCallees(componentAlias(name.text), collectBraceIdentifiers(tokens, call.closeTokenIndex + 1));
+			}
+		}
+		if (call.callee === "MipsAtomComp_Proc_") {
+			const functionName = findFunctionNameBefore(tokens, call.calleeTokenIndex);
+			let braceIndex = -1;
+			for (let tokenIndex = call.openTokenIndex + 1; tokenIndex < call.closeTokenIndex; tokenIndex += 1) {
+				if (tokens[tokenIndex].text === "{") {
+					braceIndex = tokenIndex;
+					break;
+				}
+			}
+			if (functionName && braceIndex >= 0) {
+				bindComponentCallees(componentAlias(functionName.text), collectBraceIdentifiers(tokens, braceIndex));
+			}
+		}
+		if (!domain) continue;
+		const name = tokens[call.calleeTokenIndex];
+		const after = tokens[call.closeTokenIndex + 1];
+		if (!name || !after || after.text !== "{") continue;
+		if (/^(?:gp0_|gp1_|gte_|mac_)/.test(name.text)) index.macros.set(name.text, domain);
+	}
+
+	return {
+		index: resolveComponentDomains(cloneIndex(index)),
+		declarations,
+		tokens,
+		contexts,
+		errors: [...lexical.errors, ...balanced.errors],
+	};
+}
+
+module.exports = {
+	createIndex,
+	domainFromPath,
+	mergeIndexes,
+	resolveComponentDomains,
+	scanSource,
+};

.vscode/tape-atom-syntax/syntaxes/tape_atom.tmLanguage.json

@@ -0,0 +1,71 @@
+{
+	"scopeName": "tape_atom.injection",
+	"injectionSelector": "L:source.c -comment -string, L:source.cpp -comment -string",
+	"patterns": [
+		{ "include": "#atom-declarations" },
+		{ "include": "#component-declarations" },
+		{ "include": "#annotation-arguments" },
+		{ "include": "#annotations" },
+		{ "include": "#delay-slots" },
+		{ "include": "#types" },
+		{ "include": "#attributes" }
+	],
+	"repository": {
+		"atom-declarations": {
+			"patterns": [
+				{
+					"match": "\\b(MipsAtom_)\\s*\\(\\s*([A-Za-z_][A-Za-z0-9_]*)",
+					"captures": {
+						"1": { "name": "keyword.control.duffle.atom" },
+						"2": { "name": "entity.name.function.duffle.atom" }
+					}
+				},
+				{ "match": "\\bMipsAtom_Proc_\\b", "name": "keyword.control.duffle.atom" },
+				{ "match": "\\b[A-Za-z_][A-Za-z0-9_]*_proc\\b", "name": "entity.name.function.duffle.atom" }
+			]
+		},
+		"component-declarations": {
+			"patterns": [
+				{
+					"match": "\\b(MipsAtomComp_)\\s*\\(\\s*(ac_[A-Za-z0-9_]*)",
+					"captures": {
+						"1": { "name": "keyword" },
+						"2": { "name": "entity.name.function.duffle.component" }
+					}
+				},
+				{ "match": "\\bMipsAtomComp_Proc_\\b", "name": "keyword" }
+			]
+		},
+		"annotation-arguments": {
+			"patterns": [
+				{
+					"match": "\\b(atom_offset)\\s*\\(\\s*([A-Za-z_][A-Za-z0-9_]*)\\s*,\\s*([A-Za-z_][A-Za-z0-9_]*)",
+					"captures": {
+						"1": { "name": "support.function.duffle.annotation" },
+						"2": { "name": "entity.name.label.duffle.atom" },
+						"3": { "name": "entity.name.label.duffle.atom" }
+					}
+				},
+				{ "match": "(?<=\\batom_bind\\()\\s*Binds_[A-Za-z0-9_]+", "name": "entity.name.type.duffle.bind" },
+				{ "match": "(?<=\\batom_phase\\()\\s*[A-Za-z_][A-Za-z0-9_]*", "name": "entity.name.tag.duffle.phase" },
+				{ "match": "(?<=\\batom_label\\()\\s*[A-Za-z_][A-Za-z0-9_]*", "name": "entity.name.label.duffle.atom" }
+			]
+		},
+		"annotations": {
+			"match": "\\b(atom_info|atom_bind|atom_reads|atom_writes|atom_label|atom_offset|atom_reg|atom_type|atom_ctx|atom_phase|atom_auto_reg|phase_auto_reg|atom_dbg_skip)\\b",
+			"name": "support.function.duffle.annotation"
+		},
+		"delay-slots": {
+			"match": "\\b(LdSlot_|BdSlot_|DmaSlot_|GteDelay_)\\b",
+			"name": "keyword.operator.duffle.delayslot"
+		},
+		"types": {
+			"match": "\\b(?:Binds_[A-Za-z0-9_]+|RegUse_[A-Za-z0-9_]+)\\b",
+			"name": "storage.type.duffle.type"
+		},
+		"attributes": {
+			"match": "\\b(?:FI_|I_|NI_|Relative_|Struct_|Enum_|Union_|Array_|Slice_|TypeR_|TypeV_|align_|internal|local_persist|global|RO_|LP_|gknown|expect_|cexpr_|asm|asm_words|asm_rpins|asm_clobber|O_|S_|C_|T_|tmpl|glue|r_|v_|tr_|tv_|rgcc|r_use|r_set|r_mod|r_imm|r_mem|u[1248]_|u[1248]_r|u[1248]_v|s[1248]_)\\b",
+			"name": "keyword"
+		}
+	}
+}

.vscode/tape-atom-syntax/test/classifier.test.js

@@ -0,0 +1,128 @@
+"use strict";
+
+const assert = require("node:assert/strict");
+const test = require("node:test");
+
+const { classifyDocument } = require("../classifier");
+const { createIndex } = require("../source-index");
+
+function byText(result, text) {
+	return result.spans.filter((span) => span.text === text);
+}
+
+test("classifyDocument distinguishes declaration, annotation, phase, bind, and label roles", () => {
+	const source = [
+		"typedef Struct_(Binds_CubeTri) { U4 PrimCursor; };",
+		"MipsAtom_(cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4),",
+		"\tatom_reads(R_PrimCursor), atom_writes(R_FaceCursor)) {",
+		"\tbranch_le_zero(R_T0, atom_offset(cull, exit)),",
+		"\tatom_label(exit)",
+		"};",
+	].join("\n");
+
+	const result = classifyDocument(source, "C:/x/code/hello_camera/hello_camera.atom.c", createIndex());
+
+	assert.equal(byText(result, "MipsAtom_")[0].type, "tapeAtomKeyword");
+	assert.deepEqual(byText(result, "cube_g4_face")[0].modifiers, ["declaration"]);
+	assert.equal(byText(result, "atom_bind")[0].type, "tapeAnnotation");
+	assert.equal(byText(result, "Binds_CubeTri").at(-1).type, "tapeBindType");
+	assert.equal(byText(result, "cube_g4")[0].type, "tapePhase");
+	assert.deepEqual(byText(result, "cube_g4")[0].modifiers, ["declaration"]);
+	assert.equal(byText(result, "cull")[0].type, "tapeLabel");
+	assert.equal(byText(result, "exit").every((span) => span.type === "tapeLabel"), true);
+});
+
+test("classifyDocument applies read and write modifiers to GPRs", () => {
+	const source = "atom_info(atom_reads(R_PrimCursor), atom_writes(R_FaceCursor))";
+	const result = classifyDocument(source, "C:/x/code/test.atom.c", createIndex());
+
+	assert.deepEqual(byText(result, "R_PrimCursor")[0].modifiers, ["tapeRead"]);
+	assert.deepEqual(byText(result, "R_FaceCursor")[0].modifiers, ["tapeWrite"]);
+});
+
+test("classifyDocument separates CPU, GTE, GPU, and component domains", () => {
+	const workspace = createIndex();
+	workspace.macros.set("load_word", "cpu");
+	workspace.macros.set("gte_cmdw_rtpt", "gte");
+	workspace.macros.set("gp1_word_DisplayOn", "gpu");
+	workspace.macros.set("mac_yield", "control");
+	workspace.componentAliases.add("mac_yield");
+
+	const source = "load_word(R_T0, R_T1, 0), gte_cmdw_rtpt, gp1_word_DisplayOn(), mac_yield(), C2_MAC0, gte_cr_OFX_Code";
+	const result = classifyDocument(source, "C:/x/code/test.c", workspace);
+
+	assert.equal(byText(result, "load_word")[0].type, "tapeCpuInstruction");
+	assert.equal(byText(result, "gte_cmdw_rtpt")[0].type, "tapeGteInstruction");
+	assert.equal(byText(result, "gp1_word_DisplayOn")[0].type, "tapeGpuInstruction");
+	assert.equal(byText(result, "mac_yield")[0].type, "tapeControlFlow");
+	assert.equal(byText(result, "C2_MAC0")[0].type, "tapeCop2Register");
+	assert.equal(byText(result, "gte_cr_OFX_Code")[0].type, "tapeCop2Register");
+});
+
+test("component invocations keep the domain resolved from their emitted instructions", () => {
+	const workspace = createIndex();
+	workspace.macros.set("mac_load_word_imm", "cpu");
+	workspace.macros.set("mac_gcmd_push", "gpu");
+	workspace.macros.set("mac_gte_store_f3", "gte");
+	workspace.macros.set("mac_load_v3s4", "cpu");
+
+	const source = "mac_load_word_imm(dst, imm), mac_gcmd_push(cmd), mac_gte_store_f3(cursor), mac_load_v3s4()";
+	const result = classifyDocument(source, "C:/x/code/hello_camera/hello_camera.atom.c", workspace);
+
+	assert.equal(byText(result, "mac_load_word_imm")[0].type, "tapeCpuInstruction");
+	assert.equal(byText(result, "mac_gcmd_push")[0].type, "tapeGpuInstruction");
+	assert.equal(byText(result, "mac_gte_store_f3")[0].type, "tapeGteInstruction");
+	assert.equal(byText(result, "mac_load_v3s4")[0].type, "tapeCpuInstruction");
+});
+
+test("utility macros without a hardware domain use the standard macro token", () => {
+	const workspace = createIndex();
+	workspace.macros.set("load_word", "cpu");
+	workspace.macros.set("assert", "utility");
+	workspace.macros.set("stringify", "utility");
+	workspace.macros.set("u4_hi", "utility");
+
+	const source = "load_word(R_T0, R_T1, 0), assert(ok), stringify(name), u4_hi(imm)";
+	const result = classifyDocument(source, "C:/x/code/hello_camera/hello_camera.c", workspace);
+
+	assert.equal(byText(result, "load_word")[0].type, "tapeCpuInstruction");
+	assert.equal(byText(result, "assert")[0].type, "macro");
+	assert.equal(byText(result, "stringify")[0].type, "macro");
+	assert.equal(byText(result, "u4_hi")[0].type, "macro");
+});
+
+test("document-local declarations override an empty workspace index", () => {
+	const source = [
+		"MipsAtomComp_(ac_new_component) { nop };",
+		"MipsAtomComp_Proc_(ab, { nop })",
+		"mac_new_component(),",
+	].join("\n");
+	const result = classifyDocument(source, "C:/x/code/duffle/math.atom.c", createIndex());
+
+	assert.equal(byText(result, "MipsAtomComp_")[0].type, "keyword");
+	assert.equal(byText(result, "MipsAtomComp_Proc_")[0].type, "keyword");
+	assert.equal(byText(result, "ac_new_component")[0].type, "tapeComponentName");
+	assert.equal(byText(result, "mac_new_component")[0].type, "tapeComponentInstruction");
+});
+
+test("delay slot markers share the tapeDelaySlot token", () => {
+	const source = "LdSlot_ nop, BdSlot_ nop, DmaSlot_ nop2, GteDelay_ nop";
+	const result = classifyDocument(source, "C:/x/code/duffle/gte.atom.c", createIndex());
+
+	assert.equal(byText(result, "LdSlot_")[0].type, "tapeDelaySlot");
+	assert.equal(byText(result, "BdSlot_")[0].type, "tapeDelaySlot");
+	assert.equal(byText(result, "DmaSlot_")[0].type, "tapeDelaySlot");
+	assert.equal(byText(result, "GteDelay_")[0].type, "tapeDelaySlot");
+});
+
+test("classifier returns ordered non-overlapping spans and partial malformed output", () => {
+	const source = "atom_reads(R_A /* broken";
+	const result = classifyDocument(source, "C:/x/code/test.atom.c", createIndex());
+
+	assert.equal(result.errors.some((error) => error.kind === "unterminated-block-comment"), true);
+	for (let spanIndex = 1; spanIndex < result.spans.length; spanIndex += 1) {
+		const previous = result.spans[spanIndex - 1];
+		const current = result.spans[spanIndex];
+		assert.equal(previous.start + previous.length <= current.start, true);
+	}
+});

.vscode/tape-atom-syntax/test/contract.test.js

@@ -0,0 +1,88 @@
+"use strict";
+
+const assert = require("node:assert/strict");
+const fs = require("node:fs");
+const path = require("node:path");
+const test = require("node:test");
+
+const { TOKEN_MODIFIERS, TOKEN_TYPES } = require("../classifier");
+
+const ROOT = path.resolve(__dirname, "..");
+
+function readJson(filePath) {
+	const raw = fs.readFileSync(filePath, "utf8");
+	const stripped = raw.replace(/\/\/.*$/gm, "").replace(/,\s*([}\]])/g, "$1");
+	return JSON.parse(stripped);
+}
+
+function collectScopeNames(value, output = new Set()) {
+	if (Array.isArray(value)) {
+		for (const entry of value) collectScopeNames(entry, output);
+		return output;
+	}
+	if (!value || typeof value !== "object") return output;
+	if (typeof value.name === "string") output.add(value.name);
+	for (const child of Object.values(value)) collectScopeNames(child, output);
+	return output;
+}
+
+test("package semantic legend matches classifier exports", () => {
+	const packageJson = readJson(path.join(ROOT, "package.json"));
+	const contributedTypes = packageJson.contributes.semanticTokenTypes.map((entry) => entry.id);
+	const contributedModifiers = packageJson.contributes.semanticTokenModifiers.map((entry) => entry.id);
+
+	assert.equal(packageJson.version, "0.3.0");
+	assert.deepEqual(contributedTypes, TOKEN_TYPES);
+	assert.deepEqual(contributedModifiers, TOKEN_MODIFIERS.filter((name) => name !== "declaration"));
+});
+
+test("package includes runtime files only and acknowledges local-only metadata", () => {
+	const packageJson = readJson(path.join(ROOT, "package.json"));
+
+	assert.deepEqual(packageJson.files, [
+		"classifier.js",
+		"extension.js",
+		"lexer.js",
+		"source-index.js",
+		"syntaxes/tape_atom.tmLanguage.json",
+	]);
+	assert.equal(packageJson.scripts.package.includes("--allow-missing-repository"), true);
+	assert.equal(packageJson.scripts.package.includes("--skip-license"), true);
+});
+
+test("every semantic token has a scope mapping; DSL-specific tokens also have grammar scopes", () => {
+	const packageJson = readJson(path.join(ROOT, "package.json"));
+	const grammar = readJson(path.join(ROOT, "syntaxes", "tape_atom.tmLanguage.json"));
+	const mappings = packageJson.contributes.semanticTokenScopes[0].scopes;
+	const grammarScopes = collectScopeNames(grammar);
+
+	const grammarRequired = new Set([
+		"tapeAtomKeyword", "tapeAtomName", "tapeComponentName",
+		"tapeAnnotation", "tapeBindType", "tapePhase", "tapeLabel",
+		"tapeDelaySlot", "tapeDuffleType", "keyword",
+	]);
+
+	for (const tokenType of TOKEN_TYPES) {
+		assert.equal(Array.isArray(mappings[tokenType]), true, `missing scope mapping: ${tokenType}`);
+		if (grammarRequired.has(tokenType)) {
+			assert.equal(mappings[tokenType].some((scope) => grammarScopes.has(scope)), true, `grammar does not emit: ${tokenType}`);
+		}
+	}
+});
+
+test("TextMate offset labels stay scoped to atom_offset calls", () => {
+	const grammar = readJson(path.join(ROOT, "syntaxes", "tape_atom.tmLanguage.json"));
+	const serialized = JSON.stringify(grammar);
+	const offsetRule = grammar.repository["annotation-arguments"].patterns
+		.find((rule) => rule.match.includes("atom_offset"));
+
+	assert.equal(serialized.includes("(?<=,)"), false);
+	assert.equal(offsetRule.captures[1].name, "support.function.duffle.annotation");
+	assert.equal(offsetRule.captures[2].name, "entity.name.label.duffle.atom");
+	assert.equal(offsetRule.captures[3].name, "entity.name.label.duffle.atom");
+});
+
+test("workspace enables semantic highlighting", () => {
+	const settings = readJson(path.resolve(ROOT, "..", "settings.json"));
+	assert.equal(settings["editor.semanticHighlighting.enabled"], true);
+});

.vscode/tape-atom-syntax/test/lexer.test.js

@@ -0,0 +1,77 @@
+"use strict";
+
+const assert = require("node:assert/strict");
+const test = require("node:test");
+
+const { buildCallContexts, lex, nearestCall } = require("../lexer");
+
+test("lex skips comments, strings, and character literals", () => {
+	const source = [
+		"MipsAtom_(visible)",
+		"// MipsAtom_(line_comment)",
+		"const char *s = \"atom_reads(R_Hidden)\";",
+		"char c = '\\''; /* gte_cmdw_hidden */",
+		"atom_reads(R_Visible)",
+	].join("\n");
+
+	const result = lex(source);
+	const identifiers = result.tokens
+		.filter((token) => token.kind === "identifier")
+		.map((token) => token.text);
+
+	assert.deepEqual(result.errors, []);
+	assert.equal(identifiers.includes("visible"), true);
+	assert.equal(identifiers.includes("R_Visible"), true);
+	assert.equal(identifiers.includes("line_comment"), false);
+	assert.equal(identifiers.includes("R_Hidden"), false);
+	assert.equal(identifiers.includes("gte_cmdw_hidden"), false);
+});
+
+test("lex reports unterminated block comments without returning comment tokens", () => {
+	const result = lex("R_Visible /* atom_reads(R_Hidden)");
+
+	assert.equal(result.tokens.some((token) => token.text === "R_Visible"), true);
+	assert.equal(result.tokens.some((token) => token.text === "R_Hidden"), false);
+	assert.deepEqual(result.errors.map((error) => error.kind), ["unterminated-block-comment"]);
+});
+
+test("line comments stop at CRLF boundaries", () => {
+	const result = lex("// atom_reads(R_Hidden)\r\natom_reads(R_Visible)\r\n");
+	const identifiers = result.tokens
+		.filter((token) => token.kind === "identifier")
+		.map((token) => token.text);
+
+	assert.equal(identifiers.includes("R_Hidden"), false);
+	assert.equal(identifiers.includes("R_Visible"), true);
+});
+
+test("balanced contexts retain multiline nesting and argument indexes", () => {
+	const source = [
+		"atom_info(",
+		"\tatom_phase(cube_g4),",
+		"\tatom_reads(R_A, nested(R_B, R_C)),",
+		"\tatom_writes(R_D)",
+		")",
+	].join("\n");
+	const lexical = lex(source);
+	const balanced = buildCallContexts(lexical.tokens);
+
+	const byText = new Map();
+	lexical.tokens.forEach((token, index) => {
+		if (token.kind === "identifier") byText.set(token.text, index);
+	});
+
+	assert.equal(nearestCall(balanced.contexts, byText.get("cube_g4")).callee, "atom_phase");
+	assert.equal(nearestCall(balanced.contexts, byText.get("R_A")).callee, "atom_reads");
+	assert.equal(nearestCall(balanced.contexts, byText.get("R_A")).argIndex, 0);
+	assert.equal(nearestCall(balanced.contexts, byText.get("R_C")).callee, "nested");
+	assert.equal(nearestCall(balanced.contexts, byText.get("R_D")).callee, "atom_writes");
+	assert.deepEqual(balanced.errors, []);
+});
+
+test("balanced contexts report unmatched parentheses", () => {
+	const lexical = lex("atom_reads(R_A");
+	const balanced = buildCallContexts(lexical.tokens);
+
+	assert.deepEqual(balanced.errors.map((error) => error.kind), ["unmatched-open-paren"]);
+});

.vscode/tape-atom-syntax/test/source-index.test.js

@@ -0,0 +1,134 @@
+"use strict";
+
+const assert = require("node:assert/strict");
+const test = require("node:test");
+
+const {
+	createIndex,
+	domainFromPath,
+	mergeIndexes,
+	scanSource,
+} = require("../source-index");
+
+test("scanSource discovers current atom and component forms", () => {
+	const source = [
+		"MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4), atom_reads(R_PrimCursor)) { mac_yield() };",
+		"MipsAtomComp_(ac_load_pair) { load_word(R_T0, R_T1, 0) };",
+		"internal MipsAtom* normalize_proc(AtomArena_R aa) MipsAtom_Proc_(aa, { mac_yield() })",
+		"FI_ void ac_store_pair(MipsAtomBuilder_R ab) atom_dbg_skip MipsAtomComp_Proc_(ab, { store_word(R_T0, R_T1, 0) })",
+	].join("\n");
+
+	const result = scanSource(source, "C:/projects/Pikuma/ps1/code/duffle/mips.atom.c");
+
+	assert.equal(result.index.atoms.has("cube_g4_face"), true);
+	assert.equal(result.index.atoms.has("normalize"), true);
+	assert.equal(result.index.components.has("ac_load_pair"), true);
+	assert.equal(result.index.components.has("ac_store_pair"), true);
+	assert.equal(result.index.componentAliases.has("mac_load_pair"), true);
+	assert.equal(result.index.componentAliases.has("mac_store_pair"), true);
+	assert.equal(result.index.macros.get("mac_store_pair"), "component");
+	assert.equal(result.index.componentCallees.get("mac_store_pair").includes("store_word"), true);
+	assert.equal(result.index.phases.has("cube_g4"), true);
+	assert.equal(result.index.registers.get("R_PrimCursor"), "gpr");
+	assert.deepEqual(result.errors, []);
+});
+
+test("scanSource discovers binds, labels, registers, typedefs, and macro domains", () => {
+	const source = [
+		"typedef Struct_(Binds_CubeTri) { U4 PrimCursor; };",
+		"typedef Enum_(U4, PadStatus) { PadStatus_Ok };",
+		"typedef U4 const MipsCode;",
+		"enum { R_PrimCursor = R_T7 atom_reg, C2_Custom = 12, gte_cr_Custom = 13 };",
+		"#define load_word(rt, base, off) enc_i(rt, base, off)",
+		"atom_bind(Binds_CubeTri)",
+		"atom_label(exit)",
+		"atom_offset(entry, exit)",
+	].join("\n");
+
+	const result = scanSource(source, "C:/projects/Pikuma/ps1/code/duffle/mips.h");
+
+	assert.equal(result.index.bindTypes.has("Binds_CubeTri"), true);
+	assert.equal(result.index.types.has("PadStatus"), true);
+	assert.equal(result.index.types.has("MipsCode"), true);
+	assert.equal(result.index.registers.get("R_PrimCursor"), "gpr");
+	assert.equal(result.index.registers.get("C2_Custom"), "cop2");
+	assert.equal(result.index.registers.get("gte_cr_Custom"), "cop2");
+	assert.equal(result.index.macros.get("load_word"), "cpu");
+	assert.equal(result.index.labels.has("entry"), true);
+	assert.equal(result.index.labels.has("exit"), true);
+});
+
+test("domainFromPath uses the declaration file rather than parent directory names", () => {
+	assert.equal(domainFromPath("C:/x/code/hello_gte/hello_gte.atom.c"), null);
+	assert.equal(domainFromPath("C:/x/code/duffle/mips.h"), "cpu");
+	assert.equal(domainFromPath("C:/x/code/duffle/gte.h"), "gte");
+	assert.equal(domainFromPath("C:/x/code/duffle/gp.h"), "gpu");
+});
+
+test("component aliases inherit the domain of the instructions they emit", () => {
+	const headers = mergeIndexes(
+		scanSource("#define load_word(a,b,c) 1\n#define store_word(a,b,c) 1\n#define shift_aright_var(a,b,c) 1\n#define jump_reg(rd) 1\n", "C:/x/code/duffle/mips.h").index,
+		scanSource("#define gte_sw(rt, base, off) 1\n", "C:/x/code/duffle/gte.h").index
+	);
+	const math = scanSource(
+		[
+			"MipsAtomComp_(ac_load_v3s4) { load_word(R_T0, R_T1, 0) };",
+			"#define mac_load_p3s4 mac_load_v3s4",
+		].join("\n"),
+		"C:/x/code/duffle/math.atom.c"
+	);
+	const shift = scanSource(
+		"MipsAtomComp_(ac_shift_aright_var_v3_self) { shift_aright_var(R_T0, R_T0, R_T1) };",
+		"C:/x/code/duffle/gte.atom.c"
+	);
+	const gte = scanSource(
+		"MipsAtomComp_(ac_gte_store_f3) { gte_sw(C2_SXY0, R_T0, 0) };",
+		"C:/x/code/duffle/gte.atom.c"
+	);
+	const yieldAtom = scanSource(
+		"MipsAtomComp_(ac_yield) { load_word(R_AtomJmp, R_TapePtr, 0), jump_reg(R_AtomJmp), nop };",
+		"C:/x/code/duffle/lottes_tape.h"
+	);
+
+	const merged = mergeIndexes(headers, math.index, shift.index, gte.index, yieldAtom.index);
+	assert.equal(merged.macros.get("mac_load_v3s4"), "cpu");
+	assert.equal(merged.macros.get("mac_load_p3s4"), "cpu");
+	assert.equal(merged.macros.get("mac_shift_aright_var_v3_self"), "cpu");
+	assert.equal(merged.macros.get("mac_gte_store_f3"), "gte");
+	assert.equal(merged.macros.get("mac_yield"), "control");
+});
+
+test("scanSource tags utility header defines as utility, not a hardware domain", () => {
+	const source = [
+		"#define assert(cond) ((void)(cond))",
+		"#define stringify(name) #name",
+		"#define u4_hi(imm) ((imm) >> 16)",
+	].join("\n");
+	const result = scanSource(source, "C:/projects/Pikuma/ps1/code/duffle/dsl.h");
+
+	assert.equal(result.index.macros.get("assert"), "utility");
+	assert.equal(result.index.macros.get("stringify"), "utility");
+	assert.equal(result.index.macros.get("u4_hi"), "utility");
+});
+
+test("mergeIndexes prefers a hardware domain over a later utility define", () => {
+	const left = createIndex();
+	left.macros.set("sub_s", "utility");
+	const right = createIndex();
+	right.macros.set("sub_s", "cpu");
+
+	assert.equal(mergeIndexes(left, right).macros.get("sub_s"), "cpu");
+	assert.equal(mergeIndexes(right, left).macros.get("sub_s"), "cpu");
+});
+
+test("mergeIndexes preserves domain-specific aliases", () => {
+	const left = createIndex();
+	left.macros.set("load_word", "cpu");
+	const right = createIndex();
+	right.componentAliases.add("mac_gte_store");
+	right.macros.set("mac_gte_store", "gte");
+
+	const merged = mergeIndexes(left, right);
+	assert.equal(merged.macros.get("load_word"), "cpu");
+	assert.equal(merged.macros.get("mac_gte_store"), "gte");
+});

LICENSE

@@ -1,24 +1,17 @@
-This is free and unencumbered software released into the public domain.
+Copyright (C) 2026 Edward R. Gonzalez
 
-Anyone is free to copy, modify, publish, use, compile, sell, or
-distribute this software, either in source code form or as a compiled
-binary, for any purpose, commercial or non-commercial, and by any
-means.
+This software is provided 'as-is', without any express or implied
+warranty.  In no event will the authors be held liable for any damages
+arising from the use of this software.
 
-In jurisdictions that recognize copyright laws, the author or authors
-of this software dedicate any and all copyright interest in the
-software to the public domain. We make this dedication for the benefit
-of the public at large and to the detriment of our heirs and
-successors. We intend this dedication to be an overt act of
-relinquishment in perpetuity of all present and future rights to this
-software under copyright law.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it
+freely, subject to the following restrictions:
 
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
-EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
-IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
-OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
-ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
-OTHER DEALINGS IN THE SOFTWARE.
-
-For more information, please refer to <https://unlicense.org>
+1. The origin of this software must not be misrepresented; you must not
+   claim that you wrote the original software. If you use this software
+   in a product, an acknowledgment in the product documentation would be
+   appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be
+   misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.

code/duffle/bios.h

@@ -0,0 +1,15 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#endif
+
+enum {
+	bios_init_pad_2  = 0x12,
+	bios_start_pad_2 = 0x13,
+	bios_flushcache  = 0x44,
+	bios_table_addr  = 0xA0,
+	bios_btable_addr = 0xB0,
+};
+
+enum {
+	bios_pad_buffer_size = 0x22,
+};

code/duffle/dsl.atom.h

@@ -1,5 +1,5 @@
 /*
- * atom_dsl.h
+ * dsl.atom.h
  * ============================================================================
  *
  * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
@@ -57,7 +57,6 @@
 
 #ifdef INTELLISENSE_DIRECTIVES
 #pragma once
-// #include <stdint.h>
 #endif
 
 /* ============================================================================
@@ -71,11 +70,31 @@
 /* ----------------------------------------------------------------------------
  * atom_reg (per-enum opt-in marker for the DWARF register-alias registry)
  *
- * The bare `atom_reg` token adjacent to an enum entry in mips.h / lottes_tape.h flags that alias as debug-visible for scan_source's register_alias_registry.
- * The C preprocessor strips it to a comment so no runtime symbol is created; the Lua scanner reads the bare token.
+ * Bare `atom_reg` token adjacent to an enum entry that alias as debug-visible for scan_source's register_alias_registry.
+ * Lua scanner reads the bare token.
  * ----------------------------------------------------------------------------*/
 #define atom_reg /* atom_reg: opt the preceding enum entry into the DWARF registry */
 
+// ----------------------------------------------------------------------------
+// atom_auto_reg(atom, sym) — per-atom auto-allocated GPR binding.
+//   enum {
+//       atom_auto_reg(cube_g4_face, R_Fwdx),                // expands to: R_Fwdx = R_Fwdx_Code /* atom_auto_reg: cube_g4_face */,
+//       atom_auto_reg(cube_g4_face, R_Eye_z) atom_type(S4), // atom_type chains after
+//   };
+// (The macro IS the entire enum entry — no separate LHS=RHS. The `atom` scope is
+//  preserved in a trailing C-comment on the RHS so the Lua scanner can recover
+//  it after preprocessing strips the macro form. R_<Sym>_Code is resolved from gen/auto_reg.h which the .c file #include's before the enum declaration.)
+#define atom_auto_reg(atom, sym) sym = sym ## _Code /* atom_auto_reg: atom */
+
+// ----------------------------------------------------------------------------
+// phase_auto_reg(phase, sym) — per-phase auto-allocated GPR binding.
+//   enum {
+//       phase_auto_reg(cube_g4, R_Temp0),     // expands to: R_Temp0 = R_Temp0_Code /* phase_auto_reg: cube_g4 */,
+//       phase_auto_reg(cube_g4, R_Temp1),
+//   };
+// (Same macro-as-enum-entry form as atom_auto_reg above; the `phase` scope is preserved in a trailing C-comment on the RHS for the Lua scanner to recover.)
+#define phase_auto_reg(phase, sym) sym = sym ## _Code /* phase_auto_reg: phase */
+
 /* ============================================================================
  *   atom_info :
  *      MipsAtom_(cube_tri) atom_info(
@@ -148,12 +167,12 @@
  *   ... body ...
  *   atom_label(bounds_chk)                   ← another anchor
  *
- *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
+ *   atom_offset(culling, bounds_chk)         ← resolved by gen/offsets.h
  *
- * The metaprogram generates gen/atom_offsets.h with one #define with the offset value per atom_offset(F, T) call.
+ * The metaprogram generates gen/offsets.h with one #define with the offset value per atom_offset(F, T) call.
  * The preprocessor then expands the call to the right immediate value.
  *
- * If gen/atom_offsets.h is stale (or atom_label(name) is undefined), `atom_offset_F_T` becomes an undefined macro and the C build fails.
+ * If gen/offsets.h is stale (or atom_label(name) is undefined), `atom_offset_F_T` becomes an undefined macro and the C build fails.
  * ============================================================================*/
 #define atom_offset(F, T) atom_offset_ ## F ## _ ## T
 // atom_label is a pure annotation for the metaprogram's offset calculations.

code/duffle/dsl.h

@@ -3,7 +3,7 @@
 #	include "assert.h"
 #endif
 
-#define offset_of(type, member)  cast(U8,__builtin_offsetof(type,member))
+#define offset_of(type, member)  cast(U8,__builtin_offsetof(type,member)) // Compiler builtin version of O_
 #define static_assert            _Static_assert
 #define typeof                   __typeof__
 #define typeof_ptr(ptr)          typeof((ptr)[0])
@@ -28,8 +28,9 @@
 #define internal static // internal
 
 #define asm           __asm__
-#define align_(value) __attribute__((aligned (value)))             // for easy alignment
 
+#define A_(data)      (& data)
+#define align_(value) __attribute__((aligned (value)))             // for easy alignment
 #define align_(value) __attribute__((aligned (value)))             // for easy alignment
 #define C_(type,data) ((type)(data))                               // for enforced precedence
 #define expect_(x, y) __builtin_expect(x, y)                       // so compiler knows the common path
@@ -43,7 +44,9 @@
 
 #define R_    restrict 
 #define V_    volatile 
-// Fictional, used for intiution.
+
+#pragma region Fictional //, used for intiution
+
 #define EUB_  restrict // Execute Unit Bound:    Data is siloed in the ALU Register File. The Load/Store Unit is bypassed. (Route to Execution Unit.  Keep in registers)
 #define ISO_  restrict // Isolated Provenance:   Alternative to Exu_. Guarantees electrical memory isolation, 
                        //                        unlocking the compiler’s ability to safely pack data across multiple parallel SIMD lanes (vectorization).
@@ -67,7 +70,8 @@
 #define latch_load_anchor(ptr)       //__atomic_load_n(ptr, ooo_anchor_)
 #define latch_store_drain(ptr, val)  //__atomic_store_n(ptr, val, ooo_drain_)
 #define pulse_xchg_weld(ptr, val)    //__atomic_exchange_n(ptr, val, ooo_weld_)
-//end of: Fictional.
+
+#pragma endreigon Fictional
 
 
 // R_ (restrict) establishes an "Eigen" or "Proprius" mapping.
@@ -87,12 +91,13 @@
 #define PtrSet_(type)  TypeR_(type); typedef TypeV_(type)
 #define TSet_(type)    type; typedef PtrSet_(type)
 
-#define array_len(a)                   (U4)(sizeof(a) / sizeof(typeof((a)[0])))
-#define array_decl(type, ...)          (type[]){__VA_ARGS__}
+#define Array_len(a)                   (U4)(sizeof(a) / sizeof(typeof((a)[0])))
+#define Array_decl(type, ...)          (type[]){__VA_ARGS__}
 #define Array_sym(type,len)            A ## len ## _ ## type
 #define Array_expand(type,len)         type Array_sym(type, len)[len]; typedef PtrSet_(Array_sym(type, len))
 #define Array_(type,len)               Array_expand(type,len)
 #define Bit_(id,b)                     id = (1 << b), tmpl(id,pos) = b
+#define Bitmask_(b)                    (1u << b)
 #define Enum_(underlying_type, symbol) underlying_type TSet_(symbol); enum   symbol
 #define Proc_(symbol)                  symbol
 #define Relative_(symbol)              // Does nothing but annotate that a symbol is associated with another.
@@ -130,20 +135,21 @@ typedef __UINT32_TYPE__ TSet_(B4);
 #define u4_v(value) C_(U4 V_*, value)
 enum { false = 0, true  = 1, true_overflow, };
 
-#define u4_lo(value) ((value) & 0xFFFFU)
-#define u4_hi(value) ((value) >> 12)
+#define u4_lo(value) (u4_(value) & 0xFFFFU)
+#define u4_hi(value) (u4_(value) >> (S_(U2) * 8))
 
 typedef void Proc_(VoidFn) (void);
 
-#define kilo(n)         (C_(U4, n) << 10)
-#define mega(n)         (C_(U4, n) << 20)
-#define giga(n)         (C_(U4, n) << 30)
-#define tera(n)         (C_(U4, n) << 40)
+#define Kilo_(n) (C_(U4, n) << 10)
+#define Mega_(n) (C_(U4, n) << 20)
+#define Giga_(n) (C_(U4, n) << 30)
+#define Tera_(n) (C_(U4, n) << 40)
+
 #define null             C_(U4,    0)
 #define nullptr          C_(void*, 0)
-#define O_(type, field) (C_(U4, & C_(type*,0)->field))
- 
-#define OT_(field)       O_(typeof_ptr(& field), filed))
+#define O_(type, field)  C_(U4, & C_(type*,0)->field)
+#define OA_(type, aexpr) C_(U4, & C_(type*,0) aexpr)
+#define OT_(field)       O_(typeof_ptr(& field), field))
 #define S_(data)         C_(U4, sizeof(data))
 
 #define sop_1(op,a,b) C_(U1, s1_(a) op s1_(b))
@@ -164,6 +170,8 @@ def_signed_ops(le, <=)
 #undef def_signed_ops
 #undef def_signed_op
 
+// Unused, we arent' doing any C-like asm since we have the asm dsl. We'll keep the non-generics if we somehow do.
+#if 0
 #define def_generic_sop(op, a, ...) _Generic((a), U1:  op ## _s1, U2: op ## _s2, U4: op ## _s4) (a, __VA_ARGS__)
 #define add_s(a,b) def_generic_sop(add,a,b)
 #define sub_s(a,b) def_generic_sop(sub,a,b)
@@ -173,11 +181,12 @@ def_signed_ops(le, <=)
 #define ge_s(a,b)  def_generic_sop(ge, a,b)
 #define le_s(a,b)  def_generic_sop(le, a,b)
 #undef def_generic_sop
+#endif
 
 #define alignas                             _Alignas
 #define alignof                             _Alignof
 #define byte_pad(amount, ...)               B1 glue(_PAD_, __VA_ARGS__) [amount]
-#define pcast(type, data)                   (C_(type*, & (data)) [0])
+#define C_ptr(type, data)                   (C_(type*, & (data)) [0])
 
 #define dbg_args(...)                      __VA_ARGS__
 
@@ -192,6 +201,8 @@ def_signed_ops(le, <=)
 #define defer_info(type,expr, ...) for(type       info= {__VA_ARGS__}; info.once!=1;++info.once,(expr))    // Defer with tracked state
 
 #define do_while(cond) for (U8 once=0; once!=1 || (cond); ++once)
+
+#define Jmp_nZero_(cond,label) if (cond) goto label;
 #pragma endregion Control Flow & Iteration
 
 #define span_iter(type, iter, m_begin, op, m_end) ( \
@@ -208,16 +219,16 @@ def_signed_ops(le, <=)
 typedef Span_(S4);
 typedef Span_(U4);
 
-#if 0
 #pragma region Debug
-#define debug_trap() __builtin_debugtrap()
+#define debug_trap() __builtin_trap()
 #if BUILD_DEBUG
-IA_ void assert(U8 cond) { if(cond){return;} else{debug_trap(); ms_exit_process(1);} }
+	#define assert(cond) if(cond == false){debug_trap();}
 #else
-#define assert(cond)
+#	ifndef assert
+#		include <assert.h>
+#	endif
 #endif
 #pragma endregion Debug
-#endif
 
 #define GCC_OPTIMIZATION_DISABLE _Pragma("GCC push_options") _Pragma("GCC optimize(\"O0\")")
 #define GCC_OPTIMIZATION_ENABLE  _Pragma("GCC pop_options")

code/duffle/gcc_asm.h

@@ -50,17 +50,13 @@
 #define asm_words(...) m_expand(glue(GCC_ASM_INL_, GCC_ASM_COUNT_ARGS(__VA_ARGS__))(__VA_ARGS__))
 // Very nasty macro expansion. See the Cruft pragma region after all the DSL defines
 
-/* reg_str(n) — Stringify an integer register id into the GCC asm
- * string form (e.g. 12 → "$12"). Use this anywhere GCC's parser
- * expects a literal string identifying a register: clobber lists,
- * asm templates, etc. The two-level macro is the standard preprocessor
- * idiom for forcing one level of expansion before stringify — without
- * it, `#n` would stringify the macro name `R_T4` to `"R_T4"` instead
- * of expanding `R_T4` to its value first.
+/* reg_str(n) — Stringify an integer register id into the GCC asm string form (e.g. 12 → "$12").
+ * Use this anywhere GCC's parser expects a literal string identifying a register: clobber lists,
+ * asm templates, etc. The two-level macro is the standard preprocessor idiom for forcing one level of expansion before stringify —
+ * without it, `#n` would stringify the macro name `R_T4` to `"R_T4"` instead of expanding `R_T4` to its value first.
  *
- * For declaring a register variable bound to a specific GPR, use the
- * `rgcc(n)` bundle from gcc_asm.h instead — it adds the `__asm__()`
- * qualifier around the string.
+ * For declaring a register variable bound to a specific GPR, use the `rgcc(n)` bundle from gcc_asm.h instead —
+ * it adds the `__asm__()` qualifier around the string.
  *
  *   register V3_S2* p0 __asm__(reg_str(R_T4)) = ...;        // verbose
  *   register V3_S2* p0 rgcc(R_T4) = ...;                    // bundled
@@ -85,7 +81,7 @@
  *  - The string "$12" is derived from it via reg_str, so they cannot drift apart.
  *  - Spelling `__asm__(reg_str(R_T4_Code))` at every call site is noise.
  *
- *  tmpl defined in dsl.h (the token-paste glue). 
+ *  tmpl defined in dsl.h (token-paste glue). 
  *  rgcc define here (gcc_asm.h) because the `__asm__` keyword is GCC-specific. 
  *  Anyone porting to a different compiler's asm dialect overrides rgcc,
  *  and the integer→string derivation in rlit can be retargeted in one place.
@@ -94,12 +90,10 @@
  * ------------------------------------------------------------------------ */
 #define rgcc(n) __asm__(rlit(n))
 
-/* rgcc_ref(n) — GCC operand-reference form "%N". Not currently used
- * by the placeholder-pun macros (the .word bodies are fully baked
- * at compile time and have no runtime operand references), but kept
- * here for completeness in case a future asm template needs to refer
- * to a runtime input by position. Mirror of rgcc but produces "%N"
- * instead of "$N". */
+/* rgcc_ref(n) — GCC operand-reference form "%N". Not currently used by the placeholder-pun macros
+ * (the .word bodies are fully baked at compile time and have no runtime operand references),
+ * but kept here for completeness in case a future asm template needs to refer to a runtime input by position.
+ * Mirror of rgcc but produces "%N" instead of "$N". */
 #define rgcc_ref_(n)   "%" #n
 #define rgcc_ref(n)    rgcc_ref_(n)
 
@@ -147,11 +141,9 @@
     9,  8,  7,  6,  5,  4,  3,  2,  1,  0))
 
 /* --- 2. String Concatenation Helpers --- *
- * NOTE: we use `%0`, `%1`, ... not `%c0`, `%c1`, ... because GCC's
- * asm-parser rejects `%cN` in this position with "invalid use of '%c'".
- * The `%cN` form is for printing *character* constants; for arbitrary
- * integer immediates (the only kind `"i"(...)` produces), the plain
- * `%N` form is the right one. Both expand to the bare immediate.
+ * NOTE: we use `%0`, `%1`, ... not `%c0`, `%c1`, ... because GCC's asm-parser rejects `%cN` in this position with "invalid use of '%c'".
+ * The `%cN` form is for printing *character* constants; for arbitrary integer immediates (the only kind `"i"(...)` produces),
+ * the plain `%N` form is the right one. Both expand to the bare immediate.
  */
 #define GCC_ASM_W1                  "%0"
 #define GCC_ASM_W2   GCC_ASM_W1   ", %1"

code/duffle/gen/duffle.macs.h

@@ -1,139 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#pragma once
-#endif
-// Auto-generated by ps1_meta.lua — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
-// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
-
-#ifndef WORD_COUNT
-#define WORD_COUNT(name, count)  enum { words_##name = (count) };
-#endif
-
-/* atom_dbg_skip */
-/* ---------------------------------------------------------------------------
- *  MACRO ATOM Components (Reusable Assembly Components)
- *  These do NOT yield. They are expanded inline inside Tape Atoms.
- * ---------------------------------------------------------------------------*/
-// The 'Yield' sequence for Tape Atoms (mac_yield).
-#define mac_yield(...) \
-	load_word(R_AtomJmp, R_TapePtr, 0) \
-,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
-,	jump_reg( R_AtomJmp) \
-,	nop
-WORD_COUNT(mac_yield, 4)
-
-/* atom_dbg_skip */
-/* Words: 3; Loads 3 S2 indices from the face array */
-#define mac_load_tri_indices(...) \
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)) \
-,	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)) \
-,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
-WORD_COUNT(mac_load_tri_indices, 3)
-
-/* atom_dbg_skip */
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-#define mac_gte_load_tri_verts(...) \
-	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY0) \
-,	gte_mv_to_data_r(R_V1, C2_VZ0) \
-,	shift_lleft(R_AT, R_T1, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY1) \
-,	gte_mv_to_data_r(R_V1, C2_VZ1) \
-,	shift_lleft(R_AT, R_T2, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY2) \
-,	gte_mv_to_data_r(R_V1, C2_VZ2)
-WORD_COUNT(mac_gte_load_tri_verts, 18)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-#define mac_insert_ot_tag_f3(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (5 - 1) << 24 = 4 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_f3, 11)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-#define mac_insert_ot_tag_g4(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (9 - 1) << 24 = 8 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_g4, 11)
-
-/* atom_dbg_skip */
-#define mac_pack_color_word(off, cmd, r, g, b) \
-	load_upper_i(R_AT, (cmd) << 8  | (b)) \
-,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
-,	store_word(  R_AT, R_PrimCursor, (off))
-WORD_COUNT(mac_pack_color_word, 3)
-
-/* atom_dbg_skip */
-#define mac_format_f3_color(r, g, b) \
-	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
-WORD_COUNT(mac_format_f3_color, 3)
-
-/* atom_dbg_skip */
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
- * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-#define mac_gte_store_f3_post_rtpt(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
-WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
-
-#define mac_format_g4_color(r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
-,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
-,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
-,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
-WORD_COUNT(mac_format_g4_color, 12)
-
-/* atom_dbg_skip */
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
- * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
- * three registers aligned with v0/v1/v2 you must store before RTPS).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-#define mac_gte_store_g4_p012_post_rtpt_pre_rtps(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
-WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
-
-/* atom_dbg_skip */
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
- * single-vertex result to SXY2; SXY0 still holds v0.screen from the
- * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
- * prevents).
- */
-#define mac_gte_store_g4_p3_post_rtps(...) \
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
-WORD_COUNT(mac_gte_store_g4_p3_post_rtps, 1)
-

code/duffle/gen/duffle.offsets.h

@@ -1,9 +0,0 @@
-// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
-#pragma once
-
-#pragma region lottes_tape
-
-
-#pragma endregion lottes_tape
-

code/duffle/gen/macs.h

@@ -0,0 +1,408 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\duffle/
+//   source: C:\projects\Pikuma\ps1\code\duffle\word_count.metadata.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\memory.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gcc_asm.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\bios.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.atom.c
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#define WORD_COUNT(name, count)  enum { words_##name = (count) };
+#endif
+
+/* atom_dbg_skip */
+/* ---------------------------------------------------------------------------
+ *  MACRO ATOM Components (Reusable Assembly Components)
+ *  These do NOT yield. They are expanded inline inside Tape Atoms.
+ * ---------------------------------------------------------------------------*/
+// The 'Yield' sequence for Tape Atoms (mac_yield).
+#define mac_yield(...) \
+	load_word(R_AtomJmp, R_TapePtr, 0) \
+	LdSlot_ \
+,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
+,	jump_reg( R_AtomJmp) \
+,	BdSlot_ nop
+WORD_COUNT(mac_yield, 4)
+
+/* atom_dbg_skip */
+#define mac_yield_load(...) \
+	load_word(R_AtomJmp, R_TapePtr, 0)
+WORD_COUNT(mac_yield_load, 1)
+
+/* atom_dbg_skip */
+#define mac_yield_tail(...) \
+	add_ui_self(R_TapePtr, S_(MipsCode)) \
+,	jump_reg( R_AtomJmp) \
+,	BdSlot_ nop
+WORD_COUNT(mac_yield_tail, 3)
+
+/* atom_dbg_skip */
+#define mac_load_half_v3(tx, ty, tz, base, offset) \
+	load_half(tx, base, offset + OA_(U2,[0])) \
+,	load_half(ty, base, offset + OA_(U2,[1])) \
+,	load_half(tz, base, offset + OA_(U2,[2]))
+WORD_COUNT(mac_load_half_v3, 3)
+
+#define mac_load_v3s2(transfer, base, offset) \
+	mac_load_half_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_load_v3s2, 3)
+
+/* atom_dbg_skip */
+#define mac_load_v2s2(rs_x, rs_y, r_base, offset) \
+	load_half(rs_x, r_base, offset + O_(V3_S2,x)) \
+,	load_half(rs_y, r_base, offset + O_(V3_S2,y))
+WORD_COUNT(mac_load_v2s2, 2)
+
+/* atom_dbg_skip */
+#define mac_store_v2s2(rt_x, rt_y, base, offset) \
+	store_half(rt_x, base, offset + O_(V2_S2,x)) \
+,	store_half(rt_y, base, offset + O_(V2_S2,y))
+WORD_COUNT(mac_store_v2s2, 2)
+
+/* atom_dbg_skip */
+#define mac_load_word_v3(tx, ty, tz, base, offset) \
+	load_word(tx, base, offset + OA_(U4,[0])) \
+,	load_word(ty, base, offset + OA_(U4,[1])) \
+,	load_word(tz, base, offset + OA_(U4,[2]))
+WORD_COUNT(mac_load_word_v3, 3)
+
+#define mac_load_v3s4(transfer, base, offset) \
+	mac_load_word_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_load_v3s4, 3)
+
+#define mac_load_p3s4(transfer, base, offset) \
+	mac_load_word_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_load_p3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_store_half_v3(tx, ty, tz, base, offset) \
+	store_half(tx, base, offset + OA_(U2,[0])) \
+,	store_half(ty, base, offset + OA_(U2,[1])) \
+,	store_half(tz, base, offset + OA_(U2,[2]))
+WORD_COUNT(mac_store_half_v3, 3)
+
+#define mac_store_v3s2(transfer, base, offset) \
+	mac_store_half_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_store_v3s2, 3)
+
+/* atom_dbg_skip */
+#define mac_store_word_v3(tx, ty, tz, base, offset) \
+	store_word(tx, base, offset + OA_(U4,[0])) \
+,	store_word(ty, base, offset + OA_(U4,[1])) \
+,	store_word(tz, base, offset + OA_(U4,[2]))
+WORD_COUNT(mac_store_word_v3, 3)
+
+#define mac_store_v3s4(transfer, base, offset) \
+	mac_store_word_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_store_v3s4, 3)
+
+#define mac_store_p3s4(transfer, base, offset) \
+	mac_store_word_v3(transfer.x, transfer.y, transfer.z, base, offset)
+WORD_COUNT(mac_store_p3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_add_si_v3s4(rt_x, rt_y, rt_z, base, offset) \
+	add_si(rt_x, base, O_(V3_S4,x)) \
+,	add_si(rt_y, base, O_(V3_S4,y)) \
+,	add_si(rt_z, base, O_(V3_S4,z))
+WORD_COUNT(mac_add_si_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_sub_s_v3(dx, dy, dz, sx, sy, sz, tx, ty, tz) \
+	sub_s(dx, sx, tx) \
+,	sub_s(dy, sy, ty) \
+,	sub_s(dz, sz, tz)
+WORD_COUNT(mac_sub_s_v3, 3)
+
+#define mac_sub_v3s4(d, s, t) \
+	mac_sub_s_v3(d.x, d.y, d.z, s.x, s.y, s.z, t.x, t.y, t.z)
+WORD_COUNT(mac_sub_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_sub_s_v3_self(ds_x, ds_y, ds_z, tx, ty, tz) \
+	sub_s(ds_x, ds_x, tx) \
+,	sub_s(ds_y, ds_y, ty) \
+,	sub_s(ds_z, ds_z, tz)
+WORD_COUNT(mac_sub_s_v3_self, 3)
+
+#define mac_sub_v3s4_self(ds, t) \
+	mac_sub_s_v3_self(ds.x, ds.y, ds.z, t.x, t.y, t.z)
+WORD_COUNT(mac_sub_v3s4_self, 3)
+
+/* atom_dbg_skip */
+#define mac_store_rects2(rt_x, rt_y, rt_width, rt_height, base, offset) \
+	store_half(rt_x,      base, offset + O_(Rect_S2,x)) \
+,	store_half(rt_y,      base, offset + O_(Rect_S2,y)) \
+,	store_half(rt_width,  base, offset + O_(Rect_S2,width)) \
+,	store_half(rt_height, base, offset + O_(Rect_S2,height))
+WORD_COUNT(mac_store_rects2, 4)
+
+/* atom_dbg_skip */
+#define mac_load_word_imm(dst, imm) \
+	load_upper_i(dst, u4_hi(imm)) \
+,	or_i_self(   dst, u4_lo(imm))
+WORD_COUNT(mac_load_word_imm, 2)
+
+#define mac_shift_aright_v3_self(dt_x, dt_y, dt_z, shift_amount) \
+	shift_aright(dt_x, dt_x, shift_amount) \
+,	shift_aright(dt_y, dt_y, shift_amount) \
+,	shift_aright(dt_z, dt_z, shift_amount)
+WORD_COUNT(mac_shift_aright_v3_self, 3)
+
+#define mac_shift_aright_v3s4_self(dt, shift) \
+	mac_shift_aright_v3_self(dt.x, dt.y, dt.z, shift)
+WORD_COUNT(mac_shift_aright_v3s4_self, 3)
+
+#define mac_shift_aright_var_v3(rd_v0, rd_v1, rd_v2, rs_v0, rs_v1, rs_v2, r_shift) \
+	shift_aright_var(rd_v0, rs_v0, r_shift) \
+,	shift_aright_var(rd_v1, rs_v1, r_shift) \
+,	shift_aright_var(rd_v2, rs_v2, r_shift)
+WORD_COUNT(mac_shift_aright_var_v3, 3)
+
+/* atom_dbg_skip */
+#define mac_shift_aright_var_v3_self(rds_v0, rds_v1, rds_v2, r_shift) \
+	shift_aright_var(rds_v0, rds_v0, r_shift) \
+,	shift_aright_var(rds_v1, rds_v1, r_shift) \
+,	shift_aright_var(rds_v2, rds_v2, r_shift)
+WORD_COUNT(mac_shift_aright_var_v3_self, 3)
+
+#define mac_shift_aright_var_v3s4_self(ds, shift) \
+	mac_shift_aright_var_v3_self(ds.x, ds.y, ds.z, shift)
+WORD_COUNT(mac_shift_aright_var_v3s4_self, 3)
+
+/* atom_dbg_skip */
+#define mac_load_tri_indices(r_face_cusor, r_i0, r_i1, r_i2) \
+	load_half_u(r_i0, r_face_cusor, 0 * S_(S2)) \
+,	load_half_u(r_i1, r_face_cusor, 1 * S_(S2)) \
+,	load_half_u(r_i2, r_face_cusor, 2 * S_(S2))
+WORD_COUNT(mac_load_tri_indices, 3)
+
+#define mac_gte_mv_to_cr_diag_v3s4(v) \
+	gte_mv_to_ctrl_r(v.y, gte_cr_RT13) \
+,	gte_mv_to_ctrl_r(v.z, gte_cr_RT22) \
+,	gte_mv_to_ctrl_r(v.x, gte_cr_RT11)
+WORD_COUNT(mac_gte_mv_to_cr_diag_v3s4, 3)
+
+#define mac_gte_ld_ir123_v3s4(v) \
+	gte_mv_to_data_r(v.x, C2_IR1) \
+,	gte_mv_to_data_r(v.y, C2_IR2) \
+,	gte_mv_to_data_r(v.z, C2_IR3)
+WORD_COUNT(mac_gte_ld_ir123_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_op_cross_v3s4(a, b) \
+	mac_gte_mv_to_cr_diag_v3s4(a) \
+	GteDelay_  /* RT diagonal: D1 = a.x, D2 = a.y, D3 = a.z */ \
+,	mac_gte_ld_ir123_v3s4(b) \
+	GteDelay_  /* IR: second operand (b.xyz) */ \
+,	gte_cmdw_cross                            /* OP: MAC1/2/3 = a × b (S12.20) */ \
+,	mac_gte_mv_from_mac123_v3s4(a) \
+	GteDelay_  /* Read MAC1/2/3 → a.xyz (overwrites source-A's load targets) */ \
+,	mac_shift_aright_v3s4_self(a, 12) /* Right-shift MAC by 12 (S12.20 → S12.0 OuterProduct12) */
+WORD_COUNT(mac_gte_op_cross_v3s4, 13)
+
+/* atom_dbg_skip */
+#define mac_gte_store_f3(r_primitive_cursor) \
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_F3,p0)) \
+,	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_F3,p1)) \
+,	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_F3,p2))
+WORD_COUNT(mac_gte_store_f3, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_load_tri_verts(r_vert_base, r_v0, r_v1, r_v2) \
+	shift_lleft(R_AT, r_v0, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY0) \
+,	gte_mv_to_data_r(R_V1, C2_VZ0) \
+,	shift_lleft(R_AT, r_v1, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY1) \
+,	gte_mv_to_data_r(R_V1, C2_VZ1) \
+,	shift_lleft(R_AT, r_v2, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY2) \
+,	gte_mv_to_data_r(R_V1, C2_VZ2)
+WORD_COUNT(mac_gte_load_tri_verts, 18)
+
+/* atom_dbg_skip */
+#define mac_gte_store_g4_p012(r_primitive_cursor) \
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_G4,p0)) \
+,	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_G4,p1)) \
+,	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p2))
+WORD_COUNT(mac_gte_store_g4_p012, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_store_g4_p3(r_primitive_cursor) \
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3))
+WORD_COUNT(mac_gte_store_g4_p3, 1)
+
+/* atom_dbg_skip */
+#define mac_gte_sqr_v3(r_sx, r_sy, r_sz, r_sq_x, r_sq_y, r_sq_z) \
+	mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, nop) \
+,	gte_mv_from_data_r(r_sq_x, C2_MAC1) \
+,	gte_mv_from_data_r(r_sq_y, C2_MAC2) \
+,	gte_mv_from_data_r(r_sq_z, C2_MAC3)
+WORD_COUNT(mac_gte_sqr_v3, 8)
+
+/* atom_dbg_skip */
+#define mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, delay_slot) \
+	gte_mv_to_data_r(r_sx, C2_IR1) \
+,	gte_mv_to_data_r(r_sy, C2_IR2) \
+,	gte_mv_to_data_r(r_sz, C2_IR3) \
+,	delay_slot \
+,	gte_cmdw_sqr
+WORD_COUNT(mac_gte_sqr_v3s4, 5)
+
+/* atom_dbg_skip */
+#define mac_gte_gpf_scale(r_sx, r_sy, r_sz, r_recip_est, r_shift, r_dx, r_dy, r_dz) \
+	gte_mv_to_data_r(r_recip_est, C2_IR0) \
+,	gte_mv_to_data_r(r_sx,        C2_IR1) \
+,	gte_mv_to_data_r(r_sy,        C2_IR2) \
+,	gte_mv_to_data_r(r_sz,        C2_IR3) \
+,	GteDelay_ nop2 /* retire IR0..IR3 → GPF input pre-fill (matches libgte 0x80016134..0x80016138) */ \
+,	gte_cmdw_gpf \
+,	gte_mv_from_data_r(r_dx, C2_MAC1) \
+,	gte_mv_from_data_r(r_dy, C2_MAC2) \
+,	gte_mv_from_data_r(r_dz, C2_MAC3) \
+,	shift_aright_var(r_dx, r_dx, r_shift) \
+,	shift_aright_var(r_dy, r_dy, r_shift) \
+,	shift_aright_var(r_dz, r_dz, r_shift)
+WORD_COUNT(mac_gte_gpf_scale, 13)
+
+#define mac_trans_mt3s3s4(r_mtx, r_off, r_t0, r_t1, r_t2) \
+	load_word( r_t0, r_off, O_(V3_S4,x)) \
+,	load_word( r_t1, r_off, O_(V3_S4,y)) \
+,	load_word( r_t2, r_off, O_(V3_S4,z)) \
+,	store_word(r_t0, r_mtx, O_(MT3_S2S4,t[0])) \
+,	store_word(r_t1, r_mtx, O_(MT3_S2S4,t[1])) \
+,	store_word(r_t2, r_mtx, O_(MT3_S2S4,t[2]))
+WORD_COUNT(mac_trans_mt3s3s4, 6)
+
+/* atom_dbg_skip */
+#define mac_lzcr_round_even_half_shift(r_shift, r_mag_sq, r_mag_sq_copy) \
+	and_i(r_shift, r_shift, gte_lzcr_even_mask) \
+,	or_u(r_mag_sq_copy, r_mag_sq, 0) \
+,	li_s(        r_mag_sq, 31) \
+,	sub_s(       r_mag_sq, r_mag_sq, r_shift) \
+,	shift_aright(r_mag_sq, r_mag_sq, 1)
+WORD_COUNT(mac_lzcr_round_even_half_shift, 5)
+
+#define mac_gte_general_purpose_interopolation(to_ir0, to_ir1, to_ir2, to_ir3, fr_mac1, fr_mac2, fr_mac3, nop_slot1, nop_slot2) \
+	gte_mv_to_data_r(to_ir0, C2_IR0) \
+,	gte_mv_to_data_r(to_ir1, C2_IR1) /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */ \
+,	gte_mv_to_data_r(to_ir2, C2_IR2) \
+,	gte_mv_to_data_r(to_ir3, C2_IR3) /* IR3 = src.z (reloaded) */ \
+,	GteDelay_ nop_slot1 \
+,	GteDelay_ nop_slot2 \
+,	gte_cmdw_gpf \
+,	gte_mv_from_data_r(fr_mac1, C2_MAC1) \
+,	gte_mv_from_data_r(fr_mac2, C2_MAC2) \
+,	gte_mv_from_data_r(fr_mac3, C2_MAC3)
+WORD_COUNT(mac_gte_general_purpose_interopolation, 10)
+
+#define mac_gte_mv_from_data_r_mac123(fr_mac1, fr_mac2, fr_mac3) \
+	gte_mv_from_data_r(fr_mac1, C2_MAC1) \
+,	gte_mv_from_data_r(fr_mac2, C2_MAC2) \
+,	gte_mv_from_data_r(fr_mac3, C2_MAC3)
+WORD_COUNT(mac_gte_mv_from_data_r_mac123, 3)
+
+#define mac_gte_mv_from_mac123_v3s4(v) \
+	mac_gte_mv_from_data_r_mac123(v.x, v.y, v.z)
+WORD_COUNT(mac_gte_mv_from_mac123_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
+	mac_load_word_imm(reg_transfer, cmd) \
+,	store_word(       reg_transfer, reg_base, port)
+WORD_COUNT(mac_gcmd_push, 3)
+
+/* atom_dbg_skip */
+#define mac_store_rgb8(rr, rg, rb, base, offset) \
+	store_byte(rr, base, offset + O_(RGB8,r)) \
+,	store_byte(rg, base, offset + O_(RGB8,g)) \
+,	store_byte(rb, base, offset + O_(RGB8,b))
+WORD_COUNT(mac_store_rgb8, 3)
+
+/* atom_dbg_skip */
+#define mac_pack_color_word(r_base, off, cmd, r, g, b) \
+	load_upper_i(R_AT, (cmd) << 8  | (b)) \
+,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
+,	store_word(  R_AT, r_base, (off))
+WORD_COUNT(mac_pack_color_word, 3)
+
+/* atom_dbg_skip */
+#define mac_format_f3_color(r_base, r, g, b) \
+	mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
+WORD_COUNT(mac_format_f3_color, 3)
+
+/* atom_dbg_skip */
+#define mac_format_g4_color(r_prim_cursor, r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c1), 0,               r1,g1,b1) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c2), 0,               r2,g2,b2) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c3), 0,               r3,g3,b3)
+WORD_COUNT(mac_format_g4_color, 12)
+
+#define mac_insert_ot_tag(r_ot_base, r_prim_cursor, poly_size) \
+	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
+,	add_u_self(  R_T1, r_ot_base)                             /* T1 = & OrderingTable[OTZ] */ \
+,	load_word(   R_AT, R_T1,          O_(PolyTag,code))       /* AT = old_ot_head */ \
+,	load_upper_i(R_V0, (poly_size/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) \
+,	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits))   /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
+,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
+,	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code))       /* prim->tag = packed(prim_length, old_addr) */ \
+,	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits))   /* AT = (prim_length << 24) | old_addr */ \
+,	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)) \
+,	store_word(  R_AT, R_T1,          O_(PolyTag,code))       /* OrderingTable[OTZ] = PrimCursor */
+WORD_COUNT(mac_insert_ot_tag, 11)
+
+/* atom_dbg_skip */
+#define mac_pad_set_centered_axes(state, scratch) \
+	load_upper_i(scratch, (PadAxis_Centered >> 16) & 0xFFFF) \
+,	or_i_self(   scratch,  PadAxis_Centered        & 0xFFFF) /* mac_load_word_imm(scratch, PadAxis_Centered), */ \
+,	store_word(  scratch, state, O_(PadState,axes))
+WORD_COUNT(mac_pad_set_centered_axes, 3)
+
+/* atom_dbg_skip */
+#define mac_pad_set_id_byte(state, r_id, id_value) \
+	add_ui(    r_id, R_0, id_value) \
+,	store_byte(r_id, state, O_(PadState,id))
+WORD_COUNT(mac_pad_set_id_byte, 2)
+
+/* atom_dbg_skip */
+#define mac_pad_set_status(r_tmp, r_state, pad_status) \
+	add_ui(    r_tmp, R_0, pad_status) \
+,	store_word(r_tmp, r_state, O_(PadState,status))
+WORD_COUNT(mac_pad_set_status, 2)
+
+/* atom_dbg_skip */
+#define mac_pad_store_inverted_buttons(r_buttons, r_pad_state) \
+	nor_u(     r_buttons, r_buttons, R_0) \
+,	store_half(r_buttons, r_pad_state, O_(PadState,buttons))
+WORD_COUNT(mac_pad_store_inverted_buttons, 2)
+

code/duffle/gen/offsets.h

@@ -0,0 +1,73 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\duffle\
+//   source: C:\projects\Pikuma\ps1\code\duffle\word_count.metadata.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\memory.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gcc_asm.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\bios.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.atom.c
+#pragma once
+
+#pragma region duffle
+
+
+// --- atom: example_atom_proc (10 words) ---
+
+#define _atom_offset_example_atom_proc_skip           2
+
+enum {
+    atom_offset_example_atom_proc_skip = _atom_offset_example_atom_proc_skip,
+};
+
+// --- atom: build_normalize_v3s4 (63 words) ---
+
+#define _atom_offset_aligned_done_srav_path           3
+#define _atom_offset_srav_path_aligned_done           4
+
+enum {
+    atom_offset_aligned_done_srav_path = _atom_offset_aligned_done_srav_path,
+    atom_offset_srav_path_aligned_done = _atom_offset_srav_path_aligned_done,
+};
+
+// --- atom: pad_bios_snapshot (84 words) ---
+
+#define _atom_offset_snap_root_skip_disconnected      10
+#define _atom_offset_disconnected_snap_end            65
+#define _atom_offset_case_2_id_dispatch               9
+#define _atom_offset_pending_snap_end                 54
+#define _atom_offset_id_dispatch_try_analog_stick     12
+#define _atom_offset_id_dispatch_snap_end             40
+#define _atom_offset_try_analog_stick_try_analog_pad  13
+#define _atom_offset_analog_stick_snap_end            25
+#define _atom_offset_try_analog_pad_try_unsupported   12
+#define _atom_offset_analog_pad_snap_end              10
+
+enum {
+    atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
+    atom_offset_disconnected_snap_end = _atom_offset_disconnected_snap_end,
+    atom_offset_case_2_id_dispatch = _atom_offset_case_2_id_dispatch,
+    atom_offset_pending_snap_end = _atom_offset_pending_snap_end,
+    atom_offset_id_dispatch_try_analog_stick = _atom_offset_id_dispatch_try_analog_stick,
+    atom_offset_id_dispatch_snap_end = _atom_offset_id_dispatch_snap_end,
+    atom_offset_try_analog_stick_try_analog_pad = _atom_offset_try_analog_stick_try_analog_pad,
+    atom_offset_analog_stick_snap_end = _atom_offset_analog_stick_snap_end,
+    atom_offset_try_analog_pad_try_unsupported = _atom_offset_try_analog_pad_try_unsupported,
+    atom_offset_analog_pad_snap_end = _atom_offset_analog_pad_snap_end,
+};
+
+#pragma endregion duffle
+

code/duffle/gp.atom.c

@@ -0,0 +1,61 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "dsl.h"
+#	include "gp.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(gp_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_gcmd_push(AtomBuilder_R ab, U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_load_word_imm(reg_transfer, cmd),
+	store_word(       reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_store_rgb8(AtomBuilder_R ab, U1 rr, U1 rg, U1 rb, U4 base, U4 offset) 
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_byte(rr, base, offset + O_(RGB8,r)),
+	store_byte(rg, base, offset + O_(RGB8,g)),
+	store_byte(rb, base, offset + O_(RGB8,b)),
+})
+
+FI_ Slice_MipsCode ac_pack_color_word(AtomBuilder_R ab, U4 r_base, U4 off, U4 cmd, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_upper_i(R_AT, (cmd) << 8  | (b)),
+	or_i_self(   R_AT, ((g)  << 8) | (r)),
+	store_word(  R_AT, r_base, (off)),
+})
+
+FI_ Slice_MipsCode ac_format_f3_color(AtomBuilder_R ab, U4 r_base, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ab, { mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+
+FI_ Slice_MipsCode ac_format_g4_color(AtomBuilder_R ab, U4 r_prim_cursor,
+	U1 r0, U1 g0, U1 b0,
+  U1 r1, U1 g1, U1 b1,
+  U1 r2, U1 g2, U1 b2,
+  U1 r3, U1 g3, U1 b3)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c1), 0,               r1,g1,b1),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c2), 0,               r2,g2,b2),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c3), 0,               r3,g3,b3),
+})
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list. */
+// TODO(Ed): Expose R_T1 as a r_t0, r_V0 as r_t2
+I_ Slice_MipsCode ac_insert_ot_tag(AtomBuilder_R ab, Reg r_ot_base, Reg r_prim_cursor, U2 poly_size) MipsAtomComp_Proc_(ab, {
+	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
+	add_u_self(  R_T1, r_ot_base),                             // T1 = & OrderingTable[OTZ]
+	load_word(   R_AT, R_T1,          O_(PolyTag,code)),       // AT = old_ot_head
+	load_upper_i(R_V0, (poly_size/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
+	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits)),   // Strip upper 8 bits (length from prev cell) → keep only low 24
+	or_u(        R_AT, R_AT, R_V0),                            // Merge length
+	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code)),       // prim->tag = packed(prim_length, old_addr)
+	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits)),   // AT = (prim_length << 24) | old_addr
+	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)),
+	store_word(  R_AT, R_T1,          O_(PolyTag,code)),       // OrderingTable[OTZ] = PrimCursor
+})
+
+#pragma endregion MACs (Mips Atom Components)

code/duffle/gp.h

@@ -21,7 +21,7 @@
  *    4. Semantic encoders      gp0_word_poly_f3(r,g,b)
  *    3. Composite encoders     enc_color_word(cmd, r, g, b)
  *    2. Per-field encoders     enc_gp0_color_r(r), enc_gp0_color_g(g), ...
- *    1. Bitfield layout consts gp0_color_red_shift = 0, gp0_color_red_mask = 0xFF
+ *    1. Bitfield layout consts gp0_color_red_shift = 0, gp0_color_red_width = 8
  *    0. Opcode IDs             gp0_cmd_poly_f3 = 0x20
  *
  *  Vendor mnemonics (gte_mtc2, gte_mfc2, etc.) are NOT in this header.
@@ -39,8 +39,8 @@
 /* ============================================================================
  *  Hardware MMIO Addresses
  *  ============================================================================
- *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2
- *  0x1F800000+. GP0 (offset 0x10) is the data port (commands + params).
+ *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2 0x1F800000+.
+ *  GP0 (offset 0x10) is the data port (commands + params).
  *  GP1 (offset 0x14) is the control port (status, ctrl writes).
  * ============================================================================ */
 /* IO base address (KSEG2 0x1F800000+ for the I/O register region).
@@ -49,18 +49,18 @@
  *   `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
  * Mirrors the `IO_BASE_ADDR equ 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s. */
 enum {
-    IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
-    IO_BASE_ADDR_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
+	IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
+	IO_BASE_ADDR_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
 
-    /* Offsets from IO_BASE_ADDR to each port. Used by tape-side macros
-     * that pin a register to IO_BASE_ADDR and access ports via offsets:
-     *   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
-     *   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
-    GPIO_PORT0_OFFSET = 0x1810,
-    GPIO_PORT1_OFFSET = 0x1814,
+	/* Offsets from IO_BASE_ADDR to each port. Used by tape-side macros
+		* that pin a register to IO_BASE_ADDR and access ports via offsets:
+		*   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
+		*   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
+	GPIO_PORT0_OFFSET = 0x1810,
+	GPIO_PORT1_OFFSET = 0x1814,
 
-    HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
-    HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
+	HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
+	HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
 };
 
 #define HW_GP0 C_(U4 V_*, HW_GP0_ADDR)
@@ -68,71 +68,69 @@ enum {
 
 #define gp0_send(word) (HW_GP0[0] = (word))
 #define gp1_send(word) (HW_GP1[0] = (word))
+#define DmaSlot_ // Annotate an instruction as filling a CPU <-> Command DMA delay slot/s
 
 /* ============================================================================
  *  GP0 command byte constants + Layer 1 (GPU bitfield shifts)
  *  ============================================================================
  *  8-bit GP0 opcodes (the upper byte of a primitive's first word). These are the BYTE only.
- *  The layer-1 bitfield-layout constants live in the same enum block so the encoder can reference them by name. 
  *  NO macro body past this point uses a raw shift or raw mask.
- *  Every shift/width/mask is named here, named once.
- *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
+ *  Mirrors the OPCODE_SHIFT / RS_SHIFT convention from mips.h.
  * ============================================================================ */
 enum {
-    gp0_cmd_Nop             = 0x00,
+	gp0_cmd_Nop             = 0x00,
 
-    /* Cache management */
-    gp0_cmd_ClearCache      = 0x01,
-    gp0_cmd_FillVram        = 0x02,
-    gp0_cmd_CopyVram        = 0x80,
-    gp0_cmd_CopyVramChained = 0x81,
-    gp0_cmd_ReadVram        = 0xC0,
+	/* Cache management */
+	gp0_cmd_ClearCache      = 0x01,
+	gp0_cmd_FillVram        = 0x02,
+	gp0_cmd_CopyVram        = 0x80,
+	gp0_cmd_CopyVramChained = 0x81,
+	gp0_cmd_ReadVram        = 0xC0,
 
-    /* Polygons */
-    gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
-    gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
-    gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
-    gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
-    gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
-    gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
-    gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
-    gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
+	/* Polygons */
+	gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
+	gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
+	gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
+	gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
+	gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
+	gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
+	gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
+	gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
 
-    /* Lines */
-    gp0_cmd_line_f2         = 0x40,
-    gp0_cmd_line_g2         = 0x50,
+	/* Lines */
+	gp0_cmd_line_f2         = 0x40,
+	gp0_cmd_line_g2         = 0x50,
 
-    /* Sprites + Tiles + Rects */
-    gp0_cmd_sprt_1          = 0x64,
-    gp0_cmd_sprt_8          = 0x74,
-    gp0_cmd_sprt_16         = 0x7C,
-    gp0_cmd_tile_1          = 0x60,
-    gp0_cmd_tile_8          = 0x68,
-    gp0_cmd_tile_16         = 0x70,
+	/* Sprites + Tiles + Rects */
+	gp0_cmd_sprt_1          = 0x64,
+	gp0_cmd_sprt_8          = 0x74,
+	gp0_cmd_sprt_16         = 0x7C,
+	gp0_cmd_tile_1          = 0x60,
+	gp0_cmd_tile_8          = 0x68,
+	gp0_cmd_tile_16         = 0x70,
 
-    /* State setters (not drawing primitives; set render context). */
-    gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
-    gp0_cmd_SetTextureWindow     = 0xE2,
-    gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
-    gp0_cmd_SetDrawArea_BotRight = 0xE4,
-    gp0_cmd_SetDrawOffset        = 0xE5,
-    gp0_cmd_SetMaskBit           = 0xE6,
+	/* State setters (not drawing primitives; set render context). */
+	gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
+	gp0_cmd_SetTextureWindow     = 0xE2,
+	gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
+	gp0_cmd_SetDrawArea_BotRight = 0xE4,
+	gp0_cmd_SetDrawOffset        = 0xE5,
+	gp0_cmd_SetMaskBit           = 0xE6,
 
-    /* bitfield shifts / widths / masks ----
-     * Generic GP0/GP1 command byte (upper 8 bits of every word sent to either port). */
-    gp0_cmd_shift = 24,
-    gp0_cmd_width =  8,
-    gp0_cmd_mask  = 0xFF,
+	/* bitfield shifts / widths ----
+		* Generic GP0/GP1 command byte (upper 8 bits of every word sent to either port). */
+	gp0_cmd_shift = 24,
+	gp0_cmd_width =  8,
 
-    /* Color word layout (lives in Poly_F3.color, Poly_G4.c0..c3, etc.):
-     *   bits 31..24 = command byte
-     *   bits 23..16 = BLUE
-     *   bits 15..08 = GREEN
-     *   bits 07..00 = RED     (PSX GPU is BGR, NOT RGB) */
-    gp0_color_cmd_shift   = 24, gp0_color_cmd_width   = 8, gp0_color_cmd_mask   = 0xFF,
-    gp0_color_blue_shift  = 16, gp0_color_blue_width  = 8, gp0_color_blue_mask  = 0xFF,
-    gp0_color_green_shift =  8, gp0_color_green_width = 8, gp0_color_green_mask = 0xFF,
-    gp0_color_red_shift   =  0, gp0_color_red_width   = 8, gp0_color_red_mask   = 0xFF,
+	/* Color word layout (lives in Poly_F3.color, Poly_G4.c0..c3, etc.):
+		*   bits 31..24 = command byte
+		*   bits 23..16 = BLUE
+		*   bits 15..08 = GREEN
+		*   bits 07..00 = RED     (PSX GPU is BGR, NOT RGB) */
+	gp0_color_cmd_shift   = 24, gp0_color_cmd_width   = 8,
+	gp0_color_blue_shift  = 16, gp0_color_blue_width  = 8,
+	gp0_color_green_shift =  8, gp0_color_green_width = 8,
+	gp0_color_red_shift   =  0, gp0_color_red_width   = 8,
 };
 
 /* ============================================================================
@@ -145,12 +143,12 @@ enum {
  * ============================================================================ */
 
 /* ---- Layer 1.5: per-field encoders ---- */
-#define enc_gp0_cmd(cmd)       (((cmd) & gp0_cmd_mask)         << gp0_cmd_shift)
+#define enc_gp0_cmd(cmd)       ((cmd) << gp0_cmd_shift)
 
-#define enc_gp0_color_cmd(cmd) (((cmd) & gp0_color_cmd_mask)   << gp0_color_cmd_shift)
-#define enc_gp0_color_r(r)     (((r)   & gp0_color_red_mask)   << gp0_color_red_shift)
-#define enc_gp0_color_g(g)     (((g)   & gp0_color_green_mask) << gp0_color_green_shift)
-#define enc_gp0_color_b(b)     (((b)   & gp0_color_blue_mask)  << gp0_color_blue_shift)
+#define enc_gp0_color_cmd(cmd) ((cmd) << gp0_color_cmd_shift)
+#define enc_gp0_color_r(r)     ((r)   << gp0_color_red_shift)
+#define enc_gp0_color_g(g)     ((g)   << gp0_color_green_shift)
+#define enc_gp0_color_b(b)     ((b)   << gp0_color_blue_shift)
 
 /* ---- Layer 2: composite encoders ---- */
 #define enc_color_word(cmd, r, g, b) (enc_gp0_color_cmd(cmd) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
@@ -171,10 +169,13 @@ enum {
 #define gp0_word_poly_gt4(r,g,b) enc_color_word(gp0_cmd_poly_gt4, (r),(g),(b))
 
 /* Cache management — bare-cmd words (no color/range payload). */
-#define gp0_word_clear_cache() enc_gp0_cmd_word(gp0_cmd_ClearCache)
-#define gp0_word_fill_vram()   enc_gp0_cmd_word(gp0_cmd_FillVram)
-#define gp0_word_copy_vram()   enc_gp0_cmd_word(gp0_cmd_CopyVram)
-#define gp0_word_read_vram()   enc_gp0_cmd_word(gp0_cmd_ReadVram)
+#define gp0_word_clear_cache()       enc_gp0_cmd_word(gp0_cmd_ClearCache)
+#define gp0_word_fill_vram()         enc_gp0_cmd_word(gp0_cmd_FillVram)
+#define gp0_word_copy_vram()         enc_gp0_cmd_word(gp0_cmd_CopyVram)
+#define gp0_word_read_vram()         enc_gp0_cmd_word(gp0_cmd_ReadVram)
+
+/* NOP — bare-cmd word (no effect; used as DR_ENV padding). */
+#define gp0_word_nop()               enc_gp0_cmd_word(gp0_cmd_Nop)
 
 /* ============================================================================
  *  GP1 command byte constants + Layer 1 (display-mode + range + draw-area bitfield shifts)
@@ -184,65 +185,64 @@ enum {
  *  (cmd byte in the upper 8 bits via `enc_gp0_cmd(cmd)`).
  * ============================================================================ */
 enum {
-    gp1_cmd_Reset                  = 0x00,
-    gp1_cmd_ResetCmdBuffer         = 0x01,
-    gp1_cmd_AcknowledgeIRQ         = 0x02,
-    gp1_cmd_DisplayEnable          = 0x03,
-    gp1_cmd_DMADirection           = 0x04,
-    gp1_cmd_StartDisplayArea       = 0x05,
-    gp1_cmd_HorizontalDisplayRange = 0x06,
-    gp1_cmd_VerticalDisplayRange   = 0x07,
-    gp1_cmd_DisplayMode            = 0x08,
-    /* Note: GP1 only has commands 0x00..0x08. 
-		 * The state-setter commands (SetTextureWindow, * SetDrawArea*, SetDrawOffset, SetMaskBit) 
-		 * live in the GP0 enum as * 0xE1..0xE6.
-		 * DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
+	gp1_cmd_Reset                  = 0x00,
+	gp1_cmd_ResetCmdBuffer         = 0x01,
+	gp1_cmd_AcknowledgeIRQ         = 0x02,
+	gp1_cmd_DisplayEnable          = 0x03,
+	gp1_cmd_DMADirection           = 0x04,
+	gp1_cmd_StartDisplayArea       = 0x05,
+	gp1_cmd_HorizontalDisplayRange = 0x06,
+	gp1_cmd_VerticalDisplayRange   = 0x07,
+	gp1_cmd_DisplayMode            = 0x08,
+	/* Note: GP1 only has commands 0x00..0x08. 
+		* The state-setter commands (SetTextureWindow, * SetDrawArea*, SetDrawOffset, SetMaskBit) 
+		* live in the GP0 enum as * 0xE1..0xE6.
+		* DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
 
-    /* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
-     * Bit positions match the encoder shifts below; values are the
-     * *payload* bits only (the cmd byte is OR'd in by enc_gp1_disp_mode_word). */
-    gp1_disp_HRes_256   = 0x0,
-    gp1_disp_HRes_320   = 0x1,
-    gp1_disp_HRes_512   = 0x2,
-    gp1_disp_HRes_640   = 0x3,
-    gp1_disp_VRes_240   = 0x0,
-    gp1_disp_VRes_480   = 0x1,
-    gp1_disp_Color15    = 0x0,
-    gp1_disp_Color24    = 0x1,
-    gp1_disp_VInterlace = 0x1,
+	/* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
+		* Bit positions match the encoder shifts below; values are the *payload* bits only (cmd byte is OR'd in by enc_gp1_disp_mode_word). */
+	gp1_disp_HRes_256   = 0x0,
+	gp1_disp_HRes_320   = 0x1,
+	gp1_disp_HRes_512   = 0x2,
+	gp1_disp_HRes_640   = 0x3,
+	gp1_disp_VRes_240   = 0x0,
+	gp1_disp_VRes_480   = 0x1,
+	gp1_disp_Color15    = 0x0,
+	gp1_disp_Color24    = 0x1,
+	gp1_disp_VInterlace = 0x1,
 
-    /* ---- Layer 1: GP1 display-mode + range + draw-area shifts/masks ---- */
-    gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2, gp1_disp_hres_mask      = 0x3,
-    gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1, gp1_disp_vres_mask      = 0x1,
-    gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1, gp1_disp_color_mask     = 0x1,
-    gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1, gp1_disp_interlace_mask = 0x1,
+	/* ---- Layer 1: GP1 display-mode + range + draw-area shifts/widths ---- */
+	gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2,
+	gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1,
+	gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1,
+	gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1,
 
-    /* GP1 horizontal display range: bits 0..11 = X2, bits 12..23 = X1 */
-    gp1_hrange_x1_shift = 12, gp1_hrange_x1_width = 12, gp1_hrange_x1_mask = 0xFFF,
-    gp1_hrange_x2_shift =  0, gp1_hrange_x2_width = 12, gp1_hrange_x2_mask = 0xFFF,
+	/* GP1 horizontal display range: bits 0..11 = X2, bits 12..23 = X1 */
+	gp1_hrange_x1_shift = 12, gp1_hrange_x1_width = 12,
+	gp1_hrange_x2_shift =  0, gp1_hrange_x2_width = 12,
 
-    /* GP1 vertical display range: bits 0..9 = Y2, bits 10..19 = Y1 */
-    gp1_vrange_y1_shift = 10, gp1_vrange_y1_width = 10, gp1_vrange_y1_mask = 0x3FF,
-    gp1_vrange_y2_shift =  0, gp1_vrange_y2_width = 10, gp1_vrange_y2_mask = 0x3FF,
+	/* GP1 vertical display range: bits 0..9 = Y2, bits 10..19 = Y1 */
+	gp1_vrange_y1_shift = 10, gp1_vrange_y1_width = 10,
+	gp1_vrange_y2_shift =  0, gp1_vrange_y2_width = 10,
 
-    /* GP1 draw area (top-left or bottom-right): bits 0..9 = X, bits 10..19 = Y
-     * (10-bit signed — caller pre-signs and masks with the named mask) */
-    gp1_draw_x_shift =  0, gp1_draw_x_width = 10, gp1_draw_x_mask = 0x3FF,
-    gp1_draw_y_shift = 10, gp1_draw_y_width = 10, gp1_draw_y_mask = 0x3FF,
+	/* GP1 draw area (top-left or bottom-right): bits 0..9 = X, bits 10..19 = Y
+		* (10-bit signed — caller pre-signs) */
+	gp1_draw_x_shift =  0, gp1_draw_x_width = 10,
+	gp1_draw_y_shift = 10, gp1_draw_y_width = 10,
 };
 
 /* ---- Layer 1.5: GP1 per-field encoders ---- */
-#define enc_gp1_disp_hres(h)      (((h) & gp1_disp_hres_mask)     << gp1_disp_hres_shift)
-#define enc_gp1_disp_vres(v)      (((v) & gp1_disp_vres_mask)     << gp1_disp_vres_shift)
-#define enc_gp1_disp_color(c)     (((c) & gp1_disp_color_mask)    << gp1_disp_color_shift)
-#define enc_gp1_disp_interlace(i) (((i) & gp1_disp_interlace_mask << gp1_disp_interlace_shift)
+#define enc_gp1_disp_hres(h)      ((h) << gp1_disp_hres_shift)
+#define enc_gp1_disp_vres(v)      ((v) << gp1_disp_vres_shift)
+#define enc_gp1_disp_color(c)     ((c) << gp1_disp_color_shift)
+#define enc_gp1_disp_interlace(i) ((i) << gp1_disp_interlace_shift)
 
-#define enc_gp1_hrange_x1(x1) (((x1) & gp1_hrange_x1_mask) << gp1_hrange_x1_shift)
-#define enc_gp1_hrange_x2(x2) (((x2) & gp1_hrange_x2_mask) << gp1_hrange_x2_shift)
-#define enc_gp1_vrange_y1(y1) (((y1) & gp1_vrange_y1_mask) << gp1_vrange_y1_shift)
-#define enc_gp1_vrange_y2(y2) (((y2) & gp1_vrange_y2_mask) << gp1_vrange_y2_shift)
-#define enc_gp1_draw_x(x)     (((x)  & gp1_draw_x_mask)    << gp1_draw_x_shift)
-#define enc_gp1_draw_y(y)     (((y)  & gp1_draw_y_mask)    << gp1_draw_y_shift)
+#define enc_gp1_hrange_x1(x1) ((x1) << gp1_hrange_x1_shift)
+#define enc_gp1_hrange_x2(x2) ((x2) << gp1_hrange_x2_shift)
+#define enc_gp1_vrange_y1(y1) ((y1) << gp1_vrange_y1_shift)
+#define enc_gp1_vrange_y2(y2) ((y2) << gp1_vrange_y2_shift)
+#define enc_gp1_draw_x(x)     ((x)  << gp1_draw_x_shift)
+#define enc_gp1_draw_y(y)     ((y)  << gp1_draw_y_shift)
 
 /* ---- Layer 2: GP1 composite encoders ---- */
 #define enc_gp1_disp_mode_word(h, v, c, i) (enc_gp0_cmd(gp1_cmd_DisplayMode)            | enc_gp1_disp_hres(h)  | enc_gp1_disp_vres(v) | enc_gp1_disp_color(c) | enc_gp1_disp_interlace(i))
@@ -255,6 +255,11 @@ enum {
 #define enc_gp0_draw_area_br_word(x, y)    (enc_gp0_cmd(gp0_cmd_SetDrawArea_BotRight) | enc_gp1_draw_x(x) | enc_gp1_draw_y(y))
 
 /* ---- Layer 3: GP1 semantic word builders ---- */
+#define gp1_word_Reset()             enc_gp0_cmd_word(gp1_cmd_Reset)
+#define gp1_word_ResetCmdBuffer()    enc_gp0_cmd_word(gp1_cmd_ResetCmdBuffer)
+#define gp1_word_AcknowledgeIRQ()    enc_gp0_cmd_word(gp1_cmd_AcknowledgeIRQ)
+#define gp1_word_StartDisplayArea()  enc_gp0_cmd_word(gp1_cmd_StartDisplayArea)
+
 #define gp1_word_display_enable(on)                         (enc_gp0_cmd(gp1_cmd_DisplayEnable) | ((on) & 1))
 #define gp1_word_display_disable()                          gp1_word_display_enable(0)
 #define gp1_word_display_mode_320x240_15bit_ntsc            enc_gp1_disp_mode_word(gp1_disp_HRes_320, gp1_disp_VRes_240, gp1_disp_Color15, 0)
@@ -279,31 +284,36 @@ enum {
 #define gp1_word_display_enabled   enc_gp0_cmd_word(gp1_cmd_DisplayEnable)
 #define gp1_word_display_disabled (enc_gp0_cmd_word(gp1_cmd_DisplayEnable) | 1)
 
+#define gp1_word_DisplayOn()       gp1_word_display_enable(0)
+#define gp1_word_DisplayOff()      gp1_word_display_enable(1)
+
 /* ---- DMA direction (2-bit payload on DMADirection cmd 0x04) ---- */
 enum {
-    gp1_dma_dir_Off            = 0,
-    gp1_dma_dir_FIFO           = 1,
-    gp1_dma_dir_CPU_to_GPU     = 2,
-    gp1_dma_dir_GPUREAD_to_CPU = 3,
+	gp1_dma_dir_Off            = 0,
+	gp1_dma_dir_FIFO           = 1,
+	gp1_dma_dir_CPU_to_GPU     = 2,
+	gp1_dma_dir_GPUREAD_to_CPU = 3,
 };
 #define gp1_word_dma_direction(dir) (enc_gp0_cmd(gp1_cmd_DMADirection) | ((dir) & 0x3))
+#define gp1_word_dma_to_gpu()       gp1_word_dma_direction(gp1_dma_dir_CPU_to_GPU)
+#define gp1_word_dma_read_cpu()     gp1_word_dma_direction(gp1_dma_dir_GPUREAD_to_CPU)
 
 /* ---- Standard display ranges (NTSC + PAL pre-baked) ---- */
 /* Horizontal range values are in video clock units (8 units/pixel); vertical range values are scanline numbers. */
 enum {
-    /* NTSC horizontal range: X1=608, X2=3168 */
-    gp1_hrange_NTSC_x1 = 0x260,
-    gp1_hrange_NTSC_x2 = 0xC60,
-    /* PAL horizontal range (same as NTSC for most CRTs) */
-    gp1_hrange_PAL_x1  = 0x260,
-    gp1_hrange_PAL_x2  = 0xC60,
+	/* NTSC horizontal range: X1=608, X2=3168 */
+	gp1_hrange_NTSC_x1 = 0x260,
+	gp1_hrange_NTSC_x2 = 0xC60,
+	/* PAL horizontal range (same as NTSC for most CRTs) */
+	gp1_hrange_PAL_x1  = 0x260,
+	gp1_hrange_PAL_x2  = 0xC60,
 
-    /* NTSC vertical range: Y1=24, Y2=264 */
-    gp1_vrange_NTSC_y1 = 24,
-    gp1_vrange_NTSC_y2 = 264,
-    /* PAL vertical range: Y1=24, Y2=504 */
-    gp1_vrange_PAL_y1  = 24,
-    gp1_vrange_PAL_y2  = 504,
+	/* NTSC vertical range: Y1=24, Y2=264 */
+	gp1_vrange_NTSC_y1 = 24,
+	gp1_vrange_NTSC_y2 = 264,
+	/* PAL vertical range: Y1=24, Y2=504 */
+	gp1_vrange_PAL_y1  = 24,
+	gp1_vrange_PAL_y2  = 504,
 };
 
 #define gp1_word_horizontal_range_ntsc enc_gp1_hrange_word(gp1_hrange_NTSC_x1, gp1_hrange_NTSC_x2)
@@ -314,14 +324,49 @@ enum {
 /* ---- Draw-mode setting (TPage / draw-area allowance) ---- */
 /* The "drawing enabled" word is the standard post-init state. */
 enum {
-    gp0_DrawMode_DrawToDispBit = 10,
+	/* Per psx-spx, the standard 0xE1 layout has dfe at bit 10. But libpsyx's PutDrawEnv
+	 * uses bit 19 (in the "unused" 14-23 range) for dfe in the DR_ENV code[0] — and the
+	 * PSX hardware honors bit 19 in the DR_ENV context (not bit 10). So we need a
+	 * separate bit definition for the DR_ENV-specific DrawMode. */
+	gp0_DrawMode_DrawToDispBit         = 10,   // standard psx-spx bit 10 (dfe)
+	gp0_DrawMode_DR_ENV_DrawToDispBit  = 19,   // libpsyx DR_ENV code[0] (dfe in DR_ENV context)
+	gp0_DrawMode_DR_ENV_isbgBit        = 19,   // libpsyx uses bit 19 for isbg too
 };
 #define gp0_word_draw_mode_drawing_allowed (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit))
 
-/* ---- DrawArea pre-baked at origin (0,0) and full screen (320x240) ---- */
+/* DR_ENV-specific DrawMode variants (libpsyx SetDrawEnv layout).
+ * The DR_ENV is a 16-word packet emitted at boot by gp_screen_init's ac_put_draw_env_demo
+ * atom component. Within the DR_ENV, the 0xE1 command is reused in three different bit
+ * configurations:
+ *   code[0] = `gp0_word_draw_mode_drawing_allowed` (dfe=1; standard post-init state)
+ *   code[6] = `gp0_word_dr_env_bg_color_cmd(isbg, r, g, b)` (initial-bg-color path)
+ *   code[7] = `gp0_word_dr_env_draw_mode(isbg)` (isbg-flag path)
+ * Bits 0-23 of the 0xE1 word are the payload; bits 24-31 are the cmd byte (0xE1). */
+#define gp0_word_dr_env_bg_color_cmd(isbg, r, g, b) (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
+#define gp0_word_dr_env_draw_mode(isbg)             (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0))
+
+/* State-setter bare-cmd words (no immediate payload; the GPU uses the current state machine already programmed). */
+#define gp0_word_set_texture_window() enc_gp0_cmd_word(gp0_cmd_SetTextureWindow)
+#define gp0_word_set_draw_offset()    enc_gp0_cmd_word(gp0_cmd_SetDrawOffset)
+#define gp0_word_set_mask_bit()       enc_gp0_cmd_word(gp0_cmd_SetMaskBit)
+
+/* DR_ENV code[5] Mask (0xE6 cmd + isbg bit). The isbg bit is set so the GPU knows the auto-clear path is active (paired with code[6] + code[7]). */
+#define gp0_word_dr_env_mask() (gp0_word_set_mask_bit() | gp0_dr_env_isbg_bit)
+
+/* DR_ENV pre-baked constants (libpsyx PutDrawEnv layout).
+ * DR_ENV is a 16-word packet: tag = (length << 24) | addr, where length = 15 (15 code words follow) and addr = 0 (chain to nothing). */
+enum {
+	PolyTag_len_bits  =  8,
+	PolyTag_addr_bits = 24,
+
+	gp0_dr_env_tag      = (15 << 24) | 0x00FFFFFF,
+	gp0_dr_env_isbg_bit = (1 << gp0_DrawMode_DR_ENV_isbgBit),
+};
+
+/* ---- DrawArea at origin (0,0) and full screen (320x240) ---- */
 #define gp0_word_draw_area_top_left_origin      enc_gp0_draw_area_tl_word(0, 0)
-#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(320, 240)
-#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(640, 480)
+#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(319, 239)
+#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(639, 479)
 
 #pragma endregion GPU Ports & Commands
 
@@ -332,9 +377,9 @@ enum {
  *  Read from HW_GP1; the lower bits are DMA-block-size (variable-width).
  * ============================================================================ */
 enum {
-    gp1_Status_BitReady          = 31,
-    gp1_Status_BitSendingDMA     = 25,
-    gp1_Status_DMABlockSizeShift =  0,
+	gp1_Status_BitReady          = 31,
+	gp1_Status_BitSendingDMA     = 25,
+	gp1_Status_DMABlockSizeShift =  0,
 };
 
 #define gp1_status_is_ready()       ((HW_GP1[0] >> gp1_Status_BitReady)      & 1)
@@ -360,10 +405,10 @@ typedef Struct_(RGB8) { B1 r; B1 g; B1 b; };
 #define rgb8(r,g,b) ((RGB8){r,g,b})
 
 /* ---------- PolyTag (the OT-link header; 1 word) ---------- */
-enum {
-    PolyTag_len_bits  =  8,
-    PolyTag_addr_bits = 24,
-};
+// enum {
+// 	PolyTag_len_bits  =  8,
+// 	PolyTag_addr_bits = 24,
+// };
 typedef Struct_(PolyTag) {
 	union {
 		U4 code;
@@ -374,108 +419,105 @@ typedef Struct_(PolyTag) {
 	};
 };
 
-/* DSL cast convention: every cast uses `C_()`, every pointer qualifier is `R_` (restrict) or `V_` (volatile).
- * No raw C-style casts. RHS values are assumed to be `U4` — caller passes a `U4` directly. */
 #define set_len(tag,v)  (C_(PolyTag_R,tag)->len  = u4_(v))
 #define set_addr(tag,v) (C_(PolyTag_R,tag)->addr = u4_(v))
-/* `set_code` is no longer in the new PolyTag design — the code byte lives in the primitive body
+/* `set_code` is no longer in the new PolyTag design
  * (e.g. `((Poly_F3*)(p))->code`), not in the tag.
- *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters,
- *  which set both the tag's length and the code. */
+ *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters, which set both the tag's length and the code. */
 #define get_len(tag)  C_(U4,C_(PolyTag_R,tag)->len)
 #define get_addr(tag) C_(U4,C_(PolyTag_R,tag)->addr)
 
 /* ---------- Poly_F3 (Flat Triangle; 5 words) ---------- */
 typedef Struct_(Poly_F3) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
-        A3_V2_S2 points;
-    };
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	union {
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
+		A3_V2_S2 points;
+	};
 };
 
 /* ---------- Poly_F4 (Flat Quad; 6 words) ---------- */
 typedef Struct_(Poly_F4) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
-        A4_V2_S2 points;
-    };
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	union {
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
+		A4_V2_S2 points;
+	};
 };
 
 /* ---------- Poly_G3 (Gouraud Triangle; 7 words) ---------- */
 typedef Struct_(Poly_G3) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;
 };
 
 /* ---------- Poly_G4 (Gouraud Quad; 9 words) ---------- */
 typedef Struct_(Poly_G4) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p3;
 };
 
 /* ---------- Poly_FT3 (Flat Textured Triangle; placeholder layout) ---------- */
 /* TODO(Ed): verify the textured-variant layout against PSX-SPX when needed. */
 typedef Struct_(Poly_FT3) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    U4   tpage;
-    U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	U4   tpage;
+	U4   clut;
+	V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p2; U1 u2; U1 v2;
 };
 
 /* ---------- Poly_FT4 (Flat Textured Quad) ---------- */
 typedef Struct_(Poly_FT4) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    U4   tpage;
-    U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
-    V2_S2 p3; U1 u3; U1 v3;
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	U4   tpage;
+	U4   clut;
+	V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p2; U1 u2; U1 v2;
+	V2_S2 p3; U1 u3; U1 v3;
 };
 
 /* ---------- Poly_GT3 (Gouraud Textured Triangle) ---------- */
 typedef Struct_(Poly_GT3) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
-    U4    tpage;
-    U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;
+	U4    tpage;
+	U4    clut;
+	V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp2; U1 u2; U1 v2;
 };
 
 /* ---------- Poly_GT4 (Gouraud Textured Quad) ---------- */
 typedef Struct_(Poly_GT4) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
-    U4    tpage;
-    U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
-    V2_S2 tp3; U1 u3; U1 v3;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p3;
+	U4    tpage;
+	U4    clut;
+	V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp2; U1 u2; U1 v2;
+	V2_S2 tp3; U1 u3; U1 v3;
 };
 
 /* ---------- Primitive setters (C-level) ----------
@@ -510,52 +552,55 @@ typedef Struct_(Poly_GT4) {
  *    bits 12..31  = reserved (zero)
  * ============================================================================ */
 enum {
-    /* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
-    gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4, gp0_tpage_x_mask            = 0xF,
-    gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1, gp0_tpage_y_mask            = 0x1,
-    gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2, gp0_tpage_semi_trans_mask   = 0x3,
-    gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2, gp0_tpage_color_depth_mask  = 0x3,
-    gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1, gp0_tpage_dither_mask       = 0x1,
-    gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1, gp0_tpage_draw_to_disp_mask = 0x1,
-    gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1, gp0_tpage_tex_disable_mask  = 0x1,
+	/* ---- Layer 1: TPage bitfield shifts / widths ---- */
+	gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4,
+	gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1,
+	gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2,
+	gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2,
+	gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1,
+	gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1,
+	gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1,
 
-    /* TPage color-depth payload values (NOT bit positions — these go in
-     * the 2-bit field at gp0_tpage_color_depth_shift). */
-    gp0_tpage_color_4bpp  = 0x0,
-    gp0_tpage_color_8bpp  = 0x1,
-    gp0_tpage_color_16bpp = 0x2,
+	/* TPage color-depth payload values (NOT bit positions — these go in
+	 * the 2-bit field at gp0_tpage_color_depth_shift). */
+	gp0_tpage_color_4bpp  = 0x0,
+	gp0_tpage_color_8bpp  = 0x1,
+	gp0_tpage_color_16bpp = 0x2,
 
-    /* TPage semi-transparency mode payload values (NOT bit positions). */
-    gp0_tpage_semi_trans_none  = 0x0,
-    gp0_tpage_semi_trans_alpha = 0x1,
-    gp0_tpage_semi_trans_add   = 0x2,
-    gp0_tpage_semi_trans_sub   = 0x3,
+	/* Default TPage value libpsyx's SetDefDrawEnv writes (matches the `li v1, 10; sh v1, 20(v0)` sequence at C11_only.elf:0x8001273C). */
+	gp0_tpage_default    = 10,
+
+	/* TPage semi-transparency mode payload values. */
+	gp0_tpage_semi_trans_none  = 0x0,
+	gp0_tpage_semi_trans_alpha = 0x1,
+	gp0_tpage_semi_trans_add   = 0x2,
+	gp0_tpage_semi_trans_sub   = 0x3,
 };
 
 /* ---- Layer 1.5: TPage per-field encoders. Mirrors enc_gte_sf/mx/v in gte.h. ---- */
-#define enc_gp0_tpage_x(x)            (((x) & gp0_tpage_x_mask)            << gp0_tpage_x_shift)
-#define enc_gp0_tpage_y(y)            (((y) & gp0_tpage_y_mask)            << gp0_tpage_y_shift)
-#define enc_gp0_tpage_semi_trans(s)   (((s) & gp0_tpage_semi_trans_mask)   << gp0_tpage_semi_trans_shift)
-#define enc_gp0_tpage_color_depth(c)  (((c) & gp0_tpage_color_depth_mask)  << gp0_tpage_color_depth_shift)
-#define enc_gp0_tpage_dither(d)       (((d) & gp0_tpage_dither_mask)       << gp0_tpage_dither_shift)
-#define enc_gp0_tpage_draw_to_disp(d) (((d) & gp0_tpage_draw_to_disp_mask) << gp0_tpage_draw_to_disp_shift)
-#define enc_gp0_tpage_tex_disable(t)  (((t) & gp0_tpage_tex_disable_mask)  << gp0_tpage_tex_disable_shift)
+#define enc_gp0_tpage_x(x)            ((x) << gp0_tpage_x_shift)
+#define enc_gp0_tpage_y(y)            ((y) << gp0_tpage_y_shift)
+#define enc_gp0_tpage_semi_trans(s)   ((s) << gp0_tpage_semi_trans_shift)
+#define enc_gp0_tpage_color_depth(c)  ((c) << gp0_tpage_color_depth_shift)
+#define enc_gp0_tpage_dither(d)       ((d) << gp0_tpage_dither_shift)
+#define enc_gp0_tpage_draw_to_disp(d) ((d) << gp0_tpage_draw_to_disp_shift)
+#define enc_gp0_tpage_tex_disable(t)  ((t) << gp0_tpage_tex_disable_shift)
 
 /* ---- Layer 2: TPage composite encoder. Mirrors enc_gte_cmdw in gte.h ---- */
 #define enc_gp0_tpage_word(x, y, semi_trans, color_depth, dither, draw_to_disp, tex_disable) \
-    (enc_gp0_tpage_x(x)                       \
-   | enc_gp0_tpage_y(y)                       \
-   | enc_gp0_tpage_semi_trans(semi_trans)     \
-   | enc_gp0_tpage_color_depth(color_depth)   \
-   | enc_gp0_tpage_dither(dither)             \
-   | enc_gp0_tpage_draw_to_disp(draw_to_disp) \
-   | enc_gp0_tpage_tex_disable(tex_disable))
+	 (enc_gp0_tpage_x(x)                       \
+	| enc_gp0_tpage_y(y)                       \
+	| enc_gp0_tpage_semi_trans(semi_trans)     \
+	| enc_gp0_tpage_color_depth(color_depth)   \
+	| enc_gp0_tpage_dither(dither)             \
+	| enc_gp0_tpage_draw_to_disp(draw_to_disp) \
+	| enc_gp0_tpage_tex_disable(tex_disable))
 
 typedef Struct_(TexturePage) { U4 raw; };
 
 /* ---- Layer 3: TPage semantic word builder ---- */
 #define gp0_word_tpage(x, y, semi_trans, color_depth, dither, draw_to_disp, tex_disable) \
-    enc_gp0_tpage_word((x), (y), (semi_trans), (color_depth), (dither), (draw_to_disp), (tex_disable))
+	enc_gp0_tpage_word((x), (y), (semi_trans), (color_depth), (dither), (draw_to_disp), (tex_disable))
 #pragma endregion TPage
 
 #pragma region CLUT
@@ -569,17 +614,17 @@ typedef Struct_(TexturePage) { U4 raw; };
  *    bits 24..31  = command byte — 0x20 (4bpp load) or 0x25 (8bpp load)
  * ============================================================================ */
 enum {
-    /* ---- Layer 1: CLUT bitfield shifts / widths / masks ---- */
-    gp0_clut_y_shift = 0, gp0_clut_y_width = 6, gp0_clut_y_mask = 0x3F,
-    gp0_clut_x_shift = 6, gp0_clut_x_width = 9, gp0_clut_x_mask = 0x1FF,
-    /* CLUT-load cmd-byte variants — the upper byte of the GP0 word. */
-    gp0_clut_cmd_Load4bpp = 0x20,
-    gp0_clut_cmd_Load8bpp = 0x25,
+	/* ---- Layer 1: CLUT bitfield shifts / widths ---- */
+	gp0_clut_y_shift = 0, gp0_clut_y_width = 6,
+	gp0_clut_x_shift = 6, gp0_clut_x_width = 9,
+	/* CLUT-load cmd-byte variants — the upper byte of the GP0 word. */
+	gp0_clut_cmd_Load4bpp = 0x20,
+	gp0_clut_cmd_Load8bpp = 0x25,
 };
 
 /* ---- Layer 1.5: CLUT per-field encoders ---- */
-#define enc_gp0_clut_x(x) (((x) & gp0_clut_x_mask) << gp0_clut_x_shift)
-#define enc_gp0_clut_y(y) (((y) & gp0_clut_y_mask) << gp0_clut_y_shift)
+#define enc_gp0_clut_x(x) ((x) << gp0_clut_x_shift)
+#define enc_gp0_clut_y(y) ((y) << gp0_clut_y_shift)
 
 /* ---- Layer 2: CLUT composite encoder ---- */
 #define enc_gp0_clut_word(cmd, x, y) (enc_gp0_cmd(cmd) | enc_gp0_clut_x(x) | enc_gp0_clut_y(y))
@@ -615,26 +660,26 @@ enum {
  *  Stoppped for now at the struct + enum level.
  * ============================================================================ */
 enum {
-    tim_file_id_magic = 0x10,
-    tim_type_4bpp     = 0x00,
-    tim_type_8bpp     = 0x01,
-    tim_type_16bpp    = 0x02,
-    tim_type_32bpp    = 0x03,
-    tim_type_mixed    = 0x04,
-    tim_flag_has_clut = 0x08,
+	tim_file_id_magic = 0x10,
+	tim_type_4bpp     = 0x00,
+	tim_type_8bpp     = 0x01,
+	tim_type_16bpp    = 0x02,
+	tim_type_32bpp    = 0x03,
+	tim_type_mixed    = 0x04,
+	tim_flag_has_clut = 0x08,
 };
 
 typedef Struct_(TIM_Header) {
-    U4 file_id;       /* always 0x10 = "TIM" magic */
-    U4 version;       /* ignored; always 0 */
-    U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
+	U4 file_id;       /* always 0x10 = "TIM" magic */
+	U4 version;       /* ignored; always 0 */
+	U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
 };
 typedef Struct_(TIM_SectionHeader) {
-    U4 section_length;  /* bytes in this section including this header */
-    U2 org_x;           /* origin in VRAM */
-    U2 org_y;
-    U2 width;           /* width in pixels */
-    U2 height;          /* height in pixels */
+	U4 section_length;  /* bytes in this section including this header */
+	U2 org_x;           /* origin in VRAM */
+	U2 org_y;
+	U2 width;           /* width in pixels */
+	U2 height;          /* height in pixels */
 };
 #pragma endregion TIM File Format
 

code/duffle/gte.atom.c

@@ -0,0 +1,395 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "gte.h"
+#	include "gp.h"
+#	include "lottes_tape.h"	
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(gte_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+/* Words: 3; Loads 3 S2 indices from the face array */
+FI_ Slice_MipsCode ac_load_tri_indices(AtomBuilder_R ab, U4 r_face_cusor, U4 r_i0, U4 r_i1, U4 r_i2)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_half_u(r_i0, r_face_cusor, 0 * S_(S2)),
+	load_half_u(r_i1, r_face_cusor, 1 * S_(S2)),
+	load_half_u(r_i2, r_face_cusor, 2 * S_(S2)),
+})
+
+FI_ Slice_MipsCode ac_gte_mv_to_cr_diag_v3s4(AtomBuilder_R ab, Reg_(V3_S4) v) MipsAtomComp_Proc_(ab, {
+	gte_mv_to_ctrl_r(v.y, gte_cr_RT13),
+	gte_mv_to_ctrl_r(v.z, gte_cr_RT22),
+	gte_mv_to_ctrl_r(v.x, gte_cr_RT11),
+})
+
+FI_ Slice_MipsCode ac_gte_ld_ir123_v3s4(AtomBuilder_R ab, Reg_(V3_S4) v) MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(v.x, C2_IR1),
+	gte_mv_to_data_r(v.y, C2_IR2),
+	gte_mv_to_data_r(v.z, C2_IR3),
+})
+
+/* ─── GTE OP cross product (a × b → a) ───
+ * Sets up RT diagonal from a.xyz, IR1/2/3 from b.xyz, fires OP,
+ * reads MAC1/2/3, shifts right 12 (S12.20 → S12.0 OuterProduct12), writes back to a.xyz.
+ * Composes the three sub-primitives (RT-load, IR-load, OP, MAC-read, shift)
+ * into one component for use by atoms that need the cross product inline.
+ *
+ * Output gpr (a) aliases source-A gpr; MAC read clobbers source-A's load targets,
+ * but by that point the RT load is complete and source A is dead.
+ * Pipeline: clobbers IR1..3, MAC1..3, RT11..33.
+ *
+ * The CPU→COP2 transfer chains (3 ctc2, 3 mtc2) require a 2-slot retirement gap,
+ * and the MFC2→GPR chain (3 mfc2) requires a 1-slot retirement gap, before the GPR can be read.
+ * The hazard nops are inlined below — same convention as ac_gte_gpf_scale — so any atom body inlining this component inherits them.
+ *
+ * Words: 18 (3 ctc2 + 2 nop + 3 mtc2 + 2 nop + 1 op + 3 mfc2 + 1 nop + 3 sra).
+ */
+FI_ Slice_MipsCode ac_gte_op_cross_v3s4(AtomBuilder_R ab, Reg_(V3_S4) a, Reg_(V3_S4) b) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_gte_mv_to_cr_diag_v3s4(a),  GteDelay_  /* RT diagonal: D1 = a.x, D2 = a.y, D3 = a.z */
+	mac_gte_ld_ir123_v3s4(b),       GteDelay_  /* IR: second operand (b.xyz) */
+	gte_cmdw_cross,                            /* OP: MAC1/2/3 = a × b (S12.20) */
+	mac_gte_mv_from_mac123_v3s4(a), GteDelay_  /* Read MAC1/2/3 → a.xyz (overwrites source-A's load targets) */
+	mac_shift_aright_v3s4_self(a, 12), /* Right-shift MAC by 12 (S12.20 → S12.0 OuterProduct12) */
+})
+
+/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
+ * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
+FI_ Slice_MipsCode ac_gte_store_f3(AtomBuilder_R ab, U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_F3,p0)),
+	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_F3,p1)),
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_F3,p2)),
+})
+
+/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
+I_ Slice_MipsCode ac_gte_load_tri_verts(AtomBuilder_R ab, U4 r_vert_base, U4 r_v0, U4 r_v1, U4 r_v2) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	shift_lleft(R_AT, r_v0, v3s2_byteoff), add_u_self(R_AT, r_vert_base), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
+	shift_lleft(R_AT, r_v1, v3s2_byteoff), add_u_self(R_AT, r_vert_base), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
+	shift_lleft(R_AT, r_v2, v3s2_byteoff), add_u_self(R_AT, r_vert_base), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), LdSlot_ gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+})
+
+/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
+ * G4 triangle portion to p0/p1/p2.
+ * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen).
+ * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
+ * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
+FI_ Slice_MipsCode ac_gte_store_g4_p012(AtomBuilder_R ab, Reg r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_G4,p0)),
+	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_G4,p1)),
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p2)),
+})
+
+/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
+ * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its single-vertex result to SXY2;
+ * SXY0 still holds v0.screen from the earlier RTPT.
+ */
+FI_ Slice_MipsCode ac_gte_store_g4_p3(AtomBuilder_R ab, U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, { gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3)) })
+
+/* ─── STAGE 1 of normalize: SQR + mfc2 MAC1/2/3 ───
+ * Emits squared magnitude per component (in MAC1/2/3) into caller-provided scratch regs. */
+FI_ Slice_MipsCode ac_gte_sqr_v3(AtomBuilder_R ab, U4 r_sx, U4 r_sy, U4 r_sz, U4 r_sq_x, U4 r_sq_y, U4 r_sq_z) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, nop),
+	gte_mv_from_data_r(r_sq_x, C2_MAC1),
+	gte_mv_from_data_r(r_sq_y, C2_MAC2),
+	gte_mv_from_data_r(r_sq_z, C2_MAC3),
+})
+
+/* ─── SQR FIRE — mtc2 3 GPRs into IR1/IR2/IR3, then fire SQR. ─── */
+FI_ Slice_MipsCode ac_gte_sqr_v3s4(AtomBuilder_R ab, Reg r_sx, Reg r_sy, Reg r_sz, MipsCode delay_slot)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(r_sx, C2_IR1),
+	gte_mv_to_data_r(r_sy, C2_IR2),
+	gte_mv_to_data_r(r_sz, C2_IR3),
+	delay_slot, gte_cmdw_sqr,
+})
+
+/* ─── STAGE 4 of normalize: mtc2 IR0..3 + GPF + mfc2 MAC + srav finalize ───
+ * Reusable standalone — given an IR0 = 1/|v| estimate (typically from a sqrtbl lookup) and a shift count 
+ * (typically (31 - LZCR)/2), multiplies IR0*IR[i] via GPF and shifts right to produce the normalized output.
+ * Used standalone for "scale vector by scalar".
+ * Words: 11.  Clobbers: IR0..3, MAC1..3.  Uses gte_cmdw_gpf (sf=0, lm=0). */
+FI_ Slice_MipsCode ac_gte_gpf_scale(AtomBuilder_R ab, 
+	U4 r_sx, U4 r_sy, U4 r_sz, 
+	U4 r_recip_est, U4 r_shift, 
+	U4 r_dx, U4 r_dy, U4 r_dz)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(r_recip_est, C2_IR0),
+	gte_mv_to_data_r(r_sx,        C2_IR1),
+	gte_mv_to_data_r(r_sy,        C2_IR2),
+	gte_mv_to_data_r(r_sz,        C2_IR3),
+	GteDelay_ nop2, /* retire IR0..IR3 → GPF input pre-fill (matches libgte 0x80016134..0x80016138) */
+	gte_cmdw_gpf,
+	gte_mv_from_data_r(r_dx, C2_MAC1),
+	gte_mv_from_data_r(r_dy, C2_MAC2),
+	gte_mv_from_data_r(r_dz, C2_MAC3),
+	shift_aright_var(r_dx, r_dx, r_shift),
+	shift_aright_var(r_dy, r_dy, r_shift),
+	shift_aright_var(r_dz, r_dz, r_shift),
+})
+
+/* ─── TRANS MATRIX (libgte TransMatrix port) ───
+ * Atom component — auto-generates mac_trans_matrix Mac composer macro.
+ * m->t = v (struct copy; libgte's TransMatrix at 0x8001a540 is just 3 store_words, no GTE, no add).
+ * Uses 1 GPR (r_t1 = off value) per axis; per-axis load-delay-slot pattern.
+ * Words: 9.  Clobbers: r_t1. */
+FI_ Slice_MipsCode ac_trans_mt3s3s4(AtomBuilder_R ab
+	,	U4 r_mtx, U4 r_off
+	,	U4 r_t0, U4 r_t1, U4 r_t2
+) MipsAtomComp_Proc_(ab, {
+	load_word( r_t0, r_off, O_(V3_S4,x)),
+	load_word( r_t1, r_off, O_(V3_S4,y)),
+	load_word( r_t2, r_off, O_(V3_S4,z)),
+	store_word(r_t0, r_mtx, O_(MT3_S2S4,t[0])),
+	store_word(r_t1, r_mtx, O_(MT3_S2S4,t[1])),
+	store_word(r_t2, r_mtx, O_(MT3_S2S4,t[2])),
+})
+
+/* ─── LZCR ROUND EVEN + HALF-SHIFT ───
+ * Takes the raw LZCR leading-zero/ones count (from mfc2 C2_LZCR, range 1..32 per PSX-SPX cop2r31) and the |v|² sum (in r_mag_sq from the MAC1+MAC2+MAC3 add).
+ * Produces:
+ *   r_shift        ← LZCR rounded down to even (clear bit 0)
+ *   r_mag_sq_copy  ← |v|² sum (moved out of r_mag_sq before it's overwritten)
+ *   r_mag_sq       ← (31 - even_LZCR) / 2 = the final srav/GPF shift amount
+ *
+ * Rounding to even ensures (31 - LZCR) is always odd, so the >> 1 division is consistent — no 0.5 loss.
+ * The caller branches on LZCR < 24 to decide left-shift vs right-shift of r_mag_sq_copy, then saves the shift count.
+ *
+ * Note: C2_LZCR (cop2r31) is a fixed read-only C2 data register — the caller must read it via mfc2 from C2_LZCR;
+ * there is no register choice at the hardware level. Only the GPR that holds the result is caller-determined. */
+FI_ Slice_MipsCode ac_lzcr_round_even_half_shift(AtomBuilder_R ab,
+	U4 r_shift,
+	U4 r_mag_sq,
+	U4 r_mag_sq_copy)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	and_i(r_shift, r_shift, gte_lzcr_even_mask),
+	or_u(r_mag_sq_copy, r_mag_sq, 0),
+	li_s(        r_mag_sq, 31),
+	sub_s(       r_mag_sq, r_mag_sq, r_shift),
+	shift_aright(r_mag_sq, r_mag_sq, 1),
+})
+
+FI_ Slice_MipsCode ac_gte_general_purpose_interopolation(AtomBuilder_R ab
+	, Reg to_ir0,  Reg to_ir1,  Reg to_ir2, Reg to_ir3
+	, Reg fr_mac1, Reg fr_mac2, Reg fr_mac3
+	, MipsCode nop_slot1, MipsCode nop_slot2)
+MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(to_ir0, C2_IR0),
+	gte_mv_to_data_r(to_ir1, C2_IR1), /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */
+	gte_mv_to_data_r(to_ir2, C2_IR2),
+	gte_mv_to_data_r(to_ir3, C2_IR3), /* IR3 = src.z (reloaded) */
+	GteDelay_ nop_slot1,
+	GteDelay_ nop_slot2,
+	gte_cmdw_gpf,
+	gte_mv_from_data_r(fr_mac1, C2_MAC1),
+	gte_mv_from_data_r(fr_mac2, C2_MAC2),
+	gte_mv_from_data_r(fr_mac3, C2_MAC3),
+})
+
+FI_ Slice_MipsCode ac_gte_mv_from_data_r_mac123(AtomBuilder_R ab
+	, Reg fr_mac1, Reg fr_mac2, Reg fr_mac3)
+MipsAtomComp_Proc_(ab, {
+	gte_mv_from_data_r(fr_mac1, C2_MAC1),
+	gte_mv_from_data_r(fr_mac2, C2_MAC2),
+	gte_mv_from_data_r(fr_mac3, C2_MAC3),
+})
+
+
+FI_ Slice_MipsCode ac_gte_mv_from_mac123_v3s4(AtomBuilder_R ab, Reg_(V3_S4) v) MipsAtomComp_ProcMap_(ab, mac_gte_mv_from_data_r_mac123(v.x, v.y, v.z))
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Atom Procs
+
+/* ─── Local copy of PSYQ's sqrtbl (1/sqrt lookup table for VectorNormal). ───
+ * Source: PSYQ 4.7 libgte sqrtbl at 0x800185B4 in hello_camera.elf.
+ *   objdump -s --start-address=0x800185B4 --stop-address=0x800185F4 hello_camera.elf → 192 entries × 16-bit signed, in 1.12 fixed-point (max value 0x1000 = 1.0).
+ *
+ * Data is identical to the libgte original (byte-for-byte verified).
+ * 
+ * ─── Per-entry semantics (decoded from libgte msc02 VectorNormal) ───
+ * Each entry is `1/sqrt(x)` in 1.12 fixed point (value / 4096).
+ * The 192 entries span 4 octaves of the input magnitude, with 48 entries per octave:
+ *   Octave 0 (entries  0- 47): mantissa in [0x8000,  0x10000)  output ~[1.000, 0.707]
+ *   Octave 1 (entries 48- 95): mantissa in [0x10000, 0x20000)  output ~[0.707, 0.500]
+ *   Octave 2 (entries 96-143): mantissa in [0x20000, 0x40000)  output ~[0.500, 0.354]
+ *   Octave 3 (entries144-191): mantissa in [0x40000, 0x80000)  output ~[0.354, 0.251]
+ * Within each octave, 8 sub-entries interpolate over the 8 fractional bits of the mantissa
+ * (the byte `(0x80 | (i mod 8))` for the lower-byte of the aligned value).
+ * Sampling the first value of each octave:
+ *   [0]   0x1000 = 1.0000                 ; 1 / sqrt(1.0000)
+ *   [48]  0x0e4f = 0.8940                 ; 1 / sqrt(1.2500)
+ *   [96]  0x0d10 = 0.8164                 ; 1 / sqrt(1.5000)
+ *   [144] 0x0c0a = 0.7520                 ; 1 / sqrt(1.7500)
+ * And representative sub-entries within octave 0 (mantissa in [0x8000, 0x8100)):
+ *   [0]   0x1000 = 1.0000                 ; 1 / sqrt(0x8000)
+ *   [1]   0x0fe0 = 0.9922                 ; 1 / sqrt(0x8100)
+ *   [2]   0x0fc1 = 0.9846                 ; 1 / sqrt(0x8200)
+ *   [3]   0x0fa3 = 0.9773                 ; 1 / sqrt(0x8300)
+ *   [4]   0x0f85 = 0.9700                 ; 1 / sqrt(0x8400)
+ *   [5]   0x0f68 = 0.9629                 ; 1 / sqrt(0x8500)
+ *   [6]   0x0f4c = 0.9561                 ; 1 / sqrt(0x8600)
+ *   [7]   0x0f30 = 0.9492                 ; 1 / sqrt(0x8700)
+ *
+ * The algorithm's `addi -64 / sll 1 / lh` selects the entry at `(aligned - 64) * 2` for the case where `aligned` has its top bit at bit 24.
+ * After the sllv/srav pair, `aligned` always lands in `[0x80, 0x100)` 
+ * (with top bit at bit 24 → after `sub $aligned - 64`, the index sits in `[0x40, 0x80) * 2 = [0x80, 0x100)` bytes = entries [64, 128) within the sqrtbl).
+ * The earlier 64 entries (octave 0) are reached when the magnitude after shifting puts the top bit below bit 24 (the `sllv` branch),
+ * and the load upper_halves of the table bracket the input range.
+ * The later 64 entries (octaves 2-3) are the `srav` branch when the magnitude's top bit is well above bit 24.
+ *
+ * Reproduced verbatim from libgte (verified against libpsn00b/psxgte/vector.s:100-123 — 24 rows × 8 halfwords, last entry 0x0804).
+ * */
+internal S2 const gte_normalize_sqr_tbl[192] align_(2) = {
+	0x1000, 0x0fe0, 0x0fc1, 0x0fa3, 0x0f85, 0x0f68, 0x0f4c, 0x0f30,
+	0x0f15, 0x0efb, 0x0ee1, 0x0ec7, 0x0eae, 0x0e96, 0x0e7e, 0x0e66,
+	0x0e4f, 0x0e38, 0x0e22, 0x0e0c, 0x0df7, 0x0de2, 0x0dcd, 0x0db9,
+	0x0da5, 0x0d91, 0x0d7e, 0x0d6b, 0x0d58, 0x0d45, 0x0d33, 0x0d21,
+	0x0d10, 0x0cff, 0x0cee, 0x0cdd, 0x0ccc, 0x0cbc, 0x0cac, 0x0c9c,
+	0x0c8d, 0x0c7d, 0x0c6e, 0x0c5f, 0x0c51, 0x0c42, 0x0c34, 0x0c26,
+	0x0c18, 0x0c0a, 0x0bfd, 0x0bef, 0x0be2, 0x0bd5, 0x0bc8, 0x0bbb,
+	0x0baf, 0x0ba2, 0x0b96, 0x0b8a, 0x0b7e, 0x0b72, 0x0b67, 0x0b5b,
+	0x0b50, 0x0b45, 0x0b39, 0x0b2e, 0x0b24, 0x0b19, 0x0b0e, 0x0b04,
+	0x0af9, 0x0aef, 0x0ae5, 0x0adb, 0x0ad1, 0x0ac7, 0x0abd, 0x0ab4,
+	0x0aaa, 0x0aa1, 0x0a97, 0x0a8e, 0x0a85, 0x0a7c, 0x0a73, 0x0a6a,
+	0x0a61, 0x0a59, 0x0a50, 0x0a47, 0x0a3f, 0x0a37, 0x0a2e, 0x0a26,
+	0x0a1e, 0x0a16, 0x0a0e, 0x0a06, 0x09fe, 0x09f6, 0x09ef, 0x09e7,
+	0x09e0, 0x09d8, 0x09d1, 0x09c9, 0x09c2, 0x09bb, 0x09b4, 0x09ad,
+	0x09a5, 0x099e, 0x0998, 0x0991, 0x098a, 0x0983, 0x097c, 0x0976,
+	0x096f, 0x0969, 0x0962, 0x095c, 0x0955, 0x094f, 0x0949, 0x0943,
+	0x093c, 0x0936, 0x0930, 0x092a, 0x0924, 0x091e, 0x0918, 0x0912,
+	0x090d, 0x0907, 0x0901, 0x08fb, 0x08f6, 0x08f0, 0x08eb, 0x08e5,
+	0x08e0, 0x08da, 0x08d5, 0x08cf, 0x08ca, 0x08c5, 0x08bf, 0x08ba,
+	0x08b5, 0x08b0, 0x08ab, 0x08a6, 0x08a1, 0x089c, 0x0897, 0x0892,
+	0x088d, 0x0888, 0x0883, 0x087e, 0x087a, 0x0875, 0x0870, 0x086b,
+	0x0867, 0x0862, 0x085e, 0x0859, 0x0855, 0x0850, 0x084c, 0x0847,
+	0x0843, 0x083e, 0x083a, 0x0836, 0x0831, 0x082d, 0x0829, 0x0824,
+	0x0820, 0x081c, 0x0818, 0x0814, 0x0810, 0x080c, 0x0808, 0x0804,
+};
+
+typedef Struct_(Binds_NormalizeV3S4) {
+	U2 src_offset;    /* offset of src V3_S4 within the BIOS scratchpad */
+	U2 dst_offset;    /* offset of dst V3_S4 within the BIOS scratchpad */
+};
+typedef Struct_(RegUse_build_normalize_v3s4) {
+	union { Reg_(V3_S4) res, src; };
+	union { Reg t0, src_ptr, mac2; };
+	union { Reg t1, dst_ptr; };
+	union { Reg t2, dst_offset, mac1, v_sqr_aligned; };
+	union { Reg t3, src_offset, btarget, shift_count, sqrtbl_index; };
+	union { Reg t4, mac3, v_sqr_sum, half_shift_tmp, inv_len; };
+	union { Reg t5, lzcr, half_shift; };
+};
+/* ─── Full normalize (all 4 stages inline) ───
+ * Generic 4-stage GTE normalize (SQR → sum+LZCR → align+sqrtbl → GPF+srav). */
+internal MipsAtom* build_normalize_v3s4(AtomArena_R aa, RegUse_build_normalize_v3s4 r)
+MipsAtom_Proc_(aa, {
+	load_half(r.src_offset, R_TapePtr, O_(Binds_NormalizeV3S4, src_offset)),
+	load_half(r.dst_offset, R_TapePtr, O_(Binds_NormalizeV3S4, dst_offset)),
+	LdSlot_ add_u(r.src_ptr, R_ScratchBase, r.src_offset),
+	LdSlot_ add_u(r.dst_ptr, R_ScratchBase, r.dst_offset),
+	LdSlot_ add_ui_self(R_TapePtr, S_(Binds_NormalizeV3S4)),
+
+	mac_load_v3s4(r.src, r.src_ptr, 0),
+
+	/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */
+	LdSlot_ mac_gte_sqr_v3s4(r.src.x, r.src.y, r.src.z, LdSlot_ nop),
+
+	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. src_ptr is dead; reuse as mac2. */
+	mac_gte_mv_from_data_r_mac123(r.mac1, r.mac2, r.mac3), LdSlot_ nop,
+	add_u_self(        r.v_sqr_sum, r.mac1),
+	add_u_self(        r.v_sqr_sum, r.mac2),
+	gte_mv_to_data_r(  r.v_sqr_sum, C2_LZCS), GteDelay_ nop2,
+	gte_mv_from_data_r(r.lzcr,      C2_LZCR), GteDelay_ nop,
+
+	/* Stage 3: even(LZCR), half-shift, align |v|² to bit 24. */
+	mac_lzcr_round_even_half_shift(r.lzcr, r.v_sqr_sum, r.v_sqr_aligned),
+	add_si(        r.btarget, r.lzcr, -24),
+	branch_lt_zero(r.btarget, atom_offset(aligned_done, srav_path)), BdSlot_ nop, /* bltz → srav_path (LZCR <  24 path) */
+		jump_rel(atom_offset(srav_path, aligned_done)),                             /* b → aligned_done (LZCR >= 24 path) */
+		BdSlot_ shift_lleft_var(r.v_sqr_aligned, r.v_sqr_aligned, r.btarget),
+		atom_label(srav_path)
+			li_s( r.shift_count, 24),
+			sub_s(r.shift_count, r.shift_count, r.lzcr),
+			shift_aright_var(r.v_sqr_aligned, r.v_sqr_aligned, r.shift_count),
+	atom_label(aligned_done)
+		or_u(r.half_shift, r.half_shift_tmp, 0),
+		add_si(     r.v_sqr_aligned, r.v_sqr_aligned, -64),
+		shift_lleft(r.v_sqr_aligned, r.v_sqr_aligned, 1),
+		mac_load_word_imm(r.sqrtbl_index, & gte_normalize_sqr_tbl), add_u_self(r.sqrtbl_index, r.v_sqr_aligned),
+		load_half(r.inv_len, r.sqrtbl_index, 0),
+		LdSlot_ nop,
+
+	mac_gte_general_purpose_interopolation(r.inv_len,
+		r.src.x, r.src.y, r.src.z,
+		r.res.x, r.res.y, r.res.z,
+		GteDelay_ load_word(R_AtomJmp, R_TapePtr, 0), LdSlot_   // ac_yield: word 1
+		GteDelay_ add_ui_self(         R_TapePtr, S_(MipsCode)) // ac_yield: word 2
+	),
+	mac_shift_aright_var_v3s4_self(r.res, r.half_shift),
+	mac_store_v3s4(r.res, r.dst_ptr, 0),
+
+	jump_reg(R_AtomJmp), BdSlot_ nop // ac_yield: word 3-4
+	// mac_yield()
+})
+
+/* ─── GTE OP cross product (a × b → out) ───
+ * Generalized V3_S4 cross product via GTE OP (OuterProduct12 libpsyx convention).
+ * The >> 12 shift converts S12.20 → S12.0 OuterProduct12. */
+typedef Struct_(Binds_gte_cross_v3s4) { V3_S4* src_a; V3_S4* src_b; V3_S4* out; };
+typedef Struct_(RegUse_gte_cross_v3s4) {
+	Reg_(V3_S4) a;
+	Reg_(V3_S4) b;
+	union { Reg out, t0; } x;
+	union { Reg src_a, t1, rt11; } y;
+	union { Reg src_b, t2, rt22; } z;
+};
+internal MipsAtom* gte_cross_v3s4(AtomArena_R aa, RegUse_gte_cross_v3s4 r)
+atom_info(atom_bind(Binds_gte_cross_v3s4)) MipsAtom_Proc_(aa, {
+	load_word(r.y.src_a,  R_TapePtr, O_(Binds_gte_cross_v3s4,src_a)),
+	load_word(r.z.src_b,  R_TapePtr, O_(Binds_gte_cross_v3s4,src_b)),
+	load_word(r.x.out,    R_TapePtr, O_(Binds_gte_cross_v3s4,out)),
+	LdSlot_ add_ui_self(  R_TapePtr, S_(Binds_gte_cross_v3s4)),
+
+	mac_load_v3s4(r.a, r.y.src_a, 0), LdSlot_
+	mac_load_v3s4(r.b, r.z.src_b, 0), LdSlot_
+	mac_gte_op_cross_v3s4(r.a, r.b), /* RT diagonal + IR + OP + MAC read + shift */
+	mac_store_v3s4(r.a, r.x.out, 0),
+
+	mac_yield()
+})
+#pragma endregion Atom Procs
+
+#pragma region Baked Atoms
+
+typedef Struct_(Binds_SetGteMT3S2S4) {
+	MT3_S2S4* transform;
+};
+internal MipsAtom_(set_gte_mt3s2s4) atom_info(
+		atom_bind(Binds_SetGteMT3S2S4)
+	,	atom_reads(R_TapePtr)
+){
+	/* Pop matrix address from tape into R_T3 ($11) */
+	load_word(R_T3, R_TapePtr, O_(Binds_SetGteMT3S2S4,transform)),
+	add_ui_self(    R_TapePtr, S_(Binds_SetGteMT3S2S4)),
+	/* Load 3x3 Rotation + 3x1 Translation from R_T3 into GTE CONTROL Regs (ctc2) */
+	load_word(R_T0, R_T3, 0), 
+	load_word(R_T1, R_T3, 4), 
+	gte_mv_to_ctrl_r(R_T0, gte_cr_RT11),
+	gte_mv_to_ctrl_r(R_T1, gte_cr_RT12),
+	load_word(R_T0, R_T3,  8),
+	load_word(R_T1, R_T3, 12),
+	load_word(R_T2, R_T3, 16),
+	gte_mv_to_ctrl_r(R_T0, gte_cr_RT13),
+	gte_mv_to_ctrl_r(R_T1, gte_cr_RT21),
+	gte_mv_to_ctrl_r(R_T2, gte_cr_RT22),
+	load_word(R_T0, R_T3, 20),
+	load_word(R_T1, R_T3, 24),
+	load_word(R_T2, R_T3, 28),
+	gte_mv_to_ctrl_r(R_T0, gte_cr_TRX),
+	gte_mv_to_ctrl_r(R_T1, gte_cr_TRY),
+	gte_mv_to_ctrl_r(R_T2, gte_cr_TRZ),
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/duffle/gte.h

@@ -16,10 +16,6 @@
  *            gte_mv_to_data_r   (gte + mv + to + data + register)
  *            gte_lw_v0_xy(base) (gte + lw + v0 + xy)
  *            load_upper_i       (load-upper + immediate, unique verb)
- *
- *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are
- *  NOT in this header. They live in the opt-in `gte_vendor_sym.h` for
- *  users who prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -34,20 +30,16 @@
  *  gte.h — Geometry Transformation Engine (COP2) for the PS1
  * ============================================================================
  *
- *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word`
- *  constants from C. No GCC inline-assembly string syntax in the code body.
+ *  Hand-rolled DSL for emitting GTE/MIPS instruction words from C.
+ *  No GCC inline-assembly string syntax in the code body.
  *
  *  STYLE NOTES
  *  -----------
- *  - Per-field encoders are named `enc_gte_<field>(value)` and each one
- *    self-masks its argument before shifting. Mirrors the `enc_op / enc_rs
- *    / enc_rt / ...` family in mips.h.
- *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of
- *    the per-field encoders, plus the COP2/CO base.
- *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined
- *    for the common cases so call sites read like assembly source.
- *  - All register/field values are enums (not `#define`s) so they show up
- *    in debugger symbol tables and IDE autocomplete.
+ *  - Per-field encoders are named `enc_gte_<field>(value)` and each one self-masks its argument before shifting.
+ *    Mirrors the `enc_op / enc_rs / enc_rt / ...` family in mips.h.
+ *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of the per-field encoders, plus the COP2/CO base.
+ *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined for the common cases so call sites read like assembly source.
+ *  - All register/field values are enums (not `#define`s) so they show up in debugger symbol tables and IDE autocomplete.
  *
  *  SEE ALSO
  *  --------
@@ -58,8 +50,7 @@
 
 /* --- GTE Data Registers (Coprocessor 2) ---
  * Preprocessor-visible integer ids for the COP2 data register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
  * Same pattern as the GPR `_Code` set in mips.h. */
 #define C2_VXY0_Code  0
 #define C2_VZ0_Code   1
@@ -107,20 +98,20 @@ enum {
 
 /* Semantic Aliases for GTE Data Registers */
 enum {
-	gte_in_v0_xy     = C2_VXY0, /* Input Vector 0 (X, Y) */
-	gte_in_v0_z      = C2_VZ0,  /* Input Vector 0 (Z) */
-	gte_in_v1_xy     = C2_VXY1, /* Input Vector 1 (X, Y) */
-	gte_in_v1_z      = C2_VZ1,  /* Input Vector 1 (Z) */
-	gte_in_v2_xy     = C2_VXY2, /* Input Vector 2 (X, Y) */
-	gte_in_v2_z      = C2_VZ2,  /* Input Vector 2 (Z) */
-	gte_in_rgb       = C2_RGB,  /* Input Color (R, G, B, MipsCode) */
-	gte_out_scr_xy0  = C2_SXY0, /* Output Screen Coord 0 (X, Y) */
-	gte_out_scr_xy1  = C2_SXY1, /* Output Screen Coord 1 (X, Y) */
-	gte_out_scr_xy2  = C2_SXY2, /* Output Screen Coord 2 (X, Y) */
-	gte_out_depth    = C2_OTZ,  /* Output Ordering Table Z (Depth) */
-	gte_math_accum0  = C2_MAC0, /* Math Accumulator 0 */
-	gte_math_accum1  = C2_MAC1, /* Math Accumulator 1 */
-	gte_math_accum2  = C2_MAC2, /* Math Accumulator 2 */
+	C2_InV0_XY    = C2_VXY0, /* Input Vector 0 (X, Y) */
+	C2_InV0_Z     = C2_VZ0,  /* Input Vector 0 (Z) */
+	C2_InV1_XY    = C2_VXY1, /* Input Vector 1 (X, Y) */
+	C2_InV1_Z     = C2_VZ1,  /* Input Vector 1 (Z) */
+	C2_InV2_XY    = C2_VXY2, /* Input Vector 2 (X, Y) */
+	C2_InV2_Z     = C2_VZ2,  /* Input Vector 2 (Z) */
+	C2_In_RGB     = C2_RGB,  /* Input Color (R, G, B, MipsCode) */
+	C2_OutSrc_XY0 = C2_SXY0, /* Output Screen Coord 0 (X, Y) */
+	C2_OutSrc_XY1 = C2_SXY1, /* Output Screen Coord 1 (X, Y) */
+	C2_OutSrc_XY2 = C2_SXY2, /* Output Screen Coord 2 (X, Y) */
+	C2_OutDepth   = C2_OTZ,  /* Output Ordering Table Z (Depth) */
+	C2_MathAccu0  = C2_MAC0, /* Math Accumulator 0 */
+	C2_MathAccu1  = C2_MAC1, /* Math Accumulator 1 */
+	C2_MathAccu2  = C2_MAC2, /* Math Accumulator 2 */
 };
 
 /* --- GTE Command Semantics (The Bitfield Meanings) --- 
@@ -167,6 +158,8 @@ enum {
 	gte_cmd_nclip     = 0x06, /* Normal Clipping (Backface culling) */
 	gte_cmd_op        = 0x0C, /* Outer Product */
 	gte_cmd_mvmva     = 0x12, /* Matrix Vector Multiply & Add (Custom math) */
+	gte_cmd_sqr       = 0x28, /* Square vector — MAC[i] = IR[i]²; IR[i] ← MAC[i] saturated */
+	gte_cmd_gpf       = 0x3D, /* General-purpose Interpolation — MAC[i] = IR0 * IR[i]  */
 
 /* --- GTE Command Bit-Field Layout ---
  * A GTE command word (sent to COP2 with RS=1) is laid out as:
@@ -177,25 +170,46 @@ enum {
  *   +------------+--+-----+------+------+------+------+---+--------+----------+
  *                                    \_____ GTE_PAYLOAD _____/       \__ GTE_CMD __/
  *
- * Shifts/masks below are the *bit positions* and *bit widths* of each
- * configurable field, used by the ENC_GTE_CMD encoder. 
+ * Shifts/masks below are the *bit positions* and *bit widths* of each configurable field, used by the ENC_GTE_CMD encoder. 
  * Mirrors the OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
  */
 
-	gte_shift_sf  = 19,  gte_width_sf  = 1,  gte_mask_sf  = 0x1,
-	gte_shift_mx  = 17,  gte_width_mx  = 2,  gte_mask_mx  = 0x3,
-	gte_shift_v   = 15,  gte_width_v   = 2,  gte_mask_v   = 0x3,
-	gte_shift_cv  = 13,  gte_width_cv  = 2,  gte_mask_cv  = 0x3,
-	gte_shift_lm  = 10,  gte_width_lm  = 1,  gte_mask_lm  = 0x1,
-	gte_shift_cmd =  0,  gte_width_cmd = 6,  gte_mask_cmd = 0x3F,
+	gte_shift_sf  = 19,  gte_width_sf  = 1,
+	gte_shift_mx  = 17,  gte_width_mx  = 2,
+	gte_shift_v   = 15,  gte_width_v   = 2,
+	gte_shift_cv  = 13,  gte_width_cv  = 2,
+	gte_shift_lm  = 10,  gte_width_lm  = 1,
+	gte_shift_cmd =  0,  gte_width_cmd = 6,
+
+	/* Fake command number (bits 24-20) — IGNORED by the GTE hardware per PSX-SPX `geometrytransformationenginegte.md` line 48.
+	 * libgte's compiler emits non-zero values in this field as a disassembly signature. */
+	gte_shift_fake_cmd = 20,
+	gte_width_fake_cmd = 5,
 };
 
+/* --- GTE Control Register Aliases (Pitfall 1) ---
+ * Three pairs of aliases map to the C2 control-register slot:
+ *   C2[24] = gte_cr_RBK (background R)         | gte_cr_OFX (screen offset X)
+ *   C2[25] = gte_cr_GBK (background G)         | gte_cr_OFY (screen offset Y)
+ *   C2[26] = gte_cr_BBK (background B)         | gte_cr_H   (projection plane distance H)
+ * Cross-alias writes inside one atom body, or across the wave-context boundary, silently clobber each other.
+ * The metaprogram's check_gte_cr_alias_writes (CHECK_RULES row) warns about each pair per source.
+ * See psx-spx docs/gte_reference.md §"Control-register alias table" for the silicon rationale and the libgte outer-product convention.
+ */
+
+/* --- RT-matrix packed-slot convention (Pitfall 4) ---
+ * The silicon packs two 16-bit RT elements per 32-bit C2 slot:
+ *   C2[2] = (RT22 << 16) | RT13       (gte_cr_RT13 writes the low  half, gte_cr_RT22 writes the high half)
+ *   C2[4] = (RT33 << 16) | RT22       (gte_cr_RT22 writes the low  half — clobbers prior RT22 value if RT13 was also written)
+ * OP and MVMVA read D1/D2/D3 from these packed slots.
+ * The libgte outer-product convention (see ac_apply_matrix_lv at gte.atom.c:108-122) writes C2[2] then C2[4] in sequence;
+ * the SECOND write's low half is RT22, not RT13. 
+ */
+
 /* --- GTE Control Register Indices (for ctc2/cfc2) ---
  * Preprocessor-visible integer ids for the COP2 control register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
- * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23
- * are reserved/unused on real hardware, so there's a gap. */
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23 are reserved/unused on real hardware, so there's a gap. */
 #define gte_cr_RT11_Code  0
 #define gte_cr_RT12_Code  1  /* packed with RT13 in bits 16..31 */
 #define gte_cr_RT13_Code  2  /* packed with RT22 in bits 16..31 */
@@ -223,8 +237,9 @@ enum {
 #define gte_cr_RFC_Code  27
 #define gte_cr_GFC_Code  28
 #define gte_cr_BFC_Code  29
-#define gte_cr_OFX_Code  30
-#define gte_cr_OFY_Code  31
+#define gte_cr_OFX_Code  24
+#define gte_cr_OFY_Code  25
+#define gte_cr_H_Code    26
 
 enum {
 	gte_cr_RT11 = gte_cr_RT11_Code, gte_cr_RT12 = gte_cr_RT12_Code, gte_cr_RT13 = gte_cr_RT13_Code,
@@ -246,21 +261,16 @@ enum { _C2_OPS_ = 0
 
 /* COP2 transfer sub-opcodes (5-bit field in the `rs` slot of enc_gte_tx).
  *
- * Spans the 2x2 {From, To} × {Data, Control} register classes that the
- * GTE exposes:
- *
+ * Spans the 2x2 {From, To} × {Data, Control} register classes that the GTE exposes:
  *   bit 1 (0x02): register class — 0 = data,  1 = control
  *   bit 2 (0x04): direction      — 0 = read,  1 = write
  *
- * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit
- * numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
- * (which target the data register file on any coprocessor). They are
- * re-aliased here so the four-way table reads like the spec mnemonics
- * (MFC2 / CFC2 / MTC2 / CTC2) and so the encoding lives next to its
- * only consumer (this header).
+ * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit numbers as general MIPS `cop_mf` / `cop_mt` defined in mips.h
+ * (which target the data register file on any coprocessor).
+ * They are re-aliased here so the four-way table reads like the spec mnemonics (MFC2 / CFC2 / MTC2 / CTC2)
+ * and so the encoding is next to its only consumer (this header).
  *
- * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2)
- * live in gte_vendor_sym.h. */
+ * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2) live in gte_vendor_sym.h. */
 enum { _C2_TX_SUBS_ = 0
 	, sub_mfc2 = 0x00 /* MFC2: Move From Coprocessor 2 data   reg */
 	, sub_cfc2 = 0x02 /* CFC2: Copy From Coprocessor 2 ctrl   reg */
@@ -270,11 +280,11 @@ enum { _C2_TX_SUBS_ = 0
 
 /* COP2 (GTE) Transfer Format: mfc2 / cfc2 / mtc2 / ctc2 rt, rd
  * Layout: [op_cop2:6][sub:5][rt:5][rd:5][0:11]
- *   - sub: one of sub_mfc2 / sub_cfc2 / sub_mtc2 / sub_ctc2
- *   - rt:  GPR source/dest
- *   - rd:  COP2 register index (0..31):
- *            data class → C2_VXY0_Code..C2_LZCR_Code (gte_in_v0_xy..gte_math_accum2 aliases)
- *            ctrl class → gte_cr_RT11_Code..gte_cr_OFY_Code */
+ *  - sub: one of sub_mfc2 / sub_cfc2 / sub_mtc2 / sub_ctc2
+ *  - rt:  GPR source/dest
+ *  - rd:  COP2 register index (0..31):
+ *    data class → C2_VXY0_Code..C2_LZCR_Code (gte_in_v0_xy..gte_math_accum2 aliases)
+ *    ctrl class → gte_cr_RT11_Code..gte_cr_OFY_Code */
 #define enc_gte_tx(sub, rt, rd) (enc_op(op_cop2) | enc_rs(sub) | enc_rt(rt) | enc_rd(rd))
 
 
@@ -282,8 +292,7 @@ enum { _C2_TX_SUBS_ = 0
 // #define gte_mv_from_data_r(rt, rd) enc_gte_tx(cop_mf, (rt), (rd)) /* Move GTE Control Register (rd) to GPR (rt) */
 
 /* GTE Data vs Control Register Transfers
- *
- * Each macro emits a single .word constant for one of MFC2/CFC2/MTC2/CTC2.
+ * Each macro emits a single instruction for one of MFC2/CFC2/MTC2/CTC2.
  *
  * `rd` is the C2 register index in the file the sub-opcode names:
  *   gte_mv_from_data_r / gte_mv_to_data_r  →  C2 data register file
@@ -298,14 +307,14 @@ enum { _C2_TX_SUBS_ = 0
 #define gte_mv_from_ctrl_r(rt, rd) enc_gte_tx(sub_cfc2, (rt), (rd)) /* Copy From ctrl reg */
 #define gte_mv_to_data_r(rt, rd)   enc_gte_tx(sub_mtc2, (rt), (rd)) /* Move To   data reg */
 #define gte_mv_to_ctrl_r(rt, rd)   enc_gte_tx(sub_ctc2, (rt), (rd)) /* Copy To   ctrl reg */
+#define GteDelay_ // Annotate an instruction as filling a CPU <-> GTE DMA delay slot/s
 
 /* COP2 Data Load (lwc2): `lwc2 rt, off(rs)`
  * Layout: [op_lwc2:6][rs:5][rt:5][imm:16]
  *   - rs: GPR base address
  *   - rt: COP2 data register index (0..31)
  *   - imm: signed 16-bit offset
- * NOTE: When `rs` is a runtime register, the encoding cannot be pre-baked
- * into a .word — use the string-style `gte_load_v0` macro below instead. */
+ * NOTE: When `rs` is a runtime register, the encoding cannot be pre-baked into a .word — use the string-style `gte_load_v0` macro below instead. */
 #define enc_gte_lw(rt, base, off) enc_i(op_lwc2, (base), (rt), (off))
 /* Store Word */
 #define enc_gte_sw(rt, base, off) enc_i(op_swc2, (base), (rt), (off))
@@ -314,31 +323,30 @@ enum { _C2_TX_SUBS_ = 0
  * `swc2` is redundant when we're already inside the `gte_` namespace.
  *   gte_lw rt, base, off  →  lwc2 rt, off(base)
  *   gte_sw rt, base, off  →  swc2 rt, off(base)
- * For the typical user-facing vector-level load (xy + z as two
- * instructions), use the higher-level `gte_load_vN` macros below. */
+ * For the typical user-facing vector-level load (xy + z as two instructions), use the higher-level `gte_load_vN` macros below. */
 #define gte_lw(rt, base, off) enc_gte_lw(rt, base, off)
 #define gte_sw(rt, base, off) enc_gte_sw(rt, base, off)
 
 /* GTE Command Format
  * Opcode is always MIPS_OP_COP2, RS is always 1 (CO).
- * The lower 25 bits are the GTE-specific command payload.
+ * Lower 25 bits are GTE-specific command payload.
  *
- * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs`
- * pattern in mips.h: each one self-masks and shifts its own field, so a
- * caller can build up a GTE command piece by piece (handy for state-driven
- * MVMVA emitters that vary one field at a time).
+ * The `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs` pattern in mips.h:
+ * Each one self-masks and shifts its own field, so a caller can build up a GTE command piece by piece
+ * (handy for state-driven MVMVA emitters that vary one field at a time).
  *
- * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command
- * word in one go. It just ORs the per-field encoders together. */
+ * `ENC_GTE_CMD` is an all-in-one convenience for emitting a full command word.
+ * It just ORs the per-field encoders together. */
 #define gte_cmd_base (enc_op(op_cop2) | (1 << 25))
 
 /* Per-field encoders. Each one does (value & mask) << shift on its own. */
-#define enc_gte_sf(sf)   (((sf)  & gte_mask_sf ) << gte_shift_sf )
-#define enc_gte_mx(mx)   (((mx)  & gte_mask_mx ) << gte_shift_mx )
-#define enc_gte_v(v)     (((v)   & gte_mask_v  ) << gte_shift_v  )
-#define enc_gte_cv(cv)   (((cv)  & gte_mask_cv ) << gte_shift_cv )
-#define enc_gte_lm(lm)   (((lm)  & gte_mask_lm ) << gte_shift_lm )
-#define enc_gte_cmd(cmd) (((cmd) & gte_mask_cmd) << gte_shift_cmd)
+#define enc_gte_sf(sf)      ((sf)  << gte_shift_sf )
+#define enc_gte_mx(mx)      ((mx)  << gte_shift_mx )
+#define enc_gte_v(v)        ((v)   << gte_shift_v  )
+#define enc_gte_cv(cv)      ((cv)  << gte_shift_cv )
+#define enc_gte_lm(lm)      ((lm)  << gte_shift_lm )
+#define enc_gte_cmd(cmd)    ((cmd) << gte_shift_cmd )
+#define enc_gte_fake_cmd(x) ((x)   << gte_shift_fake_cmd)
 
 /* Composite: all six GTE fields + the COP2/CO base. */
 #define enc_gte_cmdw(sf, mx, v, cv, lm, cmd) ( \
@@ -359,12 +367,13 @@ enum { _C2_TX_SUBS_ = 0
  * Decomposition (per the `enc_gte_<field>` definitions above):
  *   gte_cmdw_<name> = gte_cmd_base | enc_gte_cmd(<cmd>)
  
- * The SF/MX/V/CV/LM fields are all zero in the common cases 
+ * The SF / MX / V / CV / LM fields are all zero in the common cases 
  * (standard rotation-matrix, no scaling factor, V0 vector, translation vector, no clamp),
  * so the only varying bits are the `cmd` field.
  *
- * Naming follows the file's convention: `gte_cmd_*` is the raw 6-bit `cmd` field id, `gte_cmdw_*`
- *  is the fully-encoded 32-bit instruction word ready to drop into a `.word` directive.
+ * Naming convention:
+ * - `gte_cmd_*` : Raw 6-bit `cmd` field id 
+ * - `gte_cmdw_* : 32-bit instruction word ready to drop into a `.word` directive.
  *
  * --------------------------------------------------------------------------
  *  PsyQ-compatibility note (RTPS/RTPT):
@@ -375,12 +384,12 @@ enum { _C2_TX_SUBS_ = 0
  *  (the perspective divide happens regardless of `sf`).
  *
  *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits clear),
- *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops — 
- *  the floor's screen coordinates come out as raw projection-of-rotation (Z never divided),
+ *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops.
+ *  The floor's screen coordinates come out as raw projection-of-rotation (Z never divided), 
  *  `nclip` ends up wrong, and the triangle is culled.
  *
- *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to match the working bit pattern everyone has shipped for 25 years.
- *  NCLIP/OP/MVMVA stay spec-clean — their reserved bits really are zero in the original PsyQ source.
+ *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to match the working bit pattern.
+ *  NCLIP / OP / MVMVA stay spec-clean — their reserved bits really are zero in the original PsyQ source.
  * --------------------------------------------------------------------------
  */
 #define gte_cmdw_psyq_compat  (1u << 21 | enc_gte_sf(gte_sf_integer))
@@ -389,12 +398,91 @@ enum { _C2_TX_SUBS_ = 0
 #define gte_cmdw_rtpt   (gte_cmd_base | enc_gte_cmd(gte_cmd_rtpt ) | gte_cmdw_psyq_compat)
 #define gte_cmdw_nclip  (gte_cmd_base | enc_gte_cmd(gte_cmd_nclip))
 #define gte_cmdw_op     (gte_cmd_base | enc_gte_cmd(gte_cmd_op   ))
-#define gte_cmdw_outer_product gte_cmdw_op      /* "outer product" -- NOCASH/Sdk terminology */
-#define gte_cmdw_wedge        gte_cmdw_op      /* "wedge product"  -- geometric-algebra terminology */
+#define gte_cmdw_outer_product gte_cmdw_op      /* "outer product" -- PSY-Q terminology */
+#define gte_cmdw_wedge         gte_cmdw_op      /* "wedge product"  -- geometric-algebra terminology. */
+#define gte_cmdw_cross         gte_cmdw_op      /* "cross product"  -- geometric-algebra terminology.
+	* RGA(Lengyel): The GTE OP is a 3D signed-16-bit D x IR cross, not a generic RGA exterior product.
+	* The wedge alias is a 3D complement interpretation of the same 3 scalars (MAC1..MAC3). */
 #define gte_cmdw_mvmva  (gte_cmd_base | enc_gte_cmd(gte_cmd_mvmva))
 
+/* MVMVA with sf=0 (no shift, full-integer), cv=3 (no translation), v=3 (IR vector input).
+ * Reads input from IR1/2/3 (loaded via mtc2 rt, C2_IRx). MAC1/2/3 = RT row · IR (full product, no >>12).
+ * Per PSX-SPX: SAR (sf*12) with sf=0 = SAR 0 = no shift. */
+#define gte_cmdw_mvmva_sf0_ir  (gte_cmd_base | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA with sf=1 (>>12 shift, 4.12 fixed-point), cv=3 (no translation), v=3 (IR): for ApplyMatrixLV.
+ * Reads input from IR1/2/3 (loaded via mtc2 rt, C2_IRx). MAC1/2/3 = (RT row · IR) >> 12.
+ * Per PSX-SPX: SAR (sf*12) with sf=1 = SAR 12 = arithmetic right-shift by 12.
+ * This matches the libgte C-side ApplyMatrixLV output (R*pos >> 12). */
+#define gte_cmdw_mvmva_ir    (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=0, mx=3 (Light matrix), v=3 (IR), cv=3 (no TR).
+ * For pass1 of the C11 two-pass decomposition. Reads L matrix.
+ * Since L matrix is typically zero, pass1 contributes 0 to the combine. */
+#define gte_cmdw_mvmva_sf0_mx3_v3_cv3  (gte_cmd_base | enc_gte_sf(0) | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=1 (>>12), mx=3 (Light matrix), v=2 (V0), cv=0 (with TR).
+ * Matches the C11 ApplyMatrixLV pass 2 command word (0x49E012) exactly.
+ * The combine is (pass1 << 3) + pass2. */
+#define gte_cmdw_mvmva_pass2_c11  (gte_cmd_base | enc_gte_sf(1) | enc_gte_v(2) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=0, mx=3, v=2, cv=0. Matches the C11 pass 1 command. */
+#define gte_cmdw_mvmva_pass1_c11  (gte_cmd_base | enc_gte_v(2) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+#define gte_cmdw_mvmva_no_tr gte_cmdw_mvmva_ir
+
+/* MVMVA pass 2 — C11 ApplyMatrixLV command.
+ * Decoded: op_cop2 | CO | fake_cmd=4 | sf=1 (>>12) | mx=0 (RT matrix) | v=3 (IR) | cv=3 (no translation) | lm=0 | cmd=MVMVA.
+ * Reads (RT row · IR) >> 12 into MAC1/2/3. Per-field composition (no opaque literal)
+ * keeps the bit layout visible at the call site + matches the libgte C-side byte-exact. */
+#define gte_cmdw_mvmva_c11_pass2  (gte_cmd_base | enc_gte_fake_cmd(4) | enc_gte_sf(1) | enc_gte_v(3) | enc_gte_mx(0) | enc_gte_cv(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=1 (>>12), mx=0 (RT matrix), v=0 (V0), cv=3 (no TR). */
+#define gte_cmdw_mvmva_sf1_mx0_v0_cv3  (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_v(0) | enc_gte_mx(0) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* RTPS with sf=1 (12-bit shift, no translation): matches the output of libgte's ApplyMatrixLV when the GTE pipeline expects R*pos >> 12. 
+ * The shift produces values like (-270, 710, 1713) which match the C11 reference path. */
+#define gte_cmdw_rtps_sf1   (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_cmd(gte_cmd_rtps))
+
+/* SQR / GPF cosmetic-bits compat helpers.
+ * Each command's `_compat` macro ORs in the `fake_cmd` field value libgte happens to emit.
+ * The hardware ignores these bits (per PSX-SPX line 48). */
+#define gte_cmdw_sqr_fake_sig  enc_gte_fake_cmd(0x0A)
+#define gte_cmdw_gpf_fake_sig  enc_gte_fake_cmd(0x19)
+
+/* SQR — Square Vector.
+ * PSX-SPX `geometrytransformationenginegte.md` §"SQR":
+ *   [MAC1,MAC2,MAC3] = [IR1*IR1, IR2*IR2, IR3*IR3] SHR (sf*12)
+ *   [IR1,IR2,IR3]    = [MAC1,MAC2,MAC3] (saturated to 0x7FFF when lm=1)
+ * Sourced verbatim from libgte msc02 VectorNormal disassembly at 0x800160b0:
+ *   0x4AA00428 = gte_cmd_base | gte_cmdw_sqr_compat | enc_gte_lm(1) | enc_gte_cmd(0x28)
+ *   bit 19     sf=0
+ *   bit 10     lm=1
+ *   bits 5-0   cmd=0x28=SQR
+ *   bits 24-20 = 0x0A (libgte "nonsense SDK command number" signature) */
+#define gte_cmdw_sqr   (gte_cmd_base | enc_gte_cmd(gte_cmd_sqr) | enc_gte_lm(1) | gte_cmdw_sqr_fake_sig)
+
+/* GPF — General-purpose Interpolation.
+ * PSX-SPX `geometrytransformationenginegte.md` §"GPF":
+ *   [MAC1,MAC2,MAC3] = (([IR1,IR2,IR3] * IR0) + [MAC1,MAC2,MAC3]) SAR (sf * 12)
+ *   [IR1,IR2,IR3]    = [MAC1,MAC2,MAC3]
+ * Sourced verbatim from libgte msc02 VectorNormal disassembly at 0x8001613c:
+ *   0x4B90003D = gte_cmd_base | gte_cmdw_gpf_compat | enc_gte_cmd(0x3D)
+ *   bit 19     sf = 0
+ *   bit 10     lm = 0
+ *   bits 5-0   cmd = 0x3D = GPF
+ *   bits 24-20 = 0x19 (libgte "nonsense SDK command number" signature) */
+#define gte_cmdw_gpf   (gte_cmd_base | enc_gte_cmd(gte_cmd_gpf) | gte_cmdw_gpf_fake_sig)
+
+/* Mask to round LZCR (leading-zero/ones count, range 1..32 per PSX-SPX cop2r31) down to even.
+ * The normalize_v3s4 half-shift logic computes (31 - LZCR) >> 1; clearing bit 0 ensures the subtraction result is always odd, so the >> 1 division is consistent (no 0.5 loss). */
+enum {
+	gte_lzcr_even_mask = 0xFFFE, /* all bits except bit 0 */
+};
+
 #define gte_cmdw_rotate_translate_perspective_single gte_cmdw_rtps
 #define gte_cmdw_rotate_translate_perspective_triple gte_cmdw_rtpt
+/* RGA(Lengyel): RTPS/RTPT consume the matrix expansion of a rigid transformation (rotation matrix + translation vector) loaded into the RT/TR control registers.
+ * For unitized points the same result equals the motor antiproduct; the GTE executes the LA form, not a symbolic antiproduct. */
 
 /* PsyQ compatibility bits for AVSZ3 (Bits 20, 22, 24 must be set) */
 #define gte_cmdw_psyq_avsz3_compat (0x15 << 20)
@@ -413,20 +501,16 @@ enum { _C2_TX_SUBS_ = 0
 
 /**
  * @brief Loads a single SVECTOR to GTE vector register V0
- *
  * @details Loads values from an SVECTOR struct to GTE data registers C2_VXY0
  * (XY at offset 0) and C2_VZ0 (Z at offset 4) using `lwc2`.
  *
- * Uses string-style GCC inline asm with `%0` substitution because the 
- * base register `r0` is a runtime GPR chosen by the compiler.
+ * Uses string-style GCC inline asm with `%0` substitution because the base register `r0` is a runtime GPR chosen by the compiler.
  * It cannot be encoded into a static `.word` constant.
  *
- * Usage:
- *   asm_gte_load_v0(svector_ptr);
+ * Usage: asm_gte_load_v0(svector_ptr);
  */
 
 /* lwc2 encoding helpers parameterized on the base GPR.
- *
  * gte_lw_v0_xy(base)  → lwc2 $0,  0(base)   ; C2_VXY0
  * gte_lw_v0_z(base)   → lwc2 $1,  4(base)   ; C2_VZ0
  * gte_lw_v1_xy(base)  → lwc2 $2,  0(base)   ; C2_VXY1
@@ -435,8 +519,7 @@ enum { _C2_TX_SUBS_ = 0
  * gte_lw_v2_z(base)   → lwc2 $5,  4(base)   ; C2_VZ2
  *
  * `base` is the GPR number to bake into the .word constant's `rs` field.
- * These are pure compile-time integers; the C compiler constant-folds
- * them into .word directives. */
+ * These are pure compile-time integers; the C compiler constant-folds them into .word directives. */
 
 enum {
 	GTE_Z_Offset = 4
@@ -450,7 +533,6 @@ enum {
 #define gte_lw_v2_z(base)   enc_gte_lw(gte_in_v2_z,  (base), GTE_Z_Offset)
 
 /* gte_load_vN(r_ptr, base) — placeholder-punned lwc2 loaders
- *
  * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen GTE vector register, where `<base>` is the GPR number you pass in
  * (typically one of R_T4..R_T9 for the standard "3-pointer" pattern).
  *
@@ -459,8 +541,8 @@ enum {
  *     gte_load_v0(p_in_12, R_T4);   // R_T4 = 12, base is $12
  *
  * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register `p_in_12` can live in),
- * which is exactly the register the .word constants expect. A `"$12"` clobber would conflict with the register-variable binding
- * ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
+ * which is exactly the register the .word constants expect.
+ * A `"$12"` clobber would conflict with the register-variable binding ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
  * The other ABI-clobbers ($2/$8/$9/$31) stay because the GTE instructions don't touch caller-saved GPRs but the kernel does treat them as volatile.
  *
  * WHICH REGISTER TO PICK
@@ -499,10 +581,8 @@ enum {
 
 /* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — prelude to gte_cmd_rtpt.
  *
- * Loads all three GTE input vectors (6 words) from three separate pointers,
- * one per GTE vector register, each loaded from its own base GPR.
+ * Loads all three GTE input vectors (6 words) from three separate pointers, one per GTE vector register, each loaded from its own base GPR.
  * Caller must bind each `pN` to `bN` via a register variable.
- *
  *   register V3_S2* p0 rgcc(R_T4) = verts[0].ptr;  // → __asm__("$12")
  *   register V3_S2* p1 rgcc(R_T5) = verts[1].ptr;  // → __asm__("$13")
  *   register V3_S2* p2 rgcc(R_T6) = verts[2].ptr;  // → __asm__("$14")
@@ -521,29 +601,20 @@ enum {
 
 /**
  * @brief Rotate, Translate and Perspective Triple (23 cycles)
- *
- * @details Performs rotation, translation and perspective calculation of three
- * vertices at once. The equation performed is the same as gte_rtps() only
- * repeated three times for each vertex. The result of the first vertex is
- * stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then
- * C2_SXY2.
+ * @details Performs rotation, translation and perspective calculation of three vertices at once.
+ * The equation performed is the same as gte_rtps() only repeated three times for each vertex.
+ * The result of the first vertex is stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then C2_SXY2.
  *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency — the GTE
- *      takes ~8 cycles per perspective divide, and the nops let any
- *      preceding lwc2/swc2 retire before RTPT starts reading its
- *      inputs from V0/V1/V2.
- *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the
- *      pre-baked encoders above) — `0x0280030` decoded as
- *      `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields
- *      zero (standard rotation, no scaling, V0 vector, translation
- *      vector, no clamp).
+ *   1. Two `nop` words fill the COP2 pipeline latency — the GTE takes ~8 cycles per perspective divide, 
+ *      and the nops let any preceding lwc2/swc2 retire before RTPT starts reading its inputs from V0/V1/V2.
+ *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the pre-baked encoders above) — 
+ *      `0x0280030` decoded as `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields zero
+ *      (standard rotation, no scaling, V0 vector, translation vector, no clamp).
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register — those have to be saved by the caller if
- * they need to survive across the call (RTPT writes SXY0..2, SZ0..3,
- * OTZ, MAC0..3, IR0..3, etc.).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI)
+ * plus the standard "memory" barrier. Does not clobber any COP2 data/control register — 
+ * those have to be saved by the caller if they need to survive across the call (RTPT writes SXY0..2, SZ0..3, OTZ, MAC0..3, IR0..3, etc.).
  */
 #define gte_rtpt()                        \
 	asm volatile(                           \
@@ -559,32 +630,24 @@ enum {
 
 /**
  * @brief Normal clipping (8 cycles)
- *
- * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for
- * backface culling. If the value of C2_MAC0 is negative, the coordinates are
- * inverted and thus the triangle is back facing.
+ * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for backface culling.
+ * If the value of C2_MAC0 is negative, the coordinates are inverted and thus the triangle is back facing.
  *
  * The following equation is performed when executing this GTE command:
- *
  *     MAC0 = SX0*SY1 + SX1*SY2 + SX2*SY0 - SX0*SY2 - SX1*SY0 - SX2*SY1
- *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency - the GTE
- *      pipeline takes a few cycles per op, and the nops let any
- *      preceding lwc2/swc2/RTPT retire before NCLIP starts reading
- *      its inputs from SXY0/SXY1/SXY2.
- *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the
- *      pre-baked encoders above) - `0x01400006` decoded as
- *      `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields
- *      zero. NCLIP is spec-clean in the original PsyQ source
- *      (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat`
- *      quirk), so `gte_cmdw_nclip` does NOT OR in any reserved bits.
+ *   1. Two `nop` words fill the COP2 pipeline latency
+ *      - the GTE pipeline takes a few cycles per op, and the nops let any preceding
+ *      lwc2/swc2/RTPT retire before NCLIP starts reading its inputs from SXY0/SXY1/SXY2.
+ *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the pre-baked encoders above)
+ *      - `0x01400006` decoded as  `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields zero.
+ *      NCLIP is spec-clean in the original PsyQ source (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat` quirk),
+ *      so `gte_cmdw_nclip` does NOT OR in any reserved bits.
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register - those have to be saved by the caller if
- * they need to survive across the call (NCLIP writes MAC0 only; it
- * is purely a sign-of-double-product computation on SXY0..2).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI) plus the standard "memory" barrier.
+ * Does not clobber any COP2 data/control register.
+ * Those have to be saved by the caller if they need to survive across the call (NCLIP writes MAC0 only;
+ * it is purely a sign-of-double-product computation on SXY0..2).
  */
 #define gte_nclip()                       \
 	asm volatile(                           \
@@ -610,14 +673,10 @@ enum {
 		"cop2 0x0158002D;")
 
 /* asm_gte_matrix_set_rotation(r0)
+ * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix control registers (RT11..RT22, indices 0..4) via ctc2.
  *
- * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix
- * control registers (RT11..RT22, indices 0..4) via ctc2.
- *
- * Memory layout at r0: five contiguous 32-bit words (offsets 0..16),
- * each holding two packed 16-bit matrix elements. The first 1.5 rows
- * of a standard PSX SDK MATRIX struct (where each row is laid out as
- * [RT_xx, RT_xy] | [RT_xz, pad] | ...).
+ * Memory layout at r0: five contiguous 32-bit words (offsets 0..16), each holding two packed 16-bit matrix elements.
+ * The first 1.5 rows of a standard PSX SDK MATRIX struct (where each row is laid out as [RT_xx, RT_xy] | [RT_xz, pad] | ...).
  *
  * Generated MIPS (mirrors the source macro):
  *   lw   $12,  0( %0 )    ; word 0
@@ -631,27 +690,22 @@ enum {
  *   ctc2 $13,  $3         ; → C2_RT21
  *   ctc2 $14,  $4         ; → C2_RT22
  *
- * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a
- * register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
- * instructions to read from the right base. The `"r"(r0)` constraint
- * alone doesn't force a specific GPR — it just lets GCC pick one.
- * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field
- * of each lw, so the lw instructions will only do the right thing
- * if $12/$13/$14 hold the matrix base at runtime.
+ * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
+ * instructions to read from the right base. The `"r"(r0)` constraint alone doesn't force a specific GPR — it just lets GCC pick one.
+ * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field of each lw, so the lw instructions will 
+ * only do the right thing if $12 / $13 / $14 hold the matrix base at runtime.
  *
  *   M3_S2* m = ...;
  *   register M3_S2* m_in_12 rgcc(R_T4) = m;
  *   asm_gte_matrix_set_rotation(m_in_12);
  *
- * We clobber $12/$13/$14 (the ones we use as scratch inside the
- * inline asm) plus the system clobbers; we don't clobber `r0` because
- * the `rgcc` binding already says "this variable lives in $12".
+ * We clobber $12/$13/$14 (the ones we use as scratch inside the inline asm)
+ * plus the system clobbers; we don't clobber `r0` because the `rgcc` binding already says "this variable lives in $12".
  *
- * WARNING: Incomplete by design. The source macro only writes RT11..RT22
- * (5 of 9 rotation elements); RT23 and the entire RT3x row are left
- * untouched. Real libpsn00b SetRotMatrix writes all 9. Use only when the
- * GTE's remaining rotation entries are already correct, or you will
- * get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
+ * WARNING: Incomplete by design. The source macro only writes RT11..RT22 (5 of 9 rotation elements);
+ * RT23 and the entire RT3x row are left untouched.
+ * Real libpsn00b SetRotMatrix writes all 9. Use only when the GTE's remaining rotation entries are already correct,
+ * or you will get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
  */
 #define asm_gte_matrix_set_rotation(r0) \
 	asm volatile(                         \

code/duffle/lottes_tape.h

@@ -1,100 +1,257 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
 #	include "dsl.h"
 #	include "gcc_asm.h"
 # include "mips.h"
 # include "gte.h"
 # include "memory.h"
-#	include "atom_dsl.h"
-#	include "gen/duffle.macs.h"
-#	include "gen/duffle.offsets.h"
+#	include "dsl.atom.h"
 #endif
 
-typedef U4 const MipsCode;
-typedef Slice_(MipsCode);
-typedef Slice_MipsCode MipsAtom;
+#pragma region Tape Drive
+/* -----------------------------------------------------------------------------------------------------------
+ *  TAPE DRIVE ABI
+ * -----------------------------------------------------------------------------------------------------------
+ * Note(Ed): One of the main purposes of this codebase is to help me learn this, 
+ * as such the information below may not* be entirely realized or finalized conceptually.
+ * -----------------------------------------------------------------------------------------------------------
+ * This ABI and its associated legos were directly inspired by researching the work of 
+ * Timothy Lottes and Onat Türkçüoğlu; along with many others. It's the simplest bootstrap of a 
+ * directly executed chain of assemby arrays (Atoms) that terminate with a yield sequence to the next atom.
+ * These eventually lead to a terminal atom for the tape which is defined below as "tape_exit".
+ * 
+ * It behaves as one of the simplest runtime harnesses ontop of a host-enviornment's execution engine 
+ * to author and compose programs with. From here various conventions can be further applied. 
+ * To make things easier  to understand it may be better to focus on what this ABI does not have. 
+ * It does not have have any branching within the tape but relative branches within atoms or between atoms.
+ * Branching nearly is always downstream. Automatic stack usage is non-existent.
+ * Push/Pop, FIFO, or Arena/Bump data structures are used by atoms explicitly.
+ * In it's current form with the C11 macro DSL, the user also has fullfill manual register allocation per atom.
+ * 
+ * One of the remarkable things about utilizing this ABI is its essentially interopable with CPUs, GPUs, FPGA, 
+ * or, basically anything from the 5th generation consoles and onward.
+ * The ABI directly reflects how all computational hardware must be architected in order to execute 
+ * digital logic effectively on current era tech.
+ * On the PS1 we don't have access to a few features like multi-threading, speculative execution, or L3 cache;
+ * but, we can set the foundation for legoing whats required for eventually expanding this ABI's paradigm 
+ * and core atoms to take those newer hardware features into account. For example, you can easily expand 
+ * this to support wave-based execution model on a PS2 or PS3. Not having a stack or 
+ * automatic register allocation means the user cannot ignore excessive argument shuffle across workload or
+ * waves and thier phases. Crossing ABI boundaries to other runtimes that do has obviouss penalties.
+ * 
+ * Learning data-oriented code becomes a natural progression. Your not fighting a stack-based procedural 
+ * paradigm that wants to argument shuffle. There is no ambiguity due to the lack of constraints, for example,
+ * on how the user may "call" a procedure in traditional random dispatch runtimes. The user does have to
+ * hammer down "rules" or patterns for massaging the compiler to dissolve those call frames; just to get 
+ * the asesmbly into its desired form. The form is obvious, and once the user gets to author these compoonents
+ * it becomes a game of tetris.
+ * 
+ * Another feature is this ABI is very compatible with bootstrapping and developing simple toolchains built off
+ * of bit-packed annotated command streams the user can directly author, maintatain, and immediately execute.
+ * That being like a color forth, or maybe something more familar like an immediate mode library 
+ * for various systems such as GUIs. This can make the tetris less of a chore with some helpful policy
+ * generation for allocation of registers, helping to choose resuable components, designing DSL on the fly, etc.
+ * -----------------------------------------------------------------------------------------------------------
+ * TODO(Ed): We need pretty ascii diagrams and proper guides, articles, etc.
+ * -----------------------------------------------------------------------------------------------------------
+ * For now this ideation has just started functioning. I'm abusing C11 & a lua metaprogram to help establish 
+ * a hybrid toolchain to ideate on a traditional text-based authoring UX for this paradigm.
+ * If pcsx-redux provides viable hot-reload and persistent data storage beyond save-states
+ * (just copying ram to filesystem), I can author a color forth to mess around with.
+ * With either an editor in-emulator or on the actual machine itself. Assembly is tedius,
+ * but I think this codebase most likely has a pretty ergonomic flavor worst case...
+ * */
+/* Register Allocation Info */
+enum {
+	R_ScratchBase = R_SP atom_reg, /* Scratchpad base address (host frame top) */
+	R_AtomJmp     = R_FP atom_reg, /* Next atom target (yield handshake scratch) */
+	R_TapePtr     = R_RA atom_reg, /* The Instruction Stream Pointer */
+/* Stringification codes for the GCC inline assembler clobber lists. */
+#define R_ScratchBase_Code R_SP_Code
+#define R_AtomJmp_Code     R_FP_Code
+#define R_TapePtr_Code     R_RA_Code
 
+// R_InCursor = R_T4,
+// #define R_InCursor_Code    R_T4_Code
+
+// Reserved Registers (Callee-saved across the host ABI transition):
+// - R_SP: Holds the scratchpad base while tape code executes.
+// - R_FP: Holds the next atom target.
+// - R_RA: Holds the tape cursor.
+// All atom-body allocations must stay out of these.
+// Atom bodies may freely use R2-R25.
+
+// All allocatable registers for atom bodies (R2-R25, 24 registers):
+
+	R_PsuedoVolatile = R_AT, // Assembler temporary; never allocate.
+
+	// Atom Allocation Pool
+	R_Atom0  = R_T0,
+	R_Atom1  = R_T1,
+	R_Atom2  = R_T2,
+	R_Atom3  = R_T3,
+	R_Atom4  = R_T4,
+	R_Atom5  = R_T5,
+	R_Atom6  = R_T6,
+	R_Atom7  = R_T7,
+	R_Atom8  = R_T8,
+	R_Atom9  = R_T9,
+	R_Atom10 = R_V0, // Tend to be used with gte DMAs
+	R_Atom11 = R_V1, // Tend to be used with gte DMAs
+	R_Atom12 = R_A0,
+	R_Atom13 = R_A1,
+	R_Atom14 = R_A2,
+	R_Atom15 = R_A3,
+	R_Atom16 = R_S0,
+	R_Atom17 = R_S1,
+	R_Atom18 = R_S2,
+	R_Atom19 = R_S3,
+	R_Atom20 = R_S4,
+	R_Atom21 = R_S5,
+	R_Atom22 = R_S6,
+	R_Atom23 = R_S7,
+};
+
+typedef U2 Reg; // Register parameter used with atom or atom component procedures
+#define Reg_(type) tmpl(Reg,type) // Just a way to template register allocations of C-struct types.
+
+typedef U4 const MipsCode; // Underlying type to mips asm words.
+typedef Slice_(MipsCode);
+
+typedef U4 const MipsAtom; // Underlying type to a mips atom defnition 
+typedef Slice_(MipsAtom);
+
+// Sometimes a user will define a bundle of atoms that represent a procedure of work as:
+//   MipsAtom* <identifier>[...];
+// Unfortuantely if using slice_from_array it will make the slice's pointer: MipsAtom** so this enforce its defined as MipsAtom*
+// TODO(Ed): Alternatively we can make the MipsAtom an opaque pointer to the atom... so that the proc returns 'MipsAtom'.
+#define atombundle_from_array(array) (Slice_MipsAtom){.ptr=array[0],.len=Array_len(array)}
+
+// Underlying type to an ptr to an array of mips asm words that must terminate with an ac_yield.
 #define MipsAtom_(sym) MipsCode sym [] align_(4) =
 
+// Used for atoms with value-args
+//   internal MipsAtom* X_proc(AtomArena_R aa, args) MipsAtom_Proc_(X, aa, { body })
+// expands to:
+//   internal MipsAtom* X_proc(AtomArena_R aa, args) { MipsCode atom_comp_code[] align_(4) = { body }; return atomarena_push(aa, slice_from_array(MipsCode, atom_comp_code)); }
+// The atom name is derived by the Lua metaprogram from the preceding
+// `MipsAtom* X_proc(...)` declaration (backward walk from the macro site,
+// strips the `_proc` suffix).
+#define MipsAtom_Proc_(aa, ...) { MipsCode atom_comp_code[] align_(4) = __VA_ARGS__; return atomarena_push(aa, slice_from_array(MipsCode, atom_comp_code)); }
+
 // Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
 //   MipsAtomComp_(ac_X) { body }
 // expands to:
 //   MipsCode ac_X[] align_(4) = { body };
 #define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
 
-// Used for components with value-args (e.g., ac_format_f3_color).
-//   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+// Used for components with value-args (mandatory `ab` (atom-builder) arg).
+//   FI_ void ac_X(MipsAtomBuilder_R ab, args) MipsAtomComp_Proc_(ab, { body })
 // expands to:
-//   FI_ MipsAtom ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
-#define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
+//   FI_ void ac_X(MipsAtomBuilder_R ab, args) {
+//       MipsCode atom_comp_code[] align_(4) = { body };
+//       atombuilder_push(ab, slice_from_array(MipsCode, atom_comp_code));
+//   }
+// The body must NOT include mac_yield() (the parent atom yields).
+// The component name is derived by the Lua metaprogram from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration (backward walk from the macro site).
+// Inline-only callers (the generated `mac_<name>` aliases) skip the `ab` arg via metaprogram filtering; escape callers (ac_<name> invoked as a function) pass a long-lived builder.
+#define MipsAtomComp_Proc_(ab, ...) { MipsCode atom_comp_code[] align_(4) = __VA_ARGS__; atombuilder_push(ab, slice_from_array(MipsCode, atom_comp_code)); }
 
-// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
+// Used for trivial mappings from one atom component proc to the command of a more baser (meant for type-mapping)
+#define MipsAtomComp_ProcMap_(ab, base_command) atom_dbg_skip MipsAtomComp_Proc_(ab, {base_command })
 
-/* Register aliases */
-enum {
-	R_AtomJmp    = R_T9 atom_reg,  /* debug-visible; tape yield handshake scratch */
-	R_TapePtr    = R_T8 atom_reg,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,
+/* Line-table anchor: gcc only adds a file to the .debug_line file table when the contains line-numbered content.
+	Files containing only atoms and atom components.
+   Place `ATOM_FILE_LINE_MARKER();` once at file scope in any `.atom.c` that defines atoms.
+	 Macro expands to a file-scope `internal U4 const` declaration keeps the file in the line table.
+	 The constant is in `.rodata` so the linker may eliminate it. */
+#define ATOM_FILE_DEBUGGER_LINE_MARKER(file_name) internal U4 const tmpl(atom_file_debugger_line_marker,file_name) = 0
 
-	R_PrimCursor = R_T7 atom_reg atom_type(U4 *),    /* VRAM output cursor (primitive buffer) */
-	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2 *), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
-	R_VertBase   = R_T5 atom_reg atom_type(V3_S2 *), /* Base address of the vertex array */
-	R_OtBase     = R_T6 atom_reg atom_type(U4 *),    /* Base address of the Ordering Table */
+typedef Slice_MipsAtom Tape;
 
-/* Stringification codes for the GCC inline assembler clobber lists. */
-#define R_TapePtr_Code     R_T8_Code
-#define R_InCursor_Code    R_T4_Code
-
-#define R_PrimCursor_Code  R_T7_Code
-#define R_FaceCursor_Code  R_T4_Code
-#define R_VertBase_Code    R_T5_Code
-#define R_OtBase_Code      R_T6_Code
+typedef Struct_(TapeHostFrame) {
+	U4 s0;
+	U4 s1;
+	U4 s2;
+	U4 s3;
+	U4 s4;
+	U4 s5;
+	U4 s6;
+	U4 s7;
+	U4 fp;
+	U4 sp;
+	U4 ra;
 };
 
-#pragma region Tape Drive
-/* ---------------------------------------------------------------------------
- *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
- * ---------------------------------------------------------------------------*/
+enum {
+	TapeHostFrame_Loc = Scratchpad_End - S_(TapeHostFrame),
+	TapeScratch_Len   = TapeHostFrame_Loc - Scratchpad_Loc,
+};
+static_assert(S_(TapeHostFrame) == 11 * S_(U4));
+static_assert(TapeHostFrame_Loc == 0x1F8003D4);
 
-/* The 'Exit' Atom */
-atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+atom_dbg_skip MipsAtom_(tape_enter) {
+	mac_load_word_imm(R_V0, u4_(TapeHostFrame_Loc)),
+	store_word(R_S0, R_V0, O_(TapeHostFrame,s0)),
+	store_word(R_S1, R_V0, O_(TapeHostFrame,s1)),
+	store_word(R_S2, R_V0, O_(TapeHostFrame,s2)),
+	store_word(R_S3, R_V0, O_(TapeHostFrame,s3)),
+	store_word(R_S4, R_V0, O_(TapeHostFrame,s4)),
+	store_word(R_S5, R_V0, O_(TapeHostFrame,s5)),
+	store_word(R_S6, R_V0, O_(TapeHostFrame,s6)),
+	store_word(R_S7, R_V0, O_(TapeHostFrame,s7)),
+	store_word(R_FP, R_V0, O_(TapeHostFrame,fp)),
+	store_word(R_SP, R_V0, O_(TapeHostFrame,sp)),
+	store_word(R_RA, R_V0, O_(TapeHostFrame,ra)),
+	add_ui(R_TapePtr, R_A0, 0),
+	load_upper_i(R_ScratchBase, u4_hi(Scratchpad_Loc)),
+	load_word(R_AtomJmp, R_TapePtr, 0),
+	add_ui_self(         R_TapePtr, S_(MipsAtom)),
+	jump_reg(R_AtomJmp), BdSlot_ nop,
+};
 
-/* Generalized Tape Engine Runner */
-NI_ void tape_run(Slice_MipsCode tape) { register U4* tp rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
-	asm_words(
-			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
-		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
-		, load_word(  R_AtomJmp, R_TapePtr, 0)         /* Bootstrap the first jump */
-		, add_ui_self(R_TapePtr, S_(MipsCode))         /* Advance tape */
-		, call_reg(   R_AtomJmp)                       /* jalr $t9 */
-		, nop                                          /* Branch delay slot */
-		, load_word(  R_RA, R_SP, 0)                   /* Restore $ra from stack */
-		, add_ui_self(R_SP, MipsStackAlignment)        /* Deallocate stack space */
-	)
-	asm_rpins, r_use(tp)
-	asm_clobber: 
-			rlit(R_AT)
-		, rlit(R_V0), rlit(R_V1)
-		, rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3)
-		/* Tell GCC the tape engine owns and destroys the workspace registers */
-		, rlit(R_PrimCursor), rlit(R_FaceCursor), rlit(R_VertBase), rlit(R_OtBase)
-		, rlit(R_T9)
-		, clb_mem_drain 
-); }
+atom_dbg_skip MipsAtom_(tape_exit) {
+	mac_load_word_imm(R_V0, u4_(TapeHostFrame_Loc)),
+	load_word(R_S0, R_V0, O_(TapeHostFrame,s0)),
+	load_word(R_S1, R_V0, O_(TapeHostFrame,s1)),
+	load_word(R_S2, R_V0, O_(TapeHostFrame,s2)),
+	load_word(R_S3, R_V0, O_(TapeHostFrame,s3)),
+	load_word(R_S4, R_V0, O_(TapeHostFrame,s4)),
+	load_word(R_S5, R_V0, O_(TapeHostFrame,s5)),
+	load_word(R_S6, R_V0, O_(TapeHostFrame,s6)),
+	load_word(R_S7, R_V0, O_(TapeHostFrame,s7)),
+	load_word(R_RA, R_V0, O_(TapeHostFrame,ra)),
+	load_word(R_FP, R_V0, O_(TapeHostFrame,fp)),
+	load_word(R_SP, R_V0, O_(TapeHostFrame,sp)),
+	jump_reg(R_RA), BdSlot_ nop,
+};
 
+typedef void Proc_(TapeEntryFn)(MipsAtom* tape_ptr);
+
+FI_ void tape_run(Tape tape) { C_(TapeEntryFn*, tape_enter)(tape.ptr); }
+
+// Procedural authoring of tapes:
 typedef Relative_(FArena) Struct_(TapeBuilder) { U4 ptr; U4 capacity; U4 used; };
 FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start; tb->used = 0; }
 FI_ TapeBuilder tb_make_old(             FArena* arena) { return (TapeBuilder){ arena->start, 0 }; }
-FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }; }
+FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ u4_(mem.ptr), mem.len, 0 }; }  /* capacity in elements (matches used units) */
 
-#define tb_emit_(tb, atom) tb_emit(tb, atom)
-FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
+FI_ void tb_emit(TapeBuilder* tb, MipsAtom* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
+#define tb_emit_(atom)        tb_emit(& tb, atom)
+#define tb_data_(field, data) tb_data(& tb, u4_(data))
 
-FI_ Slice_MipsCode tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_MipsCode){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Slice_MipsCode){ C_(U4*,tb.ptr),  tb.used }; }
-#define tb_scope(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,tape_exit))
+FI_ void tb_emit_bundle(TapeBuilder_R tb, Slice_MipsAtom atoms) { mem_copy(u4_(tb->ptr), u4_(atoms.ptr), S_slice(atoms)); tb->used += atoms.len; }
 
+FI_ Tape tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Tape){ C_(U4*,tb->ptr), tb->used }; }
+FI_ Tape tb_slice(TapeBuilder  tb) {                        return (Tape){ C_(U4*,tb.ptr),  tb.used }; }
+#define tb_scope(tb)     for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,tape_exit))
+
+FI_ void tb_scope_run_end(TapeBuilder* tb) { tb_emit(tb,tape_exit); tape_run(tb_slice(tb[0])); }
+#define tb_scope_run(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_scope_run_end(tb))
 #pragma endregion Tape Drive
 
 #pragma region Macro Mips Atom Components
@@ -104,253 +261,190 @@ FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Sl
  * ---------------------------------------------------------------------------*/
 
 // The 'Yield' sequence for Tape Atoms (mac_yield).
+
 atom_dbg_skip MipsAtomComp_(ac_yield) {
-	load_word(R_AtomJmp, R_TapePtr, 0),
+	load_word(R_AtomJmp, R_TapePtr, 0), LdSlot_
 	add_ui_self(         R_TapePtr, S_(MipsCode)),
-	jump_reg( R_AtomJmp), nop,
+	jump_reg( R_AtomJmp), BdSlot_ nop,
 };
 
-/* Words: 3; Loads 3 S2 indices from the face array */
-atom_dbg_skip MipsAtomComp_(ac_load_tri_indices) {
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
-	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
-	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+atom_dbg_skip MipsAtomComp_(ac_yield_load) {
+	load_word(R_AtomJmp, R_TapePtr, 0),
 };
 
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-atom_dbg_skip MipsAtomComp_(ac_gte_load_tri_verts) {
-	shift_lleft(R_AT, R_T0, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-	shift_lleft(R_AT, R_T1, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-	shift_lleft(R_AT, R_T2, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+atom_dbg_skip MipsAtomComp_(ac_yield_tail) {
+	add_ui_self(R_TapePtr, S_(MipsCode)),
+	jump_reg( R_AtomJmp), BdSlot_ nop,
 };
 
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-MipsAtomComp_(ac_insert_ot_tag_f3) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (5 - 1) << 24 = 4 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-MipsAtomComp_(ac_insert_ot_tag_g4) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (9 - 1) << 24 = 8 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
- * byte offset. Internal helper used by the *_format_*_color macros. */
-FI_ MipsAtom ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
-atom_dbg_skip MipsAtomComp_Proc_(ac_pack_color_word, {
-	load_upper_i(R_AT, (cmd) << 8  | (b)),
-	or_i_self(   R_AT, ((g)  << 8) | (r)),
-	store_word(  R_AT, R_PrimCursor, (off)),
-})
-
-/* Words: 3; Emits the F3 command+color word (cmd byte | BLUE | GREEN | RED)
- * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields). */
-FI_ MipsAtom ac_format_f3_color(U1 r, U1 g, U1 b)
-atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
- * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
-};
-
-/* Words: 12; Emits the four (code|color) words of a Poly_G4.
- * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-FI_ MipsAtom ac_format_g4_color(
-	U1 r0, U1 g0, U1 b0,
-  U1 r1, U1 g1, U1 b1,
-  U1 r2, U1 g2, U1 b2,
-  U1 r3, U1 g3, U1 b3)
-MipsAtomComp_Proc_(ac_format_g4_color, {
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
-	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1),
-	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2),
-	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3),
-})
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
- * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
- * three registers aligned with v0/v1/v2 you must store before RTPS).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
-};
-
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
- * single-vertex result to SXY2; SXY0 still holds v0.screen from the
- * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
- * prevents).
- */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
-
 #pragma endregion Macro Atom Components
 
-#pragma region Mips Atom Builder
-// This allows for runtime procedural authoring of mips atoms.
+#pragma region Atom Builder
+// This helps with runtime procedural authoring of mips atoms.
+typedef Relative_(FArena) Struct_(AtomBuilder) { U4 start; U4 capacity; U4 used; };
 
-typedef Struct_(FMipsAtom512) { U4 data[512]; U4 used; };
+// Usual way to resolve an atom after the bulder is done.
+#define atom_from_atombuilder(ab) C_(MipsAtom*, (ab).start)
 
-// FArena Related
-typedef Relative_(FArena) Struct_(MipsAtomBuilder) { U4 start; U4 capacity; U4 used; };
-// Whatever the builder is writting to should most likely coresspond
-// to something that can fit within instruction cache?
-
-FI_ void atombuilder_unroll(MipsAtomBuilder_R ab, Slice_MipsCode_R code) {
-	assert(ab->capacity - ab->used - code->len);
-	mem_copy(ab->start, u4_(code->ptr), code->len);
-	mem_bump(ab->start, ab->capacity, & ab->used, code->len);
+FI_ void atombuilder_push(AtomBuilder_R ab, Slice_MipsCode code) {
+	assert(ab->capacity - ab->used - code.len);
+	U4 dest = ab->start + ab->used * S_(MipsCode); U4 size = S_slice(code);
+	mem_copy(dest, u4_(code.ptr), size); ab->used += size;
 }
-#define atombuilder_unroll_mac(ab, mac) atombuilder_unroll(ab, slice_arg_from_array(Slice_MipsCode, mac))
+#define atombuilder_push_mac(ab, mac) atombuilder_push(ab, slice_arg_from_array(Slice_MipsCode, mac))
 
-// When done authoring, utilize this to cap-off the atom
-FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
-	mem_copy(ab->start, u4_(ac_yield), S_(ac_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
-}
-
-#define mipsatom_from_builder(ab) (MipsAtom){ab.start, ab.used}
+// When done authoring, utilize this to cap-off the atom (if not utilizing a MipsAtom_Proc).
+FI_ void atombuilder_end(AtomBuilder_R ab) { atombuilder_push(ab, slice_from_array(MipsCode, ac_yield)); }
 
+FI_ void tb_emit_atombuilder(TapeBuilder_R tb, AtomBuilder_R ab) { tb_emit(tb, atom_from_atombuilder(ab[0])); }
 #pragma endregion Mips Atom Builder
 
+#pragma region Atom Arena
+// Just a dedicated FArena that is meant to mem_copy and return atom definitions made with MipsAtom_Proc_
+typedef Relative_(FArena) Struct_(AtomArena) { U4 start; U4 capacity; U4 used; };
+
+#define atomarena_unused_start(ab) ((ab).start + (ab).used)
+FI_ void atomarena_init(AtomArena_R arena, Slice mem) {  assert(arena != nullptr);
+	arena->start    = u4_(mem.ptr);
+	arena->capacity = mem.len;
+	arena->used     = 0;
+}
+FI_ AtomArena atomarena_make(Slice mem) { AtomArena a; atomarena_init(& a, mem); return a; }
+FI_ MipsAtom* atomarena_push(AtomArena_R aa, Slice_MipsCode code) {
+	assert(aa->capacity - aa->used - code.len);
+	U4 dest = atomarena_unused_start(aa[0]); U4 size = S_slice(code);
+	mem_copy(dest, u4_(code.ptr), size); aa->used += size;
+	return C_(MipsAtom*, dest);
+}
+FI_ void atomarena_reset(AtomArena_R aa) { aa->used = 0; }
+#pragma endregion Atom Arena
+
+#pragma region RegFile (Register File Allocator)
+// A specialized allocator utilized to help the user track which registers are bound to values
+// that must be preserved for the arena's bounds.
+// TODO(Ed): Technically we can do this at comp-time with the metaprogram, but we may have namespace conflicts.
+// Unless we follow a convention for #define <Scope_Prefix> or something per register allocation boundary.
+
+/* ABI reserves that are never handed out by alloc.
+ * R_AT is the assembler temporary (per the MIPS O32 ABI). 
+ * R_K0/K1 are kernel reserves.
+ * R_GP stays the host global pointer.
+ * R_SP/R_FP/R_RA are tape runtime carriers between tape_enter and tape_exit. */
+U4 const regfile_abi_mask =
+	(1u << R_0)  | (1u << R_AT) |
+	(1u << R_K0) | (1u << R_K1) |
+	(1u << R_GP) | (1u << R_SP) |
+	(1u << R_FP) | (1u << R_RA);
+
+internal Reg const regfile_alloc_order[] = {
+	R_V0, R_V1,
+	R_A0, R_A1, R_A2, R_A3,
+	R_T0, R_T1, R_T2, R_T3, R_T4, R_T5, R_T6, R_T7,
+	R_S0, R_S1, R_S2, R_S3, R_S4, R_S5, R_S6, R_S7,
+	R_T8, R_T9,
+};
+
+typedef Struct_(RegFile) {
+	A2_U2 GPR;
+	A2_U2 GTE;
+};
+#define regfile(pin_mask) {.GPR={u4_lo(pin_mask), u4_hi(pin_mask)} }
+FI_ void regfile_init(RegFile_R rf) {
+	/* pack the 32-bit ABI mask into the two U2s */
+	rf->GPR[0] = u4_lo(regfile_abi_mask);
+	rf->GPR[1] = u4_hi(regfile_abi_mask);
+	rf->GTE[0] = rf->GTE[1] = 0;
+}
+FI_ RegFile regfile_make(void) { RegFile rf; regfile_init(& rf); return rf; }
+
+typedef Struct_(RegFile_RInfo) {
+	U2_R section;
+	U2   mask;
+	B2   occupied;
+};
+FI_ RegFile_RInfo regfile_rinfo(A2_U2 file, Reg r_id) {
+	U2   s_id     = r_id >> 4;
+	U2_R section  = & file[s_id];
+	U2   mask     = u2_(1u << (r_id & 15));
+	B2   occupied = (section[0] & mask) != 0;
+	return (RegFile_RInfo){section, mask, occupied};
+}
+FI_ Reg regfile__alloc_helper(A2_U2 file, Reg r_id) {
+	Reg result = 0; RegFile_RInfo info = regfile_rinfo(file, r_id);
+	if (info.occupied == false) {
+		info.section[0] |= info.mask;
+		result           = r_id;
+	}
+	return result; 
+}
+/* regfile_alloc picks the next free GPR from regfile_alloc_order.
+ * The table is the first-fit allocation order: T0..T7, V0..V1, A0..A3,
+ * S0..S7, T8..T9. The 24 entries leave room for the tape program to use
+ * any of them while R0, R1, R26-R31 remain reserved. */
+I_ Reg regfile_alloc(RegFile_R rf) {
+	Reg allocated = 0;
+	for index_iter(U4, r_id, R_V0, <, R_T9) {
+		allocated = regfile__alloc_helper(rf->GPR, r_id);
+		Jmp_nZero_(allocated,resolved);
+	}
+	assert(allocated != 0);
+resolved: return allocated;
+}
+FI_ Reg regfile_pin(RegFile_R rf, Reg r_id) {
+	RegFile_RInfo info = regfile_rinfo(rf->GPR, r_id);
+	assert(info.occupied == false);
+	info.section[0] |= info.mask;
+	return r_id;
+}
+FI_ void regfile_pin_mask(RegFile_R rf, U4 mask) {
+	B4 occupied = u4_r(rf->GPR)[0] & mask;
+	assert(occupied == false);
+	u4_r(rf->GPR)[0] |= mask;
+}
+FI_ void regfile_free_mask(RegFile_R rf, U4 mask) {
+	if (regfile_abi_mask & mask) return;
+	u4_r(rf->GPR)[0] &= ~mask;
+}
+FI_ void regfile_free_reg(RegFile_R rf, Reg r_id) {
+	/* never free the ABI set */
+	if (regfile_abi_mask & (1u << r_id)) return;
+	RegFile_RInfo info = regfile_rinfo(rf->GPR, r_id);
+	info.section[0] &= ~info.mask;
+}
+FI_ void regfile_reset(RegFile_R rf) {
+	rf->GPR[0] = u4_lo(regfile_abi_mask);
+	rf->GPR[1] = u4_hi(regfile_abi_mask);
+}
+FI_ void regfile_reset_to_mask(RegFile_R rf, U4 mask) {
+	rf->GPR[0] = u4_lo(mask);
+	rf->GPR[1] = u4_hi(mask);
+}
+#pragma endregion RegFileArena (Register File Allocator)
+
+#pragma region Mips Atom Procs
+/* RegUse structs are a convention to organize register allocations for a mips atom procedure.
+	Unlike the usual enum-based declarations, they provide a namespaced scope and have view types via union declarations. */
+#define RegUse_(proc_name) (tmpl(RegUse,proc_name))
+
+	typedef Struct_(RegUse_example_atom_proc) {
+		Reg const ro_register; // Scratch base carrier.
+		Reg usual_modifiable;
+		union { Reg view_1, view_2, view_3; } t1;
+	};
+	internal MipsAtom* example_atom_proc(AtomArena_R aa, U2 offset, RegUse_example_atom_proc r) 
+	MipsAtom_Proc_(aa, {
+		add_si(r.usual_modifiable, r.ro_register, offset),
+		or_u(r.t1.view_1, r.ro_register, 0),
+		branch_lt_zero(r.t1.view_1, atom_offset(example_atom_proc, skip)), BdSlot_ nop,
+		li_s(r.t1.view_2, 100),
+	atom_label(skip)
+		add_si(r.t1.view_3, r.usual_modifiable, 10),
+		mac_yield(),
+	})
+
+#pragma endregion Mips Atom Procs
+
 #pragma region Baked Mips Atoms
 // These atoms are resolved at compile time and are (usually) statically linked readonly data.
 
-enum {
-	bios_flushcache = 0x44,
-	bios_table_addr = 0xA0,
-};
-
-/* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
- * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
- *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
- *   2. $a0 = bios_flushcache (arg0)
- *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
- *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
- *      nop                             ; branch delay slot
- *   5. lw $ra, 4($sp);  jr $ra         ; restore & return
- *   6. sp += 8
- */
-internal MipsAtom_(mips_flush_icache) {
-	add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment), // sp -= 8
-	store_word(rret_addr, rstack_ptr, S_(U4)),           // sw  $ra,   4($sp) 
-	add_ui(rret_0, rdiscard, bios_flushcache),           // addiu $a0, $0, 0x44 
-	add_ui(rtmp_0, rdiscard, bios_table_addr),           // addiu $t0, $0, 0xA0 
-	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
-	load_word(rret_addr, rstack_ptr, S_(U4)),            // lw   $ra, 4($sp) 
-	jump_reg(rret_addr),                                 // jr   $ra 
-	add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment),  // sp += 8 (BD) 
-	mac_yield(),
-};
-
-typedef Struct_(Binds_SetGteWorld) {
-	M3_S2* transform;
-};
-internal MipsAtom_(set_gte_world) atom_info(
-		atom_bind(Binds_SetGteWorld)
-	,	atom_reads(R_TapePtr)
-){
-	/* Pop matrix address from tape into R_T3 ($11) */
-	load_word(R_T3, R_TapePtr, O_(Binds_SetGteWorld,transform)),
-	add_ui_self(    R_TapePtr, S_(Binds_SetGteWorld)),
-	/* Load 3x3 Rotation + 3x1 Translation from R_T3 into GTE CONTROL Regs (ctc2) */
-	load_word(R_T0, R_T3, 0),  load_word(R_T1, R_T3,  4),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT11), gte_mv_to_ctrl_r(R_T1, gte_cr_RT12),
-	load_word(R_T0, R_T3, 8),  load_word(R_T1, R_T3, 12), load_word(R_T2, R_T3, 16),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT13), gte_mv_to_ctrl_r(R_T1, gte_cr_RT21), gte_mv_to_ctrl_r(R_T2, gte_cr_RT22),
-	load_word(R_T0, R_T3, 20), load_word(R_T1, R_T3, 24), load_word(R_T2, R_T3, 28),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_TRX),  gte_mv_to_ctrl_r(R_T1, gte_cr_TRY),  gte_mv_to_ctrl_r(R_T2, gte_cr_TRZ),
-	mac_yield()
-};
-
-/* DIAGNOSTIC 1: Pure tape loop test */
-internal MipsAtom_(diag_yield) { mac_yield() };
-
-/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
-internal MipsAtom_(diag_color) {
-	store_word(  R_0, R_T7, 0), 
-	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
-	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
-	store_word(  R_AT, R_T7, 4),
-	
-	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
-	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
-
-	add_ui(          R_T1, R_0,  10),
-	shift_lleft_self(R_T1,       S_(U4)/2), 
-	add_u_self(      R_T1,       R_T6),         
-	
-	load_word(   R_AT, R_T1, 0),        
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
-	store_word(  R_AT, R_T7, 0),       
-	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
-	or_u_self(  R_AT, R_V0),          
-	store_word( R_AT, R_T1, 0),       
-
-	add_ui(R_T7, R_T7, 20),          
-
-	mac_yield()
-};
-
-/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
-internal MipsAtom_(diag_gte) {
-	/* Load 3 indices */
-	load_half_u(R_T0, R_T4, 0),
-	load_half_u(R_T1, R_T4, 2),
-	load_half_u(R_T2, R_T4, 4),
-
-	/* Load Vertices into GTE */
-	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-
-	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-
-	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
-
-	/* Run Math */
-	nop2, gte_cmdw_rtpt,
-	nop2, gte_cmdw_nclip,
-	nop2,
-
-	/* Advance Face Cursor and Yield */
-	add_ui(R_T4, R_T4, 8),
-
-	mac_yield()
-};
-
 #pragma endregion Baked Mips Atoms

code/duffle/math.atom.h

@@ -0,0 +1,96 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "math.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(math_atom_c);
+
+#define v3s4_R_0() ((Reg_(V3_S4)){R_0,R_0,R_0})
+
+typedef Struct_(Reg_V3_S2) { Reg x, y, z; };
+typedef Struct_(Reg_V3_S4) { Reg x, y, z; }; // Register allocation of a V3_S4
+typedef Struct_(Reg_P3_S4) { Reg x, y, z; }; // Register allocation of a P3_S4
+
+#pragma region MACs (Mips Atom Component)
+
+FI_ Slice_MipsCode ac_load_half_v3(AtomBuilder_R ab, Reg tx, Reg ty, Reg tz, Reg base, U2 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_half(tx, base, offset + OA_(U2,[0])),
+	load_half(ty, base, offset + OA_(U2,[1])),
+	load_half(tz, base, offset + OA_(U2,[2])),
+})
+
+FI_ Slice_MipsCode ac_load_v3s2(AtomBuilder_R ab, Reg_(V3_S2) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_load_half_v3(transfer.x, transfer.y, transfer.z, base, offset))
+
+FI_ Slice_MipsCode ac_load_v2s2(AtomBuilder_R ab, U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_half(rs_x, r_base, offset + O_(V3_S2,x)),
+	load_half(rs_y, r_base, offset + O_(V3_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_v2s2(AtomBuilder_R ab, U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_half(rt_x, base, offset + O_(V2_S2,x)),
+	store_half(rt_y, base, offset + O_(V2_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_load_word_v3(AtomBuilder_R ab, Reg tx, Reg ty, Reg tz, Reg base, U2 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_word(tx, base, offset + OA_(U4,[0])),
+	load_word(ty, base, offset + OA_(U4,[1])),
+	load_word(tz, base, offset + OA_(U4,[2])),
+})
+
+FI_ Slice_MipsCode ac_load_v3s4(AtomBuilder_R ab, Reg_(V3_S4) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_load_word_v3(transfer.x, transfer.y, transfer.z, base, offset))
+FI_ Slice_MipsCode ac_load_p3s4(AtomBuilder_R ab, Reg_(P3_S4) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_load_word_v3(transfer.x, transfer.y, transfer.z, base, offset))
+
+FI_ Slice_MipsCode ac_store_half_v3(AtomBuilder_R ab, Reg tx, Reg ty, Reg tz, Reg base, U2 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_half(tx, base, offset + OA_(U2,[0])),
+	store_half(ty, base, offset + OA_(U2,[1])),
+	store_half(tz, base, offset + OA_(U2,[2])),
+})
+
+FI_ Slice_MipsCode ac_store_v3s2(AtomBuilder_R ab, Reg_(V3_S2) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_store_half_v3(transfer.x, transfer.y, transfer.z, base, offset))
+
+FI_ Slice_MipsCode ac_store_word_v3(AtomBuilder_R ab, Reg tx, Reg ty, Reg tz, Reg base, U2 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_word(tx, base, offset + OA_(U4,[0])),
+	store_word(ty, base, offset + OA_(U4,[1])),
+	store_word(tz, base, offset + OA_(U4,[2])),
+})
+
+FI_ Slice_MipsCode ac_store_v3s4(AtomBuilder_R ab, Reg_(V3_S4) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_store_word_v3(transfer.x, transfer.y, transfer.z, base, offset))
+FI_ Slice_MipsCode ac_store_p3s4(AtomBuilder_R ab, Reg_(P3_S4) transfer, Reg base, U2 offset) MipsAtomComp_ProcMap_(ab, mac_store_word_v3(transfer.x, transfer.y, transfer.z, base, offset))
+
+FI_ Slice_MipsCode ac_add_si_v3s4(AtomBuilder_R ab, Reg rt_x,  Reg rt_y,  Reg rt_z, Reg base, U2 offset)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	add_si(rt_x, base, O_(V3_S4,x)),
+	add_si(rt_y, base, O_(V3_S4,y)),
+	add_si(rt_z, base, O_(V3_S4,z)),
+})
+
+FI_ Slice_MipsCode ac_sub_s_v3(AtomBuilder_R ab
+	, Reg dx, Reg dy, Reg dz
+	, Reg sx, Reg sy, Reg sz
+	, Reg tx, Reg ty, Reg tz
+) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	sub_s(dx, sx, tx),
+	sub_s(dy, sy, ty),
+	sub_s(dz, sz, tz),
+})
+
+FI_ Slice_MipsCode ac_sub_v3s4(AtomBuilder_R ab, Reg_(V3_S4) d, Reg_(V3_S4) s, Reg_(V3_S4) t) MipsAtomComp_ProcMap_(ab, mac_sub_s_v3(d.x, d.y, d.z, s.x, s.y, s.z, t.x, t.y, t.z))
+
+FI_ Slice_MipsCode ac_sub_s_v3_self(AtomBuilder_R ab, Reg ds_x, Reg ds_y, Reg ds_z, Reg tx,  Reg ty,  Reg tz) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	sub_s(ds_x, ds_x, tx),
+	sub_s(ds_y, ds_y, ty),
+	sub_s(ds_z, ds_z, tz),
+})
+
+FI_ Slice_MipsCode ac_sub_v3s4_self(AtomBuilder_R ab, Reg_(V3_S4) ds, Reg_(V3_S4) t) MipsAtomComp_ProcMap_(ab, mac_sub_s_v3_self(ds.x, ds.y, ds.z, t.x, t.y, t.z))
+
+FI_ Slice_MipsCode ac_store_rects2(AtomBuilder_R ab, U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_half(rt_x,      base, offset + O_(Rect_S2,x)),
+	store_half(rt_y,      base, offset + O_(Rect_S2,y)),
+	store_half(rt_width,  base, offset + O_(Rect_S2,width)),
+	store_half(rt_height, base, offset + O_(Rect_S2,height)),
+})
+
+#pragma endregion MACs (Mips Atom Component)

code/duffle/math.h

@@ -7,10 +7,24 @@
 #define max(A, B)       (((A) > (B)) ? (A) : (B))
 #define clamp_bot(X, B) max(X, B)
 
+/* Convention
+<Type> ## <Width> _ <Component Type> ## <Component Width>
+For types with compound data (Ex: Rotation Matrix & Translation):
+<TypeA> ## <TypeB> ## <Width> _ <ComponentTypeA> ## <ComponentWidthA> ## <ComponentTypeB> ## <ComponentWidthB>
+
+A: Array
+V: Vector
+R: Range
+M: Matrix
+T: Translation
+*/
+
 enum { 
 	v3s2_byteoff = 3, // log2(8), used with shift_left_logical op for index via byte offset.
 };
 
+typedef Array_(U1, 2);
+typedef Array_(U2, 2);
 typedef Array_(U4, 2);
 typedef Array_(S2, 2);
 typedef Array_(S2, 3);
@@ -22,24 +36,46 @@ typedef S2 A3x3_S2[3][3];
 typedef Struct_(Extent2_S2) { S2 width; S2 height; };
 typedef Struct_(Extent2_S4) { S4 width; S4 height; };
 
+typedef Struct_(V2_U1)      { U1 x; U1 y; };
 typedef Struct_(V2_S2)      { S2 x; S2 y; };
 typedef Struct_(V2_S4)      { S4 x; S4 y; };
-typedef Struct_(V3_S2)      { S2 x; S2 y; S2 z; S2 pad; };
-typedef Struct_(V3_S4)      { S4 x; S4 y; S4 z; S4 pad; };
+typedef Struct_(V3_S2)      { S2 x; S2 y; S2 z; S2 pad; }; // PSY-Q: SVECTOR
+typedef Struct_(V3_S4)      { S4 x; S4 y; S4 z; S4 pad; }; // PSY-Q: VECTOR. RGA(Lengyel): Euclidean vector or direction. A zero-weight RGA point is stored as a V3_S4 with the implicit weight dropped.
 typedef Struct_(V4_S2)      { S2 x; S2 y; S2 z; S2 w; };
 typedef Struct_(V4_S4)      { S4 x; S4 y; S4 z; S4 w; };
 
-typedef Struct_(R2_S2)      { V2_S2 p0; V2_S2 p1; };
-typedef Struct_(R2_S4)      { V2_S4 p0; V2_S4 p1; };
+// typedef Struct_(P3_S4)      { S4 x; S4 y; S4 z; S4 w1; }; // RGA(Lengyel): Affine point with implicit weight one. Storage alias of V3_S4. Use P3_S4 when the value is a point.
+typedef V3_S4 P3_S4;
 
-typedef Struct_(Rect_S2) { S2 x; S2 y; S2 width; S2 height; };
-typedef Struct_(Rect_S4) { S4 x; S4 y; S4 width; S4 height; };
+typedef Struct_(R1_U2) { U2 p0; U2 p1; };
+typedef Struct_(R1_S2) { S2 p0; S2 p1; };
 
-typedef Struct_(M3_S2) { A3x3_S2 m; A3_S4 t; };
+typedef Struct_(R2_S2)      { V2_S2 p0; V2_S2 p1; }; // Range-2 Signed 2-Byte (16-bit)
+typedef Struct_(R2_S4)      { V2_S4 p0; V2_S4 p1; }; // Range-2 Signed 4-Byte (32-bit)
 
+typedef Struct_(Rect_S2)    { S2 x; S2 y; S2 width; S2 height; };
+typedef Struct_(Rect_S4)    { S4 x; S4 y; S4 width; S4 height; };
+
+typedef Struct_(MT3_S2S4)   { A3x3_S2 m; A3_S4 t; }; // PSY-Q: MATRIX. RGA(Lengyel): Matrix expansion of a rigid transformation. GTE utilizes this representation; corresponding motor not constructed here.
+
+/* RGA(Lengyel) reserved names (deferred):
+ *   P4_S4  - future flat point with explicit weight (Lengyel/TML FlatPoint3D analog).
+ *   B3_S4  - future 3D bivector (callers store a Complement(Wedge(...)) as a V3_S4).
+ *   Mo8_S4 - future motor. Not introduced until a course operation actually needs composition, interpolation, or inversion. */
+
+typedef Array_(V2_U1, 2);
+typedef Array_(V2_S2, 2);
 typedef Array_(V2_S2, 3);
 typedef Array_(V2_S2, 4);
 
+#define r1u2(p0,p1) (R1_U2){p0,p1}
+
+enum {
+	fp_one = (1 << 12),
+};
+
+#define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)
+
 #define v2s2(x,y)     (V2_S2){x,y}
 #define v3s2(x,y,z)   (V3_S2){x,y,z,0}
 #define v3s4(x,y,z)   (V3_S4){x,y,z,0}
@@ -58,10 +94,28 @@ FI_ void add_a3s4_fp(A3_S4_R out_a, A3_S4 b) {
 	(out_a[0])[2] += b[2] >> 1;
 }
 
-FI_ void add_v3s4(V3_S4_R out_a, V3_S4 b) { 
-	add_a3s4(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+FI_ void sub_a3s4(A3_S4_R out_a, A3_S4 b) {
+	(out_a[0])[0] -= b[0];
+	(out_a[0])[1] -= b[1];
+	(out_a[0])[2] -= b[2];
 }
 
-FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) { 
-	add_a3s4_fp(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+FI_ void sub_a3s4_fp(A3_S4_R out_a, A3_S4 b) {
+	(out_a[0])[0] -= b[0] >> 1;
+	(out_a[0])[1] -= b[1] >> 1;
+	(out_a[0])[2] -= b[2] >> 1;
 }
+
+FI_ void mul_a3s4(A3_S4_R out_a, A3_S4 b) {
+	(out_a[0])[0] *= b[0];
+	(out_a[0])[1] *= b[1];
+	(out_a[0])[2] *= b[2];
+}
+
+FI_ void add_v3s4   (V3_S4_R out_a, V3_S4 b) {  add_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) {  add_a3s4_fp(C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+
+FI_ void sub_v3s4   (V3_S4_R out_a, V3_S4 b) {  sub_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+FI_ void sub_v3s4_fp(V3_S4_R out_a, V3_S4 b) {  sub_a3s4_fp(C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+
+FI_ void mul_v3s4   (V3_S4_R out_a, V3_S4 b) {  mul_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }

code/duffle/memory.h

@@ -18,7 +18,7 @@ I_ U4 align_pow2(U4 x, U4 b) {
 
 #define align_struct(type_width) ((U4)(((type_width) + 3) & ~3))
 
-FI_ void mem_bump(U4 start, U4 cap, U4*R_ used, U4 amount) {
+FI_ void mem_bump(U4 cap, U4*R_ used, U4 amount) {
 	assert(amount <= (cap - used[0])); 
 	used[0] += amount; 
 }
@@ -58,37 +58,44 @@ typedef Struct_(Str8)        { UTF8* ptr; U4 len; };
 typedef Struct_(Slice_Str8)  { Str8* ptr; U4 len; };
 #define slit(string_literal) (Str8){ (UTF8*) string_literal, S_(string_literal) - 1 }
 
-typedef Struct_(Slice) { U4 ptr, len; }; // Untyped Slice
-FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){ptr, len}; }
+typedef Struct_(Slice) { B1* ptr; U4 len; }; // Untyped Slice (byte-addressable; .len in elements)
+FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){(B1*)ptr, len}; }
 
 #define Slice_(type)       Struct_(tmpl(Slice,type)) { type* ptr; U4 len; }
 typedef Slice_(B1);
 #define slice_assert(s)    do { assert((s).ptr != 0); assert((s).len > 0); } while(0)
-#define slice_end(slice)   ((slice).ptr + (slice).len)
+#define slice_end(slice)   ((slice).ptr + S_slice(slice) / S_(B1))  /* byte-ptr arithmetic; .len is in elements per slice convention */
 #define S_slice(s)         ((s).len * S_((s).ptr[0]))
 
-#define slice_ut(ptr,len)  slice_ut_(u4_(ptr), u4_(len))
-#define slice_ut_arr(a)    slice_ut_(u4_(a),   S_(a))
+#define slice_ut(ptr,len)  slice_ut_(u4_(ptr),     u4_(len))
+#define slice_ut_arr(a)    slice_ut_(u4_(a),       S_(a))
 #define slice_to_ut(s)     slice_ut_(u4_((s).ptr), S_slice(s))
 
 #define slice_iter(container, iter)     (T_((container).ptr) iter = (container).ptr; iter != slice_end(container); ++ iter)
-#define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = array_decl(type,__VA_ARGS__), .len = array_len( array_decl(type,__VA_ARGS__)) }
-#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = S_(array) }
+#define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = Array_decl(type,__VA_ARGS__), .len = Array_len( Array_decl(type,__VA_ARGS__)) }
+#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = Array_len(array) }
 
-FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(s.ptr, s.len); }
+FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(u4_(s.ptr), s.len); }
 #define  slice_zero(s)        slice_zero_(slice_to_ut(s))
 
 FI_ void slice_copy_(Slice dest, Slice src) {
-	assert(dest.len >= src.len);
+	assert(S_slice(dest) >= S_slice(src));
 	slice_assert(dest);
 	slice_assert(src);
-	mem_copy(dest.ptr, src.ptr, src.len);
+	mem_copy(u4_(dest.ptr), u4_(src.ptr), S_slice(src));
 }
 #define slice_copy(dest, src) do {  \
 	static_assert(T_same(dest, src)); \
 	slice_copy_(slice_to_ut(dest), slice_to_ut(src)); \
 } while(0)
 
+FI_ Slice slice_bump(U4_R used, U4 start, U4 len, U4 amount) {
+	assert(len - used[0] - amount);
+	U4 ptr = start + used[0]; used[0] += amount;
+	return slice_ut(ptr, amount);
+}
+
+typedef Slice_(U1);
 typedef Slice_(U4);
 
 #pragma endregion Slice
@@ -98,18 +105,19 @@ typedef Slice_(U4);
 typedef Opt_(farena)    { U4 alignment, type_width; };
 typedef Struct_(FArena) { U4 start, capacity, used; };
 FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
-	arena->start    = mem.ptr;
+	arena->start    = u4_(mem.ptr);
 	arena->capacity = mem.len;
 	arena->used     = 0;
 }
 FI_ FArena farena_make(Slice mem) { FArena a; farena_init(& a, mem); return a; }
-I_  Slice  farena_push(FArena_R arena, U4 amount, Opt_farena o) {
+FI_ Slice farena_bump(FArena_R a, U4 amount) { return slice_bump(& a->used, a->start, a->capacity, amount); }
+I_  Slice farena_push(FArena_R arena, U4 amount, Opt_farena o) {
 	if (amount == 0) { return (Slice){}; }
 	U4 desired   = amount * (o.type_width == 0 ? 1 : o.type_width);
 	U4 to_commit = align_pow2(desired, o.alignment ?  o.alignment : MEM_ALIGNMENT_DEFAULT);
 	U4 ptr       = arena->start + arena->used;
-	mem_bump(arena->start, arena->capacity, & arena->used, to_commit);
-	return (Slice){ ptr, to_commit };
+	mem_bump(arena->capacity, & arena->used, to_commit);
+	return (Slice){ (B1*)ptr, to_commit };
 }
 FI_ void farena_reset (FArena_R arena) { arena->used = 0; }
 FI_ void farena_rewind(FArena_R arena, U4 save_point) {
@@ -117,8 +125,21 @@ FI_ void farena_rewind(FArena_R arena, U4 save_point) {
 	arena->used -= save_point - arena->start;
 }
 FI_ U4 farena_save(FArena arena) { return arena.used; }
+FI_ U4 farena_unused_start(FArena arena) { return arena.start + arena.used; }
 #define farena_push_(arena, amount, ...)                                          farena_push((arena), (amount), opt_(farena, __VA_ARGS__))
 #define farena_push_type(arena, type, ...)                              C_(type*, farena_push((arena), 1,        opt_(farena, .type_width=S_(type), __VA_ARGS__)).ptr)
 #define farena_push_array(arena, type, amount, ...) (tmpl(Slice,type)){ C_(type*, farena_push((arena), (amount), opt_(farena, .type_width=S_(type), __VA_ARGS__)).ptr), (amount) }
 
 #pragma endregion FArena
+
+#pragma region BIOS Scratchpad
+/* BIOS scratchpad location. 1 KB at 0x1F800000.
+ * TapeHostFrame occupies the final 44 bytes while tape code executes.
+ * Atom scratch is bounded by the TapeHostFrame_Loc declaration in lottes_tape.h. */
+enum {
+	Scratchpad_Loc = 0x1F800000,
+	Scratchpad_Len = 0x400,                             /* 1 KB */
+	Scratchpad_End = Scratchpad_Loc + Scratchpad_Len,   /* 0x1F800400 */
+};
+#define C_scratch(type) C_(type, Scratchpad_Loc)
+#pragma endregion BIOS Scratchpad

code/duffle/mips.atom.c

@@ -0,0 +1,80 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#		include "gen/macs.h"
+#		include "gen/offsets.h"
+#	include "bios.h"
+#	include "mips.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(mips_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_load_word_imm(AtomBuilder_R ab, Reg dst, U4 imm)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_upper_i(dst, u4_hi(imm)),
+	or_i_self(   dst, u4_lo(imm)),
+})
+
+FI_ Slice_MipsCode ac_shift_aright_v3_self(AtomBuilder_R ab, Reg dt_x, Reg dt_y, Reg dt_z, U2 shift_amount)
+MipsAtomComp_Proc_( ab, {
+	shift_aright(dt_x, dt_x, shift_amount),
+	shift_aright(dt_y, dt_y, shift_amount),
+	shift_aright(dt_z, dt_z, shift_amount),	
+})
+
+FI_ Slice_MipsCode ac_shift_aright_v3s4_self(AtomBuilder_R ab, Reg_(V3_S4) dt, U2 shift) MipsAtomComp_ProcMap_(ab, mac_shift_aright_v3_self(dt.x, dt.y, dt.z, shift))
+
+FI_ Slice_MipsCode ac_shift_aright_var_v3(AtomBuilder_R ab
+	, Reg rd_v0, Reg rd_v1, Reg rd_v2
+	, Reg rs_v0, Reg rs_v1, Reg rs_v2
+	, Reg r_shift)
+MipsAtomComp_Proc_(ab, {
+	shift_aright_var(rd_v0, rs_v0, r_shift),
+	shift_aright_var(rd_v1, rs_v1, r_shift),
+	shift_aright_var(rd_v2, rs_v2, r_shift),
+})
+
+FI_ Slice_MipsCode ac_shift_aright_var_v3_self(AtomBuilder_R ab, Reg rds_v0, Reg rds_v1, Reg rds_v2, Reg r_shift) 
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	shift_aright_var(rds_v0, rds_v0, r_shift),
+	shift_aright_var(rds_v1, rds_v1, r_shift),
+	shift_aright_var(rds_v2, rds_v2, r_shift),
+})
+
+FI_ Slice_MipsCode ac_shift_aright_var_v3s4_self(AtomBuilder_R ab, Reg_(V3_S4) ds, Reg shift) MipsAtomComp_ProcMap_(ab, mac_shift_aright_var_v3_self(ds.x, ds.y, ds.z, shift))
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Baked Atoms
+
+/* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
+ * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
+ *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
+ *   2. $a0 = bios_flushcache (arg0)
+ *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
+ *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
+ *      nop                             ; branch delay slot
+ *   5. lw $ra, 4($sp)
+ *   6. sp += 8                         ; load-delay
+ *   7. jr $ra
+ *      nop                             ; BD
+ */
+
+#if 0
+// Note: Can't do this without having a way to do C-Runtime frame call from Tape ABI.
+// Don't support this without adjusting scratchpad to save tape frame in some way.
+internal MipsAtom_(mips_flush_icache) {
+	add_ui(R_SP, R_SP, -MipsStackAlignment), // sp -= 8
+	store_word(R_RA, R_SP, S_(U4)),           // sw  $ra,   4($sp) 
+	add_ui(R_V0, R_0, bios_flushcache),           // addiu $a0, $0, 0x44 
+	add_ui(R_T0, R_0, bios_table_addr),           // addiu $t0, $0, 0xA0 
+	jump_link(R_T0, R_RA), nop,                   // jalr  $t0, $ra, BD slot
+	load_word(R_RA, R_SP, S_(U4)),            // lw   $ra, 4($sp) 
+	add_ui(R_SP, R_SP, MipsStackAlignment),  // sp += 8 (load-delay)
+	jump_reg(R_RA), nop,                          // jr   $ra, BD slot
+	// mac_yield(),
+};
+#endif
+
+#pragma endregion Baked Atoms

code/duffle/mips.h

@@ -136,31 +136,31 @@ enum {
 
 /* Semantic Aliases for MIPS Registers (O32 ABI) */
 
-	, rdiscard      = R_0  /* Hardwired to 0 */
-	, rasm_tmp      = R_AT /* Assembler temporary (destroyed by some assembler pseudoinstructions!) */
-	, rret_0        = R_V0 /* Function return value */
-	, rret_1        = R_V1 /* Second return value (e.g., 64-bit) */
-	, rarg_0        = R_A0 /* First function argument */
-	, rarg_1        = R_A1 /* Second function argument */
-	, rarg_2        = R_A2 /* Third function argument */
-	, rarg_3        = R_A3 /* Fourth function argument */
-	, rtmp_0        = R_T0 /* Temporary (Caller saved) */
-	, rtmp_1        = R_T1 /* Temporary (Caller saved) */
-	, rtmp_2        = R_T2 /* Temporary (Caller saved) */
-	, rtmp_3        = R_T3 /* Temporary (Caller saved) */
-	, rtmp_4        = R_T4 /* Temporary (Caller saved) — common GTE base pointer */
-	, rtmp_9        = R_T9 /* Temporary (Caller saved) — common GTE base pointer */
-	, rstatic_0     = R_S0 /* Static (Callee saved, preserved across calls) */
-	, rstatic_1     = R_S1
-	, rstatic_2     = R_S2
-	, rstatic_3     = R_S3
-	, rstatic_4     = R_S4
-	, rstatic_5     = R_S5
-	, rstatic_6     = R_S6
-	, rstatic_7     = R_S7
-	, rsaved_0      = R_S0 /* Alias for rstatic_0 (alternate vocabulary) */
-	, rstack_ptr    = R_SP /* Stack Pointer */
-	, rret_addr     = R_RA /* Return Address (populated by JAL) */
+	// , rdiscard      = R_0  /* Hardwired to 0 */
+	// , rasm_tmp      = R_AT /* Assembler temporary (destroyed by some assembler pseudoinstructions!) */
+	// , rret_0        = R_V0 /* Function return value */
+	// , rret_1        = R_V1 /* Second return value (e.g., 64-bit) */
+	// , rarg_0        = R_A0 /* First function argument */
+	// , rarg_1        = R_A1 /* Second function argument */
+	// , rarg_2        = R_A2 /* Third function argument */
+	// , rarg_3        = R_A3 /* Fourth function argument */
+	// , rtmp_0        = R_T0 /* Temporary (Caller saved) */
+	// , rtmp_1        = R_T1 /* Temporary (Caller saved) */
+	// , rtmp_2        = R_T2 /* Temporary (Caller saved) */
+	// , rtmp_3        = R_T3 /* Temporary (Caller saved) */
+	// , rtmp_4        = R_T4 /* Temporary (Caller saved) — common GTE base pointer */
+	// , rtmp_9        = R_T9 /* Temporary (Caller saved) — common GTE base pointer */
+	// , rstatic_0     = R_S0 /* Static (Callee saved, preserved across calls) */
+	// , rstatic_1     = R_S1
+	// , rstatic_2     = R_S2
+	// , rstatic_3     = R_S3
+	// , rstatic_4     = R_S4
+	// , rstatic_5     = R_S5
+	// , rstatic_6     = R_S6
+	// , rstatic_7     = R_S7
+	// , rsaved_0      = R_S0 /* Alias for rstatic_0 (alternate vocabulary) */
+	// , rstack_ptr    = R_SP /* Stack Pointer */
+	// , rret_addr     = R_RA /* Return Address (populated by JAL) */
 
 /* --- MIPS CPU Opcodes (Bits 31-26) --- */
 
@@ -259,22 +259,22 @@ enum { _BitOffsets = 0
 	, SHAMT_SHIFT  = 6  /* Shift Amount */
 	, FC_SHIFT     = 0
 
-/* Bit Masks to prevent overflow into adjacent fields */
+/* IMM_MASK is the 16-bit two's-complement truncation for the immediate field.
+ * It is NOT a range guard — it is load-bearing for negative branch offsets
+ * (the metaprogram emits raw signed offsets; the mask truncates them to the
+ * 16-bit representation the hardware expects). The static analysis
+ * `immediate_field_width` check validates ranges at build time. */
 
-	, OPCODE_MASK  = 0x3F
-	, REG_MASK     = 0x1F
-	, SHAMT_MASK   = 0x1F /* Shift Amount */
-	, FC_MASK      = 0x3F
 	, IMM_MASK     = 0xFFFF
 };
 
-#define enc_op(op)       (((op)    & OPCODE_MASK) << OPCODE_SHIFT)
-#define enc_rs(rs)       (((rs)    & REG_MASK)    << RS_SHIFT)
-#define enc_rt(rt)       (((rt)    & REG_MASK)    << RT_SHIFT)
-#define enc_rd(rd)       (((rd)    & REG_MASK)    << RD_SHIFT)
-#define enc_shamt(shamt) (((shamt) & SHAMT_MASK)  << SHAMT_SHIFT)
-#define enc_fc(fc)       (((fc)    & FC_MASK)     << FC_SHIFT)
-#define enc_imm(imm)     (((imm)   & IMM_MASK))
+#define enc_op(op)       ((op)    << OPCODE_SHIFT)
+#define enc_rs(rs)       ((rs)    << RS_SHIFT)
+#define enc_rt(rt)       ((rt)    << RT_SHIFT)
+#define enc_rd(rd)       ((rd)    << RD_SHIFT)
+#define enc_shamt(shamt) ((shamt) << SHAMT_SHIFT)
+#define enc_fc(fc)       ((fc)    << FC_SHIFT)
+#define enc_imm(imm)     ((imm) & IMM_MASK)
 
 /* MIPS R-Type Instruction Format (Register-to-Register) */
 #define enc_r(op, rs, rt, rd, shamt, fc) (enc_op(op) | enc_rs(rs) | enc_rt(rt) | enc_rd(rd) | enc_shamt(shamt) | enc_fc(fc))
@@ -318,7 +318,10 @@ enum { _BitOffsets = 0
 #define load_half(rt, base, off)   enc_i(op_lh,    (base), (rt), (off))
 #define load_byte_u(rt, base, off) enc_i(op_lbu,   (base), (rt), (off))
 #define load_half_u(rt, base, off) enc_i(op_lhu,   (base), (rt), (off))
+#define LdSlot_
+
 #define store_word(rt, base, off)  enc_i(op_sw,    (base), (rt), (off))
+
 #define add_ui(rt, rs, imm)        enc_i(op_addiu, (rs),   (rt), (imm))
 #define and_i(rt, rs, imm)         enc_i(op_andi,  (rs),   (rt), (imm))
 // #define and_si                     and_i
@@ -336,10 +339,10 @@ enum { _BitOffsets = 0
 
 /* Logic Opcodes */
 
-#define and_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_and)
-#define or_u(rd, rs, rt)           enc_r(op_special, (rs), (rt), (rd), 0, fc_or)
-#define xor_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_xor)
-#define nor_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_nor)
+#define and_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_and)
+#define or_u(rd, rs, rt)         enc_r(op_special, (rs), (rt), (rd), 0, fc_or)
+#define xor_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_xor)
+#define nor_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_nor)
 
 #define or_u_self(rd_rs, rt)       enc_r(op_special, (rd_rs), (rt), (rd_rs), 0, fc_or)
 
@@ -348,6 +351,12 @@ enum { _BitOffsets = 0
 #define shift_lright(rd, rt, shamt) enc_r(op_special, R_0,  (rt), (rd), (shamt), fc_srl)
 #define shift_aright(rd, rt, shamt) enc_r(op_special, R_0,  (rt), (rd), (shamt), fc_sra)
 
+/* Shift Variable — register-shift forms.
+ * shift_lleft_var(rd, rt, rs)  → sllv rd, rt, rs   (shamt in low 5 bits of rs)
+ * shift_aright_var(rd, rt, rs) → srav rd, rt, rs */
+#define shift_lleft_var(rd, rt, rs)  enc_r(op_special, (rs), (rt), (rd), 0, fc_sllv)
+#define shift_aright_var(rd, rt, rs) enc_r(op_special, (rs), (rt), (rd), 0, fc_srav)
+
 #define shift_lleft_self(rd_rt, shamt) enc_r(op_special, R_0, (rd_rt), (rd_rt), (shamt), fc_sll)
 
 #define mask_upper(rd, rt, shamt) shift_lleft(rd, rt, shamt), shift_lright(rd, rt, shamt)
@@ -362,10 +371,29 @@ enum { _BitOffsets = 0
 /* call_reg rs — jump-and-link to register-held address; link in $ra. */
 #define call_reg(rs)               jump_link((rs), R_RA)
 
-/* j target — absolute jump within the current 256MB region. */
+/* j target — absolute jump within the current 256MB region.
+ * WARNING: `jump(off)` CANNOT BE USED for within-atom jumps in the current pipeline.
+ * The MIPS j opcode encodes `(target_addr >> 2)` in its 26-bit immediate field; an  ABSOLUTE byte address, not a relative word offset.
+ * The metaprogram computes `off` as a relative word offset (`target_word_idx - branch_word_idx - 1`), which the assembler/linker does NOT resolve.
+ * `jump(off)` is only safe when the BUILD PIPELINE owns the absolute position of the emitted code — i.e. when: s
+ *   - the build emits a symbol-relative `.word` expression that the linker resolvess via `R_MIPS_26`, OR
+ *   - the code is hand-assembled with explicit absolute targets, OR a custom post-build patcher resolves the 26-bit field.
+ * TODO(Ed): Review this.. technically we can resolve aboslute jumps on baked atoms? (Even proedurally generated ones...)
+ */
 #define jump(off)                  enc_i(op_j,     R_0, R_0, (off))
 
-/* call_addr off — jump-and-link to immediate address. */
+// Annotate an instruction as filling a branch-delay slot.
+#define BdSlot_
+
+/* jump_rel off — unconditional relative jump (the within-atom-safe `jump`).
+ * MIPS I R3000A has no "branch always" opcode. The idiom for an unconditional relative jump is `beq $0, $0, off`. */
+#define jump_rel(off)              branch_equal(R_0, R_0, (off))
+
+/* call_addr off — jump-and-link to immediate address.
+ * Same WARNING as `jump(off)` above: the jal opcode also encodes an absolute 26-bit target.
+ * For within-atom calls, the current pipeline has no equivalent always-taken call-and-link idiom.
+ * Workaround: `branch_link` (always-taken branch + explicit `la $ra, next_word_addr; jr $ra`), or just use `call_reg($tmp)` after loading the target into a register.
+ */
 #define call_addr(off)             enc_i(op_jal,   R_0, R_0, (off))
 
 /* --- Store family (mirrors the load family) --- */
@@ -379,13 +407,7 @@ enum { _BitOffsets = 0
  * sub_s / sub_u  → sub / subu
  * mult_s / mult_u → mult / multu   (writes HI/LO; result in LO)
  * div_s / div_u   → div / divu     (LO = quot, HI = rem)
- *
- * NOTE: dsl.h defines `add_s`/`sub_s`/`mut_s`/`gt_s`/etc. as _Generic-based signed integer-arithmetic helpers for U1/U2/U4.
- * Those live in a different conceptual layer (generic arithmetic on DSL types) and would collide with the instruction encoders here.
- * The `#undef` below lets the gas-style names below win; if a file needs both, the dsl.h versions can be reached via their long forms
- * (e.g. `def_signed_op`-style or the underlying `add_s1/s2/s4`). */
-#undef add_s
-#undef sub_s
+ */
 #define add_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_add)
 #define add_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_addu)
 #define sub_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_sub)
@@ -395,6 +417,7 @@ enum { _BitOffsets = 0
 #define div_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_div)
 #define div_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_divu)
 
+// TODO(Ed): Change convention of 'self' to ds for (destination is source)?
 #define add_u_self(rd_rs, rt)      add_u(rd_rs, rd_rs, rt)
 
 /* --- Arithmetic I-type (immediate) --- */
@@ -437,25 +460,26 @@ enum { _BitOffsets = 0
 #define shift_amount(rd, rt, n)     shift_lleft(rd, rt, n)
 
 /* nop — sll $0, $0, 0 */
-#define nop  shift_lleft(rdiscard, rdiscard, 0)
+#define nop  shift_lleft(R_0, R_0, 0)
 #define nop2 nop, nop
 
+// li_s — load signed 16-bit immediate into GPR (addiu rt, $0, imm — sign-extends).
+#define li_s(rt, imm)       add_ui((rt), R_0, (imm))
+// #define load_imm_s(rt, imm) add_ui((rt), R_0, (imm))
+
 #define load_imm_1w(rt, imm)    add_ui((rt),  R_0, (imm))
 #define load_imm_1w_s0(rt, imm) add_si((rt)), R_0, (imm))
 
 /* load_imm_2w — unconditional 2-word `li` form: `lui` + (ori | addi).
- *
  * Granular companion to `load_imm`: skips the compile-time range checks and always emits 2 .words. Use this when:
  *   - you know `imm` is > 0xFFFF (otherwise you're wasting a word), OR
- *   - `imm` is not a compile-time constant and you want predictable
- *     2-word emission without the `__builtin_constant_p` branches.
+ *   - `imm` is not a compile-time constant and you want predictable 2-word emission without the `__builtin_constant_p` branches.
  *
  * The lo16 strategy is still chosen at expansion time on the lo half:
  *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui already cleared bits 15..0)
  *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the intended bit pattern)
  *
- * For situations where you need to bypass even this choice 
- * (e.g. to force a specific encoding for a known discontiguous high/low pair),
+ * For situations where you need to bypass even this choice (e.g. to force a specific encoding for a known discontiguous high/low pair),
  * see `load_imm_2w_ori_forced` and `load_imm_2w_addi_forced` below.
  * Statement-level (not expression-level): emits its own `asm volatile(...)`.
  */
@@ -562,6 +586,7 @@ enum { _BitOffsets = 0
 	, jump_link(rtmp_0, rret_addr)              \
 	, nop                                       \
 	, load_word(rret_addr, rstack_ptr, 4)       \
-	, jump_reg(rret_addr)                       \
 	, add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment) \
+	, jump_reg(rret_addr)                       \
+	, nop                                       \
 ) asm_clobber: clbr_volatile_gprs )

code/duffle/pad.atom.c

@@ -0,0 +1,196 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#		include "gen/macs.h"
+#		include "gen/offsets.h"
+#	include "mips.h"
+#	include "dsl.atom.h"
+#	include "lottes_tape.h"
+#	include "pad.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(pad_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_pad_set_centered_axes(AtomBuilder_R ab, Reg state, Reg scratch) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+		load_upper_i(scratch, (PadAxis_Centered >> 16) & 0xFFFF),
+    or_i_self(   scratch,  PadAxis_Centered        & 0xFFFF), // mac_load_word_imm(scratch, PadAxis_Centered),
+    store_word(  scratch, state, O_(PadState,axes)),
+})
+
+FI_ Slice_MipsCode ac_pad_set_id_byte(AtomBuilder_R ab, Reg state, Reg r_id, U1 id_value) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    add_ui(    r_id, R_0, id_value),
+    store_byte(r_id, state, O_(PadState,id)),
+})
+
+FI_ Slice_MipsCode ac_pad_set_status(AtomBuilder_R ab, U4 r_tmp, U1 r_state, U4 pad_status) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    add_ui(    r_tmp, R_0, pad_status),
+    store_word(r_tmp, r_state, O_(PadState,status)),
+})
+
+/* Invert r_buttons (active-low → active-high) and store to PadState.buttons.
+ * r_buttons must already be loaded (the caller is responsible for filling the load-delay slot of
+ * the preceding load_half_u with an instruction that doesn't read r_buttons). */
+FI_ Slice_MipsCode ac_pad_store_inverted_buttons(AtomBuilder_R ab, U1 r_buttons, U1 r_pad_state) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    nor_u(     r_buttons, r_buttons, R_0),
+    store_half(r_buttons, r_pad_state, O_(PadState,buttons)),
+})
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Baked Atoms
+
+/* ----- pad_bios_snapshot -----
+ * Per-frame snapshot of one BIOS pad buffer into PadState.
+ * Decoder (branch ladder on raw[0] status + raw[1] id):
+ *   1. raw[0] == 0xFF         -> Disconnected (buttons=0, axes=0x80)
+ *   2. raw[0]==0 && raw[1]==0 -> Pending      (buttons=0, axes=0x80)
+ *   3. raw[1] == 0x41         -> Digital      (buttons normalized; axes=0x80)
+ *   4. raw[1] == 0x53         -> AnalogStick  (buttons normalized; axes from raw[4..7])
+ *   5. raw[1] in 0x7x         -> AnalogPad    (buttons normalized; axes from raw[4..7])
+ *   6. else                   -> Unsupported  (buttons=0, axes=0x80)
+ *
+ * Buttons normalization: byte_swap16((~raw_buttons) & 0xFFFF).
+ *   raw_buttons    = load_half_u(raw, 2) = raw[2] | (raw[3] << 8).
+ *   byte_swap16(x) = (x >> 8) | (x << 8); nor(x, R_0) = ~x. store_half truncates to 16 bits so the upper-16 mask is implicit in the store.
+ *
+ * Register use (atom-local; no wave-context touched):
+ *   R_T0 = raw base      : Kept throughout; axes loads read raw[4..7] from R_T0.
+ *   R_T1 = state base    : Kept throughout; all stores go through R_T1.
+ *   R_T2 = raw[0] status : Alive across the disc/pending/id dispatch, then dead.
+ *   R_T3 = raw[1] id     : Alive across the id dispatch, then dead.
+ *   R_T4 = scratch       : Shifts, compares, immediate loads, store values.
+ *   R_T5 = scratch       : Parallel lui + ori for the 0x80808080 axes constant + byte-swap target.
+ */
+enum {
+	R_PadRaw    = R_T0 atom_reg atom_type(U1),
+	R_PadState  = R_T1 atom_reg atom_type(PadState*),
+	R_RawStatus = R_T2 atom_reg,
+	R_RawId     = R_T3 atom_reg,
+};
+typedef Struct_(Binds_PadBiosSnapshot) {
+    PadBiosRaw* raw;
+    PadState*   state;
+};
+internal MipsAtom_(pad_bios_snapshot) atom_info(atom_bind(Binds_PadBiosSnapshot)
+, atom_reads( R_PadRaw, R_PadState, R_RawStatus, R_RawId)
+, atom_writes(R_PadRaw, R_PadState, R_RawStatus, R_RawId)
+) {
+	/* === Bind consumption: T0 = raw, T1 = state, advance R_TapePtr by 8. */
+	load_word(R_PadRaw,   R_TapePtr, O_(Binds_PadBiosSnapshot,raw)),
+	load_word(R_PadState, R_TapePtr, O_(Binds_PadBiosSnapshot,state)),
+	add_ui_self(          R_TapePtr, S_(Binds_PadBiosSnapshot)),
+
+	/* === Read raw[0] (status) + raw[1] (id) */
+	load_byte_u(R_RawStatus, R_PadRaw, O_(PadBiosRaw,status)),
+	load_byte_u(R_RawId,     R_PadRaw, O_(PadBiosRaw,id)),
+
+atom_label(snap_root) /* === Case 1: Disconnected (status == 0xFF). */
+	add_ui(R_T4, R_0, PadRawStatus_Timeout), branch_ne(R_RawStatus, R_T4, atom_offset(snap_root, skip_disconnected)),
+	/* BD-slot: pre-compute PadStatus_Disconnected. Branch reads R_T4=0xFF in EX before this WB completes.
+	 * If branch NOT taken (fall through to pending/id_dispatch), R_T4 is overwritten by the next case body's add_ui — harmless. */
+
+atom_label(disconnected) /* === Disconnected body. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_Disconnected),
+	store_half(        R_0,  R_PadState, O_(PadState,buttons)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	mac_pad_set_id_byte(R_PadState, R_RawId, PadRawStatus_Timeout),
+	jump_rel(atom_offset(disconnected, snap_end)),
+		/* BD-slot: load next atom's entry point (replaces the nop).
+		 * Always jumps to snap_end, where mac_yield_tail() transfers control to R_AtomJmp without re-loading it. */
+		mac_yield_load(),
+atom_label(skip_disconnected)
+
+	/* === Case 2: Pending (status == 0 && id == 0)
+	 * Combined check: if (status | id) != 0 then skip to id_dispatch. Falls through to the Pending case only when both are zero. */
+	or_u_self(R_RawStatus, R_RawId), branch_ne(R_RawStatus, R_0, atom_offset(case_2, id_dispatch)),
+	/* BD-slot: pre-compute PadStatus_Pending. Branch reads R_RawStatus in EX before this WB completes.
+	 * If branch NOT taken (fall through to id_dispatch), R_T4 is overwritten by the digital/analog body add_ui - harmless. */
+
+atom_label(pending) /* === Pending body (status=0, id=0 — pre-IRQ-empty buffer). */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_Pending),
+	store_half(        R_0,  R_PadState, O_(PadState,buttons)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	store_byte(R_RawId, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(pending, snap_end)),
+		mac_yield_load(),
+
+atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
+	add_ui(R_T4, R_0, PadRawId_Digital), branch_ne(R_RawId, R_T4, atom_offset(id_dispatch, try_analog_stick)),
+	/* BD-slot: pre-compute PadStatus_Digital. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_stick), R_T4 is overwritten by the analog body add_ui. */
+
+	/* === Digital body (status, buttons normalize, axes=0x80, id, branch.
+	 * R_T5 holds the 0x80808080 axes constant (loaded into the load-delay slot of the buttons-load).
+	 * R_T5 is then "dead" — only consumed at the analog_pad range check downstream. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_Digital),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw, buttons)), /* R_T4 = raw_buttons; */
+	mac_load_word_imm( R_T5, PadAxis_Centered), /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	// load_upper_i(R_T5, PadAxis_Centered_Hi), or_i_self(R_T5, PadAxis_Centered_Lo), 
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),            /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	store_word(R_T5,    R_PadState, O_(PadState, axes)),         /* single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y) */
+	mac_pad_set_id_byte(R_PadState, R_T4, PadRawId_Digital),
+
+	jump_rel(atom_offset(id_dispatch, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_stick) /* === Case 4: AnalogStick (id == 0x53)*/
+	add_ui(R_T4, R_0, PadRawId_AnalogStick), branch_ne(R_RawId, R_T4, atom_offset(try_analog_stick, try_analog_pad)),
+	/* BD-slot: pre-compute PadStatus_AnalogStick. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_pad), R_T4 is overwritten by the analog_pad body add_ui. */
+
+atom_label(analog_stick) /* === AnalogStick body
+	* R_T5 holds left_xy (loaded into the load-delay slot of the buttons-load via the left-axis load_half_u).
+	* R_T4 holds right_xy (loaded into the load-delay slot of the left-load).
+	* R_T5 is then "dead" — reused for the id-byte value load in mac_pad_write_id_byte.
+	* The buttons invert+store happens BEFORE R_T4 is overwritten by the right_xy load. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_AnalogStick),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw,buttons)),  /* R_T4 = raw_buttons; delay slot at the next instruction */
+	load_half_u(       R_T5, R_PadRaw, O_(PadBiosRaw,left)),     /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),            /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	load_half_u( R_T4,  R_PadRaw,   O_(PadBiosRaw,right)),       /* fills R_T5's load-delay slot (doesn't read R_T5); overwrites R_T4 (was buttons) with right_xy */
+	store_half(  R_T5,  R_PadState, O_(PadState,  left)),
+	store_half(  R_T4,  R_PadState, O_(PadState,  right)),
+	mac_pad_set_id_byte(R_PadState, R_T5, PadRawId_AnalogStick),
+	jump_rel(atom_offset(analog_stick, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_pad) /* === Case 5-6: AnalogPad (id & 0xF0 == 0x70) */
+	and_i(    R_T4, R_RawId, PadRawId_AnalogPadMask),
+	add_ui(   R_T5, R_0,     PadRawId_AnalogPadValue),
+	branch_ne(R_T4, R_T5, atom_offset(try_analog_pad, try_unsupported)),
+	/* BD-slot: pre-compute PadStatus_AnalogPad. Branch reads R_T4 in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_unsupported), R_T4 is overwritten by the unsupported body add_ui. */
+
+atom_label(analog_pad) /* === AnalogPad body
+	* Same shape as AnalogStick with AnalogPad status. R_T5 holds left_xy (it's dead on this path).
+	* The id byte is raw id from the BIOS buffer (R_RawId already holds raw[1]).
+	* Buttons invert + store happens before R_T4 is overwritten by the right_xy load. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_AnalogPad),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw,buttons)), /* R_T4 = raw_buttons; delay slot at the next instruction */
+	load_half_u(       R_T5, R_PadRaw, O_(PadBiosRaw,left)),    /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),           /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	load_half_u(R_T4,    R_PadRaw,   O_(PadBiosRaw,right)),     /* fills R_T5's load-delay slot (doesn't read R_T5); overwrites R_T4 with right_xy */
+	store_half( R_T5,    R_PadState, O_(PadState,  left)),
+	store_half( R_T4,    R_PadState, O_(PadState,  right)),
+	store_byte( R_RawId, R_PadState, O_(PadState,  id)),
+
+	jump_rel(atom_offset(analog_pad, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the AnalogPad range-check miss. */
+	add_ui(    R_T4, R_0, PadStatus_Unsupported),
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	mac_pad_set_id_byte(R_PadState, R_RawId, PadUnknownId_Sentinel),
+	/* Fall through to snap_end. */
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
+atom_label(snap_end)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the case-exit branch. */
+	mac_yield_tail(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/pad.c

@@ -0,0 +1,78 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "dsl.h"
+#	include "gcc_asm.h"
+#	include "mips.h"
+#	include "bios.h"
+#	include "pad.h"
+#endif
+
+/* Uses ONE 8-byte frame allocated via the compiler's standard prologue.
+ * 4 wasted-arg words for B(12h) InitPAD2 are at [SP+0..15] but are not explicitly allocated.
+ * Compiler handles the MIPS O32 "wasted stack" convention for us by treating the B-call as a 4-arg call.
+ *
+ * The buffer pointers are passed as arguments so the compiler keeps them in callee-saved registers;
+ * The B(12h) asm volatile block does NOT clobber those registers (it clobbers only the volatile GPRs + B-table arg registers explicitly).
+ * The C-level writes after the call re-load the pointers from their callee-saved homes.
+ *
+ * The clobber list for both B-calls names the full BIOS destroy set documented in kernelbios.md:167-174 (R1..R15, R24..R25, R31, HI/LO).
+ * The kernel-ABI "volatile GPRs" subset is clb_mem_drain; the rest of the destroy set is enumerated explicitly here. */
+NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
+{
+	/* Pin raw0 + raw1 to $a0 + $a1 via rgcc; the B(12h) call uses these directly.
+	 * The `(void)` casts mark them as unread after the call so the compiler doesn't need to move them back. */
+	register PadBiosRaw* p0 rgcc(R_A0) = raw0;
+	register PadBiosRaw* p1 rgcc(R_A1) = raw1;
+	(void)p0; (void)p1;
+
+	// TODO(Ed): Properly annotate the raw values in the inline asm instructions.
+	// Use enums.
+
+	/* B(12h) InitPAD2(raw0, 0x22, raw1, 0x22)
+	 *   $a0 = raw0 (rgcc-bound; survives the sequence below)
+	 *   $a1 = raw1 (preserved into $a2 before $a1 is overwritten)
+	 *   $a2 = raw1 (moved from $a1; survives $a1's overwrite)
+	 *   $a3 = 0x22 (immediate)
+	 *   $t1 = 0x12 (function number)
+	 *   $t2 = 0xB0 (BIOS B-table address) */
+	asm volatile(
+		asm_words(
+			or_u(    R_A2, R_A1, R_0),                 /* $a2 = $a1 = raw1 */
+			add_ui(  R_A1, R_0, bios_pad_buffer_size), /* $a1 = 0x22 */
+			add_ui(  R_A3, R_0, bios_pad_buffer_size), /* $a3 = 0x22 */
+			add_ui(  R_T1, R_0, bios_init_pad_2),      /* $t1 = 0x12 */
+			add_ui(  R_T2, R_0, bios_btable_addr),     /* $t2 = 0xB0 */
+			call_reg(R_T2),                            /* jalr $t2, $ra */
+			nop                                        /* BD slot */
+		)
+		asm_rpins, r_use(p0), r_use(p1)
+		asm_clobber: 
+			rlit(R_AT), 
+			rlit(R_V0), rlit(R_V1), 
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+
+	/* The C-level writes re-load the pointers via the parameter names and write 0xFF to each 
+	 * buffer's status byte to mark the initial-state hazard documented in kernelbios.md:1621-1624. */
+	u1_v(raw0)[0] = 0xFF;
+	u1_v(raw1)[0] = 0xFF;
+
+	/* B(13h) StartPAD2() — no args. The BIOS preserves $sp. */
+	asm volatile(
+		asm_words(
+			add_ui(  R_T1, R_0, bios_start_pad_2), /* $t1 = 0x13 */
+			add_ui(  R_T2, R_0, bios_btable_addr), /* $t2 = 0xB0 (re-load) */
+			call_reg(R_T2),                        /* jalr $t2, $ra */
+			nop                                    /* BD slot */
+		)
+		asm_clobber:
+			rlit(R_AT),
+			rlit(R_V0), rlit(R_V1),
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+}

code/duffle/pad.h

@@ -0,0 +1,115 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "dsl.h"
+#	include "math.h"
+#endif
+
+/* PSX button bit positions — 1:1 with PSX-SPX docs at docs/psx-spx/docs/controllersandmemorycards.md:405-421.
+ * Wire is active-low (0 = pressed).
+ * The decoder atom computes buttons = (~raw_buttons) & 0xFFFF; 
+ * active-low-to-active-high inversion is applied bit-by-bit. */
+typedef Enum_(U2, PadBtns) {
+	Bit_(Pad_Select,    0),
+	Bit_(Pad_L3,        1),
+	Bit_(Pad_R3,        2),
+	Bit_(Pad_Start,     3),
+	Bit_(Pad_Up,        4),
+	Bit_(Pad_Right,     5),
+	Bit_(Pad_Down,      6),
+	Bit_(Pad_Left,      7),
+	Bit_(Pad_L2,        8),
+	Bit_(Pad_R2,        9),
+	Bit_(Pad_L1,       10),
+	Bit_(Pad_R1,       11),
+	Bit_(Pad_Triangle, 12),
+	Bit_(Pad_Circle,   13),
+	Bit_(Pad_Cross,    14),
+	Bit_(Pad_Square,   15),
+};
+
+enum {
+	PadId_Offset = 4,
+
+	Pad0 = 0 << PadId_Offset,
+	Pad1 = 1 << PadId_Offset,
+};
+
+/* =============================================================================
+ * BIOS pad-buffer subsystem: docs/psx-spx/docs/kernelbios.md (B(12h) + B(13h))
+ * ============================================================================= */
+
+enum {
+	PAD_BIOS_RAW_SIZE = 0x22,
+};
+// BIOS pad buffer layout (docs/psx-spx/docs/kernelbios.md (InitPAD2 returns 0x22 = 34 bytes per port)).
+// Bytes 0..7 are the named snapshot region; bytes 8..33 are reserved (the BIOS writes the buffer raw; we only read bytes 0..7 via O_(PadBiosRaw, ...)).
+typedef Struct_(PadBiosRaw) {
+	U1    status;  /* offset 0   (PadRawStatus_Ok / PadRawStatus_Timeout) */
+	U1    id;      /* offset 1   (PadRawId_Digital / PadRawId_AnalogStick / 0x7x AnalogPad) */
+	U2    buttons; /* offset 2-3 (active-low 16-bit button map) */
+	V2_U1 right;   /* offset 4-5 (right stick x, y) */
+	V2_U1 left;    /* offset 6-7 (left  stick x, y) */
+	U1    reserved[PAD_BIOS_RAW_SIZE - 8]; /* offset 8..33 */
+};
+
+typedef Enum_(U4, PadStatus) {
+	PadStatus_Disconnected,
+	PadStatus_Digital,
+	PadStatus_AnalogStick,
+	PadStatus_AnalogPad,
+	PadStatus_Unsupported,
+	PadStatus_Pending,
+	PadStatus_Invalid,
+};
+
+/* Distinct from the game-facing PadStatus enum: PadRawStatus_Ok and PadRawStatus_Timeout are raw BIOS values;
+ * PadStatus_* are game-facing post-decode states. PadUnknownId_Sentinel is written by the decoder 
+ * when the controller id does not match any known controller type.
+ * PadAxisCentered_Word: Four-byte 0x80 pattern used to clear / center 
+ * four byte axes at PadState.left_x through PadState.right_y. */
+typedef Enum_(U1, PadRawStatus) {
+	PadRawStatus_Ok      = 0x00,
+	PadRawStatus_Timeout = 0xFF,
+};
+typedef Enum_(U1, PadRawId) {
+	PadRawId_Digital           = 0x41,
+	PadRawId_AnalogStick       = 0x53,
+	PadRawId_AnalogPadMask     = 0xF0,
+	PadRawId_AnalogPadValue    = 0x70,
+};
+typedef Enum_(U1, PadUnknownId) {
+	PadUnknownId_Sentinel      = 0xFF,
+};
+typedef Enum_(U4, PadAxisCentered) {
+	PadAxis_Centered_Hi = 0x8080,
+	PadAxis_Centered_Lo = 0x8080,
+	PadAxis_Centered    = 0x80808080U,
+};
+typedef Enum_(U1, PadDeadZone) {
+	PadDeadZone_LowBound  = 0x70,  /* left_x <  LowBound  → active;  delta = 0x80 - left_x > 0 (rightward pull) */
+	PadDeadZone_Center    = 0x80,  /* analog rest position; left_x == Center → delta = 0 (no rotation) */
+	PadDeadZone_HighBound = 0x90,  /* left_x >  HighBound → active;  delta = 0x80 - left_x < 0 (leftward pull) */
+};
+
+
+typedef Struct_(PadAxes) {
+	V2_U1  left;  /* offset 8-9 */
+	V2_U1  right; /* offset 10-11 */
+};
+// Field order is chosen so that the 4 axes (left_x, left_y, right_x, right_y)
+// form a contiguous 4-byte block at offset 8, allowing a single `store_word` to clear-or-write all 4 axes in one MIPS instruction.
+typedef Struct_(PadState) {
+	PadStatus status;   /* offset 0,    (U4) */
+	PadBtns   buttons;  /* offset 4,    */
+	U1        id;       /* offset 6,    */
+	byte_pad(1);        /* offset 7,  explicit pad to align the axes block */
+	union {
+		A2_V2_U1 axes;    /* offset 8-11 store_target (4-byte aligned)*/
+		struct {
+			V2_U1  left;    /* offset 8-9 */
+			V2_U1  right;   /* offset 10-11 */
+		};
+	};
+};
+
+internal void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1);

code/duffle/psyq.atom.c

@@ -0,0 +1,7 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "psyq.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(pysq_atom_c);

code/duffle/psyq.h

@@ -0,0 +1,121 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "dsl.h"
+#	include "math.h"
+#	include "gp.h"
+#endif
+
+typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };
+typedef Struct_(DrawEnv) {
+	Rect_S2 clip_area;
+	V2_S2   drawing_offset[2];
+	Rect_S2 texture_window;
+	S2      texture_page;
+	B1      flag_dither;
+	B1      flag_draw_on_display;
+	B1      enable_auto_clear;
+	RGB8    initial_bg_color;
+	DrawEnv_Packed dr_env; // reserved
+};
+typedef Struct_(DisplayEnv) {
+	Rect_S2 display_area;
+	Rect_S2 screen;
+	B1      vinterlace;
+	B1      color24;
+	B1      pad0;
+	B1      pad1;
+};
+typedef Array_(DrawEnv,    2);
+typedef Array_(DisplayEnv, 2);
+
+typedef Struct_(DoubleBuffer) {
+	A2_DrawEnv    draw;
+	A2_DisplayEnv display;
+};
+
+DisplayEnv* displayenv_init(DisplayEnv* env, S4 x, S4 y, S4 w, S4 h) asm("SetDefDispEnv");
+DrawEnv*    drawenv_init   (DrawEnv*    env, S4 x, S4 y, S4 w, S4 h) asm("SetDefDrawEnv");
+
+DisplayEnv* displayenv_put(DisplayEnv* env) asm("PutDispEnv");
+DrawEnv*    drawenv_put   (DrawEnv*    env) asm("PutDrawEnv");
+
+U4   geom_init(void)             asm("InitGeom");
+void geom_set_offset(U4 x, U4 y) asm("SetGeomOffset");
+void geom_set_screen(U4 h)       asm("SetGeomScreen");
+
+U4* orderingtbl_clear_reverse(U4* ot, U4 len) asm("ClearOTagR");
+
+U4 reset_graph(U4 mode)           asm("ResetGraph");
+void set_display_enabled(U4 mask) asm("SetDispMask");
+
+U4 draw_sync(U4 mode) asm("DrawSync");
+U4 vsync(U4 mode)     asm("VSync");
+
+void draw_orderingtbl(U4* buf) asm("DrawOTag");
+
+typedef Struct_(Tile) {
+	U4      tag;
+	RGB8    color;
+	B1      code;
+	Rect_S2 rect;
+};
+
+/*
+	Linear Algebra
+*/
+
+MT3_S2S4* mt3s2s4_rotation   (V3_S2*    vec, MT3_S2S4* mat) asm("RotMatrix");
+MT3_S2S4* mt3s2s4_translation(MT3_S2S4* mat, V3_S4*    vec) asm("TransMatrix");
+MT3_S2S4* mt3s2s4_scale      (MT3_S2S4* mat, V3_S4*    vec) asm("ScaleMatrix");
+
+// Rotation, Translation, Perspective
+
+S4 rtp_v3s2_raw(V3_S2* vec, S4* xy, S4* pp, S4* flag) asm("RotTransPers");
+FI_ S4 rtp_v3s2(V3_S2* vec, V2_S2* xy, A2_S2* pp, S4* flag) { return rtp_v3s2_raw(vec, C_(S4*R_, & xy->x), C_(S4*R_, pp), r_(flag)); }
+
+S4 rtp_avg_nclip_a3_v3s2_raw(V3_S2* v0, V3_S2* v1, V3_S2* v2, S4* xy1, S4* xy2, S4* xy3, S4* pp, S4* otz, S4* flag) asm("RotAverageNclip3");
+FI_  S4 rtp_avg_nclip_a3_v3s2(
+	V3_S2* v0,  V3_S2* v1,  V3_S2* v2, 
+	V2_S2* xy0, V2_S2* xy1, V2_S2* xy2, 
+	A2_S2* pp, S4* otz, S4* flag
+){
+	return rtp_avg_nclip_a3_v3s2_raw(
+		v0, v1, v2, 
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
+	);
+}
+
+S4 rtp_avg_nclip_a4_v3s2_raw(V3_S2* v0, V3_S2* v1, V3_S2* v2, V3_S2* v3, S4* xy1, S4* xy2, S4* xy3, S4* xy4, S4* pp, S4* otz, S4* flag) asm("RotAverageNclip4");
+FI_ S4 rtp_avg_nclip_a4_v3s2(
+	V3_S2* v0,  V3_S2* v1,  V3_S2* v2,  V3_S2* v3,
+	V2_S2* xy0, V2_S2* xy1, V2_S2* xy2, V2_S2* xy3,
+	A2_S2* pp,  S4* otz,    S4* flag
+){
+	return rtp_avg_nclip_a4_v3s2_raw(
+		v0, v1, v2, v3,
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2), C_(S4*R_, xy3),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
+	);
+}
+
+void gte_matrix_set_rotation   (MT3_S2S4* mat) asm("SetRotMatrix");
+void gte_matrix_set_translation(MT3_S2S4* mat) asm("SetTransMatrix");
+
+// Einheit, Metrication to unit vector. "Normalization", not Orthogonal "Normal, Normalis". Directionalization.
+// RGA(Lengyel): Normalize the bulk of a zero-weight direction. This is not finite-point unitization (which forces w=1).
+S4 normalize_v3s4(V3_S4* v0, V3_S4* v1) asm("VectorNormal");
+
+// RGA(Lengyel): Apply the matrix expansion of a rigid transformation.
+// Motor antiproduct is equivalent for unitized points; LA form is what GTE consumes.
+V3_S4* mul_m3s2_v3s4(MT3_S2S4* m, V3_S4* v, V3_S4* result) asm("ApplyMatrixLV");
+
+// RGA(Lengyel): Store the full translation column. The motor translator would store half this displacement in m.xyz.
+MT3_S2S4* trans_m3s2(MT3_S2S4* m, V3_S4* off) asm("TransMatrix");
+
+MT3_S2S4* gte_comp_coord_m3s2(MT3_S2S4* m0, MT3_S2S4* m1, MT3_S2S4* result) asm("CompMatrixLV");
+
+// RGA(Lengyel): Complement(Wedge(a,b)), i.e. the Euclidean 3D complement of the exterior product, stored as a V3_S4.
+// The underlying GTE OP is a specialized signed-16-bit D x IR command; the wedge interpretation is a 3D dual of the same 3 scalars.
+void cross_v3s4(V3_S4* v0, V3_S4* v1, V3_S4* result) asm("OuterProduct12");
+

code/duffle/word_count.metadata.h

@@ -6,7 +6,7 @@
 // One line per macro that appears in your atom sources.
 //
 // This file is encoding-macros-only. 
-// The auto-generated component macros (mac_X) live in duffle/gen/<dir>.macs.h (included separately by the unity build).
+// The auto-generated component macros (mac_X) live in the source directory's own gen/macs.h (per-directory aggregation; included separately by the unity build).
 // The unity build should include THIS file and the .macs.h file in the same TU, with both wrapped 
 // (or the include guard order handled) to avoid WORD_COUNT redeclaration.
 //
@@ -15,6 +15,8 @@
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
 
 WORD_COUNT(nop,                              1)
+WORD_COUNT(atom_label,                        0)
+WORD_COUNT(atom_offset,                       0)
 WORD_COUNT(load_upper_i,                     1)
 WORD_COUNT(jump_reg,                         1)
 WORD_COUNT(jump_link,                        1)
@@ -22,6 +24,7 @@ WORD_COUNT(call_reg,                         1)
 WORD_COUNT(call_addr,                        1)
 WORD_COUNT(branch_le_zero,                   1)
 WORD_COUNT(branch_equal,                     1)
+WORD_COUNT(branch_ne,                        1)
 WORD_COUNT(add_ui,                           1)
 WORD_COUNT(set_lt_u,                         1)
 WORD_COUNT(set_lt_s,                         1)
@@ -29,7 +32,9 @@ WORD_COUNT(set_lt_si,                        1)
 WORD_COUNT(set_lt_ui,                        1)
 WORD_COUNT(load_word,                        1)
 WORD_COUNT(load_half_u,                      1)
+WORD_COUNT(load_byte_u,                      1)
 WORD_COUNT(store_word,                       1)
+WORD_COUNT(store_byte,                       1)
 WORD_COUNT(add_ui_self,                      1)
 WORD_COUNT(add_u_self,                       1)
 WORD_COUNT(add_u,                            1)
@@ -37,6 +42,7 @@ WORD_COUNT(or_i,                             1)
 WORD_COUNT(or_i_self,                        1)
 WORD_COUNT(or_u,                             1)
 WORD_COUNT(or_u_self,                        1)
+WORD_COUNT(nor_u,                            1)
 WORD_COUNT(shift_lleft,                      1)
 WORD_COUNT(shift_lleft_self,                 1)
 WORD_COUNT(shift_lright,                     1)
@@ -49,7 +55,17 @@ WORD_COUNT(gte_mv_to_ctrl_r,                 1)
 WORD_COUNT(gte_sw,                           1)
 WORD_COUNT(gte_cmdw_rtpt,                    1)
 WORD_COUNT(gte_cmdw_nclip,                   1)
+WORD_COUNT(gte_cmdw_op,                      1)
 WORD_COUNT(gte_avg_sort_z3,                  1)
+WORD_COUNT(gte_cmdw_sqr,                     1)
+WORD_COUNT(gte_cmdw_gpf,                     1)
+WORD_COUNT(shift_lleft_var,                  1)
+WORD_COUNT(shift_aright_var,                 1)
+WORD_COUNT(li_s,                             1)
+WORD_COUNT(and_i,                            1)
+WORD_COUNT(add_si,                           1)
+WORD_COUNT(branch_lt_zero,                   1)
+WORD_COUNT(sub_s,                            1)
 WORD_COUNT(sub_u,                            1)
 WORD_COUNT(nop2,                             2)
 

code/gte_hello/hello_gte_tape.c

@@ -1,156 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
-#	include "duffle/gen/duffle.offsets.h"
-#	include "duffle/atom_dsl.h"
-#	include "duffle/lottes_tape.h"
-#	include "duffle/word_count.metadata.h"
-#	include "gen/gte_hello.offsets.h"
-#	include "hello_gte.h"
-#endif
-
-#pragma region MACs (Mips Atom components)
-
-
-
-#pragma endregion MACs
-
-#pragma region Baked Atoms
-
-typedef Struct_(Binds_CubeTri) {
-	U4     PrimCursor;
-	V4_S2* FaceCursor;
-	V3_S2* VertBase;
-	U4*    OtBase;
-};
-internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
-,	atom_reads(R_TapePtr)
-,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-){
-	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
-	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
-	mac_yield()
-};
-
- // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
-internal
-MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
-		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-		atom_writes(R_PrimCursor, R_FaceCursor)
-){
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
-	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
-	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
-	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
-
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-	nop2, gte_cmdw_nclip,
-
-	nop2,  gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
-
-	store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
-	mac_format_g4_color(
-		/* c0 magenta */ 0xFF, 0x00, 0xFF,
-		/* c1 yellow  */ 0xFF, 0xFF, 0x00,
-		/* c2 cyan    */ 0x00, 0xFF, 0xFF,
-		/* c3 green   */ 0x00, 0xFF, 0x00),
-	mac_gte_store_g4_p012_post_rtpt_pre_rtps(),
-
-	shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
-	load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
-	gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
-
-	nop2, gte_cmdw_rotate_translate_perspective_single,
-	mac_gte_store_g4_p3_post_rtps(),
-
-	nop2, gte_cmdw_avg_sort_z4,
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
-
-	add_ui(      R_AT, R_0,  OrderingTbl_Len),
-	set_lt_u(    R_AT, R_T1, R_AT),
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
-	mac_insert_ot_tag_g4(),
-
-atom_label(cube_g4_face_exit)
-	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
-	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
-	mac_yield()
-};
-
-typedef Struct_(Binds_FloorTri) {
-	U4     PrimCursor;
-	V3_S2* FaceCursor;
-	V3_S2* VertBase;
-	U4*    OtBase;
-};
-internal
-MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
-	, atom_reads(R_TapePtr)
-	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-){
-	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
-	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
-	mac_yield()
-};
-
-atom_dbg_skip
-internal
-MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
-	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-	, atom_writes(R_PrimCursor, R_FaceCursr)
-) {
-	mac_load_tri_indices(  R_T0, R_T1, R_T2),
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
-	gte_cmdw_nclip,
-
-	/* Culling (Branch forward if Backface) */
-	gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
-	/* Format Primitive */
-	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-	mac_gte_store_f3_post_rtpt(),
-
-	/* Calculate Depth */
-	gte_avg_sort_z3,
-	gte_mv_from_data_r(R_T1, C2_OTZ),
-	/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
-	add_ui(      R_AT, R_0,  OrderingTbl_Len),
-	set_lt_u(    R_AT, R_T1, R_AT),
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
-	/* Insert into Ordering Table Linked List */
-	mac_insert_ot_tag_f3(),
-	add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
-		// Note(Ed): No bounds checking, should be checked before atom runs.
-
-/* Advance Input Cursor & Yield (Both branch targets land here) */
-atom_label(floor_f3_face_exit)
-	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
-	mac_yield()
-};
-
-typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
-internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
-	, atom_reads( R_TapePtr, R_PrimCursor)
-	, atom_writes(R_TapePtr)
-){
-	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
-	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
-	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
-	/* Calculate byte offset and store directly back to RAM */
-	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
-	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
-	mac_yield()
-};
-
-#pragma endregion Baked Atoms

code/hello_camera/gen/auto_reg.h

@@ -0,0 +1,13 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua (passes/auto_reg.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.c
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.h
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.atom.c
+// Per-phase register allocations resolved by the lua pass.
+// R_<Sym>_Code = <chosen GPR's _Code constant> for every marker in this directory.
+
+#define R_GpTmp_Code R_V0_Code
+

code/hello_camera/gen/macs.h

@@ -0,0 +1,41 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera/
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.c
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.h
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.atom.c
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#define WORD_COUNT(name, count)  enum { words_##name = (count) };
+#endif
+
+#define mac_put_disp_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_disp_env, 5)
+
+#define mac_put_draw_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port) /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */ \
+,	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port) /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port) /* code[1]  TextureWindow (tw=(0,0)) */ \
+,	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port) /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */ \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port) /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */ \
+,	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port) /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */ \
+,	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port) /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */ \
+,	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port) /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */ \
+,	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port) /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port)	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_draw_env, 16)
+

code/hello_camera/gen/offsets.h

@@ -0,0 +1,68 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera\
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.c
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.h
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.atom.c
+#pragma once
+
+#pragma region hello_camera
+
+
+// --- atom: pad_input_cube_rotation (61 words) ---
+
+#define _atom_offset_dpad_left_exit_dpad_left         6
+#define _atom_offset_dpad_right_exit_dpad_right       6
+#define _atom_offset_dead_zone_low_check_dead_low_active  8
+#define _atom_offset_dead_zone_high_check_dead_high_active  15
+#define _atom_offset_dead_zone_skip_exit_stick        24
+#define _atom_offset_end_low_exit_stick               12
+
+enum {
+    atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
+    atom_offset_dpad_right_exit_dpad_right = _atom_offset_dpad_right_exit_dpad_right,
+    atom_offset_dead_zone_low_check_dead_low_active = _atom_offset_dead_zone_low_check_dead_low_active,
+    atom_offset_dead_zone_high_check_dead_high_active = _atom_offset_dead_zone_high_check_dead_high_active,
+    atom_offset_dead_zone_skip_exit_stick = _atom_offset_dead_zone_skip_exit_stick,
+    atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
+};
+
+// --- atom: pad_input_cam (40 words) ---
+
+#define _atom_offset_left_x_exit_left_x               3
+#define _atom_offset_right_x_exit_right_x             3
+#define _atom_offset_up_y_exit_up_y                   3
+#define _atom_offset_down_y_exit_down_y               3
+#define _atom_offset_cross_z_exit_cross_z             3
+#define _atom_offset_circle_z_exit_circle_z           3
+
+enum {
+    atom_offset_left_x_exit_left_x = _atom_offset_left_x_exit_left_x,
+    atom_offset_right_x_exit_right_x = _atom_offset_right_x_exit_right_x,
+    atom_offset_up_y_exit_up_y = _atom_offset_up_y_exit_up_y,
+    atom_offset_down_y_exit_down_y = _atom_offset_down_y_exit_down_y,
+    atom_offset_cross_z_exit_cross_z = _atom_offset_cross_z_exit_cross_z,
+    atom_offset_circle_z_exit_circle_z = _atom_offset_circle_z_exit_circle_z,
+};
+
+// --- atom: cube_g4_face (75 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           41
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+#pragma endregion hello_camera
+

code/hello_camera/hello_camera.atom.c

@@ -0,0 +1,709 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#		include "duffle/gen/macs.h"
+#		include "duffle/gen/offsets.h"
+#	include "duffle/dsl.atom.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/mips.h"
+#	include "duffle/gte.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#	include "duffle/word_count.metadata.h"
+#	include "duffle/psyq.h"
+#	include "duffle/math.atom.h"
+#	include "duffle/mips.atom.c"
+#	include "duffle/gte.atom.c"
+#	include "duffle/gp.atom.c"
+#	include "duffle/psyq.atom.c"
+#		include "gen/offsets.h"
+#		include "gen/macs.h"
+#		include "gen/auto_reg.h"
+#	include "hello_camera.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(hello_joypad_atom_c);
+
+#pragma region MACs (Mips Atom components)
+
+FI_ Slice_MipsCode ac_put_disp_env(AtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
+	// Sequence per libpsyx PutDispEnv: DrawArea TL → DrawArea BR → Mask → DrawArea TL → DrawArea BR
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+})
+
+I_ Slice_MipsCode ac_put_draw_env(AtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	/*
+	* ORIGIN: each code word corresponds to the EXACT value libpsyx's PutDrawEnv function would compute for the same DrawEnv settings.
+	* References:
+	*   - libpsyx source: `toolchain/psyq-4_7/lib/libgpu.a` (binary, function `PutDrawEnv`)
+	*   - PSX-SPX doc:     https://problemkaputt.de/psx-spx.htm#gputdrawingcommands
+	*   - PSYQ SDK:        `setdrawenv` / `makelongdr_env` source
+	*   - NOCASH PSX spec: §"GP0(E1h) Draw Mode setting" through §"DR_ENV"
+	*
+	* The 16-word format is documented in the PSYQ SDK manual and on NOCASH's PSX-spec.txt. The libpsyx reference is at:
+	*   ./toolchain/psyq-4_7/lib/libgpu.a
+	*   (binary; the PutDrawEnv implementation builds the 16-word DR_ENV from the user's DRAWENV struct and emits it via GP0 GPU commands.)
+	*
+	* Word indices (libpsyx PutDrawEnv / SetDrawEnv order):
+	*   tag          = (length << 24) | addr                — 16-word packet (1 tag + 15 code)
+	*   code[0]      = DrawMode (dfe=1, dtd=0, tpage=0)     — must come first per libpsyx
+	*   code[1]      = TextureWindow (tw=(0,0))             — bare-cmd word; GPU uses current state
+	*   code[2]      = DrawArea top-left (clip.x=0, clip.y=240)
+	*   code[3]      = DrawArea bottom-right (clip.x+w=320, clip.y+h=480)
+	*   code[4]      = DrawOffset (ofs=(0,0))               — bare-cmd word
+	*   code[5]      = Mask (dtd=0, dfe=1, isbg=1)          — 0xE6 cmd + isbg bit
+	*   code[6]      = Initial-bg-color (isbg=1, r=7, g=7, b=7)
+	*   code[7]      = DrawMode (isbg=1, tpage=0)           — re-asserts DrawMode with isbg
+	*   code[8..10]  = padding (NOP)                        — 3 words to fill the packet
+	*   code[11..12] = TextureWindow bottom-right           — defaults to (0,0,0,0)
+	*   code[13..14] = padding (NOP)                        — completes the 16-word packet
+	*/
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port), /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */
+	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port), /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */
+	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port), /* code[1]  TextureWindow (tw=(0,0)) */
+	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port), /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port), /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */
+
+	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port), /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */
+	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port), /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */
+	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port), /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */
+	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port), /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */
+
+	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+
+	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+
+	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+})
+
+#pragma endregion MACs
+
+#pragma region Atom Procs
+// Modular Atoms
+
+#define AtomBundle_(name)              Struct_(tmpl(AtomBundle,name))
+#define AtomBundle_Len(name)           S_(tmpl(AtomBundle,name))/S_(MipsAtom*)
+#define AtomBundleEntry_(bundle,entry) tmpl(bundle,entry)
+
+#pragma region resolve_look_at
+/* ─── resolve_look_at bundle chain atoms ──────────────────────────── */
+
+typedef AtomBundle_(resolve_look_at) { MipsAtom
+	*input_and_sub,
+	*normalize_fwd_uz,
+	*cross_to_right,
+	*normalize_right_ux,
+	*cross_to_up,
+	*normalize_up_uy,
+	*pop_mv_trans;
+};
+
+typedef Struct_(ResolveLookAtScratch) {
+	V3_S4 fwd;
+	V3_S4 uz;
+	V3_S4 right;
+	V3_S4 ux;
+	V3_S4 up;
+	V3_S4 uy;
+	P3_S4 eye;
+	P3_S4 target;
+	V3_S4 up_in;
+};
+
+/* Binds_ResolveLookAtSub — what the C side pushes onto the tape before input_and_sub.
+ * The scratchpad base is no longer pushed because R_ScratchBase (= R_SP) is a tape carrier
+ * preserved across atoms; the atom body reads 0x1F800000 directly from R_SP. */
+typedef Struct_(Binds_ResolveLookAt) {
+	MT3_S2S4* look_at;
+	P3_S4*    eye;
+	P3_S4*    target;
+	V3_S4*    up_in;
+};
+typedef Struct_(Binds_ResolveLookAtSub) {
+	P3_S4* target;
+	P3_S4* eye;
+	V3_S4* up_in;
+};
+typedef Struct_(RegUse_resolve_look_at_input_and_sub) {
+	Reg target; Reg eye; Reg up_in;
+	Reg t0; Reg t1; Reg t2; Reg t3; Reg t4;
+};
+/* Atom 0 in the bundle: input_and_sub. Stages C-side inputs into the scratchpad and computes fwd = target - eye. */
+internal MipsAtom* AtomBundleEntry_(resolve_look_at,input_and_sub)(AtomArena_R aa, RegUse_resolve_look_at_input_and_sub r)
+atom_info(atom_bind(Binds_ResolveLookAtSub)) MipsAtom_Proc_(aa, {
+	load_word(r.target, R_TapePtr, O_(Binds_ResolveLookAtSub,target)),
+	load_word(r.eye,    R_TapePtr, O_(Binds_ResolveLookAtSub,eye)),
+	load_word(r.up_in,  R_TapePtr, O_(Binds_ResolveLookAtSub,up_in)),
+	LdSlot_ add_ui_self(R_TapePtr, S_(Binds_ResolveLookAtSub)),
+
+	/* Stage up_in.x/y/z into the scratchpad. R_ScratchBase = R_SP = 0x1F800000. */
+	mac_load_word_v3( r.t0, r.t1, r.t2, r.up_in,   0), LdSlot_
+	mac_store_word_v3(r.t0, r.t1, r.t2, R_ScratchBase, O_(ResolveLookAtScratch,up_in)),
+
+	// Stage eye.x/y/z into the scratchpad (atom 6 reads these for the translation column).
+	mac_load_word_v3( r.t0, r.t1, r.t2, r.eye,      0), LdSlot_
+	mac_store_word_v3(r.t0, r.t1, r.t2, R_ScratchBase, O_(ResolveLookAtScratch,eye)),
+
+	/* Compute fwd = target - eye. */
+	// mac_load_p3s4(t3, R_AT, t4, r.eye,    0),
+	mac_load_word_v3(r.t3, R_AT, r.t4, r.target, 0), LdSlot_
+	mac_sub_s_v3_self(
+		r.t3, R_AT, r.t4,
+		r.t0, r.t1, r.t2),
+	mac_store_word_v3(r.t3, R_AT, r.t4, R_ScratchBase, O_(ResolveLookAtScratch,fwd)),
+
+	mac_yield()
+})
+
+
+typedef Struct_(Binds_ResolveLookAtPopMvTrans) {
+	U4 look_at;  /* MT3_S2S4* — destination matrix address */
+};
+typedef Struct_(RegUse_resolve_look_at__pop_mv_trans) {
+	Reg         look_at;
+	Reg_(V3_S4) row;             /* populate phase: load ux/uy/uz */
+	union { Reg ux, v_x; } t6;   /* populate addr (canonical) → matrix_vector v_x */
+	union { Reg uy, v_y; } t7;   /* populate uy → matrix_vector v_y */
+	union { Reg uz, v_z; } t8;   /* populate uz → matrix_vector v_z */
+	Reg         eye;             /* matrix_vector phase: load -eye */
+};
+/* Atom 6 (fused): write look_at->m[][] from ux/uy/uz as packed S2 (populate),
+ * ctc2 RT chain into C2[0..4] (matrix_vector), MVMVA RT*(-eye)>>12, store off
+ * directly to look_at->t[] (trans_matrix). Replaces the previous 3 separate atoms
+ * (populate + matrix_vector + trans_matrix).
+ *
+ * MT3_S2S4 { A3x3_S2 m; A3_S4 t; }
+ *   m[][] is S2 packed (9 × 2 = 18 bytes at offset 0)
+ *   t[0..2]   is S4       (3 × 4 = 12 bytes at offset 18)
+ *
+ * C11 ApplyMatrixLV semantics (gte.atom.c ac_apply_matrix_lv; libgte reference):
+ *   1. ctc2 RT matrix (5 ctc2s to C2[0..4])
+ *   2. lw -eye from memory
+ *   3. S15 decomposition (eliminated here — the fused body takes the >>12 path
+ *      directly via mtc2 IR + MVMVA pass2, matching the libgte canonical output)
+ *   4. mtc2 to IR1/2/3, nop2, MVMVA pass2 (sf=1, mx=0, v=3, cv=3)
+ *   5. mfc2 MACs → off
+ *   6. store off to look_at->t[] (skip scratch.eye intermediate)
+ *
+ * GPR codes (assigned by resolve_look_at_init):
+ *   r_scratch : R_ResolveScratch (R_T4 carrier)
+ *   r_look_at : ralloc() — also serves as the off-dst in the trans_matrix phase
+ *   r_row     : V3_S4, reused for ux/uy/uz loads in populate phase
+ *   r_eye     : ralloc() — &scratch.eye, used for -eye load in matrix_vector phase
+ *   r_v_x/v_y/v_z : ralloc() — populate scratch addrs (ux/uy/uz), reused as
+ *                    ctc2 transfer + MVMVA -eye temp in matrix_vector phase
+ *   (v_x/v_y/v_z alias ux/uy/uz via the union; lifetime ends for ux/uy/uz after
+ *    populate's mac_load_v3s4, so reusing for v.x/v.y/v.z is safe)
+ * Pool cost: 1 carrier + 1 look_at + 3 row + 1 eye + 3 aliased = 9 GPRs
+ *
+ * Net word savings vs the previous 3-atom flow: ~15 words + 2 mac_yields + 1 tape pop.
+ *   - 2 mac_yields (trans_matrix's + matrix_vector's) → fused into one yield
+ *   - 1 redundant tb_data (look_at was pushed 2x; now once)
+ *   - mac_trans_mt3s3s4 (6 words) → replaced by direct mac_store_v3s4
+ *   - mac_store_v3s4 to scratch.eye (3 words intermediate) → eliminated
+ *   - add_si for r_off_ptr (2 words) → eliminated
+ *   - mac_store_v3s4 zero-store of t[] (3 words) → eliminated (matrix_vector writes
+ *     off directly; no consumer needed the zero first)
+ *   - 1 set_gte_mt3s2s4 ctc2 chain (13 baked words) → eliminated (matrix_vector
+ *     has its own ctc2 RT chain; cube rendering atoms reload C2 state themselves)
+ */
+internal MipsAtom* resolve_look_at__pop_mv_trans(AtomArena_R aa,
+	RegUse_resolve_look_at__pop_mv_trans r
+) MipsAtom_Proc_(aa, {
+	/* --- Tape pop: look_at pointer --- */
+	load_word(r.look_at, R_TapePtr, O_(Binds_ResolveLookAtPopMvTrans,look_at)),
+	LdSlot_ add_ui_self(R_TapePtr, S_(Binds_ResolveLookAtPopMvTrans)),
+
+	/* --- Scratch addresses for ux/uy/uz/eye (populate phase; t6/t7/t8 alias ux/uy/uz).
+	 * R_ScratchBase (= R_SP) holds 0x1F800000; no per-atom bake is required because
+	 * R_SP is a tape carrier preserved across atoms. --- */
+	add_si(r.t6.ux, R_ScratchBase, O_(ResolveLookAtScratch, ux)),  LdSlot_
+	add_si(r.t7.uy, R_ScratchBase, O_(ResolveLookAtScratch, uy)),
+	add_si(r.t8.uz, R_ScratchBase, O_(ResolveLookAtScratch, uz)),
+	add_si(r.eye,   R_ScratchBase, O_(ResolveLookAtScratch, eye)),
+
+	/* --- POPULATE phase: write look_at->m[][] from ux/uy/uz as packed S2 --- */
+	mac_load_v3s4(r.row, r.t6.ux, 0), LdSlot_ mac_store_v3s2(r.row, r.look_at, O_(MT3_S2S4, m[0])),
+	mac_load_v3s4(r.row, r.t7.uy, 0), LdSlot_ mac_store_v3s2(r.row, r.look_at, O_(MT3_S2S4, m[1])),
+	mac_load_v3s4(r.row, r.t8.uz, 0), LdSlot_ mac_store_v3s2(r.row, r.look_at, O_(MT3_S2S4, m[2])),
+
+	/* --- MATRIX-VECTOR phase: ctc2 RT chain + MVMVA RT*(-eye)>>12 --- */
+	/* RT packing (per libgte ApplyMatrixLV convention; see gte.h:217-220 +
+	 * atom_6b_disasm_comparison.md:28-32):
+	 *   C2[0] = (RT12<<16)|RT11  ← ctc2 RT11 from m[0][0..1] packed word
+	 *   C2[1] = (RT21<<16)|RT13  ← ctc2 RT12 from m[0][2..3] packed word
+	 *   C2[2] = (RT23<<16)|RT22  ← ctc2 RT13 from m[1][1..2] packed word
+	 *   C2[3] = (RT32<<16)|RT31  ← ctc2 RT21 from m[2][0..1] packed word
+	 *   C2[4] = (RT33<<16)|junk  ← ctc2 RT22 from m[2][2] (half)
+	 * Each ctc2 writes a WHOLE 32-bit C2 slot; the "macro name" identifies
+	 * which C2 register, not which 16-bit half. */
+	load_word(  r.t6.v_x, r.look_at, O_(MT3_S2S4, m[0][0])), /* RT11|RT12 */ LdSlot_
+	load_word(  r.t7.v_y, r.look_at, O_(MT3_S2S4, m[0][2])), /* RT13|RT21 */ LdSlot_ gte_mv_to_ctrl_r(r.t6.v_x, gte_cr_RT11),
+	load_word(  r.t8.v_z, r.look_at, O_(MT3_S2S4, m[1][1])), /* RT22|RT23 */ LdSlot_ gte_mv_to_ctrl_r(r.t7.v_y, gte_cr_RT12),
+	load_word(  r.t6.v_x, r.look_at, O_(MT3_S2S4, m[2][0])), /* RT31|RT32 */ LdSlot_ gte_mv_to_ctrl_r(r.t8.v_z, gte_cr_RT13),
+	load_half_u(r.t7.v_y, r.look_at, O_(MT3_S2S4, m[2][2])), /* RT33 */      LdSlot_ gte_mv_to_ctrl_r(r.t6.v_x, gte_cr_RT21),
+	/* pos = -eye. The three loads also retire the last CTC2. */                  gte_mv_to_ctrl_r(r.t7.v_y, gte_cr_RT22),
+	GteDelay_ mac_load_word_v3(r.t6.v_x, r.t7.v_y, r.t8.v_z, r.eye, 0), LdSlot_ 
+	mac_sub_s_v3(r.t6.v_x, r.t7.v_y, r.t8.v_z, R_0, R_0, R_0, r.t6.v_x, r.t7.v_y, r.t8.v_z),
+	/* mtc2 pos (as S16) to IR1/2/3. The GTE takes low 16 bits. pos fits in S16. For negative pos, the 32-bit sign-extended value's low 16 bits = correct S16. */
+	gte_mv_to_data_r(r.t6.v_x, C2_IR1),
+	gte_mv_to_data_r(r.t7.v_y, C2_IR2),
+	gte_mv_to_data_r(r.t8.v_z, C2_IR3),
+	GteDelay_ nop2,
+
+	/* MVMVA pass 2 — C11 ApplyMatrixLV command.
+	 * sf=1, mx=0 (RT), v=3 (IR), cv=3. Reads RT × IR >> 12. */
+	gte_cmdw_mvmva_c11_pass2,                                      GteDelay_ nop,
+	mac_gte_mv_from_data_r_mac123(r.t6.v_x, r.t7.v_y, r.t8.v_z),   GteDelay_ nop,
+
+	/* --- TRANS-MATRIX phase: store off directly to look_at->t[] (skip scratch.eye intermediate) --- */
+	mac_store_word_v3(r.t6.v_x, r.t7.v_y, r.t8.v_z, r.look_at, O_(MT3_S2S4, t)),
+
+	mac_yield()
+})
+#pragma endregion resolve_look_at
+
+#pragma endregion Atom Procs
+
+#pragma region Baked Atoms
+
+enum {
+	R_ScreenX   = R_T5 atom_reg atom_type(U2),
+	R_ScreenY   = R_T6 atom_reg atom_type(U2),
+	R_ScreenBuf = R_T7 atom_reg, /* Caller-pinned: & smem.screen_buf */
+#define R_ScreenBuf_Code R_T7_Code
+};
+//screen_env_init. Mirrors the libpsyx's SetDefDispEnv + SetDefDrawEnv + the manual enable_auto_clear / initial_bg_color writes.
+internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
+,	atom_reads(R_T0, R_ScreenX, R_ScreenY, R_ScreenBuf)
+, atom_writes(R_T0, R_ScreenX, R_ScreenY)
+) {
+	/* display[0] = (0, 0, 320, 240); rest of struct zeroed. */
+	add_ui(R_ScreenX, R_0, ScreenRes_X), add_ui(R_ScreenY, R_0, ScreenRes_Y),
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area.width) + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,display_area) + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)       + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace)   + O_(DoubleBuffer,display[0])),
+
+	/* display[1] = (0, 240, 320, 240); rest of struct zeroed. */
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area) + O_(DoubleBuffer,display[1])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)     + O_(DoubleBuffer,display[1])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace) + O_(DoubleBuffer,display[1])),
+
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area)         + O_(DoubleBuffer,draw[0])), /* draw[0].clip_area = (0, 240, 320, 240). C11's SetDefDrawEnv writes clip.y = y_arg. */
+	mac_store_v2s2(  R_0, R_ScreenY,                       R_ScreenBuf, O_(DrawEnv,drawing_offset[0]) + O_(DoubleBuffer,draw[0])), /* draw[0].drawing_offset[0] = (0, 240); C11 passes y_arg as ofs. */
+
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area.width) + O_(DoubleBuffer,draw[1])),
+
+	/* draw[0].texture_window = (0, 0, 0, 0); two word-zeroes cover the full 8-byte tw field. */
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + O_(DoubleBuffer,draw[0])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + O_(DoubleBuffer,draw[0])),
+
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,drawing_offset[0].x)  + O_(DoubleBuffer,draw[1])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + O_(DoubleBuffer,draw[1])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + O_(DoubleBuffer,draw[1])),
+
+	/* draw[0].texture_page = 10 (gp0_tpage_default). C11 SetDefDrawEnv at C11_only.elf:0x8001273C writes the same 0x0A. . */
+	add_ui(R_T0, R_0, gp0_tpage_default),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + O_(DoubleBuffer,draw[0])),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + O_(DoubleBuffer,draw[1])),
+
+	/* draw[0] control bytes: flag_dither=1, flag_draw_on_display=1 (the dfe bit per psx-spx; libpsyx sets it via `SetDefDrawEnv`'s conditional at C11_only.elf:0x80012728), enable_auto_clear=1. Each byte is named;
+	 * the previous `store_word(R_0, ..., +20)` overwrote all four with zero. */
+	add_ui(R_T0, R_0, 1),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + O_(DoubleBuffer,draw[1])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + O_(DoubleBuffer,draw[1])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + O_(DoubleBuffer,draw[1])),
+
+	/* draw[0].initial_bg_color = (r=7, g=7, b=7). */
+	add_ui(R_T0, R_0, 7),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + O_(DoubleBuffer,draw[0])),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + O_(DoubleBuffer,draw[1])),
+
+	mac_yield(),
+};
+
+/* gp_screen_init's GPR setup. Tests the mixed user-pinning + auto-reg pattern:
+ *   - R_IO_BaseAddr = R_T4 (user-pinned via atom_reg; pre-existing)
+ *   - R_GP1_Offset  = R_T2 (user-pinned via atom_reg; NEW -- for GPIO_PORT1_OFFSET)
+ *   - R_ScreenX     = R_T5 (user-pinned via atom_reg; used as a transfer and GTE setup reg)
+ *   - R_GpTmp       = auto-allocated by the lua pass and used for several GPU transfers;
+ *                    the C preprocessor resolves it to the chosen free pool GPR.
+ *
+ * For gp_screen_init, the auto-reg pool exclusions are:
+ *   user_pinned (from the corpus register_alias_registry) : R_T0..R_T7 (all 8 user-pinned across hello_camera.atom.c)
+ *   body-parsed physical registers                         : aliases resolve through the registry;
+ *                                                           the body uses R_ScreenX, not raw R_T5
+ *   source_pool after both subtractions                   : {R_V0, R_V1} only
+ *   R_GpTmp gets R_V0 (the first-fit choice). Its repeated GPU-transfer use proves that the
+ *   auto-reg allocation is active while the R_ScreenX references prove the pinned alias is used.
+ *   R_TapePtr (R_T9), R_AtomJmp (R_T8), R_AT are excluded from the POOL by construction in
+ *   passes/auto_reg.lua -- see the "obvious exclusions" comment block at the top of that file.
+ */
+enum {
+	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+	R_GP1_Offset  = R_T2 atom_reg, /* Caller-pinned: GPIO_PORT1_OFFSET = 0x10 */
+	atom_auto_reg(gp_screen_init, R_GpTmp),  /* Auto-allocated scratch; resolved to a free pool GPR by the lua pass. C-preprocessor expands to R_GpTmp = R_GpTmp_Code with an atom_auto_reg trailing comment. */
+#define R_IO_BaseAddr_Code R_T4_Code
+#define R_GP1_Offset_Code  R_T2_Code
+};
+internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads(R_IO_BaseAddr)) {
+	store_word(R_0, R_IO_BaseAddr, GPIO_PORT1_OFFSET),                                   /* GP1(00h) Reset */
+	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(01h) ClearFIFO; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(02h) AckIRQ; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(03h) Display ON; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_GpTmp,   R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(04h) DMADirection=2 (CPU->GPU). libpsyx's per-frame PutDrawEnv/DrawOTag use DMA2; without this the DMA queue never drains. Uses auto-allocated R_GpTmp. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_GpTmp,   R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(05h) StartDisplayArea (X=0, Y=0); uses auto-allocated R_GpTmp. */
+
+	/* GP1: DisplayMode + Display Ranges. */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+
+	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
+	load_upper_i(R_ScreenX, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFX_Code),
+	load_upper_i(R_ScreenX, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFY_Code),
+
+	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
+	add_ui(R_ScreenX, R_0, ScreenZ), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_H_Code),
+
+	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
+	mac_gcmd_push(gp1_word_DisplayOn(), R_GpTmp, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* Uses auto-allocated R_GpTmp. */
+	mac_yield(),
+};
+
+typedef Struct_(Binds_PadApplyInput) {
+	PadState* state;
+	V3_S2*    cube_rot;
+	V3_S2*    floor_rot;
+};
+enum {
+	R_PadStateT5 = R_T5 atom_reg,
+	R_CubeRot    = R_T1 atom_reg,
+	R_FloorRot   = R_T2 atom_reg,
+};
+internal MipsAtom_(pad_input_cube_rotation) atom_info(atom_bind(Binds_PadApplyInput)
+, atom_reads(R_T0, R_CubeRot, R_FloorRot, R_T3, R_T4, R_PadStateT5, R_TapePtr)
+, atom_writes(     R_CubeRot, R_FloorRot)
+) {
+	/* Pop Binds from tape (state, cube_rot, floor_rot) */
+	load_word(R_PadStateT5, R_TapePtr, O_(Binds_PadApplyInput,state)),
+	load_word(R_CubeRot,    R_TapePtr, O_(Binds_PadApplyInput,cube_rot)),
+	load_word(R_FloorRot,   R_TapePtr, O_(Binds_PadApplyInput,floor_rot)),
+	LdSlot_ add_ui_self(    R_TapePtr, S_(Binds_PadApplyInput)),
+
+	/* Load pad[0].buttons into R_T0. */
+	load_word(R_T0, R_PadStateT5, O_(PadState,buttons)), LdSlot_ nop,
+	// Note(Ed): Potential op with delay slot?
+
+	/* D-pad Left: cube_rot.y += 30, floor_rot.y += 5.  */
+	and_i(R_T3, R_T0, Pad_Left), branch_le_zero(R_T3, atom_offset(dpad_left, exit_dpad_left)), BdSlot_
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)), LdSlot_
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, 30),
+		add_si(    R_T3, R_T3, 5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_left)
+
+	/* D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5. */
+	and_i(R_T3, R_T0, Pad_Right), branch_le_zero(R_T3, atom_offset(dpad_right, exit_dpad_right)), BdSlot_
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)), LdSlot_
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, -30),
+		add_si(    R_T3, R_T3, -5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_right)
+
+	/* Analog left-stick X: dead zone 0x70..0x90.
+	 * Cube delta = (0x80 - left_x) >> 2; floor delta = (0x80 - left_x) >> 5. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left.x)), LdSlot_ //?
+
+	/* Dead-zone check: skip analog if left_x in [0x70, 0x90] inclusive. Outside dead zone on LOW side: left_x < 0x70 (strictly).
+	 * set_lt_u(R_T4, R_T3, R_T4=0x70) → R_T4 = (left_x < 0x70) ? 1 : 0. */
+	add_ui(R_T4, R_0, PadDeadZone_HighBound), set_lt_u(R_T4, R_T3, R_T4), branch_ne(R_T4, R_0, atom_offset(dead_zone_low_check, dead_low_active)),
+	add_ui(R_T4, R_0, PadDeadZone_Center), /* BD-slot: pre-load 0x80 for dead_low_active */
+
+atom_label(dead_check_upper)
+	/* left_x >= 0x70 → check upper bound. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left.x)), /* reload */ LdSlot_ //? 
+	add_ui(     R_T4, R_0, PadDeadZone_HighBound),
+
+	/* R_T4 = (0x90 < left_x) ? 1 : 0 → (left_x > 0x90) ? 1 : 0 */
+	set_lt_u(R_T4, R_T4, R_T3), branch_ne(R_T4, R_0, atom_offset(dead_zone_high_check, dead_high_active)), BdSlot_
+	add_ui(  R_T4, R_0, PadDeadZone_Center), /* BD-slot: pre-load 0x80 for dead_high_active */
+	jump_rel(atom_offset(dead_zone_skip, exit_stick)),
+		BdSlot_ mac_yield_load(), LdSlot_
+
+atom_label(dead_low_active)
+	/* R_T3 = left_x (from line 632 lbu; not clobbered between dead_zone_low_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_low_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3), /* R_T3 = 0x80 - left_x */
+	/* delta = 0x80 - left_x (positive). */
+
+	/* R_T4 = cube_delta */
+	shift_aright(R_T4, R_T3, 2),
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)), LdSlot_ nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+	/* R_T4 = floor_delta — moved into the load-delay slot of the floor load below (fills the 1-instruction gap;
+	 * doesn't read R_T0; R_T4 settles by the subsequent add_u). */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)), LdSlot_
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+	jump_rel(atom_offset(end_low, exit_stick)),
+		BdSlot_ mac_yield_load(), LdSlot_
+
+atom_label(dead_high_active)
+	/* R_T3 = left_x (from line 641 lbu in dead_check_upper; not clobbered between dead_zone_high_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_high_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3),
+	/* delta = 0x80 - left_x (signed negative). */
+
+	shift_aright(R_T4, R_T3, 2),                      /* R_T4 = cube_delta (signed) */
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)), LdSlot_ nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+
+	/* R_T4 = floor_delta (signed) — moved into the load-delay slot of the floor load below. */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)), LdSlot_
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(), LdSlot_
+
+atom_label(exit_stick)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the dead-zone/exit branch. */
+	mac_yield_tail(),
+};
+
+enum {
+	R_Cam         = R_T4 atom_reg,
+	R_CamPadState = R_T5 atom_reg,
+};
+typedef Struct_(Binds_PadInputCam) {
+	PadState* state;
+	Camera*   cam;
+};
+internal MipsAtom_(pad_input_cam) atom_info(atom_bind(Binds_PadInputCam)
+, atom_reads( R_Cam, R_CamPadState, R_TapePtr)
+, atom_writes(R_Cam)
+) {
+	/* Bind pop: state → R_CamPadState (R_T5), cam → R_Cam (R_T4), advance R_TapePtr by 8. */
+	load_word(R_CamPadState, R_TapePtr, O_(Binds_PadInputCam,state)),
+	load_word(R_Cam,         R_TapePtr, O_(Binds_PadInputCam,cam)),
+	LdSlot_ add_ui_self(     R_TapePtr, S_(Binds_PadInputCam)),
+
+	/* Load pad[0].buttons into R_T0; nop fills the load-delay slot. */
+	load_word(R_T0, R_CamPadState, O_(PadState,buttons)), LdSlot_
+	load_word(R_T1, R_Cam, O_(Camera,pos.x)),
+
+	// D-pad Left → cam.pos.x -= 50.  and_i fulfills BD-slot for load on R_Cam.
+	LdSlot_ and_i(R_T3, R_T0, Pad_Left), branch_le_zero(R_T3, atom_offset(left_x, exit_left_x)), BdSlot_ mac_yield_load(), LdSlot_
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.x)), 
+	atom_label(exit_left_x)
+
+	/* D-pad Right → cam.pos.x += 50. Reuses R_T1 from Left. */
+	and_i(R_T3, R_T0, Pad_Right), branch_le_zero(R_T3, atom_offset(right_x, exit_right_x)), BdSlot_ nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.x)),
+	atom_label(exit_right_x)
+
+	/* D-pad Up → cam.pos.y -= 50. Load pos.y BEFORE the andi. */
+	load_word(R_T1, R_Cam, O_(Camera,pos.y)), LdSlot_ 
+	and_i(R_T3, R_T0, Pad_Up), branch_le_zero(R_T3, atom_offset(up_y, exit_up_y)), BdSlot_ nop,
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.y)),
+	atom_label(exit_up_y)
+
+	/* D-pad Down → cam.pos.y += 50. Reuses R_T1 from Up. */
+	and_i(R_T3, R_T0, Pad_Down), branch_le_zero(R_T3, atom_offset(down_y, exit_down_y)), BdSlot_ nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.y)),
+	atom_label(exit_down_y)
+
+	/* D-pad Cross → cam.pos.z -= 50. Load pos.z BEFORE the andi. */
+	load_word(R_T1, R_Cam, O_(Camera,pos.z)), LdSlot_
+	and_i(R_T3, R_T0, Pad_Cross), branch_le_zero(R_T3, atom_offset(cross_z, exit_cross_z)), BdSlot_ nop,
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.z)),
+	atom_label(exit_cross_z)
+
+	/* D-pad Circle → cam.pos.z += 50. Reuses R_T1 from Cross. */
+	and_i(R_T3, R_T0, Pad_Circle), branch_le_zero(R_T3, atom_offset(circle_z, exit_circle_z)), BdSlot_ nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.z)),
+	atom_label(exit_circle_z)
+
+	mac_yield_tail(),
+};
+
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* Output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
+#define R_PrimCursor_Code  R_T7_Code
+#define R_FaceCursor_Code  R_T4_Code
+#define R_VertBase_Code    R_T5_Code
+#define R_OtBase_Code      R_T6_Code
+};
+typedef Struct_(Binds_CubeTri) {
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
+	LdSlot_ add_ui_self(    R_TapePtr, S_(Binds_CubeTri)),
+	mac_yield()
+};
+
+// cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+	// load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
+
+	LdSlot_ mac_gte_load_tri_verts(R_VertBase, R_T0, R_T1, R_T2), GteDelay_ load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)), LdSlot_
+	GteDelay_ nop, gte_cmdw_rotate_translate_perspective_triple,
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), GteDelay_ nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
+		/* BD-slot: Write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
+		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded —
+		 * harmless because the OT entry that points to this prim is created later. */
+		BdSlot_ store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)), LdSlot_
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(R_PrimCursor),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(R_PrimCursor),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), BdSlot_ nop,
+			mac_insert_ot_tag(R_OtBase, R_PrimCursor, S_(Poly_G4)),
+			mac_format_g4_color(R_PrimCursor,
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
+	LdSlot_ add_ui_self(    R_TapePtr, S_(Binds_FloorTri)),
+	mac_yield()
+};
+
+// atom_dbg_skip
+internal
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(R_FaceCursor, R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase,   R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot.
+		/* Format Primitive */
+		mac_gte_store_f3(R_PrimCursor),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
+			mac_format_f3_color(R_PrimCursor, 0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag(R_OtBase, R_PrimCursor, S_(Poly_F3)),   /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+	// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)), LdSlot_
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/hello_camera/hello_camera.c

@@ -0,0 +1,454 @@
+#pragma region Vendors
+#include <stdio.h>
+#include <stdlib.h>
+// #include <assert.h>
+// #include "libgpu.h"
+// #include "libetc.h"
+// #include "libgte.h"
+#pragma endregion Vendors
+
+#pragma region Duffle Headers
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+
+#include "duffle/word_count.metadata.h"
+
+#include "duffle/dsl.h"
+#include "duffle/memory.h"
+#include "duffle/math.h"
+
+#include "duffle/gcc_asm.h"
+#include "duffle/mips.h"
+#include "duffle/gp.h"
+#include "duffle/gte.h"
+#include "duffle/pad.h"
+
+#include "duffle/dsl.atom.h"
+#include "duffle/lottes_tape.h"
+
+#include "duffle/bios.h"
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/pad.c"
+#include "duffle/math.atom.h"
+#include "duffle/mips.atom.c"
+#include "duffle/gte.atom.c"
+#include "duffle/gp.atom.c"
+#include "duffle/pad.atom.c"
+#include "duffle/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Hello Camera Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "gen/auto_reg.h"
+
+#include "hello_camera.h"
+#pragma endregion Hello Camera Headers
+
+#pragma region Hello Joypad TUs
+#include "hello_camera.atom.c"
+#pragma endregion Hello Joypad TUs
+
+enum {
+	MemTape_Len              = 512,
+
+	ResolveLookAtArena_Words = 1024,
+	ResolveLookAtArena_Size  = ResolveLookAtArena_Words * S_(MipsCode),
+
+	CT_InitAtomMem_Words     = Kilo_(4),
+	CT_InitAtomMem_Size       = CT_InitAtomMem_Words * S_(MipsCode),
+};
+typedef Struct_(SMemory) {
+	PrimitiveArena          primitives;
+	A2_OrderingTable_Buffer ordering_tbl;
+	DoubleBuffer            screen_buf;
+	S4                      active_buf_id;
+
+	U4 MemTape[MemTape_Len];
+
+	MT3_S2S4 tform_world;
+	MT3_S2S4 tform_view;
+
+	Camera cam;
+
+	Ent_Cube  cube;
+	Ent_Floor floor;
+
+	PadBiosRaw pad_raw[2];
+	PadState   pad[2];
+
+	// TODO(Ed): We don't need this we can just cast at any point an address to a desired view of scratchpad, we have the address.
+	U4_V scratchpad; // d-cache
+
+	U1 ct_init_atom_mem[CT_InitAtomMem_Size];
+	MipsAtom* normalize_v3s4;
+	MipsAtom* gte_cross_v3s4;
+
+	U1        resolve_look_at_mem[ResolveLookAtArena_Size];
+	MipsAtom* resolve_look_at_bundle[AtomBundle_Len(resolve_look_at)];
+};
+global SMemory smem;
+extern SMemory smem;
+
+#define pad0_btn_(btn) btn & smem.pad[0].buttons
+#define pad1_btn_(btn) btn & smem.pad[1].buttons
+
+I_ B1* prim__alloc(U4 type_width, Str8 type_name) {
+	gknown PrimitiveArena* pa  = & smem.primitives;
+	gknown B1*             buf = (B1*) r_(smem.primitives.buf)[smem.active_buf_id];
+	assert(pa->used + type_width < PrimitiveBuff_Len);
+	B1* next  = buf + pa->used;
+	pa->used += type_width;
+	return next;
+}
+#define prim_alloc(type) (type*)prim__alloc(S_(type), slit( stringify(type)))
+
+I_ void resolve_look_at_c11(MT3_S2S4* look_at, P3_S4* eye, P3_S4* target, V3_S4* up_in) {
+// RGA(Lengyel): Build matrix expansion of a rigid transformation. Corresponding motor is not constructed; we write the LA form for GTE.
+// Preconditions: eye != target, up_in not collinear with (target - eye).
+	V3_S4 right, up, forward;
+	V3_S4 ux, uy, uz;
+	V3_S4 pos, off;
+
+	forward = target[0]; sub_v3s4(& forward, eye[0]); // RGA(Lengyel): Affine point - point = zero-weight direction.
+	normalize_v3s4(& forward, & uz);                  // RGA(Lengyel): Normalize the direction bulk. Not finite-point unitization.
+
+	cross_v3s4(& uz,   up_in, & right); normalize_v3s4(& right, & ux); // RGA(Lengyel): Complement(Wedge(forward, up_in)) -> right axis.
+	cross_v3s4(& uz, & ux,    & up);    normalize_v3s4(& up,    & uy); // RGA(Lengyel): Complement(Wedge(forward, right)) -> up axis.
+
+	// RGA(Lengyel): matrix expansion of the world-to-camera rotation (basis rows).
+	look_at->m[0][0] = ux.x; look_at->m[0][1] = ux.y; look_at->m[0][2] = ux.z;
+	look_at->m[1][0] = uy.x; look_at->m[1][1] = uy.y; look_at->m[1][2] = uy.z;
+	look_at->m[2][0] = uz.x; look_at->m[2][1] = uz.y; look_at->m[2][2] = uz.z;
+
+	pos = eye[0]; mul_v3s4(& pos, v3s4(-1,-1,-1)); // RGA(Lengyel): -eye in world coordinates (spatial bulk only; implicit weight is dropped).
+
+	// RGA(Lengyel): R * (-eye) is the full matrix translation column. 
+	// Motor translator would store half this displacement in m.xyz; GTE consumes full column.
+	mul_m3s2_v3s4(look_at, & pos, & off);
+	trans_m3s2(   look_at,        & off);
+}
+FI_ void camera_look_at_c11(Camera* c, P3_S4* target, V3_S4* up_in) { resolve_look_at_c11(& c->look_at, & c->pos, target, up_in); }
+
+internal void compile_init_atoms(void) {
+	AtomArena ab = atomarena_make(slice_ut_arr(smem.ct_init_atom_mem));
+	RegFile   rf = regfile(regfile_abi_mask);
+#define ralloc() regfile_alloc(& rf)
+#define ralloc_v3() { ralloc(), ralloc(), ralloc() }
+
+	smem.gte_cross_v3s4 = gte_cross_v3s4(& ab,
+		RegUse_(gte_cross_v3s4) {
+			.a  = ralloc_v3(),
+			.b  = ralloc_v3(),
+			.x  = ralloc(),
+			.y  = ralloc(),
+			.z  = ralloc(),
+		});
+	regfile_reset(& rf);
+
+	smem.normalize_v3s4 = build_normalize_v3s4(& ab,
+		RegUse_(build_normalize_v3s4) {
+			.res = ralloc_v3(),
+			.t0     = ralloc(),
+			.t1     = ralloc(),
+			.t2     = ralloc(),
+			.t3     = ralloc(),
+			.t4     = ralloc(),
+			.t5     = ralloc(),
+		});
+	regfile_reset(& rf);
+
+	assert(ab.used <= CT_InitAtomMem_Size);
+#undef ralloc
+#undef ralloc_v3
+}
+
+internal void compile_resolve_look_at(void) {
+	AtomArena ab = atomarena_make(slice_ut_arr(smem.resolve_look_at_mem));
+	AtomBundle_resolve_look_at_R bundle = C_(void*, smem.resolve_look_at_bundle);
+
+	/* R_ScratchBase (= R_SP) is a tape carrier preserved across atoms; no carrier
+	 * pin is needed in the regfile. The standard 24-register pool is sufficient. */
+	RegFile rf = regfile(regfile_abi_mask);
+#define ralloc() regfile_alloc(& rf)
+#define ralloc_v3() { ralloc(), ralloc(), ralloc() }
+
+	bundle->input_and_sub = AtomBundleEntry_(resolve_look_at, input_and_sub)(& ab,
+		RegUse_(resolve_look_at_input_and_sub) {
+			.target  = ralloc(),
+			.eye     = ralloc(),
+			.up_in   = ralloc(),
+			.t0      = ralloc(),
+			.t1      = ralloc(),
+			.t2      = ralloc(),
+			.t3      = ralloc(),
+			.t4      = ralloc(),
+		});
+	regfile_reset(& rf);
+
+	bundle->normalize_fwd_uz   = smem.normalize_v3s4;
+	bundle->cross_to_right     = smem.gte_cross_v3s4;
+	bundle->normalize_right_ux = smem.normalize_v3s4;
+	bundle->cross_to_up        = smem.gte_cross_v3s4;
+	bundle->normalize_up_uy    = smem.normalize_v3s4;
+
+	bundle->pop_mv_trans = resolve_look_at__pop_mv_trans(& ab,
+		RegUse_(resolve_look_at__pop_mv_trans){
+			.look_at = ralloc(),
+			.eye     = ralloc(),
+			.row     = ralloc_v3(),
+			.t6      = ralloc(),
+			.t7      = ralloc(),
+			.t8      = ralloc(),
+		});
+
+	/* Sanity check: arena didn't overflow. */
+	assert(ab.used <= ResolveLookAtArena_Size);
+#undef ralloc
+}
+
+/* Emit the resolve_look_at bundle into the tape. Called once per frame from update(). */
+I_ void resolve_look_at(TapeBuilder_R tb
+	,	MT3_S2S4* look_at
+	,	P3_S4*    eye
+	,	P3_S4*    target
+	,	V3_S4*    up_in
+){
+	/* Typed view of the scratchpad for field-address arithmetic. */
+	ResolveLookAtScratch* sp = C_scratch(ResolveLookAtScratch*);
+	AtomBundle_resolve_look_at_R bundle = C_(void*, smem.resolve_look_at_bundle);
+
+	tb_emit(tb, bundle->input_and_sub); {
+		tb_data(tb, u4_(target));
+		tb_data(tb, u4_(eye));
+		tb_data(tb, u4_(up_in));
+	}
+	tb_emit(tb, bundle->normalize_fwd_uz); {
+		tb_data(tb, u4_(O_(ResolveLookAtScratch, fwd) | (O_(ResolveLookAtScratch, uz) << 16)));
+	}
+	tb_emit(tb, bundle->cross_to_right); {
+		tb_data(tb, u4_(& sp->uz));
+		tb_data(tb, u4_(& sp->up_in));
+		tb_data(tb, u4_(& sp->right));
+	}
+	tb_emit(tb, bundle->normalize_right_ux); {
+		tb_data(tb, u4_(O_(ResolveLookAtScratch, right) | (O_(ResolveLookAtScratch, ux) << 16)));
+	}
+	tb_emit(tb, bundle->cross_to_up); {
+		tb_data(tb, u4_(& sp->uz));
+		tb_data(tb, u4_(& sp->ux));
+		tb_data(tb, u4_(& sp->up));
+	}
+	tb_emit(tb, bundle->normalize_up_uy); {
+		tb_data(tb, u4_(O_(ResolveLookAtScratch, up) | (O_(ResolveLookAtScratch, uy) << 16)));
+	}
+	tb_emit(tb, bundle->pop_mv_trans); {
+		tb_data(tb, u4_(look_at));
+	}
+}
+
+GCC_OPTIMIZATION_DISABLE
+void update(PrimitiveArena* pa, U4* ordering_buf) 
+{
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
+	// Pad Input
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			// Grab latest state from bios.
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw,   & smem.pad_raw[0]);
+				tb_data_(state, & smem.pad[0]);
+			// tb_emit_(pad_bios_snapshot);
+			// 	tb_data_(raw,   & smem.pad_raw[1]);
+			// 	tb_data_(state, & smem.pad[1]);
+
+			tb_emit_(pad_input_cam);
+				tb_data_(state, & smem.pad[0]);
+				tb_data_(cam, & smem.cam);
+
+			// tb_emit_(pad_input_cube_rotation);
+			// 	tb_data_(state,     & smem.pad[0]);
+			// 	tb_data_(cube_rot,  & smem.cube.rot);
+			// 	tb_data_(floor_rot, & smem.floor.rot);
+		}
+	}
+
+	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
+
+	// Update the position based on acceleration and velocity
+	gknown V3_S4_R pos = & smem.cube.pos;
+	gknown V3_S4_R vel = & smem.cube.vel;
+	gknown V3_S4_R acc = & smem.cube.accel;
+	add_v3s4(vel, acc[0]);
+	add_v3s4_fp(pos, vel[0]);
+
+	if (pos->y + 150 > smem.floor.pos.y) vel->y *= -1;
+
+	// Prep
+	S4 nclip = 0;
+	S4 orderingtbl_z = 0;
+	A2_S2 p; //???
+	S4 flag; //????
+
+	B4 use_c11_path = false;
+	if (use_c11_path) {
+		camera_look_at_c11(& smem.cam, & smem.cube.pos, & v3s4(0, -fp_one, 0));
+	}
+	if (use_c11_path == false)
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			resolve_look_at(& tb, & smem.cam.look_at, & smem.cam.pos, & smem.cube.pos, & v3s4(0, -fp_one, 0));
+		}
+	}
+
+	// Draw cube
+	if (1)
+	{
+		mt3s2s4_rotation   (& smem.cube.rot,    & smem.tform_world);
+		mt3s2s4_translation(& smem.tform_world, & smem.cube.pos);
+		mt3s2s4_scale      (& smem.tform_world, & smem.cube.scale);
+
+		// Combine world and look_at matrix.
+		gte_comp_coord_m3s2(& smem.cam.look_at, & smem.tform_world, & smem.tform_view);
+		gte_matrix_set_rotation   (& smem.tform_view);
+		gte_matrix_set_translation(& smem.tform_view);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.cube.faces));
+				tb_data(& tb, u4_(smem.cube.verts));
+				tb_data(& tb, u4_(ordering_buf));
+
+			for (U4 i = 0; i < Cube_num_faces; i++) {
+				// Two triangles per quad face: (x,y,z) and (x,z,w)
+				tb_emit(& tb, cube_g4_face);
+			}
+
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));// Fire off the tape (bigger-clobber variant).
+
+		// smem.cube.rot.y += 30;
+	}
+	// Draw floor
+	if (1)
+	{
+		mt3s2s4_rotation   (& smem.floor.rot,   & smem.tform_world);
+		mt3s2s4_translation(& smem.tform_world, & smem.floor.pos);
+		mt3s2s4_scale      (& smem.tform_world, & smem.floor.scale);
+
+		// Combine world and look_at matrix.
+		gte_comp_coord_m3s2(& smem.cam.look_at, & smem.tform_world, & smem.tform_view);
+
+		gte_matrix_set_rotation   (& smem.tform_view);
+		gte_matrix_set_translation(& smem.tform_view);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris?
+		// The tape atoms in-flight should not need to care.
+
+		// Prepare the tape. (Push protocol to tape)
+		tb.used = 0; tb_scope(& tb) {
+			// tb_emit(& tb, set_gte_mt3s2s4);
+			// 	tb_data(& tb, u4_(& smem.tform_view));
+
+			tb_emit(& tb, rbind_floor_f3_face);
+			// TODO(Ed): Just use a single context struct ref?
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.floor.faces));
+				tb_data(& tb, u4_(smem.floor.verts));
+				tb_data(& tb, u4_(ordering_buf));
+			for (U4 i = 0; i < Floor_num_faces; i++) {
+				tb_emit(& tb, floor_f3_face);
+			}
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));// Fire off the tape (bigger-clobber variant).
+
+		// C-side state (pa->used) has already been updated by the tape!
+		// smem.floor.rot.y += 5;
+	}
+}
+GCC_OPTIMIZATION_ENABLE
+
+void render(void) {
+}
+
+void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
+	draw_sync(0);
+	vsync(0);
+	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
+	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
+	{
+		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
+		pa->used = 0;
+	}
+	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
+}
+
+GCC_OPTIMIZATION_DISABLE
+int main(void)
+{
+	smem = (SMemory){0};
+	// TODO(Ed): remove this field we don't need it in smem.
+	smem.scratchpad = C_(U4_V, Scratchpad_Loc);
+	// smem.primitives.used = 0;
+	// smem.active_buf_id   = 0;
+	smem.cam.pos = v3s4(500, -1000, -1500);
+	/*Persistent Entity Setup*/{
+		ent_cube128_init(& smem.cube.verts, & smem.cube.faces); {
+			Ent_Cube* cube = & smem.cube;
+			cube->rot    = v3s2(0, 0, 0);
+			cube->scale  = v3s4_fp_one();
+			cube->accel  = v3s4(0, 1, 0);
+			cube->pos    = v3s4(0, -400, 1800);
+		}
+		ent_floor_init(& smem.floor.verts, & smem.floor.faces); {
+			Ent_Floor* floor = & smem.floor;
+			floor->rot   = v3s2(0, 0, 0);
+			floor->pos   = v3s4(0, 450, 1800);
+			floor->scale = v3s4_fp_one();
+		}
+	}
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape)); {
+		reset_graph(0);
+		/* Direct BIOS: poll both ports during VBlank. */
+		pad_bios_init_start(& smem.pad_raw[0], & smem.pad_raw[1]);
+
+		compile_init_atoms();
+		compile_resolve_look_at();
+
+		/* Pinned registers for the GPU init atom. */
+		register U4*           io_base_addr rgcc(R_IO_BaseAddr) = u4_r(IO_BASE_ADDR);
+		register DoubleBuffer* screen_buf   rgcc(R_ScreenBuf)   = & smem.screen_buf;
+		tb.used = 0; tb_scope_run(& tb) {
+			tb_emit(& tb, screen_env_init);
+			tb_emit(& tb, gp_screen_init);
+		}
+	}
+	while (1) {
+		gknown S4* active_buf_id  = & smem.active_buf_id;
+		gknown U4* ordering_buf   = r_(smem.ordering_tbl)[active_buf_id[0]];
+		gknown PrimitiveArena* pa = & smem.primitives;
+		update(pa, ordering_buf);
+		render();
+		gp_display_frame(& smem.screen_buf, active_buf_id, ordering_buf, pa);
+	};
+	return 0;
+}
+GCC_OPTIMIZATION_ENABLE

code/hello_camera/hello_camera.h

@@ -0,0 +1,102 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "duffle/dsl.h"
+#	include "duffle/math.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#endif
+
+enum {
+	// PrimitiveBuff_Len = 4096,
+	// OrderingTbl_Len   = 2048,
+	PrimitiveBuff_Len = 131072,
+	OrderingTbl_Len   = 8192,
+};
+
+enum {
+	ScreenRes_X = 320,
+	ScreenRes_Y = 240,
+	ScreenZ     = 320,
+	ScreenRes_CenterX = (ScreenRes_X >> 1),
+	ScreenRes_CenterY = (ScreenRes_Y >> 1),
+};
+
+typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
+typedef Array_(OrderingTable_Buffer, 2);
+
+typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
+typedef Array_(PrimitiveBuffer, 2);
+typedef Struct_(PrimitiveArena) {
+	A2_PrimitiveBuffer buf;
+	U4                 used;
+};
+
+#define Cube_num_verts 8
+typedef Array_(V3_S2, Cube_num_verts);
+#define Cube_num_faces 6
+typedef Array_(V4_S2, Cube_num_faces);
+I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
+	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
+		{ -128, -128, -128 },
+		{  128, -128, -128 },
+		{  128, -128,  128 },
+		{ -128, -128,  128 },
+		{ -128,  128, -128 },
+		{  128,  128, -128 },
+		{  128,  128,  128 },
+		{ -128,  128,  128 }
+	};
+	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
+		{ 3, 2, 0, 1 },
+		{ 0, 1, 4, 5 },
+		{ 4, 5, 7, 6 },
+		{ 1, 2, 5, 6 },
+		{ 2, 3, 6, 7 },
+		{ 3, 0, 7, 4 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
+	return;
+}
+typedef Struct_(Ent_Cube) {
+	V3_S4 accel;
+	V3_S4 vel;
+	V3_S4 pos; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	V3_S4 scale;
+	V3_S2 rot;
+	A8_V3_S2 verts;
+	A6_V4_S2 faces;
+};
+
+#define Floor_num_verts 4
+typedef Array_(V3_S2, Floor_num_verts);
+#define Floor_num_faces 2
+typedef Array_(V3_S2, Floor_num_faces);
+I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
+	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
+		{ -900, 0, -900 },
+		{ -900, 0,  900 },
+		{  900, 0, -900 },
+		{  900, 0,  900 },
+	};
+	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
+		{ 0, 1, 2 },
+		{ 1, 3, 2 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
+};
+typedef Struct_(Ent_Floor) {
+	V3_S4 accel;
+	V3_S4 pos; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};
+
+typedef Struct_(Camera) {
+	P3_S4    pos; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	V3_S2    rot;
+	MT3_S2S4 look_at;
+};

code/hello_gte/gen/gte_hello.offsets.h

@@ -5,10 +5,10 @@
 #pragma region hello_gte_tape
 
 
-// --- atom: cube_g4_face (87 words) ---
+// --- atom: cube_g4_face (77 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           48
-#define _atom_offset_bounds_chk_cube_g4_face_exit     12
+#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
 
 enum {
     atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
@@ -18,7 +18,7 @@ enum {
 // --- atom: floor_f3_face (58 words) ---
 
 #define _atom_offset_culling_floor_f3_face_exit       25
-#define _atom_offset_bounds_chk_floor_f3_face_exit    13
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
 
 enum {
     atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,

code/hello_gte/gen/hello_gte.offsets.h

@@ -0,0 +1,29 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\hello_gte\hello_gte.tape.c
+#pragma once
+
+#pragma region hello_gte.tape
+
+
+// --- atom: cube_g4_face (77 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+#pragma endregion hello_gte.tape
+

code/hello_gte/hello_gte.c

@@ -14,16 +14,16 @@
 #include "duffle/gp.h"
 #include "duffle/gte.h"
 
-#	include "duffle/gen/duffle.macs.h"
-#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
 #include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
 #include "duffle/word_count.metadata.h"
 
-#	include "gen/gte_hello.offsets.h"
+#	include "gen/offsets.h"
 #include "hello_gte.h"
 
-#include "hello_gte_tape.c"
+#include "hello_gte.tape.c"
 
 typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
 typedef Array_(OrderingTable_Buffer, 2);
@@ -122,7 +122,7 @@ global SMemory smem;
 extern SMemory smem;
 
 // TODO(Ed):
-FI_ U4* spad_warm(MipsAtom atom) {
+FI_ U4* spad_warm(Slice_MipsCode atom) {
 	return nullptr;
 }
 

code/hello_gte/hello_gte.tape.c

@@ -0,0 +1,218 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+#	include "duffle/atom_dsl.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/word_count.metadata.h"
+#	include "gen/offsets.h"
+#	include "hello_gte.h"
+#endif
+
+#pragma region MACs (Mips Atom components)
+
+
+
+#pragma endregion MACs
+
+#pragma region Baked Atoms
+
+/* DIAGNOSTIC 1: Pure tape loop test */
+internal MipsAtom_(diag_yield) { mac_yield() };
+
+/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
+internal MipsAtom_(diag_color) {
+	store_word(  R_0, R_T7, 0), 
+	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
+	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
+	store_word(  R_AT, R_T7, 4),
+	
+	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
+	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
+
+	add_ui(          R_T1, R_0,  10),
+	shift_lleft_self(R_T1,       S_(U4)/2), 
+	add_u_self(      R_T1,       R_T6),         
+	
+	load_word(   R_AT, R_T1, 0),        
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
+	store_word(  R_AT, R_T7, 0),       
+	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
+	or_u_self(  R_AT, R_V0),          
+	store_word( R_AT, R_T1, 0),       
+
+	add_ui(R_T7, R_T7, 20),          
+
+	mac_yield()
+};
+
+/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
+internal MipsAtom_(diag_gte) {
+	/* Load 3 indices */
+	load_half_u(R_T0, R_T4, 0),
+	load_half_u(R_T1, R_T4, 2),
+	load_half_u(R_T2, R_T4, 4),
+
+	/* Load Vertices into GTE */
+	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
+
+	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
+
+	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+
+	/* Run Math */
+	nop2, gte_cmdw_rtpt,
+	nop2, gte_cmdw_nclip,
+	nop2,
+
+	/* Advance Face Cursor and Yield */
+	add_ui(R_T4, R_T4, 8),
+
+	mac_yield()
+};
+
+typedef Struct_(Binds_CubeTri) {
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
+	mac_yield()
+};
+
+ // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
+
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
+		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+
+		add_ui(      R_AT, R_0, OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT), 
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+			mac_insert_ot_tag_g4(),
+			mac_format_g4_color(
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
+	mac_yield()
+};
+
+// atom_dbg_skip
+internal
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(  R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
+		/* Format Primitive */
+		mac_gte_store_f3(),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < OrderingTbl_Len (Branch forward to skip insertion) */
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
+			mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(),                 /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/hello_joypad/gen/macs.h

@@ -0,0 +1,41 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_joypad/
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.c
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.h
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.atom.c
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#define WORD_COUNT(name, count)  enum { words_##name = (count) };
+#endif
+
+#define mac_put_disp_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_disp_env, 5)
+
+#define mac_put_draw_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port) /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */ \
+,	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port) /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port) /* code[1]  TextureWindow (tw=(0,0)) */ \
+,	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port) /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */ \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port) /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */ \
+,	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port) /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */ \
+,	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port) /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */ \
+,	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port) /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */ \
+,	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port) /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port)	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_draw_env, 16)
+

code/hello_joypad/gen/offsets.h

@@ -0,0 +1,76 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_joypad\
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.c
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.h
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.atom.c
+#pragma once
+
+#pragma region hello_joypad
+
+
+// --- atom: cube_g4_face (76 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           41
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+// --- atom: pad_bios_snapshot (78 words) ---
+
+#define _atom_offset_snap_root_skip_disconnected      8
+#define _atom_offset_disconnected_snap_end            61
+#define _atom_offset_case_2_id_dispatch               8
+#define _atom_offset_pending_snap_end                 51
+#define _atom_offset_id_dispatch_try_analog_stick     11
+#define _atom_offset_id_dispatch_snap_end             38
+#define _atom_offset_try_analog_stick_try_analog_pad  12
+#define _atom_offset_analog_stick_snap_end            24
+#define _atom_offset_try_analog_pad_try_unsupported   11
+#define _atom_offset_analog_pad_snap_end              10
+
+enum {
+    atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
+    atom_offset_disconnected_snap_end = _atom_offset_disconnected_snap_end,
+    atom_offset_case_2_id_dispatch = _atom_offset_case_2_id_dispatch,
+    atom_offset_pending_snap_end = _atom_offset_pending_snap_end,
+    atom_offset_id_dispatch_try_analog_stick = _atom_offset_id_dispatch_try_analog_stick,
+    atom_offset_id_dispatch_snap_end = _atom_offset_id_dispatch_snap_end,
+    atom_offset_try_analog_stick_try_analog_pad = _atom_offset_try_analog_stick_try_analog_pad,
+    atom_offset_analog_stick_snap_end = _atom_offset_analog_stick_snap_end,
+    atom_offset_try_analog_pad_try_unsupported = _atom_offset_try_analog_pad_try_unsupported,
+    atom_offset_analog_pad_snap_end = _atom_offset_analog_pad_snap_end,
+};
+
+// --- atom: pad_apply_input (60 words) ---
+
+#define _atom_offset_dpad_left_exit_dpad_left         6
+#define _atom_offset_dpad_right_exit_dpad_right       6
+#define _atom_offset_dead_zone_low_check_dead_low_active  8
+#define _atom_offset_dead_zone_high_check_dead_high_active  15
+#define _atom_offset_dead_zone_skip_exit_stick        24
+#define _atom_offset_end_low_exit_stick               12
+
+enum {
+    atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
+    atom_offset_dpad_right_exit_dpad_right = _atom_offset_dpad_right_exit_dpad_right,
+    atom_offset_dead_zone_low_check_dead_low_active = _atom_offset_dead_zone_low_check_dead_low_active,
+    atom_offset_dead_zone_high_check_dead_high_active = _atom_offset_dead_zone_high_check_dead_high_active,
+    atom_offset_dead_zone_skip_exit_stick = _atom_offset_dead_zone_skip_exit_stick,
+    atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
+};
+
+#pragma endregion hello_joypad
+

code/hello_joypad/hello_joypad.atom.c

@@ -0,0 +1,638 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#		include "duffle/gen/macs.h"
+#		include "duffle/gen/offsets.h"
+#	include "duffle/dsl.atom.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/mips.h"
+#	include "duffle/gte.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#	include "duffle/word_count.metadata.h"
+#	include "duffle/psyq.h"
+#	include "duffle/math.atom.c"
+#	include "duffle/mips.atom.c"
+#	include "duffle/gte.atom.c"
+#	include "duffle/gp.atom.c"
+#	include "duffle/psyq.atom.c"
+#		include "gen/offsets.h"
+#		include "gen/macs.h"
+#	include "hello_joypad.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(hello_joypad_atom_c);
+
+#pragma region MACs (Mips Atom components)
+
+FI_ Slice_MipsCode ac_put_disp_env(MipsAtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
+	// Sequence per libpsyx PutDispEnv: DrawArea TL → DrawArea BR → Mask → DrawArea TL → DrawArea BR
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_put_draw_env(MipsAtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	/*
+	* ORIGIN: each code word corresponds to the EXACT value libpsyx's PutDrawEnv function would compute for the same DrawEnv settings.
+	* References:
+	*   - libpsyx source: `toolchain/psyq-4_7/lib/libgpu.a` (binary, function `PutDrawEnv`)
+	*   - PSX-SPX doc:     https://problemkaputt.de/psx-spx.htm#gputdrawingcommands
+	*   - PSYQ SDK:        `setdrawenv` / `makelongdr_env` source
+	*   - NOCASH PSX spec: §"GP0(E1h) Draw Mode setting" through §"DR_ENV"
+	*
+	* The 16-word format is documented in the PSYQ SDK manual and on NOCASH's PSX-spec.txt. The libpsyx reference is at:
+	*   ./toolchain/psyq-4_7/lib/libgpu.a
+	*   (binary; the PutDrawEnv implementation builds the 16-word DR_ENV from the user's DRAWENV struct and emits it via GP0 GPU commands.)
+	*
+	* Word indices (libpsyx PutDrawEnv / SetDrawEnv order):
+	*   tag      = (length << 24) | addr                — 16-word packet (1 tag + 15 code)
+	*   code[0]  = DrawMode (dfe=1, dtd=0, tpage=0)     — must come first per libpsyx
+	*   code[1]  = TextureWindow (tw=(0,0))             — bare-cmd word; GPU uses current state
+	*   code[2]  = DrawArea top-left (clip.x=0, clip.y=240)
+	*   code[3]  = DrawArea bottom-right (clip.x+w=320, clip.y+h=480)
+	*   code[4]  = DrawOffset (ofs=(0,0))               — bare-cmd word
+	*   code[5]  = Mask (dtd=0, dfe=1, isbg=1)          — 0xE6 cmd + isbg bit
+	*   code[6]  = Initial-bg-color (isbg=1, r=7, g=7, b=7)
+	*   code[7]  = DrawMode (isbg=1, tpage=0)           — re-asserts DrawMode with isbg
+	*   code[8..10]  = padding (NOP)                    — 3 words to fill the packet
+	*   code[11..12] = TextureWindow bottom-right       — defaults to (0,0,0,0)
+	*   code[13..14] = padding (NOP)                    — completes the 16-word packet
+	*/
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port), /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */
+	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port), /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */
+	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port), /* code[1]  TextureWindow (tw=(0,0)) */
+	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port), /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port), /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */
+
+	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port), /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */
+	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port), /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */
+	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port), /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */
+	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port), /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */
+
+	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+
+	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+
+	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+})
+
+#pragma endregion MACs
+
+#pragma region Baked Atoms
+
+enum {
+	R_ScreenX   = R_T5 atom_reg atom_type(U2),
+	R_ScreenY   = R_T6 atom_reg atom_type(U2),
+	R_ScreenBuf = R_T7 atom_reg, /* Caller-pinned: & smem.screen_buf */
+#define R_ScreenBuf_Code R_T7_Code
+};
+//screen_env_init. Mirrors the libpsyx's SetDefDispEnv + SetDefDrawEnv + the manual enable_auto_clear / initial_bg_color writes.
+internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
+,	atom_reads(R_T0, R_ScreenX, R_ScreenY, R_ScreenBuf)
+, atom_writes(R_T0, R_ScreenX, R_ScreenY)
+) {
+	/* display[0] = (0, 0, 320, 240); rest of struct zeroed. */
+	add_ui(R_ScreenX, R_0, ScreenRes_X), add_ui(R_ScreenY, R_0, ScreenRes_Y),
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area.width) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)       + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace)   + OA_(DoubleBuffer,display,0)),
+
+	/* display[1] = (0, 240, 320, 240); rest of struct zeroed. */
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)     + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace) + OA_(DoubleBuffer,display,1)),
+
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area)         + OA_(DoubleBuffer,draw,0)), /* draw[0].clip_area = (0, 240, 320, 240). C11's SetDefDrawEnv writes clip.y = y_arg. */
+	mac_store_v2s2(R_0, R_ScreenY,                       R_ScreenBuf, O_(DrawEnv,drawing_offset[0]) + OA_(DoubleBuffer,draw,0)), /* draw[0].drawing_offset[0] = (0, 240); C11 passes y_arg as ofs. */
+
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_window = (0, 0, 0, 0); two word-zeroes cover the full 8-byte tw field. */
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,0)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,0)),
+
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,drawing_offset[0].x)  + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_page = 10 (gp0_tpage_default). C11 SetDefDrawEnv at C11_only.elf:0x8001273C writes the same 0x0A. . */
+	add_ui(R_T0, R_0, gp0_tpage_default),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,0)),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0] control bytes: flag_dither=1, flag_draw_on_display=1 (the dfe bit per psx-spx; libpsyx sets it via `SetDefDrawEnv`'s conditional at C11_only.elf:0x80012728), enable_auto_clear=1. Each byte is named;
+	 * the previous `store_word(R_0, ..., +20)` overwrote all four with zero. */
+	add_ui(R_T0, R_0, 1),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].initial_bg_color = (r=7, g=7, b=7). */
+	add_ui(R_T0, R_0, 7),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,0)),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,1)),
+
+	mac_yield(),
+};
+
+enum {
+	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+#define R_IO_BaseAddr_Code R_T4_Code
+};
+internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads(R_IO_BaseAddr)) {
+	store_word(R_0, R_IO_BaseAddr, GPIO_PORT1_OFFSET),                                   /* GP1(00h) Reset */
+	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(01h) ClearFIFO */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(02h) AckIRQ */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(03h) Display ON */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(04h) DMADirection=2 (CPU→GPU). libpsyx's per-frame PutDrawEnv/DrawOTag use DMA2; without this the DMA queue never drains. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(05h) StartDisplayArea (X=0, Y=0) */
+
+	/* GP1: DisplayMode + Display Ranges */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+
+	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
+	load_upper_i(R_T5, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_T5, gte_cr_OFX_Code),
+	load_upper_i(R_T5, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_T5, gte_cr_OFY_Code),
+
+	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
+	add_ui(R_T5, R_0, ScreenZ), gte_mv_to_ctrl_r(R_T5, gte_cr_H_Code),
+
+	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
+	mac_gcmd_push(gp1_word_DisplayOn(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_yield(),
+};
+
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
+#define R_PrimCursor_Code  R_T7_Code
+#define R_FaceCursor_Code  R_T4_Code
+#define R_VertBase_Code    R_T5_Code
+#define R_OtBase_Code      R_T6_Code
+};
+
+typedef Struct_(Binds_CubeTri) {
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
+	mac_yield()
+};
+
+ // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
+
+	mac_gte_load_tri_verts(R_VertBase, R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
+		/* BD-slot: write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
+		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded —
+		 * harmless because the OT entry that points to this prim is created later, only on the body path. */
+		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(R_PrimCursor),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(R_PrimCursor),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+			mac_insert_ot_tag_g4(R_OtBase, R_PrimCursor),
+			mac_format_g4_color(R_PrimCursor,
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
+	mac_yield()
+};
+
+// atom_dbg_skip
+internal
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(R_FaceCursor, R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase,   R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot.
+		/* Format Primitive */
+		mac_gte_store_f3(R_PrimCursor),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
+			mac_format_f3_color(R_PrimCursor, 0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(R_OtBase, R_PrimCursor),         /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+	// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+/* ----- pad_bios_snapshot -----
+ * Per-frame snapshot of one BIOS pad buffer into PadState.
+ * Decoder (branch ladder on raw[0] status + raw[1] id):
+ *   1. raw[0] == 0xFF         -> Disconnected (buttons=0, axes=0x80)
+ *   2. raw[0]==0 && raw[1]==0 -> Pending      (buttons=0, axes=0x80)
+ *   3. raw[1] == 0x41         -> Digital      (buttons normalized; axes=0x80)
+ *   4. raw[1] == 0x53         -> AnalogStick  (buttons normalized; axes from raw[4..7])
+ *   5. raw[1] in 0x7x         -> AnalogPad    (buttons normalized; axes from raw[4..7])
+ *   6. else                   -> Unsupported  (buttons=0, axes=0x80)
+ *
+ * Buttons normalization: byte_swap16((~raw_buttons) & 0xFFFF).
+ *   raw_buttons    = load_half_u(raw, 2) = raw[2] | (raw[3] << 8).
+ *   byte_swap16(x) = (x >> 8) | (x << 8); nor(x, R_0) = ~x. store_half truncates to 16 bits so the upper-16 mask is implicit in the store.
+ *
+ * Register use (atom-local; no wave-context touched):
+ *   R_T0 = raw base      (kept throughout; axes loads read raw[4..7] from R_T0)
+ *   R_T1 = state base    (kept throughout; all stores go through R_T1)
+ *   R_T2 = raw[0] status (alive across the disc/pending/id dispatch, then dead)
+ *   R_T3 = raw[1] id     (alive across the id dispatch, then dead)
+ *   R_T4 = scratch       (shifts, compares, immediate loads, store values)
+ *   R_T5 = scratch       (parallel lui+ori for the 0x80808080 axes constant + byte-swap target)
+ */
+enum {
+	R_PadRaw    = R_T0 atom_reg atom_type(U1),
+	R_PadState  = R_T1 atom_reg,
+	R_RawStatus = R_T2 atom_reg,
+	R_RawId     = R_T3 atom_reg,
+};
+typedef Struct_(Binds_PadBiosSnapshot) {
+    PadBiosRaw* raw;
+    PadState*   state;
+};
+internal MipsAtom_(pad_bios_snapshot) atom_info(atom_bind(Binds_PadBiosSnapshot)
+, atom_reads( R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+, atom_writes(R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+) {
+	/* === Bind consumption: T0 = raw, T1 = state, advance R_TapePtr by 8. */
+	load_word(R_PadRaw,   R_TapePtr, O_(Binds_PadBiosSnapshot,raw)),
+	load_word(R_PadState, R_TapePtr, O_(Binds_PadBiosSnapshot,state)),
+	add_ui_self(          R_TapePtr, S_(Binds_PadBiosSnapshot)),
+
+	/* === Read raw[0] (status) + raw[1] (id) */
+	load_byte_u(R_RawStatus, R_PadRaw, 0),
+	load_byte_u(R_RawId,     R_PadRaw, 1),
+
+atom_label(snap_root) /* === Case 1: Disconnected (status == 0xFF). */
+	add_ui(R_T4, R_0, 0xFF), branch_ne(R_RawStatus, R_T4, atom_offset(snap_root, skip_disconnected)),
+	/* BD-slot: pre-compute PadStatus_Disconnected. Branch reads R_T4=0xFF in EX before this WB completes.
+	 * If branch NOT taken (fall through to pending/id_dispatch), R_T4 is overwritten by the next case body's add_ui — harmless. */
+
+atom_label(disconnected) /* === Disconnected body. */
+	/* R_T4 = PadStatus_Disconnected from snap_root BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(disconnected, snap_end)),
+		/* BD-slot: load next atom's entry point (replaces the nop).
+		 * The unconditional branch always jumps to snap_end, where mac_yield_tail()
+		 * transfers control to R_AtomJmp without re-loading it. */
+		mac_yield_load(),
+atom_label(skip_disconnected)
+
+	/* === Case 2: Pending (status == 0 && id == 0)
+	 * Combined check: if (status | id) != 0 then skip to id_dispatch.
+	 * Falls through to the Pending case only when both are zero. */
+	or_u_self(R_RawStatus, R_RawId), branch_ne(R_RawStatus, R_0, atom_offset(case_2, id_dispatch)),
+	/* BD-slot: pre-compute PadStatus_Pending. Branch reads R_RawStatus in EX before this WB completes.
+	 * If branch NOT taken (fall through to id_dispatch), R_T4 is overwritten by the digital/analog body add_ui — harmless. */
+
+atom_label(pending) /* === Pending body */
+	/* R_T4 = PadStatus_Pending from case_2 BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(pending, snap_end)),
+		mac_yield_load(),
+
+atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
+	add_ui(R_T4, R_0, 0x41), branch_ne(R_RawId, R_T4, atom_offset(id_dispatch, try_analog_stick)),
+	/* BD-slot: pre-compute PadStatus_Digital. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_stick), R_T4 is overwritten by the analog body add_ui. */
+
+	/* === Digital body (status, buttons normalize, axes=0x80, id, branch. */
+	/* R_T4 = PadStatus_Digital from id_dispatch BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),
+	/* Fill R_T4's load-delay slot with the 0x80808080 axes constant into R_T5
+	 * (R_T5 is dead on this path; it's only consumed at the analog_pad range check). */
+	load_upper_i(R_T5, 0x8080), or_i_self(R_T5, 0x8080),
+	nor_u(      R_T4, R_T4, R_0), /* raw_buttons is already in host bit order; no swap needed */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	store_word(  R_T5, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0x41),
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+
+	jump_rel(atom_offset(id_dispatch, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_stick) /* === Case 4: AnalogStick (id == 0x53)*/
+	add_ui(R_T4, R_0, 0x53), branch_ne(R_RawId, R_T4, atom_offset(try_analog_stick, try_analog_pad)),
+	/* BD-slot: pre-compute PadStatus_AnalogStick. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_pad), R_T4 is overwritten by the analog_pad body add_ui. */
+
+atom_label(analog_stick) /* === AnalogStick body
+	* Axes are loaded as two halfwords: raw[6..7] → left_xy (sh at offset 8), raw[4..5] → right_xy (sh at offset 10).
+	* R_T5 holds left_xy / id-value in turn (it's dead on this path — only consumed at the analog_pad range check). */
+	/* R_T4 = PadStatus_AnalogStick from try_analog_stick BD-slot. */
+	store_word(  R_T4, R_PadState, O_(PadState,status)),
+	load_half_u( R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u( R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot (doesn't read R_T4) */
+	nor_u(       R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half(  R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u( R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half(  R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half(  R_T4, R_PadState, O_(PadState,right_x)),
+	add_ui(      R_T5, R_0, 0x53),                        /* R_T5 = id value (clobbers left_xy, already stored) */
+	store_byte(  R_T5, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(analog_stick, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_pad) /* === Case 5-6: AnalogPad (id & 0xF0 == 0x70) */
+	and_i(    R_T4, R_RawId, 0xF0),
+	add_ui(   R_T5, R_0,     0x70),
+	branch_ne(R_T4, R_T5, atom_offset(try_analog_pad, try_unsupported)),
+	/* BD-slot: pre-compute PadStatus_AnalogPad. Branch reads R_T4 in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_unsupported), R_T4 is overwritten by the unsupported body add_ui. */
+
+atom_label(analog_pad) /* === AnalogPad body
+	* Same shape as AnalogStick with AnalogPad status. R_T5 holds left_xy (it's dead on this path). */
+	/* R_T4 = PadStatus_AnalogPad from try_analog_pad BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u(R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot */
+	nor_u(      R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u(R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half( R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half( R_T4, R_PadState, O_(PadState,right_x)),
+	store_byte( R_RawId, R_PadState, O_(PadState,id)),
+
+	jump_rel(atom_offset(analog_pad, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the AnalogPad range-check miss. */
+	add_ui(    R_T4, R_0, PadStatus_Unsupported),
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0xFF),                      /* 0xFF sentinel: "unknown id" */
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+	/* Fall through to snap_end. */
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
+atom_label(snap_end)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the case-exit branch. */
+	mac_yield_tail(),
+};
+
+/* ----- pad_apply_input -----
+ * Reads pad[0].buttons + pad[0].left_x; 
+ * Applies the input-semantics deltas to cube_rot.y + floor_rot.y:
+ *   - D-pad Left:  cube_rot.y += 30, floor_rot.y += 5
+ *   - D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5
+ *   - Analog stick X (dead zone 0x70..0x90):
+ *       cube delta  = (0x80 - left_x) >> 2  (range approx -32..+32)
+ *       floor delta = (0x80 - left_x) >> 5  (range approx -4..+4)
+ *   - D-pad + analog deltas add when used together.
+ *
+ * Convention:
+ *   pad_state = 0 means no buttons active. 
+ *   The fail-safe zero-button value flows through unchanged, so a disconnected/fresh pad produces no rotation.
+ *   The branch_le_zero pattern below matches the existing pad_input_demo convention (atom body lines 248/257).
+ *
+ * Signed-delta trick:
+ * load_byte_u zero-extends left_x to 32 bits; sub_u from 0x80 wraps to a SIGNED two's-complement value in the negative range;
+ * shift_aright (sra) then correctly sign-extends the shift for both positive (left_x < 0x80) and negative (left_x > 0x80) cases.
+ * Digital pads publish left_x = 0x80 → delta = 0 → no rotation, so the analog step is naturally a no-op for digital controllers.
+ */
+typedef Struct_(Binds_PadApplyInput) {
+    PadState* state;
+    V3_S2*    cube_rot;
+    V3_S2*    floor_rot;
+};
+enum {
+	R_PadStateT5 = R_T5 atom_reg,
+	R_CubeRot    = R_T1 atom_reg,
+	R_FloorRot   = R_T2 atom_reg,
+};
+internal MipsAtom_(pad_apply_input) atom_info(atom_bind(Binds_PadApplyInput)
+, atom_reads(R_T0, R_CubeRot, R_FloorRot, R_T3, R_T4, R_PadStateT5, R_TapePtr)
+, atom_writes(     R_CubeRot, R_FloorRot)
+) {
+	/* Pop Binds from tape (state, cube_rot, floor_rot) */
+	load_word(R_PadStateT5, R_TapePtr, O_(Binds_PadApplyInput,state)),
+	load_word(R_CubeRot,    R_TapePtr, O_(Binds_PadApplyInput,cube_rot)),
+	load_word(R_FloorRot,   R_TapePtr, O_(Binds_PadApplyInput,floor_rot)),
+	add_ui_self(            R_TapePtr, S_(Binds_PadApplyInput)),
+
+	/* Load pad[0].buttons into R_T0. */
+	load_word(R_T0, R_PadStateT5, O_(PadState,buttons)), nop,
+	// Note(Ed): Potential op with delay slot?
+
+	/* D-pad Left: cube_rot.y += 30, floor_rot.y += 5.  */
+	and_i(R_T3, R_T0, pad0_(Pad_Left)), branch_le_zero(R_T3, atom_offset(dpad_left, exit_dpad_left)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, 30),
+		add_si(    R_T3, R_T3, 5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_left)
+
+	/* D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5. */
+	and_i(R_T3, R_T0, pad0_(Pad_Right)), branch_le_zero(R_T3, atom_offset(dpad_right, exit_dpad_right)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, -30),
+		add_si(    R_T3, R_T3, -5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_right)
+
+	/* Analog left-stick X: dead zone 0x70..0x90.
+	 * Cube delta = (0x80 - left_x) >> 2; floor delta = (0x80 - left_x) >> 5. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)),
+
+	/* Dead-zone check: skip analog if left_x in [0x70, 0x90] inclusive. Outside dead zone on LOW side: left_x < 0x70 (strictly).
+	 * set_lt_u(R_T4, R_T3, R_T4=0x70) → R_T4 = (left_x < 0x70) ? 1 : 0. */
+	add_ui(R_T4, R_0, 0x70), set_lt_u(R_T4, R_T3, R_T4), branch_ne(R_T4, R_0, atom_offset(dead_zone_low_check, dead_low_active)),
+	add_ui(R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_low_active */
+
+atom_label(dead_check_upper)
+	/* left_x >= 0x70 → check upper bound. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)), /* reload */
+	add_ui(     R_T4, R_0, 0x90),
+
+	/* R_T4 = (0x90 < left_x) ? 1 : 0 → (left_x > 0x90) ? 1 : 0 */
+	set_lt_u(R_T4, R_T4, R_T3), branch_ne(R_T4, R_0, atom_offset(dead_zone_high_check, dead_high_active)),
+	add_ui(  R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_high_active */
+	jump_rel(atom_offset(dead_zone_skip, exit_stick)),
+		mac_yield_load(),
+
+atom_label(dead_low_active)
+	/* R_T3 = left_x (from line 632 lbu; not clobbered between dead_zone_low_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_low_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3), /* R_T3 = 0x80 - left_x */
+	/* delta = 0x80 - left_x (positive). */
+
+	/* R_T4 = cube_delta */
+	shift_aright(R_T4, R_T3, 2),
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)),
+	nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+	/* R_T4 = floor_delta — moved into the load-delay slot of the floor load below (fills the 1-instruction gap;
+	 * doesn't read R_T0; R_T4 settles by the subsequent add_u). */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)),
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+	jump_rel(atom_offset(end_low, exit_stick)),
+		mac_yield_load(),
+
+atom_label(dead_high_active)
+	/* R_T3 = left_x (from line 641 lbu in dead_check_upper; not clobbered between dead_zone_high_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_high_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3),
+	/* delta = 0x80 - left_x (signed negative). */
+
+	shift_aright(R_T4, R_T3, 2),                      /* R_T4 = cube_delta (signed) */
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)),
+	nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+
+	/* R_T4 = floor_delta (signed) — moved into the load-delay slot of the floor load below. */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)),
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
+atom_label(exit_stick)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the dead-zone/exit branch. */
+	mac_yield_tail(),
+};
+
+#pragma endregion Baked Atoms

code/hello_joypad/hello_joypad.c

@@ -0,0 +1,441 @@
+#pragma region Vendors
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+// #include "libgpu.h"
+// #include "libetc.h"
+// #include "libgte.h"
+#pragma endregion Vendors
+
+#pragma region Duffle Headers
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+
+#include "duffle/word_count.metadata.h"
+
+#include "duffle/dsl.h"
+#include "duffle/memory.h"
+#include "duffle/math.h"
+
+#include "duffle/gcc_asm.h"
+#include "duffle/mips.h"
+#include "duffle/gp.h"
+#include "duffle/gte.h"
+#include "duffle/pad.h"
+
+#include "duffle/dsl.atom.h"
+#include "duffle/lottes_tape.h"
+
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/math.atom.c"
+#include "duffle/mips.atom.c"
+#include "duffle/gte.atom.c"
+#include "duffle/gp.atom.c"
+#include "duffle/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Joypade Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
+#include "hello_joypad.h"
+#pragma region Joypad Headers
+
+#pragma region Hello Joypad TUs
+#include "hello_joypad.atom.c"
+#pragma endregion Hello Joypad TUs
+
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
+typedef Struct_(SMemory) {
+	PrimitiveArena          primitives;
+	A2_OrderingTable_Buffer ordering_tbl;
+	DoubleBuffer            screen_buf;
+	S4                      active_buf_id;
+
+	U4 MemTape[MemTape_Len];
+
+	M3_S2 tform_world;
+
+	Ent_Cube  cube;
+	Ent_Floor floor;
+
+	PadBiosRaw pad_raw[2];
+	PadState   pad[2];
+
+	U4_V scratchpad; // d-cache
+};
+global SMemory smem;
+extern SMemory smem;
+
+I_ B1* prim__alloc(U4 type_width, Str8 type_name) {
+	gknown PrimitiveArena* pa  = & smem.primitives;
+	gknown B1*             buf = (B1*) r_(smem.primitives.buf)[smem.active_buf_id];
+	assert(pa->used + type_width < PrimitiveBuff_Len);
+	B1* next  = buf + pa->used;
+	pa->used += type_width;
+	return next;
+}
+#define prim_alloc(type) (type*)prim__alloc(S_(type), slit( stringify(type)))
+
+/* Uses ONE 8-byte frame allocated via the compiler's standard prologue.
+ * The 4 wasted-arg words for B(12h) InitPAD2 live at [SP+0..15] but are not explicitly allocated.
+ * The compiler handles the MIPS O32 "wasted stack" convention for us by treating the B-call as a 4-arg call.
+ *
+ * The buffer pointers are passed as arguments so the compiler keeps them in callee-saved registers;
+ * The B(12h) asm volatile block does NOT clobber those registers (it clobbers only the volatile GPRs + the B-table arg registers explicitly).
+ * The C-level writes after the call re-load the pointers from their callee-saved homes.
+ *
+ * The clobber list for both B-calls names the full BIOS destroy set documented in kernelbios.md:167-174 (R1..R15, R24..R25, R31, HI/LO).
+ * The kernel-ABI "volatile GPRs" subset is clb_system; the rest of the destroy set is enumerated explicitly here. */
+NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
+{
+	/* Pin raw0 + raw1 to $a0 + $a1 via rgcc; the B(12h) call uses these directly.
+	 * The `(void)` casts mark them as unread after the call so the compiler doesn't need to move them back. */
+	register PadBiosRaw* p0 rgcc(R_A0) = raw0;
+	register PadBiosRaw* p1 rgcc(R_A1) = raw1;
+	(void)p0; (void)p1;
+
+	// TODO(Ed): Properly annotate the raw values in the inline asm instructions.
+	// Use enums.
+
+	/* B(12h) InitPAD2(raw0, 0x22, raw1, 0x22)
+	 *   $a0 = raw0 (rgcc-bound; survives the sequence below)
+	 *   $a1 = raw1 (preserved into $a2 before $a1 is overwritten)
+	 *   $a2 = raw1 (moved from $a1; survives $a1's overwrite)
+	 *   $a3 = 0x22 (immediate)
+	 *   $t1 = 0x12 (function number)
+	 *   $t2 = 0xB0 (BIOS B-table address) */
+	asm volatile(
+		asm_words(
+			or_u(    rarg_2, rarg_1, rdiscard), /* $a2 = $a1 = raw1 */
+			add_ui(  rarg_1, rdiscard, 0x22),   /* $a1 = 0x22 */
+			add_ui(  rarg_3, rdiscard, 0x22),   /* $a3 = 0x22 */
+			add_ui(  rtmp_1, rdiscard, 0x12),   /* $t1 = 0x12 */
+			add_ui(  rtmp_2, rdiscard, 0xB0),   /* $t2 = 0xB0 */
+			call_reg(rtmp_2),                   /* jalr $t2, $ra */
+			nop                                 /* BD slot */
+		)
+		asm_rpins, r_use(p0), r_use(p1)
+		asm_clobber: 
+			rlit(R_AT), 
+			rlit(R_V0), rlit(R_V1), 
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+
+	/* The C-level writes re-load the pointers via the parameter names and write 0xFF to each 
+	 * buffer's status byte to mark the initial-state hazard documented in kernelbios.md:1621-1624. */
+	u1_v(raw0)[0] = 0xFF;
+	u1_v(raw1)[0] = 0xFF;
+
+	/* B(13h) StartPAD2() — no args. The BIOS preserves $sp. */
+	asm volatile(
+		asm_words(
+			add_ui(  rtmp_1, rdiscard, 0x13), /* $t1 = 0x13 */
+			add_ui(  rtmp_2, rdiscard, 0xB0), /* $t2 = 0xB0 (re-load) */
+			call_reg(rtmp_2),                 /* jalr $t2, $ra */
+			nop                               /* BD slot */
+		)
+		asm_clobber:
+			rlit(R_AT),
+			rlit(R_V0), rlit(R_V1),
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+}
+
+GCC_OPTIMIZATION_DISABLE
+void update(PrimitiveArena* pa, U4* ordering_buf) 
+{
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
+	if (0) // Pad Input (dead — kept for the source-as-written record; references the deleted `pad_state` field)
+	{
+		(void)Pad_Left; (void)Pad_Right; /* suppress unused-token warnings */
+		if (false) {
+			smem.cube.rot.y  += 30;
+			smem.floor.rot.y += 5;
+		}
+		if (false) {
+			smem.cube.rot.y  -= 30;
+			smem.floor.rot.y -= 5;
+		}
+	}
+	if (1) // Pad Input (Tape version)
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			/* BIOS-owned polling: per-frame snapshot of both ports. */
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw, & smem.pad_raw[0]);
+				tb_data_(state, & smem.pad[0]);
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw,   & smem.pad_raw[1]);
+				tb_data_(state, & smem.pad[1]);
+			/* Per-frame rotation apply: consume pad[0].buttons + pad[0].left_x */
+			tb_emit_(pad_apply_input);
+				tb_data_(state,     & smem.pad[0]);
+				tb_data_(cube_rot,  & smem.cube.rot);
+				tb_data_(floor_rot, & smem.floor.rot);
+		}
+	}
+
+	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
+
+	// Update the position based on acceleration and velocity
+	gknown V3_S4_R pos = & smem.cube.pos;
+	gknown V3_S4_R vel = & smem.cube.vel;
+	gknown V3_S4_R acc = & smem.cube.accel;
+	add_v3s4(vel, acc[0]);
+	add_v3s4_fp(pos, vel[0]);
+	// vel->x += acc->x;
+	// vel->y += acc->y;
+	// vel->z += acc->z;
+	// pos->x += vel->x;
+	// pos->y += vel->y;
+	// pos->z += vel->z;
+
+	if (pos->y + 150 > smem.floor.pos.y) vel->y *= -1;
+
+	// Prep
+	S4 nclip = 0;
+	S4 orderingtbl_z = 0;
+	A2_S2 p;    //???
+	S4 flag; //????
+
+
+	// Draw Cube
+	if (0)
+	{
+		m3s2_rotation   (& smem.cube.rot,    & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.cube.pos);
+		m3s2_scale      (& smem.tform_world, & smem.cube.scale);
+		// gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+		for (U4 face_id = 0; face_id < Cube_num_faces; face_id += 1)
+		{
+			Poly_G4* quad = prim_alloc(Poly_G4); set_poly_g4(quad);
+			quad->c0 = rgb8(255,   0, 255);
+			quad->c1 = rgb8(255, 255,   0);
+			quad->c2 = rgb8(  0, 255, 255);
+			quad->c3 = rgb8(  0, 255,   0);
+
+			V4_S2* face = & smem.cube.faces[face_id];
+			V3_S2* p0   = & smem.cube.verts[face->x];
+			V3_S2* p1   = & smem.cube.verts[face->y];
+			V3_S2* p2   = & smem.cube.verts[face->z];
+			V3_S2* p3   = & smem.cube.verts[face->w];
+
+			nclip = rtp_avg_nclip_a4_v3s2(
+				p0, p1, p2, p3,
+				& quad->p0, & quad->p1, & quad->p2, & quad->p3,
+				& p, & orderingtbl_z, & flag
+			);
+			if (nclip <= 0) {
+				continue;
+			}
+
+			if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+				orderingtbl_add_primitive(ordering_buf[orderingtbl_z], quad);
+			}
+		}
+		// smem.cube.rot.x +=  6;
+		// smem.cube.rot.y +=  8;
+		// smem.cube.rot.z += 12;
+		smem.cube.rot.y += 30;
+	}
+	// Draw cube (tape method) - two triangles per face
+	if (1)
+	{
+		m3s2_rotation   (& smem.cube.rot,    & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.cube.pos);
+		m3s2_scale      (& smem.tform_world, & smem.cube.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.cube.faces));
+				tb_data(& tb, u4_(smem.cube.verts));
+				tb_data(& tb, u4_(ordering_buf));
+
+			for (U4 i = 0; i < Cube_num_faces; i++) {
+				// Two triangles per quad face: (x,y,z) and (x,z,w)
+				tb_emit(& tb, cube_g4_face);
+			}
+
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));
+
+		// smem.cube.rot.y += 30;
+	}
+	// Draw Floor
+	if (0)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+		for (U4 face_id = 0; face_id < Floor_num_faces; face_id += 1)
+		{
+			Poly_F3* tri = prim_alloc(Poly_F3); set_poly_f3(tri);
+			tri->color   = rgb8(255, 255, 255);
+
+			V3_S2* face = & smem.floor.faces[face_id];
+			register V3_S2* p0 rgcc(R_T4) = & smem.floor.verts[face->x];
+			register V3_S2* p1 rgcc(R_T5) = & smem.floor.verts[face->y];
+			register V3_S2* p2 rgcc(R_T6) = & smem.floor.verts[face->z];
+
+			gte_load_v0(p0, R_T4);
+			/*
+			asm volatile( ".word " "%0" ", %1" : : 
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_xy & REG_MASK) << RT_SHIFT) | (0            & IMM_MASK)),
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_z  & REG_MASK) << RT_SHIFT) | (GTE_Z_Offset & IMM_MASK)), 
+				"r"(p0) : 
+				"$2", "$8", "$9", "$31", "memory" 
+			);
+			*/
+			gte_load_v1(p1, R_T5);
+			gte_load_v2(p2, R_T6);
+
+			gte_rtpt();
+			gte_nclip();
+			gte_stotz(& nclip);
+
+			// nclip = rtp_avg_nclip_a3_v3s2(p0, p1, p2
+			// 	, & tri->p0, & tri->p1, & tri->p2
+			// 	, & p, & orderingtbl_z, & flag
+			// );
+			// if (nclip <= 0) {
+			// 	continue;
+			// }
+
+			if (nclip > 0 ) {
+				gte_stsxy3(& tri->p0, & tri->p1, & tri->p2);
+				gte_avsz3();
+				gte_stotz(& orderingtbl_z);
+
+				if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+					orderingtbl_add_primitive(ordering_buf[orderingtbl_z], tri);
+				}
+			}
+		}
+		smem.floor.rot.y += 5;
+	}
+	// Draw floor tape method
+	if (1)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris?
+		// The tape atoms in-flight should not need to care.
+
+		// Prepare the tape. (Push protocol to tape)
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, set_gte_world);
+				tb_data(& tb, u4_(& smem.tform_world));
+
+			tb_emit(& tb, rbind_floor_f3_face);
+			// TODO(Ed): Just use a single context struct ref
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.floor.faces));
+				tb_data(& tb, u4_(smem.floor.verts));
+				tb_data(& tb, u4_(ordering_buf));
+			for (U4 i = 0; i < Floor_num_faces; i++) {
+				tb_emit(& tb, floor_f3_face);
+			}
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));// Fire off the tape.
+
+		// C-side state (pa->used) has already been updated by the tape!
+		// smem.floor.rot.y += 5;
+	}
+}
+GCC_OPTIMIZATION_ENABLE
+
+void render(void) {
+}
+
+void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
+	draw_sync(0);
+	vsync(0);
+	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
+	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
+	{
+		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
+		pa->used = 0;
+	}
+	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
+}
+
+GCC_OPTIMIZATION_DISABLE
+int main(void)
+{
+	smem = (SMemory){0};
+	smem.scratchpad = C_(U4_V, 0x1F800000);
+	// smem.primitives.used = 0;
+	// smem.active_buf_id   = 0;
+	/*Persistent Entity Setup*/{
+		ent_cube128_init(& smem.cube.verts, & smem.cube.faces); {
+			Ent_Cube* cube = & smem.cube;
+			cube->rot    = v3s2(0, 0, 0);
+			cube->scale  = v3s4_fp_one();
+			cube->accel  = v3s4(0, 1, 0);
+			cube->pos    = v3s4(0, -400, 1800);
+		}
+		ent_floor_init(& smem.floor.verts, & smem.floor.faces); {
+			Ent_Floor* floor = & smem.floor;
+			floor->rot   = v3s2(0, 0, 0);
+			floor->pos   = v3s4(0, 450, 1800);
+			floor->scale = v3s4_fp_one();
+		}
+	}
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape)); {
+		reset_graph(0);
+		/* Direct BIOS: poll both ports during VBlank. */
+		pad_bios_init_start(& smem.pad_raw[0], & smem.pad_raw[1]);
+		/* Pinned registers for the GPU init atom. */
+		register U4*           io_base_addr rgcc(R_IO_BaseAddr) = u4_r(IO_BASE_ADDR);
+		register DoubleBuffer* screen_buf   rgcc(R_ScreenBuf)   = & smem.screen_buf;
+		tb.used = 0; tb_scope_run(& tb) {
+			tb_emit(& tb, screen_env_init);
+			tb_emit(& tb, gp_screen_init);
+		}
+	}
+	while (1) {
+		gknown S4* active_buf_id  = & smem.active_buf_id;
+		gknown U4* ordering_buf   = r_(smem.ordering_tbl)[active_buf_id[0]];
+		gknown PrimitiveArena* pa = & smem.primitives;
+		update(pa, ordering_buf);
+		render();
+		gp_display_frame(& smem.screen_buf, active_buf_id, ordering_buf, pa);
+	};
+	return 0;
+}
+GCC_OPTIMIZATION_ENABLE

code/hello_joypad/hello_joypad.h

@@ -0,0 +1,102 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "duffle/dsl.h"
+#	include "duffle/math.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#endif
+
+enum {
+	// PrimitiveBuff_Len = 4096,
+	// OrderingTbl_Len   = 2048,
+	PrimitiveBuff_Len = 131072,
+	OrderingTbl_Len   = 8192,
+};
+
+enum {
+	ScreenRes_X = 320,
+	ScreenRes_Y = 240,
+	ScreenZ     = 320,
+	ScreenRes_CenterX = (ScreenRes_X >> 1),
+	ScreenRes_CenterY = (ScreenRes_Y >> 1),
+};
+
+enum {
+	fp_one = (1 << 12),
+};
+
+#define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)
+
+typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
+typedef Array_(OrderingTable_Buffer, 2);
+
+typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
+typedef Array_(PrimitiveBuffer, 2);
+typedef Struct_(PrimitiveArena) {
+	A2_PrimitiveBuffer buf;
+	U4                 used;
+};
+
+#define Cube_num_verts 8
+typedef Array_(V3_S2, Cube_num_verts);
+#define Cube_num_faces 6
+typedef Array_(V4_S2, Cube_num_faces);
+I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
+	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
+		{ -128, -128, -128 },
+		{  128, -128, -128 },
+		{  128, -128,  128 },
+		{ -128, -128,  128 },
+		{ -128,  128, -128 },
+		{  128,  128, -128 },
+		{  128,  128,  128 },
+		{ -128,  128,  128 }
+	};
+	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
+		{ 3, 2, 0, 1 },
+		{ 0, 1, 4, 5 },
+		{ 4, 5, 7, 6 },
+		{ 1, 2, 5, 6 },
+		{ 2, 3, 6, 7 },
+		{ 3, 0, 7, 4 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
+	return;
+}
+typedef Struct_(Ent_Cube) {
+	V3_S4 accel;
+	V3_S4 vel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A8_V3_S2 verts;
+	A6_V4_S2 faces;
+};
+
+#define Floor_num_verts 4
+typedef Array_(V3_S2, Floor_num_verts);
+#define Floor_num_faces 2
+typedef Array_(V3_S2, Floor_num_faces);
+I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
+	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
+		{ -900, 0, -900 },
+		{ -900, 0,  900 },
+		{  900, 0, -900 },
+		{  900, 0,  900 },
+	};
+	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
+		{ 0, 1, 2 },
+		{ 1, 3, 2 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
+};
+typedef Struct_(Ent_Floor) {
+	V3_S4 accel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};

code/hello_joypad/sio.tape.c

@@ -0,0 +1,659 @@
+
+#if 0 /* ac_pad_sio_write_pad_state — superseded by pad_bios_snapshot */
+
+/* ============================================================
+ * raw_sio_pad_poll_20260802 — superseded by bios_pad_buffer_snapshot_20260803.
+ * The doomed raw-SIO production atoms (ac_pad_sio_write_pad_state,
+ * pad_sio_init, pad_sio_step, pad_sio_diag_pin, pad_sio_diag_byte_exchange)
+ * reference symbols that were removed from code/duffle/pad.h during
+ * Phase 1. Each is wrapped in a narrow `#if 0` so the C compile skips
+ * the body while the source-as-written text stays in place for the
+ * Phase 5.1 deletion pass. The wrap is removed (and the bodies are
+ * deleted) by Phase 5.1 of this track.
+ * ============================================================ */
+
+ * Writes the per-port PadState in 5 instructions plus 4 store_word calls (status,
+ * buttons, left_x/y/right_x/right_y packed, attempt). The provisional decode publishes
+ * 0x0000FFFF buttons + centered axes on every path until response-byte decode lands.
+ *
+ * Args:
+ *   status_val    - the PadSioStatus enum value to publish
+ *   state_ptr_reg - the PadState* base (R_PadState at the call site)
+ *   scratch_reg   - scratch register for the value being stored (e.g., R_T0)
+ *
+ * Emits 9 instructions (status/buttons/axes/attempt stores plus the
+ * two-instruction zero-extended buttons load).
+ */
+FI_ Slice_MipsCode ac_pad_sio_write_pad_state(MipsAtomBuilder_R ab, U4 status_val, U4 state_ptr_reg, U4 scratch_reg)
+MipsAtomComp_Proc_(ac_pad_sio_write_pad_state, ab, {
+	add_ui(scratch_reg, R_0, status_val),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,status)),
+	/* FIX 2026-08-02: buttons = 0x0000FFFF = "no buttons pressed" in
+	 * libetc convention. Build it with LUI + ORI so addiu does not
+	 * sign-extend 0xFFFF to 0xFFFFFFFF. */
+	load_upper_i(scratch_reg, 0x0000),
+	or_i(scratch_reg, scratch_reg, 0xFFFF),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,buttons)),
+	add_ui(scratch_reg, R_0, 0x80808080),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,left_x)),
+	add_ui(scratch_reg, R_0, 0),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,attempt))
+})
+#endif /* end ac_pad_sio_write_pad_state wrap */
+
+/* ----- pad_sio_init -----
+ * Boot-time SIO0 init. Caller pins R_T6 = sio_base_addr0.
+ * Issues SIO CTRL=0x0040 (reset), MODE=0x000D, BAUD=0x0088.
+ * (Phase 2 fills the body.)
+ */
+#if 0 /* pad_sio_init — superseded by pad_bios_init_start (Phase 1.3) */
+internal MipsAtom_(pad_sio_init) atom_info(atom_phase(pad_init)
+, atom_reads(R_T5, R_T6)
+, atom_writes(R_T5, R_T6)
+) {
+	/* FIX 2026-08-02: explicitly load the KSEG1 base into R_T6 at the top of
+	 * the atom body. The rgcc(R_PadSioBase) binding in main() pins R_T6 = base
+	 * when main() runs, but $12 is caller-saved per the O32 ABI — when tape_run
+	 * is invoked, R_T6 is fair game. The atom body cannot rely on the value. */
+	load_upper_i(R_T6, pad_IO_KSEG1_BASE >> 16),   /* R_T6 high 16 = 0xBF80 */
+	or_i(R_T6, R_T6, pad_IO_KSEG1_BASE & 0xFFFF),   /* R_T6 = 0xBF800000 */
+
+	/* SIO CTRL = 0x0040 (reset) */
+	add_ui(R_T5, R_0, pad_SIO_CTRL_RESET),
+	store_half(R_T5, R_T6, pad_SIO_CTRL_OFFSET),
+	/* SIO MODE = 0x000D (MUL1, 8-bit, no parity, idle-high) */
+	add_ui(R_T5, R_0, pad_SIO_MODE_INIT),
+	store_half(R_T5, R_T6, pad_SIO_MODE_OFFSET),
+	/* SIO BAUD = 0x0088 (~250 kHz) */
+	add_ui(R_T5, R_0, pad_SIO_BAUD_INIT),
+	store_half(R_T5, R_T6, pad_SIO_BAUD_OFFSET),
+	mac_yield(),
+};
+#endif /* end pad_sio_init wrap */
+
+/* ----- pad_sio_step -----
+ * Per-frame bounded raw-SIO transaction. Reads PadState pointers + SIO
+ * base addresses from Binds_PadSioStep; writes per-port status +
+ * buttons + axes into smem.pad[0..1].
+ * Body shape (per spec §"Transaction model (per port, per pad_sio_step)"):
+ *   port 0: CTRL=CLEANUP → settle → CTRL=port-select → settle → exchange 5
+ *           bytes (addr + 0x42 0x00 0x00 0x00) → decode → write PadState[0]
+ *           → CTRL=CLEANUP.
+ *   port 1: swap scratch regs (sio_base_addr1 → R_PadSioBase, state1 →
+ *           R_PadState) → mirror port 0 sequence.
+ *
+ * Bounded-loop semantics: every countdown is wrapped in
+ *   add_ui_self(R_T1, -1) + branch_ne(R_T1, R_0, ...)
+ * with a known maximum (pad_SIO_SETTLE_BEFORE_TX=1000, pad_SIO_SETTLE_AFTER_TX=2000,
+ * pad_SIO_WAIT_BUDGET=4096). The static-analysis pass currently reports
+ * has_loops = true; the follow-up metaprogram track that learns modeled-bounded
+ * loops is out of scope here (per spec §"Risks").
+ *
+ * Scratch register strategy:
+ *   R_PadStatus    = R_T4 — RESERVED for port-1 swap (holds state1)
+ *   R_PadCountdown = R_T5 — RESERVED for port-1 swap (holds sio_base_addr1)
+ *   R_T0           — byte-exchange value + STAT read (clobbered freely)
+ *   R_T1           — countdown budget (clobbered freely)
+ *   R_PadState     = R_T7 — PadState* (preserved for PadState writes)
+ *   R_PadSioBase   = R_T6 — SIO base (preserved through the port)
+ *
+ * Response decode (Task 3.1 teaching scope):
+ *   - status   = PadSioStatus_Digital (hardcoded)
+ *   - buttons  = 0xFFFF (no buttons pressed in the provisional libetc
+ *                         convention; full response-byte decode is follow-up)
+ *   - axes     = 0x80808080 (centered: left_x=0x80, left_y=0x80,
+ *                             right_x=0x80, right_y=0x80)
+ *   - attempt  = 0
+ *   - DualShock handshake (0x43 0x01 → 0x44 0x01 0x03 → 0x43 0x00) is
+ *     follow-up scope; the hardcoded digital decode is a placeholder.
+ *
+ * Both ports raise /CS (CTRL = pad_SIO_CTRL_CLEANUP) before exit. Both ports
+ * treat response timeout as PadSioStatus_Disconnected per the spec §"Failure
+ * handling" + the canonical per-port timeout semantics.
+ */
+#if 0 /* pad_sio_step — superseded by pad_bios_snapshot (Phase 2.1) */
+internal MipsAtom_(pad_sio_step) atom_info(atom_bind(Binds_PadSioStep)
+, atom_reads(R_TapePtr, R_PadSioBase, R_PadState, R_PadStatus, R_PadCountdown)
+, atom_writes(R_PadStatus, R_PadCountdown)
+) {
+	/* FIX 2026-08-02: explicitly load KSEG1 base into R_PadSioBase (R_T6) at the
+	 * top. The rgcc() binding in main() does NOT survive the tape_run call
+	 * because R_T6 is caller-saved per the O32 ABI. The pad_sio_init atom
+	 * (also in the per-frame tape) reloads R_T6 separately. */
+	load_upper_i(R_PadSioBase, pad_IO_KSEG1_BASE >> 16),
+	or_i(R_PadSioBase, R_PadSioBase, pad_IO_KSEG1_BASE & 0xFFFF),
+
+	/* Pop Binds from tape (in Binds_PadSioStep declaration order) */
+	load_word(R_PadState,     R_TapePtr, O_(Binds_PadSioStep,state0)),
+	load_word(R_PadStatus,    R_TapePtr, O_(Binds_PadSioStep,state1)),    /* reserved for port-1 swap */
+	load_word(R_PadSioBase,   R_TapePtr, O_(Binds_PadSioStep,sio_base_addr0)),
+	load_word(R_PadCountdown, R_TapePtr, O_(Binds_PadSioStep,sio_base_addr1)),  /* reserved for port-1 swap */
+	add_ui_self(R_TapePtr, S_(Binds_PadSioStep)),
+
+	/* ============== PORT 0 TRANSACTION ============== */
+	/* Use R_T0 (byte value / STAT read) + R_T1 (countdown) as scratch.
+	 * R_PadStatus (state1) + R_PadCountdown (sio_base_addr1) are preserved
+	 * through the port-0 body and swapped into R_PadSioBase + R_PadState
+	 * at atom_offset(port1_start, ...) below. */
+
+	/* 1. Cleanup: CTRL = 0x0010 (raise /CS, clear stale status) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_BEFORE_TX = 1000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_BEFORE_TX),
+atom_label(settle_pre_port0)
+	nop,  /* BD slot */
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_pre_port0, settle_pre_port0)),
+
+	/* 2. Port-select: CTRL = 0x0003 (TX enable + DTR /CS) for port 0 */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),  /* set /CS line low */
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_AFTER_TX = 2000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_AFTER_TX),
+atom_label(settle_post_port0)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_post_port0, settle_post_port0)),
+
+	/* 3. Address byte (0x01) — send + RX-ready wait + read response + RX-drain confirmation */
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack0_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack0_port0, ack0_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack0_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack0_port0, wait_ack0_port0)),
+	/* RX timeout → mark disconnected; skip to port 1 */
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack0)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack0, port1_start)),
+
+atom_label(ack0_received_port0)
+	/* Read open-bus response byte 0 — discard per docs/psx-spx §controllersandmemorycards.md */
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Confirm RX FIFO drained before sending byte 1. Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel0_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel0_port0, ack_released_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel0_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel0_port0, wait_ackrel0_port0)),
+	/* RX-drain timeout → disconnected; skip to port 1 */
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel0)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel0, port1_start)),
+
+atom_label(ack_released_port0)
+
+	/* === Byte 1 (port 0): send 0x42 (cmd read) + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack1_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack1_port0, ack1_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack1_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack1_port0, wait_ack1_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack1)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack1, port1_start)),
+
+atom_label(ack1_received_port0)
+	/* Read response ID byte — discarded for teaching scope (decode hardcoded). */
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* RX FIFO drain wait. Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel1_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel1_port0, ack_released1_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel1_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel1_port0, wait_ackrel1_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel1)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel1, port1_start)),
+
+atom_label(ack_released1_port0)
+
+	/* === Byte 2 (port 0): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack2_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack2_port0, ack2_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack2_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack2_port0, wait_ack2_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack2)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack2, port1_start)),
+
+atom_label(ack2_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel2_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel2_port0, ack_released2_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel2_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel2_port0, wait_ackrel2_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel2)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel2, port1_start)),
+
+atom_label(ack_released2_port0)
+
+	/* === Byte 3 (port 0): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack3_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack3_port0, ack3_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack3_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack3_port0, wait_ack3_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack3)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack3, port1_start)),
+
+atom_label(ack3_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel3_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel3_port0, ack_released3_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel3_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel3_port0, wait_ackrel3_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel3)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel3, port1_start)),
+
+atom_label(ack_released3_port0)
+
+	/* === Byte 4 (FINAL, port 0): send 0x00 + RX-not-empty wait + read final byte === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_rx4_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_rx4_port0, rx4_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_rx4_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_rx4_port0, wait_rx4_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_rx4)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_rx4, port1_start)),
+
+atom_label(rx4_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),  /* discard final byte */
+
+	/* === RESPONSE DECODE (hardcoded for teaching scope) ===
+	 * Per the plan §"Phase 3 task 3.1" + spec §"Architecture":
+	 *   - Full decode (buttons/axes from response bytes) is follow-up scope.
+	 *   - Teaching scope: hardcode digital poll response.
+	 *     status = PadSioStatus_Digital
+	 *     buttons = 0x0000FFFF (no buttons pressed — placeholder)
+	 *     axes = 0x80808080 (left_x=0x80, left_y=0x80, right_x=0x80, right_y=0x80)
+	 *     attempt = 0
+	 */
+atom_label(decode_port0)
+	mac_pad_sio_write_pad_state(PadSioStatus_Digital, R_PadState, R_T0),
+
+	/* /CS cleanup: raise /CS, clear stale status before exiting port 0. */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+	/* ============== PORT 1 SETUP ============== */
+	/* Swap: R_PadCountdown holds sio_base_addr1; R_PadStatus holds state1. */
+atom_label(port1_start)
+	add_u(R_PadSioBase, R_0, R_PadCountdown),  /* sio_base_addr1 → R_PadSioBase */
+	add_u(R_PadState,   R_0, R_PadStatus),     /* state1 → R_PadState */
+
+	/* ============== PORT 1 TRANSACTION (mirror of port 0) ============== */
+	/* R_PadStatus + R_PadCountdown are no longer reserved (port 1 is the
+	 * last transaction); we still use R_T0/R_T1 as scratch to match port 0. */
+
+	/* 1. Cleanup: CTRL = 0x0010 (raise /CS, clear stale status) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_BEFORE_TX = 1000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_BEFORE_TX),
+atom_label(settle_pre_port1)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_pre_port1, settle_pre_port1)),
+
+	/* 2. Port-select: CTRL = 0x0003 | (1 << 13) (port 1 select) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),
+	or_i(R_T0, R_T0, 1 << 13),  /* port 1 select bit (CTRL bit 13 = port select) */
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_AFTER_TX = 2000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_AFTER_TX),
+atom_label(settle_post_port1)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_post_port1, settle_post_port1)),
+
+	/* 3. Address byte (0x01) — send + RX-ready wait + read response + RX-drain confirmation */
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack0_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack0_port1, ack0_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack0_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack0_port1, wait_ack0_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack0)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack0, end_atom)),
+
+atom_label(ack0_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel0_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel0_port1, ack_released_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel0_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel0_port1, wait_ackrel0_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel0)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel0, end_atom)),
+
+atom_label(ack_released_port1)
+
+	/* === Byte 1 (port 1): send 0x42 (cmd read) + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack1_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack1_port1, ack1_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack1_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack1_port1, wait_ack1_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack1)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack1, end_atom)),
+
+atom_label(ack1_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel1_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel1_port1, ack_released1_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel1_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel1_port1, wait_ackrel1_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel1)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel1, end_atom)),
+
+atom_label(ack_released1_port1)
+
+	/* === Byte 2 (port 1): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack2_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack2_port1, ack2_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack2_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack2_port1, wait_ack2_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack2)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack2, end_atom)),
+
+atom_label(ack2_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel2_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel2_port1, ack_released2_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel2_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel2_port1, wait_ackrel2_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel2)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel2, end_atom)),
+
+atom_label(ack_released2_port1)
+
+	/* === Byte 3 (port 1): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack3_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack3_port1, ack3_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack3_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack3_port1, wait_ack3_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack3)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack3, end_atom)),
+
+atom_label(ack3_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel3_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel3_port1, ack_released3_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel3_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel3_port1, wait_ackrel3_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel3)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel3, end_atom)),
+
+atom_label(ack_released3_port1)
+
+	/* === Byte 4 (FINAL, port 1): send 0x00 + RX-not-empty wait + read final byte === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_rx4_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_rx4_port1, rx4_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_rx4_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_rx4_port1, wait_rx4_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_rx4)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_rx4, end_atom)),
+
+atom_label(rx4_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),  /* discard final byte */
+
+	/* === RESPONSE DECODE (port 1) === */
+atom_label(decode_port1)
+	mac_pad_sio_write_pad_state(PadSioStatus_Digital, R_PadState, R_T0),
+
+	/* /CS cleanup: raise /CS, clear stale status before exiting port 1. */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+atom_label(end_atom)
+	mac_yield(),
+};
+#endif /* end pad_sio_step wrap */
+
+/* ----- pad_sio_diag_pin -----
+ * Per-frame diagnostic counter. The caller binds R_DiagPinScratch to
+ * scratch_for_atom_diag_pin for temporary gdb verification.
+ */
+#if 0 /* pad_sio_diag_pin — superseded (raw-SIO phase removed) */
+internal MipsAtom_(pad_sio_diag_pin) atom_info(atom_phase(pad_init)
+, atom_reads(R_T0, R_T1, R_DiagPinScratch)
+, atom_writes(R_T0, R_T1, R_DiagPinScratch)
+) {
+	/* FIX 2026-08-02: explicitly reload R_DiagPinScratch (R_T3 = $t3). Caller-saved
+	 * per O32 ABI; the rgcc binding in main() does not survive tape_run. */
+	load_upper_i(R_DiagPinScratch, 0x8001),
+	or_i(R_DiagPinScratch, R_DiagPinScratch, 0xC800),
+
+	/* High half = 0xD1A6; low half increments once per atom invocation. */
+	load_word(R_T1, R_DiagPinScratch, 0),
+	nop,
+	add_ui(R_T1, R_T1, 1),
+	and_i(R_T0, R_T1, 0xFFFF),
+	load_upper_i(R_T1, 0xD1A6),
+	or_i(R_T1, R_T1, 0),
+	or_u(R_T1, R_T1, R_T0),
+	store_word(R_T1, R_DiagPinScratch, 0),
+	mac_yield(),
+};
+#endif /* end pad_sio_diag_pin wrap */
+
+/* ----- pad_sio_diag_byte_exchange -----
+ * Temporary two-byte wire probe: sends 0x01 and 0x42, then stores the
+ * open-bus byte and response ID in scratch_for_atom_diag_pin.
+ */
+#if 0 /* pad_sio_diag_byte_exchange — superseded (raw-SIO phase removed) */
+internal MipsAtom_(pad_sio_diag_byte_exchange) atom_info(atom_phase(pad_init)
+, atom_reads(R_T0, R_T1, R_T2, R_PadSioBase, R_DiagPinScratch)
+, atom_writes(R_T0, R_T1, R_T2, R_PadSioBase, R_DiagPinScratch)
+) {
+	/* FIX 2026-08-02: explicitly reload R_DiagPinScratch (R_T3 = $t3). Caller-saved
+	 * per O32 ABI; the rgcc binding in main() does not survive tape_run. */
+	load_upper_i(R_DiagPinScratch, 0x8001),
+	or_i(R_DiagPinScratch, R_DiagPinScratch, 0xC800),
+
+	/* FIX 2026-08-02: explicitly load KSEG1 base into R_PadSioBase (R_T6) at the
+	 * top. The rgcc() binding in main() does NOT survive the tape_run call
+	 * because R_T6 is caller-saved per the O32 ABI. */
+	load_upper_i(R_PadSioBase, pad_IO_KSEG1_BASE >> 16),
+	or_i(R_PadSioBase, R_PadSioBase, pad_IO_KSEG1_BASE & 0xFFFF),
+
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(diag_wait_ack0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(diag_wait_ack0, diag_ack0_done)),
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(diag_wait_ack0, diag_wait_ack0)),
+	add_ui(R_T0, R_0, 0xDEADAC01),
+	store_word(R_T0, R_DiagPinScratch, 0),
+	branch_equal(R_0, R_0, atom_offset(diag_timeout_ack0, diag_timeout)),
+atom_label(diag_ack0_done)
+	load_byte_u(R_T2, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(diag_wait_ack1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(diag_wait_ack1, diag_ack1_done)),
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(diag_wait_ack1, diag_wait_ack1)),
+	add_ui(R_T0, R_0, 0xDEADAC02),
+	store_word(R_T0, R_DiagPinScratch, 0),
+	branch_equal(R_0, R_0, atom_offset(diag_timeout_ack1, diag_timeout)),
+atom_label(diag_ack1_done)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	nop,
+	shift_lleft(R_T0, R_T0, 8),
+	or_u(R_T2, R_T2, R_T0),
+	store_word(R_T2, R_DiagPinScratch, 0),
+atom_label(diag_success)
+	branch_equal(R_0, R_0, atom_offset(diag_success, diag_done)),
+	nop,
+atom_label(diag_timeout_ack0)
+	add_ui(R_T0, R_0, 0xDEADAC01),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_timeout_ack1)
+	add_ui(R_T0, R_0, 0xDEADAC02),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_timeout)
+	add_ui(R_T0, R_0, 0xDEADACFF),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_done)
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	mac_yield(),
+};
+#endif /* end pad_sio_diag_byte_exchange wrap */

readme.md

@@ -7,12 +7,12 @@ A rest from the usual.
 ## Dependencies
 
 I will be programming from a Windows 11 machine (may eventually try this on the Steam Deck...):
-![system_info](./docs/assets/system_info.png)
 
 [armips](https://github.com/Kingcom/armips)
 
 * Supports doing bare-metal assembly for the ps1
 * `scoop install armips` or just clone and build..
+* Was used early in the course. Now I just use an macro asm dsl in C11.
 
 [luajit-2.1](https://github.com/LuaJIT/LuaJIT.git)
 
@@ -73,3 +73,9 @@ scoop install luajit
 ![hello_psyq!](./docs/assets/pcsx-redux_2025-08-05_23-01-19.png)
 ![cube!](./docs/assets/pcsx-redux_2025-10-11_03-04-01.png)
 ![cube and floor!](./docs/assets/pcsx-redux_2026-07-10_22-47-02.png)
+
+Win 11 machine:
+
+![system_info](./docs/assets/system_info.png)
+
+Still haven't gotten around to trying this on linux...

scripts/build_psyq.ps1

@@ -180,29 +180,15 @@ function link-modules { param([string[]]$link_modules, [string]  $elf, [string[]
 	$link_args += ($f_link_pass_through_prefix + $f_link_mapfile + $map)
 
 	$link_args += ($f_link_pass_through_prefix + $f_link_start_group)
+	# 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math, mcrd, mcx, press, sio, snd, spu, tap)
+	# had LOAD lines in the map but ZERO .o files pulled in — they were unused.
+	# 5 kept libraries (api, c, etc, gpu, gte) are required by the C-side calls in hello_joypad.c (reset_graph, draw_sync, vsync, etc.).
 	$libraries = @(
 		"api",
 		"c",
-		"c2",
-		"card",
-		"cd",
-		"comb",
-		"ds",
 		"etc",
 		"gpu",
-		"gs",
-		"gte",
-		"gun",
-		"hmd",
-		"math",
-		"mcrd",
-		"mcx",
-		"pad",
-		"press",
-		"sio",
-		"snd",
-		"spu",
-		"tap"
+		"gte"
 	)
 	foreach ($lib in $libraries) {
 		$link_args += ($f_link_lib + $lib)
@@ -238,9 +224,7 @@ function ps1-meta { param(
 		[string[]]$passes = @('--pre-link'),
 		[string[]]$extra_args = @()
 	)
-	# `--unity-root` and `--source` are
-	# mutually exclusive. Exactly one of `$unity_root` / `$sources` must
-	# be supplied; the other must be absent.
+	# `--unity-root` and `--source` are mutually exclusive. Exactly one of `$unity_root` / `$sources` must be supplied; the other must be absent.
 	if ($null -ne $unity_root -and $unity_root -ne '') 
 	{
 		if ($null -ne $sources -and $sources.Count -gt 0) {
@@ -276,6 +260,73 @@ function ps1-meta { param(
 	}
 }
 
+function inject-dwarf { param(
+	[string]$elf,
+	[string]$path_gen
+)
+	$base_name               = [System.IO.Path]::GetFileNameWithoutExtension($elf)
+	$path_dwarf_line_bin     = join-path $path_gen  "$base_name.dwarf_line.bin"
+	$path_dwarf_aranges_bin  = join-path $path_gen  "$base_name.dwarf_aranges.bin"
+	$path_dwarf_rnglists_bin = join-path $path_gen  "$base_name.dwarf_rnglists.bin"
+	$path_dwarf_info_bin     = join-path $path_gen  "$base_name.dwarf_info.bin"
+	$path_dwarf_abbrev_bin   = join-path $path_gen  "$base_name.dwarf_abbrev.bin"
+	$path_dwarf_str_bin      = join-path $path_gen  "$base_name.dwarf_str.bin"
+	$path_dwarf_loc_bin      = join-path $path_gen  "$base_name.dwarf_loc.bin"
+	$path_dwarf_loclists_bin = join-path $path_gen  "$base_name.dwarf_loclists.bin"
+	$path_inject_elf         = join-path $path_build "$base_name.dwarf-injected.elf"
+
+	if (-not (Test-Path $path_dwarf_line_bin))     { return }
+	if (-not (Test-Path $path_dwarf_aranges_bin))  { return }
+	if (-not (Test-Path $path_dwarf_rnglists_bin)) { return }
+
+	Write-Host "[build] DWARF-injecting $elf -> $path_inject_elf"
+	Copy-Item -LiteralPath $elf -Destination $path_inject_elf -Force
+
+	# Objcopy call 1: 3x --update-section for the PC-mapping tables (line, aranges, rnglists).
+	$objcopy_args_dwarf_pc = @(
+		"--update-section=.debug_line=$path_dwarf_line_bin",
+		"--update-section=.debug_aranges=$path_dwarf_aranges_bin",
+		"--update-section=.debug_rnglists=$path_dwarf_rnglists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_pc $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-pc splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Objcopy call 2: 3x --update-section + 2x --add-section for the debug-data tables (info, abbrev, str, loc, loclists).
+	$objcopy_args_dwarf_info = @(
+		"--update-section=.debug_info=$path_dwarf_info_bin",
+		"--update-section=.debug_abbrev=$path_dwarf_abbrev_bin",
+		"--update-section=.debug_str=$path_dwarf_str_bin",
+		"--add-section=.debug_loc=$path_dwarf_loc_bin",
+		"--add-section=.debug_loclists=$path_dwarf_loclists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_info $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-info splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
+	# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
+	# The original ELF and PS-EXE remain byte/flag unchanged.
+	& $Objcopy `
+		--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
+		--set-section-flags ".data=alloc,load,data,code,contents" `
+		$path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+	}
+	else {
+		Write-Host "[build] DWARF-injected ELF: $path_inject_elf"
+	}
+}
+# inject-dwarf
+
 function build-hello_psyqo {
 	$includes += @()
 
@@ -350,10 +401,10 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function build-gte_hello {
+function build-hello_gte {
 	$includes += @()
 
-	$path_module        = join-path $path_code   'gte_hello'
+	$path_module        = join-path $path_code   'hello_gte'
 	$path_duffle        = join-path $path_code   'duffle'
 	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
 	$path_build_gen     = join-path $path_build   'gen'
@@ -402,64 +453,111 @@ function build-gte_hello {
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
 	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
 
-	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
-	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
-	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
-	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
-	{
-		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
-		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		$f_args = @(
-			"--update-section=.debug_line=$dwarfLineBin",
-			"--update-section=.debug_aranges=$dwarfArangesBin",
-			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
-		)
-		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		$dwarfInfoBin     = join-path $path_build_gen 'hello_gte.dwarf_info.bin'
-		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_gte.dwarf_abbrev.bin'
-		$dwarfStrBin      = join-path $path_build_gen 'hello_gte.dwarf_str.bin'
-		$dwarfLocBin      = join-path $path_build_gen 'hello_gte.dwarf_loc.bin'
-		$dwarfLoclistsBin = join-path $path_build_gen 'hello_gte.dwarf_loclists.bin'
-		$g_args = @(
-			"--update-section=.debug_info=$dwarfInfoBin",
-			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
-			"--update-section=.debug_str=$dwarfStrBin",
-			"--add-section=.debug_loc=$dwarfLocBin",
-			"--add-section=.debug_loclists=$dwarfLoclistsBin"
-		)
-		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
-		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
-		# The original ELF and PS-EXE remain byte/flag unchanged.
-		& $Objcopy `
-			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
-			--set-section-flags ".data=alloc,load,data,code,contents" `
-			$injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-		} 
-		else {
-			Write-Host "[build] DWARF-injected ELF: $injectElf"
-		}
-	}
+	inject-dwarf $elf $path_build_gen
 }
-build-gte_hello
+# build-hello_gte
 
+function build-hello_joypad {
+	$includes += @()
+
+	$path_module        = join-path $path_code   'hello_joypad'
+	$path_duffle        = join-path $path_code   'duffle'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build  'gen'
+
+	$src_c = join-path $path_module 'hello_joypad.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
+
+	$assemble_args = @()
+	$assemble_args += $f_debug
+	$assemble_args += $f_optimize_none
+	$assemble_args += ($f_include + $path_code)
+
+	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
+	$module_asm_crt = join-path $path_build         'crt0.o'
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+
+	$module_c = join-path $path_build  'hello_joypad_c.o'
+
+	$compile_args = @()
+	$compile_args += $f_debug
+	$compile_args += $f_optimize_none
+	# $compile_args += $f_optimize_intrinsics
+	# $compile_args += $f_optimize_size
+	# $compile_args += $f_optimize_debug
+	$compile_args += ($f_include + $path_code)
+	compile-unit $src_c $module_c $includes $compile_args
+
+	$elf = join-path $path_build 'hello_joypad.elf'
+	$exe = join-path $path_build 'hello_joypad.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
+
+	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
+
+	inject-dwarf $elf $path_build_gen
+}
+# build-hello_joypad
+
+function build-hello_camera {
+	$includes += @()
+
+	$path_module        = join-path $path_code   'hello_camera'
+	$path_duffle        = join-path $path_code   'duffle'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build  'gen'
+
+	$src_c = join-path $path_module 'hello_camera.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--pre-link')
+
+	$assemble_args = @()
+	$assemble_args += $f_debug
+	$assemble_args += $f_optimize_none
+	$assemble_args += ($f_include + $path_code)
+
+	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
+	$module_asm_crt = join-path $path_build         'crt0.o'
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+
+	$module_c = join-path $path_build  'hello_camera_c.o'
+
+	$compile_args = @()
+	$compile_args += $f_debug
+	$compile_args += ($f_define + 'BUILD_DEBUG')
+	$compile_args += $f_optimize_none
+	# $compile_args += $f_optimize_intrinsics
+	# $compile_args += $f_optimize_size
+	# $compile_args += $f_optimize_debug
+	$compile_args += ($f_include + $path_code)
+	compile-unit $src_c $module_c $includes $compile_args
+
+	$elf = join-path $path_build 'hello_camera.elf'
+	$exe = join-path $path_build 'hello_camera.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
+
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
+
+	inject-dwarf $elf $path_build_gen
+}
+build-hello_camera
 
 # NO idea if this works yet...
 function Send-ToEmulator { param( [string]$exePath )

scripts/duffle_emit.lua

@@ -0,0 +1,936 @@
+--- duffle_emit.lua — project_emission + decl finders.
+local scan = require("duffle_scan")
+local isa  = require("duffle_isa")
+local M = {}
+for k, v in pairs(scan) do M[k] = v end
+for k, v in pairs(isa)  do M[k] = v end
+
+-- Section 8: Cross-source component-body index + word-event expansion
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Shared, memoized helpers: a single emitted-word event stream that every downstream pass reads from,
+-- built once from the pre-tokenized bodies.
+
+--- @class ComponentBodyEntry
+--- @field body_tokens    table        -- pre-tokenized {{tok=string, rel=integer}, ...}
+--- @field body_off       integer      -- byte offset of body[1] in `source`
+--- @field line_of        fun(pos:integer):integer  -- byte-offset → 1-based line number in `source`
+--- @field source         string       -- absolute path of the source containing the declaration
+--- @field declaration    integer      -- 1-based line number of the MipsAtomComp_(ac_X) declaration
+--- @field kind           string       -- "comp_bare" | "comp_proc"
+
+-- The cross-source component-body index is owned by the corpus (`corpus.component_body_index`, populated by `passes/components.lua`).
+-- Consumers (`passes/static_analysis.lua`, `passes/emission_model.lua`) read it directly; per-pass memoization helpers stay out of scope.
+
+-- ASCII byte constants used by split_call_args (kept local to keep Section 8 self-contained).
+local E_BYTE_OPEN_PAREN  = 0x28
+local E_BYTE_OPEN_BRACE  = 0x7B
+local E_BYTE_OPEN_BRACK  = 0x5B
+local E_BYTE_DQUOTE      = 0x22
+local E_BYTE_SQUOTE      = 0x27
+local E_BYTE_COMMA       = 0x2C
+
+-- Map an open-delimiter byte to its matching close string for read_balanced.
+local E_OPEN_CLOSE = {
+	[E_BYTE_OPEN_PAREN] = ")",
+	[E_BYTE_OPEN_BRACE] = "}",
+	[E_BYTE_OPEN_BRACK] = "]",
+}
+
+--- Split the INSIDE of a `f(...)` call on top-level commas.
+--- Honors nested parens / braces / brackets and skips strings / comments.
+--- Returns a list of trimmed argument strings in source order.
+--- (Mirrors split_top_level_commas but for paren-body args; intentionally distinct so a caller's brace-body split isn't confused with an arg list.)
+--- @param inner string
+--- @return string[]
+local function split_call_args(inner)
+	local args = {}
+	if not inner or inner == "" then return args end
+	local pos   = 1
+	local len   = #inner
+	local start = 1
+	while pos <= len do
+		local c     = inner:byte(pos)
+		local close = E_OPEN_CLOSE[c]
+		if close then
+			local _, after = M.read_balanced(inner, string.char(c), close, pos)
+			pos = after
+		elseif c == E_BYTE_DQUOTE or c == E_BYTE_SQUOTE then
+			pos = M.skip_str_or_cmt(inner, pos)
+		elseif c == E_BYTE_COMMA then
+			args[#args + 1] = M.trim(inner:sub(start, pos - 1))
+			start = pos + 1
+			pos   = pos + 1
+		else
+			pos = pos + 1
+		end
+	end
+	if start <= len then args[#args + 1] = M.trim(inner:sub(start, len)) end
+	return args
+end
+
+--- Extract the leading identifier + top-level args list from a token string.
+--- Returns (ident, args). For tokens without a `(...)` call, args is `{}`.
+--- @param tok string
+--- @return string, string[]
+local function token_ident_and_args(tok)
+	local  ident, after = M.read_ident(tok, 1)
+	if not ident then return "?", {} end
+	local paren_pos = M.skip_ws_and_cmt(tok, after)
+	if tok:sub(paren_pos, paren_pos) ~= "(" then return ident, {} end
+	local  inner = M.read_parens(tok, paren_pos)
+	if not inner then return ident, {} end
+	return ident, split_call_args(inner)
+end
+
+-- The macro-name prefix that marks a `mac_X(...)` component invocation.
+local E_MAC_PREFIX     = "mac_"
+local E_MAC_PREFIX_LEN = 4
+
+--- Expand a body entry into the flat sequence of emitted machine-word events.
+---
+--- Semantics (one event per emitted machine word):
+---   * Direct one-word encoders `load_word`, `add_ui`, `nop`, `gte_lw`, ...: One event with `ident` = leading ident, `args` = parsed top-level args.
+---   * `nop2` (2-word pseudo-instruction): Two events, both with `ident = "nop"` so the recognized "this slot is a no-op" semantic is visible to downstream analyses.
+---   * Any other N-word token in `word_counts`: N events sharing the same `ident` + `args` so useful CPU words retire slots in the cycle budget.
+---   * Known `mac_X(...)` calls: Recursively expand the indexed component body, including nested components. Every event from the expansion carries:
+---     - `source` / `line` = the COMPONENT'S source path + the line of the token within the component body (i.e. "definition site").
+---     - `call_source` / `call_line` = the ROOT atom's source path + call-site line, PRESERVED across recursion so nested events still point at the original root.
+---   * Unknown `mac_X` (not in `component_index`): fall back to `word_counts[ident]` if present; otherwise emit one opaque event so the cycle budget accounts for the word.
+---   * Marker Tokens (`atom_label(...)` / `atom_offset(...)`): Zero events (they are pure metaprogram hints).
+---
+--- Cycle protection: a per-expansion `visiting` set tracks components currently on the expansion stack;
+--- a re-entry produces a deterministic `{kind = "cycle", ...}` error and aborts that branch (does NOT hang, does NOT recurse).
+---
+--- Pure: reads `body_entry` / `component_index` / `word_counts`. Memoization is the caller's responsibility.
+--- Callers wanting `word_events` / `word_event_errors` precomputed for many atoms should memoize them per atom.
+--- @param body_entry       table -- `{body_tokens, body_off, line_of, source, declaration}` (declaration = root atom's atom.line)
+--- @param component_index  table -- the bare-name → ComponentBodyEntry map from M.get_component_body_index
+--- @param word_counts      table -- macro name → emitted-word count (from `ctx.shared.word_counts`)
+--- @return WordEvent[], WordEventError[]
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Section 11: project_emission (per-atom emission projection)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Per-atom emission projection is owned by `passes/emission_model.lua`.
+-- The projection is built from the root atom body only; invocation ancestry recursively expands nested components.
+-- The items stream is the single ordered source of truth; `word_events` and `markers` are dense views over it.
+--
+-- The helper below operates on a body string (not a body_entry) so the pass can call it without depending on the older SourceScan / body_off conventions.
+-- component_index argument is reserved for recursive component expansion.
+-- word_counts table is authored-metadata + current-component count table.
+
+--- @class EmissionProjection
+--- @field items       table[] -- Ordered stream of word|label|offset|invoke_begin|invoke_end
+--- @field word_events table[] -- Dense view of items where kind == "word"
+--- @field markers     table[] -- Dense view of items where kind == "label"|"offset"
+--- @field invocations InvocationRecord[] -- dense view of items where kind == "invoke_begin"|"invoke_end"
+--- @field errors      table[] -- Token-resolution failures surfaced without fail-loud
+--- @field warnings    table[] -- Opaque warnings (e.g. unknown uncounted macro)
+
+--- @class InvocationRecord
+--- Lives at `atom.paths.invocations[*]`. Constructed once at the single invocation-construction site
+--- (`emit_invoke_begin` inside `_project_emission_inner`); `invoke_begin` / `invoke_end` markers in the items stream share the same `id`.
+--- @field id              integer  -- 1-based, monotonic per-atom invocation id (0 is reserved for "no open invocation")
+--- @field parent_id       integer  -- 0 for the outermost (root) call; otherwise the id of the immediately enclosing invocation
+--- @field kind            string   -- "comp_bare" | "comp_proc" (component form that triggered the expansion)
+--- @field component_name  string   -- Bare component name without the `mac_` prefix
+--- @field call_text       string   -- Immediate `mac_X(...)` token text (or root call text for the outermost entry)
+--- @field root_call_text  string   -- IMMUTABLE outermost `mac_X(...)` token text for every word emitted in this call's expansion
+--- @field call_path       string   -- Source path of the call site (root atom source for direct calls, component source for nested expansions)
+--- @field call_line       integer  -- Source line of the call site
+--- @field def_path        string   -- Source path of the component definition
+--- @field def_line        integer  -- Source line of the component declaration
+--- @field start_pos       integer  -- 0-based emitted-word position of the FIRST word inside this invocation (the value of `word_idx` AT `emit_invoke_begin` time, BEFORE the first word is emitted). Words emitted inside this invocation occupy `start_pos..start_pos+#body_lines-1` (inclusive, 0-based). Downstream DWARF/provenance consumers MUST read this; do NOT reconstruct it from `start_word` (which is the 1-based items index including `invoke_begin`/`invoke_end` markers).
+--- @field end_pos         integer  -- 0-based position of the LAST word inside this invocation (set by `emit_invoke_end` to `word_idx - 1` AFTER all body words are emitted).
+--- @field start_word      integer  -- 1-based items index of the `invoke_begin` item
+--- @field end_word        integer  -- 1-based items index of the `invoke_end` item (set by `emit_invoke_end`)
+--- @field word_count      integer  -- Number of `word` items emitted between `start_word` and `end_word` (inclusive)
+--- @field debug_skip      boolean  -- `debug_skip` stamp; true iff `corpus.components[name].debug_skip` is true at construction. Always boolean (never `nil`).
+--- @field errors          table[]  -- Per-invocation construction errors (cycle / count_mismatch); does not include pass-level errors
+
+-- Internal recursive walker. The items stream holds every emitted event in order; `word_events`, `markers`,
+-- `invocations`, `errors`, `warnings` are dense views / side outputs appended alongside.
+--
+-- Output rules:
+--   * `word` items record: `invocation_ids` (innermost last) and `outermost_invocation_id` (0 if no invocation is open).
+--   * `invoke_begin` / `invoke_end` items are zero-width at the current word index; the same `word_index` is recorded on both.
+--   * `root_call_text` is the outermost `mac_X(...)` token text for every word emitted inside a component expansion;
+--     it is `nil` for direct words emitted from the root atom body.
+--   * `call_text` is the IMMEDIATE top-level token spelling for the word (for nested words this is the inner `mac_X(...)` token; 
+--     for direct words it is the trimmed encoder token).
+--   * `def_path` / `def_line` are the definition site of the current body (component source for nested words; root atom source for direct words, filled in by the pass caller).
+--   * Unknown uncounted macros emit one opaque word + one warning. Unknown metadata-backed macros (entry in `word_counts`) emit the declared word count, no warning.
+--   * Cycle detection uses an active DFS stack (`visiting`); a cycle appends a construction error to BOTH the projection errors and the cycle invocation's own errors,
+--     then breaks out without recursing (the cycle entry still receives an invocation ID + paired `invoke_begin` / `invoke_end` items, so the boundary invariant is preserved).
+--   * Component declared-count mismatch (declared vs. measured) is a construction error (kind = "count_mismatch"); recorded on the invocation record and pass-level errors list.
+--   * Final boundary check: if any invocation is still open at end of walk, surface a "unbalanced" construction error.
+local function _project_emission_inner(root_body_entry, ctx_table)
+	local items       = {}
+	local word_events = {}
+	local markers     = {}
+	local invocations = {}
+	local errors      = {}
+	local warnings    = {}
+
+	local word_idx         = 0
+	local invocation_stack = {}   -- stack of currently-open invocation records
+	local next_inv_id      = 0
+
+	local reg_use_schema = ctx_table.reg_use_schema
+	local reg_use_param  = ctx_table.reg_use_param
+	local atom_name      = ctx_table.atom_name
+
+	local slot_readonly = {}
+	if reg_use_schema then
+		for _, slot in ipairs(reg_use_schema.slots or {}) do
+			slot_readonly[slot.name] = slot.readonly == true
+		end
+	end
+
+	local function apply_sub(sub_map, operand)
+		if not (sub_map and type(operand) == "string") then return operand end
+		if sub_map[operand] then return sub_map[operand] end
+		local dot = operand:find(".", 1, true)
+		if dot then
+			local head = operand:sub(1, dot - 1)
+			local mapped = sub_map[head]
+			if type(mapped) == "string" then
+				return mapped .. operand:sub(dot)
+			end
+		end
+		return operand
+	end
+
+	local function resolve_gpr_key(operand)
+		if type(operand) ~= "string" then return nil end
+		if operand:sub(1, 2) == "R_" then return operand end
+		if not (reg_use_schema and reg_use_param) then return nil end
+		local prefix = reg_use_param .. "."
+		if operand:sub(1, #prefix) ~= prefix then return nil end
+		local member_path = operand:sub(#prefix + 1)
+		local slot = reg_use_schema.alias_to_slot[member_path]
+		if not slot then return nil, member_path end
+		return "reguse:" .. atom_name .. ":" .. slot, nil, slot
+	end
+
+	local function open_invocation_ids_snapshot()
+		local ids = {}
+		for _, inv in ipairs(invocation_stack) do
+			ids[#ids + 1] = inv.id
+		end
+		return ids
+	end
+
+	local function emit_word(encoder, args, line, word_call_text,
+		def_source_now, def_line_now,
+		immediate_call_text, root_call_text_w, sub_map)
+		local inv_ids   = open_invocation_ids_snapshot()
+		local outermost = inv_ids[1] or 0
+		-- For words emitted at the root atom body, `immediate_call_text` is nil and the walker's `word_call_text` (the word's own token, e.g. "nop") becomes the effective call_text.
+		-- For words emitted inside a component expansion, `immediate_call_text` is the immediate outer `mac_X(...)` token text;
+		-- The call that triggered the body expansion we're currently walking.
+		local eff_call_text      = immediate_call_text or word_call_text
+		local eff_root_call_text = root_call_text_w
+		local gpr_keys = nil
+		if reg_use_schema or sub_map then
+			gpr_keys = {}
+			for pos, arg in ipairs(args or {}) do
+				local effective = apply_sub(sub_map, arg)
+				local key, unresolved, slot = resolve_gpr_key(effective)
+				gpr_keys[pos] = key
+				if unresolved then
+					errors[#errors + 1] = {
+						kind = "reguse_unresolved",
+						line = line,
+						msg  = string.format("RegUse operand %q does not resolve in schema %q",
+							effective, (reg_use_schema and reg_use_schema.name) or "?"),
+					}
+				end
+				if key and slot and slot_readonly[slot] then
+					local row = M.instr(encoder)
+					if row and row.writes then
+						for _, wpos in ipairs(row.writes) do
+							if wpos == pos then
+								errors[#errors + 1] = {
+									kind = "reguse_const_write",
+									line = line,
+									msg  = string.format("RegUse slot %q is Reg const; %s writes it",
+										slot, encoder),
+								}
+							end
+						end
+					end
+				end
+			end
+		end
+		if not reg_use_schema then
+			gpr_keys = nil
+		end
+		local isa = M.instr(encoder)
+		local isa_kind = isa and isa.kind or "unknown"
+		local nop_words = (encoder == "nop" and 1) or (encoder == "nop2" and 2) or 0
+		local is_yield = (encoder == "mac_yield" or encoder == "mac_yield_tail")
+		local gp0_shape = type(encoder) == "string"
+			and encoder:match("^mac_format_([%w_]+)_color$")
+			or nil
+		items[#items + 1] = {
+			kind                     = "word",
+			encoder                  = encoder,
+			args                     = args,
+			i                        = word_idx,
+			word_count               = 1,
+			line                     = line,
+			call_text                = eff_call_text,
+			root_call_text           = eff_root_call_text,
+			invocation_ids           = inv_ids,
+			outermost_invocation_id  = outermost,
+			gpr_keys                 = gpr_keys,
+			ident                    = encoder,
+			isa_kind                 = isa_kind,
+			nop_words                = nop_words,
+			is_yield                 = is_yield,
+			gp0_shape                = gp0_shape,
+		}
+		word_events[#word_events + 1] = {
+			i                        = word_idx,
+			encoder                  = encoder,
+			args                     = args,
+			def_path                 = def_source_now or "",
+			def_line                 = def_line_now or 0,
+			call_text                = eff_call_text,
+			root_call_text           = eff_root_call_text,
+			invocation_ids           = inv_ids,
+			outermost_invocation_id  = outermost,
+			word_count               = 1,
+			gpr_keys                 = gpr_keys,
+			ident                    = encoder,
+			kind                     = isa_kind,
+			nop_words                = nop_words,
+			is_yield                 = is_yield,
+			gp0_shape                = gp0_shape,
+		}
+		word_idx = word_idx + 1
+	end
+
+	local function emit_marker(kind, name, target, line,
+		immediate_call_text, root_call_text_w,
+		consuming_encoder, consuming_arg_pos)
+		local inv_ids   = open_invocation_ids_snapshot()
+		local outermost = inv_ids[1] or 0
+		-- Markers carry the open invocation stack snapshot. `call_text` / `root_call_text` belong to words, not markers — markers are zero-width and skip per-word call-site attribution.
+		-- `consuming_encoder` + `consuming_arg_pos` carry the surrounding control-transfer instruction context
+		-- (e.g. `branch_le_zero` consuming its 3rd argument, or `jump` / `call_addr` consuming their only argument).
+		-- `passes/offsets.lua` reads these to dispatch per-consuming-instruction offset encoding.
+		-- nil for top-level markers (where the marker is the entire token — no surrounding consuming instruction).
+		local it = {
+			kind                     = kind,
+			name                     = name,
+			line                     = line,
+			word_index               = word_idx,
+			invocation_ids           = inv_ids,
+			outermost_invocation_id  = outermost,
+		}
+		if target ~= nil then it.target = target end
+		if consuming_encoder then it.consuming_encoder = consuming_encoder end
+		if consuming_arg_pos  then it.consuming_arg_pos  = consuming_arg_pos  end
+		items[#items + 1] = it
+		markers[#markers + 1] = {
+			kind              = kind,
+			name              = name,
+			line              = line,
+			word_index        = word_idx,
+			target            = target,
+			consuming_encoder = consuming_encoder,
+			consuming_arg_pos = consuming_arg_pos,
+		}
+	end
+
+	-- Count top-level commas in `tok` between position `from_pos` (inclusive) and `to_pos` (exclusive).
+	-- Tracks paren depth so commas inside nested () don't count. Skips string literals + comments.
+	-- Used by `emit_embedded_markers` to compute `consuming_arg_pos` for each embedded marker.
+	local function count_top_level_commas(tok, from_pos, to_pos)
+		local depth  = 0
+		local count  = 0
+		local i      = from_pos
+		while i < to_pos do
+			local c = tok:sub(i, i)
+			if c == "'" or c == '"' then
+				local next_pos = M.skip_str_or_cmt(tok, i)
+				i = (next_pos > i) and next_pos or (i + 1)
+			elseif c == "/" and tok:sub(i + 1, i + 1) == "/" then
+				-- line comment: skip to end of line
+				local nl = tok:find("\n", i, true)
+				i = (nl and nl + 1) or (#tok + 1)
+			elseif c == "/" and tok:sub(i + 1, i + 1) == "*" then
+				-- block comment: skip to matching */
+				local close = tok:find("*/", i + 2, true)
+				i = (close and close + 2) or (#tok + 1)
+			elseif c == "(" then
+				depth = depth + 1
+				i = i + 1
+			elseif c == ")" then
+				depth = depth - 1
+				i = i + 1
+			elseif c == "," and depth == 0 then
+				count = count + 1
+				i = i + 1
+			else
+				i = i + 1
+			end
+		end
+		return count
+	end
+
+	-- Find the position of the consuming instruction's open paren (the `(` that starts the consuming instruction's argument list).
+	-- Returns nil if the token's leading text isn't an ident followed by `(` (e.g. the ident is at the start of a non-instruction token).
+	local function find_consuming_paren(tok)
+		local i = 1
+		while i <= #tok do
+			local c = tok:sub(i, i)
+			if    c == "(" then return i end
+			if not c:match("[%w_]") and c ~= " " then return nil end
+			i = i + 1
+		end
+		return nil
+	end
+
+	local function emit_embedded_markers(tok, tok_line, consuming_encoder)
+		-- When called with a non-nil `consuming_encoder`, the marker is nested inside that instruction's argument list.
+		-- We compute each marker's arg position by counting top-level commas between the consuming instruction's `(` and the marker's start.
+		local consuming_paren = nil
+		if consuming_encoder then consuming_paren = find_consuming_paren(tok) end
+		local pos = 1
+		while pos <= #tok do
+			-- Trim leading whitespace and comments before each scan.
+			pos = M.skip_ws_and_cmt(tok, pos)
+			if pos > #tok then break end
+			local  ident, after = M.read_ident(tok, pos)
+			if not ident then
+				-- Not an ident: token is a string or comment; skip or one-step.
+				local next_pos = M.skip_str_or_cmt(tok, pos)
+				pos = (next_pos > pos) and next_pos or (pos + 1)
+				goto continue_loop
+			end
+			if M.DELAY_MARKERS[ident] then
+				local arg_pos = nil
+				if consuming_encoder and consuming_paren then
+					arg_pos = count_top_level_commas(tok, consuming_paren + 1, pos) + 1
+				end
+				emit_marker("delay", ident, nil, tok_line, nil, nil, consuming_encoder, arg_pos)
+				pos = after
+				goto continue_loop
+			end
+			if ident ~= "atom_label" and ident ~= "atom_offset" then
+				-- Ordinary ident; nothing to emit, step past the ident only.
+				pos = after
+				goto continue_loop
+			end
+			-- Marker ident: parse the (...) arguments.
+			local open               = M.skip_ws_and_cmt(tok, after)
+			local inner, after_paren = M.read_parens(tok, open)
+			if not inner then
+				-- (...) Unreadable: fall back to non-marker behavior.
+				pos = after
+				goto continue_loop
+			end
+			-- Commit: label takes 1 arg, offset takes 2.
+			-- For embedded markers, propagate the consuming_encoder + the marker's arg position
+			-- (1-based) so `passes/offsets.lua` can dispatch per-consuming-instruction offset encoding.
+			-- Top-level markers (no consuming_encoder) get nil for both — the offsets pass treats
+			-- them as branch-equivalent for backward compatibility.
+			local arg_pos = nil
+			if consuming_encoder and consuming_paren then
+				arg_pos = count_top_level_commas(tok, consuming_paren + 1, pos) + 1
+			end
+			local args = split_call_args(inner)
+			if ident == "atom_label" then emit_marker("label",  args[1] or "", nil,           tok_line, nil, nil, consuming_encoder, arg_pos)
+			else                          emit_marker("offset", args[1] or "", args[2] or "", tok_line, nil, nil, consuming_encoder, arg_pos)
+			end
+			pos = after_paren
+			::continue_loop::
+		end
+	end
+
+	local function emit_invoke_begin(inv_kind, component_name, call_text,
+		root_call_text, call_path, call_line)
+		next_inv_id = next_inv_id + 1
+		-- Invocation-level debug_skip stamp: Emission pass owns `atom.paths.invocations[*].debug_skip`.
+		-- The stamp is resolved from the `corpus.components[name]` registry (passed in via `ctx_table.components` by `emission_model.run`), 
+		-- Unmarked components stamp `false` (not `nil`) so consumers can dispatch on the boolean without nil checks.
+		--
+		-- The walker has already found the component body in `ctx_table.component_index[component_name]`, so the matching entry MUST exist in `ctx_table.components[component_name]`
+		-- (both registries are populated from the same source by the components pass).
+		-- A missing entry is a corpus-plumbing bug; we fail loudly here rather than silently stamp `false` and mask the regression.
+		local  components    = ctx_table.components
+		local  component_def = components and components[component_name] or nil
+		if not component_def then
+			error("duffle.emit_invoke_begin: component " .. string.format("%q", component_name)
+				.. " is present in `component_index` (the walker matched a `mac_" .. component_name .. "()` call) but absent from `components` (the canonical corpus.components registry). "
+				.. "This is a corpus-plumbing bug — the components pass must populate corpus.components[name] for every component it puts in corpus.component_body_index[name]. " 
+				.. "The emission pass refuses to silently stamp `debug_skip = false` for a missing registry entry."
+				, 0
+			)
+		end
+		local debug_skip_stamp = component_def.debug_skip == true
+		local inv = {
+			id              = next_inv_id,
+			parent_id       = 0,         -- patched below by caller
+			kind            = inv_kind,
+			component_name  = component_name,
+			call_text       = call_text,
+			root_call_text  = root_call_text,
+			call_path       = call_path,
+			call_line       = call_line,
+			def_path        = nil,       -- patched below after component lookup
+			def_line        = nil,
+			-- 0-based emitted-word position. `word_idx` is the monotonic 0-based counter of `word` items emitted so far in this walk — 
+			-- BEFORE this invocation's first word is emitted, it equals the position of the first word inside the invocation. 
+			-- `start_word` (1-based items index of `invoke_begin`) is kept for items-walking consumers (Annotation pass bounds checks),
+			-- but DWARF / provenance rows MUST read `start_pos` because those rows are 1-based over the dense `word_events` stream (which has no `invoke_begin` items).
+			start_pos       = word_idx,
+			start_word      = #items + 1, -- 1-based items index of invoke_begin
+			end_pos         = nil,        -- patched by emit_invoke_end
+			end_word        = nil,        -- patched by emit_invoke_end
+			word_count      = 0,
+			debug_skip      = debug_skip_stamp,
+			errors          = {},
+		}
+		invocations[#invocations + 1] = inv
+		items      [#items       + 1] = {
+			kind           = "invoke_begin",
+			invocation_id  = inv.id,
+			word_index     = word_idx,
+			invocation_ids = open_invocation_ids_snapshot(),
+		}
+		invocation_stack[#invocation_stack + 1] = inv
+		return inv
+	end
+
+	local function emit_invoke_end(inv)
+		-- 0-based emitted-word position of the LAST word inside this invocation.
+		-- After the last body word was emitted, `word_idx` was incremented past it, so `word_idx - 1` is the 0-based position of the last word.
+		inv.end_pos  = word_idx - 1
+		inv.end_word = #items   + 1  -- 1-based items index of invoke_end
+		items[#items + 1] = {
+			kind           = "invoke_end",
+			invocation_id  = inv.id,
+			word_index     = word_idx,
+			invocation_ids = open_invocation_ids_snapshot(),
+		}
+		for i = #invocation_stack, 1, -1 do
+			if invocation_stack[i] == inv then
+				table.remove(invocation_stack, i)
+				break
+			end
+		end
+	end
+
+	-- Resolve the per-token word count. If unresolved, surface ONE warning
+	-- and fall back to 1 opaque word so the cycle budget still accounts for the slot.
+	local function resolve_count(ident, tok_line)
+		local wc = ctx_table.word_counts
+		if    wc and wc[ident] then return wc[ident] end
+		local canon = M.gte_canon(ident)
+		if canon ~= ident and wc and wc[canon] then return wc[canon] end
+		warnings[#warnings + 1] = {
+			kind = "uncounted",
+			line = tok_line,
+			msg  = string.format("project_emission: opaque word emitted for %q (no entry in word_counts or component_index)", 
+			ident),
+		}
+		return 1
+	end
+
+	-- Recursive walker: walk one body entry, possibly descending into components.
+	-- walk_parent_inv_id:       Invocation ID of the enclosing call (0 for the root call).
+	-- walk_root_call_text:      Outermost `mac_X(...)` token text (preserved across recursion).
+	-- walk_immediate_call_text: IMMEDIATE outer `mac_X(...)` token text for words emitted in this body — nil for the root atom body.
+	-- Two trackers are propagated as separate parameters so words deep inside nested expansions correctly identify both their immediate call site and the outermost call site.
+	local function walk_body_entry(body_entry, walk_parent_inv_id,
+		walk_root_call_text, walk_immediate_call_text)
+		local tokens     = body_entry.body_tokens or {}
+		local body_off   = body_entry.body_off or 0
+		local line_of    = body_entry.line_of or M.LineIndex("")
+		local def_source = body_entry.source or ""
+		local def_line   = body_entry.declaration or 0
+		local sub_map    = body_entry.sub_map
+		-- Per-token dispatch: each matched branch returns; only the fall-through
+		-- "opaque word" emit handles direct encoders + mac_X-without-component.
+		local function process_token(bt)
+			local tok = M.trim(bt.tok or "")
+			if tok == "" then return end
+			local ident, after = M.read_ident(tok, 1)
+			if not ident then ident = "?" end
+			local _, args  = token_ident_and_args(tok)
+			local tok_line = line_of(body_off + bt.rel) or 0
+			if M.DELAY_MARKERS[ident] then
+				emit_marker("delay", ident, nil, tok_line)
+				local rest = tok:sub(after or (#tok + 1))
+				while true do
+					rest = M.trim(rest)
+					if rest:sub(1, 2) ~= "/*" then break end
+					local close = rest:find("*/", 3, true)
+					if not close then rest = ""; break end
+					rest = rest:sub(close + 2)
+				end
+				if rest ~= "" then
+					process_token({ tok = rest, rel = bt.rel })
+				end
+				return
+			end
+			-- embedded markers live only in non-marker tokens.
+			-- Pass `ident` as the consuming instruction so `emit_embedded_markers` can compute each marker's arg position + record the consuming_encoder for the offsets pass.
+			-- Canonicalize `jump_rel` to `branch_equal` (its preprocessor-expanded form) so the `consuming_encoder` metadata in marker records is canonical. 
+			-- `jump_rel`: unconditional jump alias from `code/duffle/mips.h`.
+			local consuming_encoder_for_markers = (ident == "jump_rel") and "branch_equal" or ident
+			if ident ~= "atom_label" and ident ~= "atom_offset" then
+				emit_embedded_markers(tok, tok_line, consuming_encoder_for_markers)
+			end
+			-- atom_label / atom_offset: terminal markers, no further descent.
+			-- Top-level markers (the marker IS the entire token) have no consuming instruction;
+			-- nil for both `consuming_encoder` and `consuming_arg_pos`.
+			-- The offsets pass treats these as branch-equivalent for backward compatibility.
+			-- TODO(Ed): Review this don't want legacy cruft here..
+			if     ident == "atom_label"  then emit_marker("label",  args[1] or "", nil,           tok_line); return
+			elseif ident == "atom_offset" then emit_marker("offset", args[1] or "", args[2] or "", tok_line); return
+			end
+			if ident:sub(1, 4) == "mac_" then
+				local bare = ident:sub(5)
+				local comp = ctx_table.component_index[bare]
+				if comp then
+					local invocation_root_call_text = walk_root_call_text or tok
+					if ctx_table.visiting[bare] then
+						-- Cycle: still allocate inv_id, emit zero-width begin/end, record the cycle error; do NOT recurse.
+						local inv = emit_invoke_begin(comp.kind or "comp_bare", bare, tok, invocation_root_call_text, def_source, tok_line)
+						inv.parent_id = walk_parent_inv_id
+						inv.call_text = tok
+						local err = {
+							kind   = "cycle",
+							msg    = string.format("project_emission: component cycle detected: %q", bare),
+							source = def_source,
+							line   = tok_line,
+						}
+						inv.errors[#inv.errors + 1] = err
+						errors    [#errors     + 1] = err
+						emit_invoke_end(inv)
+						return
+					end
+					-- First visit: descend + count + count_mismatch-check below.
+					ctx_table.visiting[bare] = true
+					local inv = emit_invoke_begin(comp.kind or "comp_bare", bare, tok, invocation_root_call_text, def_source, tok_line)
+					inv.parent_id = walk_parent_inv_id
+					inv.call_text = tok
+					inv.def_path  = comp.source
+					inv.def_line  = comp.declaration
+					-- Propagate trackers into the recursive walk:
+					--   immediate_call_text = this call's tok (the IMMEDIATE outer call for words emitted in this body)
+					--   root_call_text      = the OUTERMOST call (immutable across the recursion)
+					local formal_names = ctx_table.component_index[bare]
+						and ctx_table.component_index[bare].arg_names
+					local child_map = nil
+					if formal_names then
+						child_map = {}
+						for i, fname in ipairs(formal_names) do
+							child_map[fname] = apply_sub(sub_map, args[i])
+						end
+					end
+					walk_body_entry({
+							body_tokens = comp.body_tokens or {},
+							body_off    = comp.body_off or 0,
+							line_of     = comp.line_of,
+							source      = comp.source,
+							declaration = comp.declaration,
+							sub_map     = child_map,
+						},
+						inv.id,
+						invocation_root_call_text,
+						tok)
+					ctx_table.visiting[bare] = nil
+					emit_invoke_end(inv)
+					-- Count `word` items inside [start_word, end_word].
+					local wc_inside = 0
+					for i = inv.start_word, inv.end_word do
+						local it = items[i]
+						if it and it.kind == "word" then
+							wc_inside = wc_inside + 1
+						end
+					end
+					inv.word_count = wc_inside
+					-- count_mismatch is a construction error: word_counts["mac_X"] is the declared count populated by the components pass; 
+					-- We compare against the measured word count.
+					local declared = ctx_table.word_counts["mac_" .. bare]
+					if declared and wc_inside ~= declared then
+						local err = {
+							kind   = "count_mismatch",
+							msg    = string.format("project_emission: mac_%s declared=%d measured=%d", bare, declared, wc_inside),
+							source = def_source,
+							line   = tok_line,
+						}
+						inv.errors[#inv.errors + 1] = err
+						errors    [#errors     + 1] = err
+					end
+					return
+				end
+				-- mac_X NOT in component_index: fall through to opaque emit.
+			end
+			-- Direct encoder, or mac_X-without-component: resolve count + emit n words.
+			-- Resolve_count may emit a warning if the count is unresolved.
+			local n         = resolve_count(ident, tok_line)
+			local out_ident = (ident == "nop2") and "nop" or ident
+			for _ = 1, n do
+				emit_word(out_ident, args, tok_line, tok, def_source, def_line, walk_immediate_call_text, walk_root_call_text, sub_map)
+			end
+		end
+		
+		for _, bt in ipairs(tokens) do
+			process_token(bt)
+		end
+	end
+
+	-- Initialize the per-walk mutable context.
+	-- `visiting` is the active DFS component stack; `root_call_path` / `root_call_line` are preserved across recursion so nested words always point at the
+	-- ORIGINAL root atom call site.
+	ctx_table.visiting       = ctx_table.visiting        or {}
+	ctx_table.root_call_path = ctx_table.root_call_path  or ""
+	ctx_table.root_call_line = ctx_table.root_call_line  or 0
+
+	-- Walk first; the pass caller stamps the root call site for direct words after the projection returns.
+	-- For nested words the def_path / def_line already point at the component source and MUST be preserved (the stamping helper checks for that).
+	walk_body_entry(root_body_entry, 0, nil, nil)
+
+	-- Boundary check: every invoke_begin must have a matching invoke_end.
+	-- If anything is still open, surface a hard error.
+	if #invocation_stack > 0 then
+		errors[#errors + 1] = {
+			kind = "unbalanced",
+			msg  = string.format("project_emission: invocation boundaries not balanced (%d unclosed invocation(s) at end of walk)", #invocation_stack),
+		}
+	end
+
+	return {
+		items       = items,
+		word_events = word_events,
+		markers     = markers,
+		invocations = invocations,
+		errors      = errors,
+		warnings    = warnings,
+	}
+end
+
+--- Project a body string into the per-atom emission projection.
+---
+--- Semantics:
+---   * Direct one-word tokens (`nop`, `add_ui`, ...): one `word` item, encoder = ident, word_count = 1.
+---   * Metadata-backed N-word tokens (`nop2`, `mask_upper`, ...): N `word` items, all sharing the same encoder + word_count = 1.
+---     `nop2` is normalized to encoder `nop` (per the spec).
+---   * `atom_label(F)` markers: one `label` item with `name = "F"`, `word_index = current word_idx`; zero-width (does NOT advance word_idx).
+---   * `atom_offset(B, T)` markers: one `offset` item with `name = "B"`, `target = "T"`, `word_index = current word_idx`; zero-width.
+---   * Delay markers (`GteDelay_` / `LdSlot_` / `BdSlot_` / `DmaSlot_`): one `delay` item; zero-width. The following encoder is the next token.
+---   * `mac_X(...)` calls: emit `invoke_begin` (zero-width), recurse into the component body, emit `invoke_end` (zero-width).
+---     The component body's words land between the begin/end pair; one invocation record is allocated per call (monotonic ID per atom).
+---   * Unknown uncounted macros emit 1 opaque word + one warning per occurrence.
+---   * Tokens whose count cannot be resolved (e.g. `mac_unknown` not in word_counts and not in component_index) surface one
+---     warning; cycle + count-mismatch + boundary violations are construction errors on `pass.errors`.
+---
+--- Every emitted `word` carries: `i` (0-based word index), `encoder`, `args` (top-level args), `def_path`, `def_line`,
+--- `call_text` (the immediate token spelling), `root_call_text` (outermost `mac_X(...)` text), `word_count` (always 1),
+--- `invocation_ids` (innermost last), `outermost_invocation_id`.
+--- Markers carry: `kind`, `name`, `line`, `word_index`, `target` (only for offset kind), plus `invocation_ids` / `outermost_invocation_id`
+--- for the open invocation stack at that word.
+---
+--- @param body_text       string  -- the raw atom body string
+--- @param component_index table   -- bare-name → component record (corpus.component_body_index)
+--- @param word_counts     table   -- macro name → emitted word count
+--- @param components      table   -- bare-name → component definition (corpus.components); REQUIRED — consumed at the invocation-construction site to stamp
+---                                   `invocation.debug_skip`. A missing or non-table `components` raises a fail-loud error rather than silently falling back.
+--- @return EmissionProjection
+function M.project_emission(body_text, component_index, word_counts, components, reg_use_ctx)
+	-- The recursive walk delegates to `_project_emission_inner` so component bodies (which arrive as
+	-- `{body_tokens, body_off, line_of, source, declaration}` records from `corpus.component_body_index`)
+	-- re-enter the same walker with the same shared output state.
+	--
+	-- The walker is body-relative: it builds `line_of` from `body_text` and stamps body-relative line numbers (1..N)
+	-- into `item.line` and `invocation.call_line`. `passes/emission_model.lua::stamp_root_provenance` performs the single
+	-- conversion from body-relative to physical source line at the close site, using the source's `line_of` closure that
+	-- the pass forwarded. One owner of the line state.
+	if type(components) ~= "table" then
+		error("duffle.project_emission: `components` is required "
+			.. "(bare-name -> component definition, e.g. corpus.components); "
+			.. "got " .. type(components) .. ". "
+			.. "The emission pass MUST forward the corpus registry "
+			.. "so the invocation-construction site can stamp `debug_skip` "
+			.. "without a second pass, source parse, or parallel lookup.",
+			0)
+	end
+
+	if type(body_text) ~= "string" or body_text == "" then
+		-- Empty body: still return a valid (empty) projection.
+		return {
+			items       = {},
+			word_events = {},
+			markers     = {},
+			invocations = {},
+			errors      = {},
+			warnings    = {},
+		}
+	end
+
+	local tokens = M.tokenize_body(body_text)
+	return _project_emission_inner({
+		body_tokens = tokens,
+		body_off    = 0,
+		line_of     = M.LineIndex(body_text),
+		source      = "",
+		declaration = 0,
+	},
+	{
+		component_index = component_index or {},
+		word_counts     = word_counts     or {},
+		components      = components,
+		reg_use_schema  = reg_use_ctx and reg_use_ctx.reg_use_schema,
+		reg_use_param   = reg_use_ctx and reg_use_ctx.reg_use_param,
+		atom_name       = reg_use_ctx and reg_use_ctx.atom_name,
+		schema_name     = reg_use_ctx and reg_use_ctx.schema_name,
+	})
+end
+
+-------------------------------------------------------------------------------
+-- find_function_decl_for — backward walk for MipsAtomComp_Proc_ name extraction.
+--
+-- After the `sym` arg was dropped from MipsAtomComp_Proc_, the component name is derived from the preceding
+-- `FI_ Slice_MipsCode ac_X(args)` function declaration. This function walks backward from `before_pos` to find it.
+--
+-- Returns (raw_name, args_inner) or (nil, nil).
+--   raw_name   — e.g. "ac_load_word_imm"
+--   args_inner — e.g. "AtomBuilder_R ab, Reg dst, U4 imm"
+--
+-- The walk finds the LAST "Slice_MipsCode" before before_pos, then skips whitespace + qualifiers
+-- (FI_, atom_dbg_skip, comments) until it finds an ident followed by "(".
+-- That ident is the function name; the parens contents are the args.
+-------------------------------------------------------------------------------
+function M.find_function_decl_for(source, before_pos, slice_mips_code_len)
+	local search_pos = 1
+	local last_match = nil
+	while true do
+		local found = source:find("Slice_MipsCode", search_pos, true)
+		if not found or found >= before_pos then break end
+		last_match = found
+		search_pos = found + slice_mips_code_len
+	end
+	if not last_match then return nil, nil end
+
+	local pos = last_match + slice_mips_code_len
+	while pos < before_pos do
+		-- skip whitespace
+		while pos <= #source do
+			local c = source:sub(pos, pos)
+			if c == " " or c == "\t" or c == "\n" or c == "\r" then
+				pos = pos + 1
+			else
+				break
+			end
+		end
+		if pos > #source then break end
+		-- skip line comments
+		if source:sub(pos, pos + 1) == "//" then
+			while pos <= #source and source:sub(pos, pos) ~= "\n" do pos = pos + 1 end
+			pos = pos + 1
+			goto continue
+		end
+		-- skip block comments
+		if source:sub(pos, pos + 1) == "/*" then
+			local  close = source:find("*/", pos + 2, true)
+			if not close then break end
+			pos = close + 2
+			goto continue
+		end
+		-- try to read an ident
+		local  ident, ident_end = M.read_ident(source, pos)
+		if not ident then break end
+		-- check if the next non-ws char after ident is "("
+		local next_pos = M.skip_ws_and_cmt(source, ident_end)
+		if source:sub(next_pos, next_pos) == "(" then
+			local inner = M.read_parens(source, next_pos)
+			if inner then
+				return ident, inner
+			end
+		end
+		-- ident not followed by "(" — it's a qualifier (FI_, atom_dbg_skip, etc); skip it
+		pos = ident_end
+		::continue::
+	end
+	return nil, nil
+end
+
+-------------------------------------------------------------------------------
+-- find_atom_proc_decl_for — backward walk for MipsAtom_Proc_ name extraction.
+--
+-- The atom name is the preceding `MipsAtom* ident(args)` function ident.
+-- This function walks backward from `before_pos` to find it.
+--
+-- Returns (raw_name, args_inner, func_ident, after_paren) or (nil, nil).
+--   raw_name    — the function ident as written
+--   args_inner  — e.g. "AtomArena_R aa, U4 r_scratch, ..."
+--   after_paren — source position after the function `)`
+--
+-- The walk finds the LAST "MipsAtom*" before before_pos, then skips whitespace + qualifiers (internal, I_, FI_, comments)
+-- until it finds an ident followed by "(".
+-------------------------------------------------------------------------------
+function M.find_atom_proc_decl_for(source, before_pos, mips_atom_ptr_len)
+	local search_pos = 1
+	local last_match = nil
+	while true do
+		-- plain=true: "*" is literal, no escaping needed
+		local  found = source:find("MipsAtom*", search_pos, true)
+		if not found or found >= before_pos then break end
+		last_match = found
+		search_pos = found + mips_atom_ptr_len
+	end
+	if not last_match then return nil, nil end
+
+	local pos = last_match + mips_atom_ptr_len
+	while pos < before_pos do
+		-- skip whitespace
+		while pos <= #source do
+			local c = source:sub(pos, pos)
+			if c == " " or c == "\t" or c == "\n" or c == "\r" then
+				pos = pos + 1
+			else
+				break
+			end
+		end
+		if pos > #source then break end
+		-- skip line comments
+		if source:sub(pos, pos + 1) == "//" then
+			while pos <= #source and source:sub(pos, pos) ~= "\n" do pos = pos + 1 end
+			pos = pos + 1
+			goto continue
+		end
+		-- skip block comments
+		if source:sub(pos, pos + 1) == "/*" then
+			local  close = source:find("*/", pos + 2, true)
+			if not close then break end
+			pos = close + 2
+			goto continue
+		end
+		-- try to read an ident
+		local  ident, ident_end = M.read_ident(source, pos)
+		if not ident then break end
+		-- check if the next non-ws char after ident is "("
+		local next_pos = M.skip_ws_and_cmt(source, ident_end)
+		if source:sub(next_pos, next_pos) == "(" then
+			local inner, after_paren = M.read_parens(source, next_pos)
+			if inner then
+				return ident, inner, ident, after_paren
+			end
+		end
+		-- ident not followed by "(" — it's a qualifier; skip it
+		pos = ident_end
+		::continue::
+	end
+	return nil, nil
+end
+
+return M

scripts/duffle_isa.lua

@@ -0,0 +1,735 @@
+--- duffle_isa.lua — encoder / GTE / hardware tables.
+local M = {}
+
+-- Section 7: domain tables
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- atom_info sub-calls: atom_bind, atom_reads, atom_writes, atom_view, atom_reg_types, atom_ctx, atom_phase.
+M.TAPE_ATOM_MACROS = {
+	["atom_info"] = { kind = "info", binds = false },
+}
+
+-- Empty C macros that prefix the next encoder. Zero words.
+-- BdSlot_ nop is one nop word. The marker is not the BD instruction.
+M.DELAY_MARKERS = {
+	["GteDelay_"] = true,
+	["LdSlot_"]   = true,
+	["BdSlot_"]   = true,
+	["DmaSlot_"]  = true,
+}
+
+-- One row per encoder. Old table names are load-time views (build_isa_views).
+M.INSTRUCTION = {
+	["BdSlot_"]            = { cycles = 0,  kind = "marker", },
+	["LdSlot_"]            = { cycles = 0,  kind = "marker", },
+	["add_s"]              = { cycles = 1,  kind = "alu", },
+	["add_si"]             = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm = { { arg = 3, signed = true, width  = 16, },}, },
+	["add_u"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["add_u_self"]         = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 },                                                       value = { dest = 1, op = "add_u", sources = { 1, 2 }, }, },
+	["add_ui"]             = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm   = { { arg = 3, signed = true, width = 16, }, }, value = { dest = 1, immediate = 3, op = "add_ui", source = 2, }, },
+	["add_ui_self"]        = { cycles = 1,  kind = "alu",        reads = { 1 },    writes = { 1 }, imm   = { { arg = 2, signed = true, width = 16, }, }, value = { dest = 1, immediate = 2, op = "add_ui", source = 1, },  },
+	["and"]                = { cycles = 1,  kind = "alu", },
+	["and_i"]              = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm   = { { arg = 3, width = 16, }, },                value = { dest = 1, immediate = 3, op = "and_i", source = 2, }, },
+	["and_u"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["atom_bind"]          = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["atom_info"]          = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["atom_label"]         = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["atom_offset"]        = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["atom_reads"]         = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["atom_writes"]        = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["branch_equal"]       = { cycles = 2,  kind = "branch",     reads = { 1, 2 }, writes = {}, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["branch_ge_zero"]     = { cycles = 2,  kind = "branch",     reads = { 1 },    writes = {}, imm = { { arg = 2, signed = true, width = 16, }, }, },
+	["branch_gt_zero"]     = { cycles = 2,  kind = "branch",     reads = { 1 },    writes = {}, imm = { { arg = 2, signed = true, width = 16, }, }, },
+	["branch_le_zero"]     = { cycles = 2,  kind = "branch",     reads = { 1 },    writes = {}, imm = { { arg = 2, signed = true, width = 16, }, }, },
+	["branch_lt_zero"]     = { cycles = 2,  kind = "branch",     reads = { 1 },    writes = {}, imm = { { arg = 2, signed = true, width = 16, }, }, },
+	["branch_ne"]          = { cycles = 2,  kind = "branch",     reads = { 1, 2 }, writes = {}, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["call_addr"]          = { cycles = 2,  kind = "call",       reads = {},       writes = { 1 }, },
+	["call_reg"]           = { cycles = 2,  kind = "call",       reads = { 1 },    writes = { 2 }, },
+	["div_s"]              = { cycles = 35, kind = "alu",        reads = { 1, 2 }, writes = {}, },
+	["div_u"]              = { cycles = 35, kind = "alu",        reads = { 1, 2 }, writes = {}, },
+	["gte_load_v0"]        = { cycles = 2,  kind = "cop2_xfer",  reads = { 2 },    writes = {}, },
+	["gte_load_v0v1v2"]    = { cycles = 6,  kind = "cop2_xfer",  reads = { 2 },    writes = {}, },
+	["gte_load_v1"]        = { cycles = 2,  kind = "cop2_xfer",  reads = { 2 },    writes = {}, },
+	["gte_load_v2"]        = { cycles = 2,  kind = "cop2_xfer",  reads = { 2 },    writes = {}, },
+	["gte_lw"]             = { cycles = 1,  kind = "load",       reads = { 2 },    writes = {}, },
+	["gte_lwc2"]           = { cycles = 1,  kind = "load", },
+	["gte_mv_from_ctrl_r"] = { cycles = 1,  kind = "cop2_xfer",  reads = {},       writes = { 1 }, },
+	["gte_mv_from_data_r"] = { cycles = 1,  kind = "cop2_xfer",  reads = {},       writes = { 1 }, },
+	["gte_mv_to_ctrl_r"]   = { cycles = 1,  kind = "cop2_xfer",  reads = { 1 },    writes = {}, },
+	["gte_mv_to_data_r"]   = { cycles = 1,  kind = "cop2_xfer",  reads = { 1 },    writes = {}, },
+	["gte_stotz"]          = { cycles = 1,  kind = "cop2_xfer",  reads = {},       writes = {}, },
+	["gte_stsxy3"]         = { cycles = 1,  kind = "cop2_xfer",  reads = {},       writes = {}, },
+	["gte_sw"]             = { cycles = 1,  kind = "store",      reads = { 2 },    writes = {}, },
+	["gte_swc2"]           = { cycles = 1,  kind = "store", },
+	["jump"]               = { cycles = 2,  kind = "jump",       reads = {},       writes = {}, },
+	["jump_link"]          = { cycles = 2,  kind = "call",       reads = { 1 },    writes = { 2 }, },
+	["jump_reg"]           = { cycles = 2,  kind = "jump",       reads = { 1 },    writes = {}, suppress_arg1 = { R_AtomJmp = "fixed mac_yield handshake", }, },
+	["jump_rel"]           = { cycles = 2,  kind = "branch", delay_slot = true, },
+	["li_s"]               = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, value = { dest = 1, immediate = 3, op = "add_ui", source = 2, }, },
+	["load_byte"]          = { cycles = 1,  kind = "load",       reads = { 2 },    writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["load_byte_u"]        = { cycles = 1,  kind = "load",       reads = { 2 },    writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["load_half"]          = { cycles = 1,  kind = "load",       reads = { 2 },    writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["load_half_u"]        = { cycles = 1,  kind = "load",       reads = { 2 },    writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["load_imm"]           = { cycles = 2,  kind = "alu",        reads = {},       writes = { 1 }, },
+	["load_ui"]            = { cycles = 1,  kind = "alu",        reads = {},       writes = { 1 }, },
+	["load_upper_i"]       = { cycles = 1,  kind = "alu",        reads = {},       writes = { 1 }, imm = { { arg = 2, width = 16, }, }, value = { dest = 1, immediate = 2, op = "load_upper_i", }, },
+	["load_word"]          = { cycles = 1,  kind = "load",       reads = { 2 },    writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["mac_yield"]          = { cycles = 0,  kind = "marker",     reads = {},       writes = {}, },
+	["mask_upper"]         = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, },
+	["mov_from_high"]      = { cycles = 2,  kind = "alu",        reads = {},       writes = { 1 }, },
+	["mov_from_low"]       = { cycles = 2,  kind = "alu",        reads = {},       writes = { 1 }, },
+	["mov_to_high"]        = { cycles = 1,  kind = "alu",        reads = { 1 },    writes = {}, },
+	["mov_to_low"]         = { cycles = 1,  kind = "alu",        reads = { 1 },    writes = {}, },
+	["mult_s"]             = { cycles = 12, kind = "alu",        reads = { 1, 2 }, writes = {}, },
+	["mult_u"]             = { cycles = 12, kind = "alu",        reads = { 1, 2 }, writes = {}, },
+	["nop"]                = { cycles = 1,  kind = "nop",        reads = {},       writes = {}, },
+	["nop2"]               = { cycles = 2,  kind = "nop",        reads = {},       writes = {}, },
+	["nor_u"]              = { cycles = 1,  kind = "alu", },
+	["or_i"]               = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm = { { arg = 3, width = 16, }, }, value = { dest = 1, immediate = 3, op = "or_i", source = 2, }, },
+	["or_i_self"]          = { cycles = 1,  kind = "alu",        reads = { 1 },    writes = { 1 }, imm = { { arg = 2, width = 16, }, }, value = { dest = 1, immediate = 2, op = "or_i", source = 1, }, },
+	["or_u"]               = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["or_u_self"]          = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, value = { dest = 1, op = "or", sources = { 1, 2 }, }, },
+	["set_lt_s"]           = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["set_lt_si"]          = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, },
+	["set_lt_u"]           = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["set_lt_ui"]          = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, },
+	["shift_aright"]       = { cycles = 1,  kind = "alu",        reads = { 2 },    writes = { 1 }, imm = { { arg = 3, width = 5, }, }, },
+	["shift_aright_var"]   = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, imm = { { arg = 3, width = 5, }, }, },
+	["shift_lleft"]        = { cycles = 1,  kind = "alu",        reads = { 2 },    writes = { 1 }, imm = { { arg = 3, width = 5, }, }, },
+	["shift_lleft_self"]   = { cycles = 1,  kind = "alu",        reads = { 1 },    writes = { 1 }, imm = { { arg = 2, width = 5, }, }, value = { dest = 1, immediate = 2, op = "shift_lleft", source = 1, }, },
+	["shift_lleft_var"]    = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["shift_lright"]       = { cycles = 1,  kind = "alu",        reads = { 2 },    writes = { 1 }, imm = { { arg = 3, width = 5, }, }, },
+	["slt_s"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["slt_si"]             = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["slt_u"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["slt_ui"]             = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm = { { arg = 3, signed = true, width = 16,}, }, },
+	["store_byte"]         = { cycles = 1,  kind = "store",      reads = { 1, 2 }, writes = {},    imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["store_half"]         = { cycles = 1,  kind = "store",      reads = { 1, 2 }, writes = {},    imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["store_word"]         = { cycles = 1,  kind = "store",      reads = { 1, 2 }, writes = {},    imm = { { arg = 3, signed = true, width = 16, }, }, },
+	["sub_s"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["sub_u"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+	["sys_mov_from_cop0"]  = { cycles = 1,  kind = "cop0_xfer",  reads = {},       writes = { 1 }, },
+	["sys_mov_to_cop0"]    = { cycles = 1,  kind = "cop0_xfer",  reads = { 1 },    writes = {}, },
+	["xor_i"]              = { cycles = 1,  kind = "alu",        reads = { 1, 2 }, writes = { 1 }, imm = { { arg = 3, width = 16, }, }, value = { dest = 1, immediate = 3, op = "xor_i", source = 2, }, },
+	["xor_u"]              = { cycles = 1,  kind = "alu",        reads = { 2, 3 }, writes = { 1 }, },
+}
+
+-- One row per GTE command. Alias cycle numbers live here, not on INSTRUCTION.
+M.GTE_COMMAND = {
+	["gte_cmdw_avsz3"] = {
+		aliases = { "gte_avg_sort_z3", "gte_avsz3", "gte_cmdw_avg_sort_z3" },
+		cycles  = 5,
+		inputs  = { "C2_SZ0", "C2_SZ1", "C2_SZ2", "C2_SZ3", "gte_cr_ZSF3" },
+		outputs = {
+			{ register = "C2_OTZ", role = "otz", },
+		},
+		latch = {
+			{ register = "C2_OTZ", required = 4, },
+		},
+	},
+	["gte_cmdw_avsz4"] = {
+		aliases = { "gte_avg_sort_z4", "gte_avsz4", "gte_cmdw_avg_sort_z4" },
+		cycles  = 6,
+		inputs  = { "C2_SZ0", "C2_SZ1", "C2_SZ2", "C2_SZ3", "gte_cr_ZSF4" },
+		outputs = {
+			{ register = "C2_OTZ", role = "otz", },
+		},
+		latch = {
+			{ register = "C2_OTZ", required = 4, },
+		},
+	},
+	["gte_cmdw_gpf"] = {
+		aliases = {},
+		cycles = 5,
+		inputs = { "C2_IR0", "C2_IR1", "C2_IR2", "C2_IR3" },
+		outputs = {
+			{ register = "C2_MAC1", role = "mac_result", },
+			{ register = "C2_MAC2", role = "mac_result", },
+			{ register = "C2_MAC3", role = "mac_result", },
+			{ register = "C2_IR1", role = "latest_color", },
+			{ register = "C2_IR2", role = "latest_color", },
+			{ register = "C2_IR3", role = "latest_color", },
+		},
+		latch = {
+			{ register = "C2_MAC1", required = 4, },
+			{ register = "C2_MAC2", required = 4, },
+			{ register = "C2_MAC3", required = 4, },
+			{ register = "C2_IR1",  required = 4, },
+			{ register = "C2_IR2",  required = 4, },
+			{ register = "C2_IR3",  required = 4, },
+		},
+	},
+	["gte_cmdw_mvmva"] = {
+		aliases = {},
+		cycles = 8,
+		inputs = { 
+			"C2_VXY0", "C2_VZ0",
+			"C2_VXY1", "C2_VZ1",
+			"C2_VXY2", "C2_VZ2",
+			"C2_IR1", "C2_IR2", "C2_IR3",
+			"gte_cr_RT11", "gte_cr_RT12", "gte_cr_RT13",
+			"gte_cr_RT21", "gte_cr_RT22", "gte_cr_RT23",
+			"gte_cr_RT31", "gte_cr_RT32", "gte_cr_RT33",
+			"gte_cr_TRX", "gte_cr_TRY", "gte_cr_TRZ"
+		},
+		outputs = {
+			{ register = "C2_IR1", role = "latest_color", },
+			{ register = "C2_IR2", role = "latest_color", },
+			{ register = "C2_IR3", role = "latest_color", },
+		},
+		latch = {
+			{ register = "C2_IR1", required = 4, },
+			{ register = "C2_IR2", required = 4, },
+			{ register = "C2_IR3", required = 4, },
+		},
+	},
+	["gte_cmdw_nclip"] = {
+		aliases = { "gte_nclip" },
+		cycles  = 8,
+		inputs  = { "C2_SXY0", "C2_SXY1", "C2_SXY2" },
+		outputs = {
+			{ register = "C2_SZ3", role = "mac_result", },
+		},
+		latch = {
+			{ register = "C2_SZ3", required = 4, },
+		},
+	},
+	["gte_cmdw_op"] = {
+		aliases = { "gte_cmdw_outer_product", "gte_cmdw_wedge" },
+		cycles  = 6,
+		inputs  = {},
+		outputs = {
+			{ register = "C2_IR1", role = "latest_color", },
+			{ register = "C2_IR2", role = "latest_color", },
+			{ register = "C2_IR3", role = "latest_color", },
+		},
+		latch = {
+			{ register = "C2_IR1", required = 4, },
+			{ register = "C2_IR2", required = 4, },
+			{ register = "C2_IR3", required = 4, },
+		},
+	},
+	["gte_cmdw_rtps"] = {
+		aliases = { "gte_cmdw_rotate_translate_perspective_single", "gte_rtps" },
+		cycles = 15,
+		inputs = { 
+			"C2_VXY0", "C2_VZ0",
+			"C2_VXY1", "C2_VZ1",
+			"C2_VXY2", "C2_VZ2",
+			"C2_RGB", "C2_OTZ",
+			"C2_IR0", "C2_IR1", "C2_IR2", "C2_IR3",
+			"C2_SZ0", "C2_SZ1", "C2_SZ2", "C2_SZ3",
+			"gte_cr_RT11", "gte_cr_RT12", "gte_cr_RT13",
+			"gte_cr_RT21", "gte_cr_RT22", "gte_cr_RT23",
+			"gte_cr_RT31", "gte_cr_RT32", "gte_cr_RT33",
+			"gte_cr_TRX", "gte_cr_TRY", "gte_cr_TRZ",
+			"gte_cr_OFX", "gte_cr_OFY",
+			"gte_cr_H",
+			"gte_cr_DQA", "gte_cr_DQB" 
+		},
+		outputs = {
+			{ register = "C2_SXY2", role = "latest_screen_xy", },
+			{ register = "C2_SZ2",  role = "latest_screen_z", },
+			{ register = "C2_OTZ",  role = "otz", },
+			{ register = "C2_IR0",  role = "latest_color", },
+		},
+		latch = {
+			{ register = "C2_SXY2", required = 4, },
+			{ register = "C2_SZ2",  required = 4, },
+			{ register = "C2_OTZ",  required = 4, },
+			{ register = "C2_IR0",  required = 4, },
+		},
+	},
+	["gte_cmdw_rtpt"] = {
+		aliases = { "gte_cmdw_rotate_translate_perspective_triple", "gte_rtpt" },
+		cycles = 23,
+		inputs = { 
+			"C2_VXY0", "C2_VZ0",
+			"C2_VXY1", "C2_VZ1",
+			"C2_VXY2", "C2_VZ2",
+			"C2_RGB",  "C2_OTZ",
+			"C2_IR0",  "C2_IR1", "C2_IR2", "C2_IR3",
+			"C2_SZ0",  "C2_SZ1", "C2_SZ2", "C2_SZ3",
+			"gte_cr_RT11", "gte_cr_RT12", "gte_cr_RT13",
+			"gte_cr_RT21", "gte_cr_RT22", "gte_cr_RT23",
+			"gte_cr_RT31", "gte_cr_RT32", "gte_cr_RT33",
+			"gte_cr_TRX",  "gte_cr_TRY",  "gte_cr_TRZ",
+			"gte_cr_OFX",  "gte_cr_OFY",
+			"gte_cr_H",
+			"gte_cr_DQA", "gte_cr_DQB"
+		},
+		outputs = {
+			{ register = "C2_SXY0", role = "screen_xy[0]", },
+			{ register = "C2_SXY1", role = "screen_xy[1]", },
+			{ register = "C2_SXY2", role = "latest_screen_xy", },
+			{ register = "C2_SZ3",  role = "latest_screen_z", },
+			{ register = "C2_OTZ",  role = "otz", },
+		},
+		latch = {
+			{ register = "C2_SXY0", required = 4, },
+			{ register = "C2_SXY1", required = 4, },
+			{ register = "C2_SXY2", required = 4, },
+			{ register = "C2_SZ3",  required = 4, },
+			{ register = "C2_OTZ",  required = 4, },
+		},
+	},
+	["gte_cmdw_sqr"] = {
+		aliases = {},
+		cycles  = 5,
+		inputs  = { "C2_IR1", "C2_IR2", "C2_IR3" },
+		outputs = {
+			{ register = "C2_MAC1", role = "mac_result", },
+			{ register = "C2_MAC2", role = "mac_result", },
+			{ register = "C2_MAC3", role = "mac_result", },
+			{ register = "C2_IR1",  role = "latest_color", },
+			{ register = "C2_IR2",  role = "latest_color", },
+			{ register = "C2_IR3",  role = "latest_color", },
+		},
+		latch = {
+			{ register = "C2_MAC1", required = 4, },
+			{ register = "C2_MAC2", required = 4, },
+			{ register = "C2_MAC3", required = 4, },
+			{ register = "C2_IR1",  required = 4, },
+			{ register = "C2_IR2",  required = 4, },
+			{ register = "C2_IR3",  required = 4, },
+		},
+	},
+}
+
+function M.instr    (ident) return M.INSTRUCTION            [ident]          end
+function M.gte_canon(ident) return M.ALIAS_TO_CANONICAL     [ident] or ident end
+function M.gte      (ident) return M.GTE_COMMAND[M.gte_canon(ident)]         end
+
+local function build_isa_views()
+	M.ALIAS_TO_CANONICAL = {}
+	for canon, row in pairs(M.GTE_COMMAND) do
+		M.ALIAS_TO_CANONICAL[canon] = canon
+		for _, alias in ipairs(row.aliases or {}) do
+			M.ALIAS_TO_CANONICAL[alias] = canon
+		end
+	end
+	M.INSTRUCTION_LATENCY     = {}
+	M.INSTRUCTION_GPR_EFFECTS = {}
+	M.IMMEDIATE_FIELD_WIDTHS  = {}
+	M.GPR_VALUE_RULES         = {}
+	M.CONTROL_TRANSFER_DELAY_SLOT_POLICIES = {}
+	for name, row in pairs(M.INSTRUCTION) do
+		M.INSTRUCTION_LATENCY[name] = row.cycles
+		if row.reads or row.writes then
+			M.INSTRUCTION_GPR_EFFECTS[name] = {
+				reads  = row.reads or {},
+				writes = row.writes or {},
+			}
+		end
+		if row.imm   then M.IMMEDIATE_FIELD_WIDTHS[name] = row.imm end
+		if row.value then M.GPR_VALUE_RULES       [name] = row.value end
+		if (row.kind == "branch" or row.kind == "jump" or row.kind == "call")
+			and row.delay_slot ~= false then
+			M.CONTROL_TRANSFER_DELAY_SLOT_POLICIES[name] = {
+				family        = row.kind,
+				suppress_arg1 = row.suppress_arg1,
+			}
+		end
+	end
+	M.GTE_COMMAND_ALIASES = {}
+	M.GTE_COMMAND_INPUTS  = {}
+	M.GTE_COMMAND_OUTPUTS = {}
+	M.GTE_COMMAND_LATCH_WINDOWS = {}
+	for canon, row in pairs(M.GTE_COMMAND) do
+		M.GTE_COMMAND_ALIASES    [canon] = canon
+		M.INSTRUCTION_LATENCY    [canon] = row.cycles
+		M.INSTRUCTION_GPR_EFFECTS[canon] = { reads = {}, writes = {} }
+		for _, alias in ipairs(row.aliases or {}) do
+			M.GTE_COMMAND_ALIASES    [alias] = canon
+			M.INSTRUCTION_LATENCY    [alias] = row.cycles
+			M.INSTRUCTION_GPR_EFFECTS[alias] = { reads = {}, writes = {} }
+		end
+		M.GTE_COMMAND_INPUTS       [canon] = row.inputs
+		M.GTE_COMMAND_OUTPUTS      [canon] = row.outputs
+		M.GTE_COMMAND_LATCH_WINDOWS[canon] = row.latch
+	end
+end
+build_isa_views()
+
+
+--- GTE control-register alias groups.
+--- Aliases within a group write to the same C2 control-register slot (the HW double-maps some C2 slots across multiple PSX SDK / libgte conventions).
+--- Aliases across groups write to distinct C2 slots.
+---
+--- Cross-alias writes inside one atom body, or across the wave-context boundary, silently clobber each other.
+--- The `check_gte_cr_alias_writes` check warns about each pair per source. See `docs/gte_reference.md` §"Control-register alias table"
+--- for the HW rationale and the libgte outer-product convention.
+M.GTE_CR_ALIAS_GROUPS = {
+	{ 24, { "gte_cr_RBK", "gte_cr_OFX" } },   -- background R vs screen offset X
+	{ 25, { "gte_cr_GBK", "gte_cr_OFY" } },   -- background G vs screen offset Y
+	{ 26, { "gte_cr_BBK", "gte_cr_H"   } },   -- background B vs projection plane distance H
+}
+
+-- Packed RT slots named by the gte.h packed-slot comment. First must be written before second.
+M.GTE_PACKED_SLOT_RELATIONS = {
+	{ slot = 2, first = "gte_cr_RT13", second = "gte_cr_RT22" },
+}
+
+-- Operand-class table for the COP2->GPR load-delay check.
+-- Maps each emitting-token ident to the set of GPR operand positions it reads.
+-- Covers the current encoder vocabulary (`code/duffle/mips.h` + `code/duffle/gte.h`); add rows here as new encoders land.
+--
+-- Semantics:
+--   * A "GPR operand position" is the textual slot in the macro's argument list, 1-based; e.g. `load_word(rt, base, off)` has positional operands 1 (rt), 2 (base), 3 (off).
+--     The table reads operands 1 + 2 + 3 to find what GPRs the macro touches.
+--   * The check tracks one entry per destination GPR per MFC2 / CFC2 event.
+--     A subsequent event counts as a "use" iff any of its read operand positions reference that destination GPR's ident (e.g. `R_T0`).
+--   * Branch delay slots are out of scope (MIPS control-flow; tracked separately).
+M.OPERAND_READ_POSITIONS = {
+	-- CPU ALU with one or two GPR operands. Reads every GPR operand.
+	["add_ui"]                 = {1, 2},
+	["li_s"]                   = {1, 2},  -- rt (write), imm16 (immediate)
+	["add_ui_self"]            = {1},
+	["add_si"]                 = {1, 2},
+	["add_u"]                  = {1, 2, 3},
+	["add_u_self"]             = {1, 2},
+	["sub_s"]                  = {1, 2, 3},
+	["sub_u"]                  = {1, 2, 3},
+	["and_i"]                  = {1, 2},
+	["and"]                    = {1, 2, 3},
+	["or_i"]                   = {1, 2},
+	["or_i_self"]              = {1},
+	["or"]                     = {1, 2, 3},
+	["or_self"]                = {1, 2},
+	["xor_i"]                  = {1, 2},
+	["xor"]                    = {1, 2, 3},
+	["slt_s"]                  = {1, 2, 3},
+	["slt_u"]                  = {1, 2, 3},
+	["slt_si"]                 = {1, 2},
+	["slt_ui"]                 = {1, 2},
+	["mult_s"]                 = {1, 2},
+	["mult_u"]                 = {1, 2},
+	["div_s"]                  = {1, 2},
+	["div_u"]                  = {1, 2},
+	-- Shifts: shift_lleft(rd, rt, shamt); the rt operand is the value, rd is dest.
+	["shift_lleft"]            = {1, 2},
+	["shift_lright"]           = {1, 2},
+	["shift_aright"]           = {1, 2},
+	["shift_lleft_self"]       = {1},
+	-- Loads: load_word(rt, base, off); the rt operand is the destination (it's written, not read) and base + off are non-GPR operands.
+	-- The check treats the rt operand as a write, so the read-positions table for `load_*` is empty.
+	["load_word"]              = {},
+	["load_half_u"]            = {},
+	["load_byte_u"]            = {},
+	["load_half"]              = {},
+	["load_byte"]              = {},
+	["load_upper_i"]           = {},
+	["load_ui"]                = {},
+	-- Stores write to memory; base + rt operands are non-read for load-delay purposes.
+	["store_word"]             = {},
+	["store_half"]             = {},
+	["store_byte"]             = {},
+	-- Branches read rs (+ rt for beq/bne). The branch delay slot is out of scope.
+	["branch_equal"]           = {1, 2},
+	["branch_ne"]              = {1, 2},
+	["branch_le_zero"]         = {1},
+	["branch_lt_zero"]         = {1},
+	["branch_ge_zero"]         = {1},
+	["branch_gt_zero"]         = {1},
+	-- Jumps / link: jr / jalr read rs only (the target). RD is the destination link.
+	["jump_reg"]               = {1},
+	["jump_link"]              = {1},
+	["call_reg"]               = {1},
+	["call_addr"]              = {},
+	["jump"]                   = {},
+	-- mask_upper is a 2-word macro: shift_lleft then shift_lright. The first reads rt.
+	["mask_upper"]             = {1, 2},
+	-- move from/to HI/LO.
+	["mov_from_high"]          = {},
+	["mov_from_low"]           = {},
+	["mov_to_high"]            = {1},
+	["mov_to_low"]             = {1},
+	-- GTE transfers / loads / stores / commands: the relevant table values live in the check itself.
+	-- `gte_mv_to_*` writes its rt operand; `gte_mv_from_*` writes its rt operand; `gte_*` commands are atomic-from-the-CPU-POV
+	-- once they issue (the CPU holds until the command completes, so load-delay violations don't surface here).
+	["gte_mv_from_data_r"]     = {},
+	["gte_mv_from_ctrl_r"]     = {},
+	["gte_mv_to_data_r"]       = {},
+	["gte_mv_to_ctrl_r"]       = {},
+	["gte_lw"]                 = {},
+	["gte_sw"]                 = {},
+	["shift_lleft_var"]        = {1, 2, 3},  -- rd, rt, rs (variable shift amount)
+	["shift_aright_var"]       = {1, 2, 3},
+}
+
+-- GP0 packet sizes (total words including the 1-word tag) per GP0 cmd byte.
+-- Per PSX-SPX `docs/psx-spx/docs/graphicsprocessingunitgpu.md` §"GPU Render Polygon Commands":
+--   Each polygon command's word count = 1 (tag/cmd) + per-vertex (vertex + optional color + optional UV).
+--   F3:  cmd + 3 vertices = 4 words; +1 tag = 5
+--   F4:  cmd + 4 vertices = 5 words; +1 tag = 6
+--   G3:  cmd + 3×(color + vertex) = 6 words; +1 tag = 7
+--   G4:  cmd + 4×(color + vertex) = 8 words; +1 tag = 9
+--   FT3: cmd + tpage + clut + 3×(vertex + UV) = 7 words; +1 tag = 8
+--   FT4: cmd + tpage + clut + 4×(vertex + UV) = 9 words; +1 tag = 10
+--   GT3: cmd + tpage + clut + 3×(color + vertex + UV) = 9 words; +1 tag = 10
+--   GT4: cmd + tpage + clut + 4×(color + vertex + UV) = 12 words; +1 tag = 13
+--
+-- Cross-checked against code/duffle/gp.h struct sizes + the set_poly_* macros
+-- (which encode "len" = "words after tag"):
+--   set_poly_f3(p)   -> set_len(p, 4)   ->  5 total GP0 0x20
+--   set_poly_ft3(p)  -> set_len(p, 7)   ->  8 total GP0 0x24
+--   set_poly_f4(p)   -> set_len(p, 5)   ->  6 total GP0 0x28
+--   set_poly_ft4(p)  -> set_len(p, 9)   -> 10 total GP0 0x2C
+--   set_poly_g3(p)   -> set_len(p, 6)   ->  7 total GP0 0x30
+--   set_poly_gt3(p)  -> set_len(p, 9)   -> 10 total GP0 0x34
+--   set_poly_g4(p)   -> set_len(p, 8)   ->  9 total GP0 0x38
+--   set_poly_gt4(p)  -> set_len(p, 12)  -> 13 total GP0 0x3C
+M.GP0_CMD_SIZE = {
+	[0x20] =  5,  -- Poly_F3
+	[0x24] =  8,  -- Poly_FT3
+	[0x28] =  6,  -- Poly_F4
+	[0x2C] = 10,  -- Poly_FT4
+	[0x30] =  7,  -- Poly_G3
+	[0x34] = 10,  -- Poly_GT3
+	[0x38] =  9,  -- Poly_G4
+	[0x3C] = 13,  -- Poly_GT4
+}
+
+-- Shape suffix (after `ac_format_` / `mac_format_` prefix) -> GP0 cmd byte.
+-- Lets the static-analysis check derive the cmd byte from a macro name like `mac_format_g4_color` -> `g4` -> 0x38 -> 9 expected words.
+M.GP0_CMD_BY_SHAPE = {
+	["f3"]  = 0x20, ["ft3"] = 0x24,
+	["f4"]  = 0x28, ["ft4"] = 0x2C,
+	["g3"]  = 0x30, ["gt3"] = 0x34,
+	["g4"]  = 0x38, ["gt4"] = 0x3C,
+}
+
+M.UNKNOWN_INSTRUCTION_CYCLES = 1
+
+-- Hardware-relation policy table.
+--
+-- The forward walker in `passes/static_analysis.lua::analyze_hardware_relations` reads every emitted word_event, matches its `encoder` against `row.token`, and:
+--   * stages the event as a producer in `atom.paths.forward_state`; or
+--   * matches it as a consumer against pending producers and records a hazard on `atom.paths.hazards` when the gap is below `visibility.required`.
+--
+-- Each row is the contract for one CPU-to-coprocessor transfer semantic (the coprocessor-to-CPU path mirrors the same shape).
+-- The `reads` / `writes` sub-tables carry the argument positions the analyzer inspects:
+--   * `writes.arg` is the destination operand (the producer's effect); the analyzer stages this register as a pending producer.
+--   * `reads` (when present) lists the operand positions the same token reads back from hardware; for MTC2 / CTC2 the producer reads the GPR source it is loading from.
+--     The `fanout_to` field (MTC2-IRGB row only) tells the consumer-match logic which downstream COP2 registers are transitively updated by the write.
+--
+-- Visibility semantics:
+--   * `kind = "post_producer_words"` means the consumer observes the producer's effect after `required` independent emitted words that are
+--     strictly between the producer and the consumer. The producer's own emitted slot is implicit (it counts as the slot of issue, not toward `required`) 
+--     per the PSX-SPX rule: "Store delays are counted in numbers of clock cycles (not in numbers of opcodes).
+--     For 3 cycle delay, one must usually insert 3 cached opcodes (or one uncached opcode)."
+--   * `required` is the minimum count of intervening emitted words between producer and consumer.
+--     `required = 0` permits the consumer on the very next slot; `required < 0` would place the consumer on the same slot as the producer
+--     and is reserved for future "self-retires" relations.
+--
+-- Evidence:
+--   * `evidence.confidence` is one of `"exact"`, `"conservative"`, `"unknown"`. The severity comes from `violation_kind`;
+--     A hardware measurement that the vendor caveats may still classify as `"conservative"` even when the underlying timing is numerically known.
+--   * `evidence.source` is the upstream reference (file + line range) the row is sourced from. New rows must carry this citation.
+--
+-- Consumers:
+--   * passes/static_analysis.lua::analyze_hardware_relations          (forward walker).
+--   * passes/static_analysis.lua::transfer_hazards CHECK_RULES reader (renders hazards onto `findings`).
+-- This table is consumed by the hardware-relation analyzer and hazard renderer.
+M.HARDWARE_RELATIONS = {
+	-- CPU → COP2 data register (MTC2). The ordinary default is 2 cached words between producer and consumer (cpuspecifications.md:407-419).
+	{
+		id             = "mtc2_gpr_visibility",
+		semantic       = "MTC2",
+		consumer       = "cop2_input",
+		token          = "gte_mv_to_data_r",
+		direction      = "gpr_to_cop2_data",
+		reads          = { domain = "gpr",       arg = 1 },
+		writes         = { domain = "cop2.data", arg = 2 },
+		visibility     = { kind = "post_producer_words", required = 2 },
+		evidence       = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:407-419",
+		},
+		violation_kind = "error",
+	},
+	-- CPU → COP2 data register when the destination is C2_IRGB (data 28).
+	-- C2_IRGB drives the IR1/IR2/IR3 color-conversion fan-out, which extends the propagation delay to 3 cached words.
+	-- `destination_match = "C2_IRGB"` is the row's filter; the analyzer consults this when the producer's destination operand equals "C2_IRGB".
+	-- C2_ORGB (data 29) is read-only and is never classified as a writable fan-out destination.
+	{
+		id              = "mtc2_irgb_visibility",
+		semantic        = "MTC2",
+		consumer        = "cop2_input",
+		token           = "gte_mv_to_data_r",
+		direction       = "gpr_to_cop2_data",
+		reads           = { domain = "gpr",       arg = 1 },
+		writes          = { domain = "cop2.data", arg = 2 },
+		destination_match = "C2_IRGB",
+		fanout_to       = { "C2_IR1", "C2_IR2", "C2_IR3" },
+		visibility      = { kind = "post_producer_words", required = 3 },
+		evidence        = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:407-419",
+		},
+		violation_kind  = "error",
+	},
+	-- CPU → COP2 control register (CTC2). Ordinary minimum 2;
+	-- no IRGB-style fan-out exists for control registers (per spec §3.6: only C2_IRGB has the 3-cycle fan-out on the data side).
+	{
+		id             = "ctc2_gpr_visibility",
+		semantic       = "CTC2",
+		consumer       = "cop2_input",
+		token          = "gte_mv_to_ctrl_r",
+		direction      = "gpr_to_cop2_control",
+		reads          = { domain = "gpr",        arg = 1 },
+		writes         = { domain = "cop2.ctrl",  arg = 2 },
+		visibility     = { kind = "post_producer_words", required = 2 },
+		evidence       = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:407-419",
+		},
+		violation_kind = "error",
+	},
+	-- COP2 data → GPR (MFC2). One cached slot between the transfer and the first GPR consumer;
+	-- the GPR is not updated until the instruction AFTER the MFC2 completes (geometrytransformationenginegte.md:29-32).
+	{
+		id             = "mfc2_gpr_visibility",
+		semantic       = "MFC2",
+		consumer       = "gpr_read",
+		token          = "gte_mv_from_data_r",
+		direction      = "cop2_data_to_gpr",
+		reads          = { domain = "cop2.data",  arg = 2 },
+		writes         = { domain = "gpr",        arg = 1 },
+		visibility     = { kind = "post_producer_words", required = 1 },
+		evidence       = {
+			confidence = "exact",
+			source     = "geometrytransformationenginegte.md:29-32",
+		},
+		violation_kind = "error",
+	},
+	-- COP2 control → GPR (CFC2). Same delay as MFC2 (cpuspecifications.md treats the two load-from-COP2 paths symmetrically).
+	{
+		id             = "cfc2_gpr_visibility",
+		semantic       = "CFC2",
+		consumer       = "gpr_read",
+		token          = "gte_mv_from_ctrl_r",
+		direction      = "cop2_control_to_gpr",
+		reads          = { domain = "cop2.ctrl",  arg = 2 },
+		writes         = { domain = "gpr",        arg = 1 },
+		visibility     = { kind = "post_producer_words", required = 1 },
+		evidence       = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:382-419",
+		},
+		violation_kind = "error",
+	},
+	-- COP0 control → GPR (MFC0).
+	-- One cached slot; the analyzer treats `sys_mov_from_cop0(rt, 12)` (the SR/CU2 transfer) as the same shape as the COP2 load-delay path.
+	-- The semantic-level SR/CU2 transition models the load delay;
+	-- SR.CU2 bounded-value propagation is modeled separately).
+	{
+		id             = "mfc0_gpr_visibility",
+		semantic       = "MFC0",
+		consumer       = "gpr_read",
+		token          = "sys_mov_from_cop0",
+		direction      = "cop0_control_to_gpr",
+		reads          = { domain = "cop0.ctrl",  arg = 2 },
+		writes         = { domain = "gpr",        arg = 1 },
+		visibility     = { kind = "post_producer_words", required = 1 },
+		evidence       = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:171-178",
+		},
+		violation_kind = "error",
+	},
+	-- Memory -> COP2 data register (LWC2).
+	-- The memory-side timing is not measured by the vendored GTE latch experiment, so this relation has no numeric retirement threshold.
+	-- The LWC2 destination has TWO retirement regimes (per PSX-SPX):
+	--   * GTE-command consumer (`gte_cmdw_*`): the GTE pipeline LATCHES the LWC2 result, so a `gte_cmdw_*`
+	--     in the very next slot uses the latched value. Gap = 0 is allowed. (Per `docs/psx-spx/docs/gtepipelinetimings.md:271-274`.)
+	--   * Any other consumer: standard MIPS load delay applies. Gap = 1 required. (Per `docs/psx-spx/docs/cpuspecifications.md:407-419`.)
+	-- Two separate relations so the walker can dispatch by consumer type and emit different severities
+	-- (the GTE-command path is `info` because the latch is intentional; the non-GTE-consumer path is `error` because the missing nop is a real bug).
+	{
+		id             = "lwc2_to_gte_command",
+		semantic       = "LWC2_to_GTE",
+		consumer       = "cop2_input",
+		token          = "gte_lw",
+		direction      = "memory_to_cop2_data",
+		reads          = { domain = "memory",    arg = 2 },
+		writes         = { domain = "cop2.data", arg = 1 },
+		required       = 0,  -- GTE-command consumer: gap = 0 OK (latched).
+		evidence       = {
+			confidence = "measured",
+			source     = "gtepipelinetimings.md:271-274",
+		},
+		violation_kind = "info",
+		clear_on_consumer = true,
+	},
+	{
+		id             = "lwc2_to_other_consumer",
+		semantic       = "LWC2_to_other",
+		consumer       = "cop2_input",
+		token          = "gte_lw",
+		direction      = "memory_to_cop2_data",
+		reads          = { domain = "memory",    arg = 2 },
+		writes         = { domain = "cop2.data", arg = 1 },
+		required       = 1,  -- Non-GTE-consumer: standard MIPS load delay.
+		evidence       = {
+			confidence = "inferred",
+			source     = "cpuspecifications.md:407-419",
+		},
+		violation_kind = "error",
+		clear_on_consumer = true,
+	},
+	-- COP2 data register -> memory (SWC2). A read of C2 state, not a CPU-to-COP2 write.
+	-- The policy row stays in for direction/provenance; staging it as a later command-input producer is suppressed.
+	{
+		id             = "swc2_memory_write",
+		semantic       = "SWC2",
+		consumer       = "gpr_read",
+		token          = "gte_sw",
+		direction      = "cop2_data_to_memory",
+		reads          = { domain = "cop2.data", arg = 1 },
+		writes         = { domain = "memory",    arg = 2 },
+		visibility     = { kind = "none", required = 0 },
+		evidence       = {
+			confidence = "exact",
+			source     = "cpuspecifications.md:79",
+		},
+		violation_kind = "info",
+		stage          = false,
+	},
+	-- MTC0 Status/SR.CU2. The ordinary COP0 store has no general store-delay relation;
+	-- this row feeds the dedicated CU2 transition logic in the same forward walk and is therefore not staged in `pending`.
+	{
+		id             = "mtc0_cu2_visibility",
+		semantic       = "MTC0",
+		consumer       = "gpr_read",
+		token          = "sys_mov_to_cop0",
+		direction      = "gpr_to_cop0_status",
+		reads          = { domain = "gpr",          arg = 1 },
+		writes         = { domain = "cop0.status",  arg = 2 },
+		status_register = 12,
+		visibility     = { kind = "post_producer_words", required = 2 },
+		evidence       = {
+			confidence = "conservative",
+			source     = "cpuspecifications.md:543,625-628",
+		},
+		violation_kind = "warning",
+		stage          = false,
+		cu2_transition = true,
+	},
+ }
+
+-- Bounded Status/SR.CU2 transition policy.
+-- The value lattice and the transition consumer both read this immutable row; no second value pass is permitted.
+-- The source says the enable/disable transition takes "2 clock cycles or so", so the boundary is conservative rather than exact.
+M.CU2_TRANSITION_POLICY = {
+	status_register = 12,
+	enable_bit      = 0x40000000,
+	required        = 2,
+	visibility_kind = "post_producer_words",
+	evidence        = {
+		confidence = "conservative",
+		source     = "cpuspecifications.md:543,625-628",
+	},
+}
+
+return M

scripts/duffle_paths.lua

@@ -22,12 +22,10 @@ local M = {}
 local CACHE_KEY = "__duffle_repo_root__"
 
 --- Resolve the repo root from this script's own path. Zero shell spawn.
---- `duffle_paths.lua` always lives at `<repo>/scripts/duffle_paths.lua`, so the repo root is the
---- parent of the directory containing this script. We derive it directly from `debug.getinfo(1, "S").source`
---- (returns `@<path>` for the currently-running chunk).
+--- `duffle_paths.lua` always lives at `<repo>/scripts/duffle_paths.lua`, so the repo root is the parent of the directory containing this script.
+--- We derive it directly from `debug.getinfo(1, "S").source` (returns `@<path>` for the currently-running chunk).
 ---
---- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source),
---- return nil and let `M.setup()` fail loud.
+--- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source), return nil and let `M.setup()` fail loud.
 --- @return string|nil
 local function find_repo_root()
 	if package.loaded[CACHE_KEY] then return package.loaded[CACHE_KEY] end
@@ -47,22 +45,17 @@ local function find_repo_root()
 	return root
 end
 
---- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
---- `package.cpath` (for `lpeg.dll`).
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and `package.cpath` (for `lpeg.dll`).
 ---
 --- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$PATH` mods. 
 --- It just sets `package.path` and `package.cpath` (the standard Lua way to register module search dirs). 
---- lpeg is built by `update_deps.ps1` to `toolchain/lpeg/`, 
---- which we wire into `package.cpath` here (so `require("lpeg")` from `duffle.lua` resolves without any global state).
+--- lpeg is built by `update_deps.ps1` to `toolchain/lpeg/`, which we wire into `package.cpath` here (so `require("lpeg")` from `duffle.lua` resolves without any global state).
 function M.setup()
-	local repo_root = find_repo_root()
+	local  repo_root = find_repo_root()
 	if not repo_root then
-		-- Unreachable in practice: find_repo_root() derives the repo root from this script's
-		-- own source path via debug.getinfo(1, "S").source (no subprocess, no git CLI, <1ms).
-		-- A nil return means the source path did not match the expected
-		-- <repo>/scripts/duffle_paths.lua layout — a packaging bug, not a "missing git repo"
-		-- condition. os.exit(2) is retained so a real failure surfaces loud rather than
-		-- silently producing an unconfigured module table.
+		-- Unreachable in practice: find_repo_root() derives the repo root from this script's own source path via debug.getinfo(1, "S").source (no subprocess, no git CLI, <1ms).
+		-- A nil return means the source path did not match the expected <repo>/scripts/duffle_paths.lua layout — a packaging bug, not a "missing git repo" condition.
+		-- os.exit(2) is retained so a real failure surfaces loud rather than silently producing an unconfigured module table.
 		os.exit(2)
 	end
 
@@ -87,6 +80,6 @@ end
 -- Run the setup as a side effect.
 M.setup()
 
--- Now that package.path includes scripts/, `require("duffle")` resolves. Return the duffle module
--- so callers can do `local duffle = dofile(...duffle_paths.lua)` in one line.
+-- Now that package.path includes scripts/, `require("duffle")` resolves.
+-- Return the duffle module so callers can do `local duffle = dofile(...duffle_paths.lua)` in one line.
 return require("duffle")

scripts/elf32.lua

@@ -0,0 +1,418 @@
+-- elf32.lua — Pure-Lua ELF32 format helpers with no lfs / no lpeg dependency.
+-- The reload helper's `parse_manifest` (scripts/pcsx_debug_helper/reload.lua) 
+-- and the metaprogram's `read_elf_sections` + `read_nm` (scripts/elf_dwarf.lua)
+-- both parsed ELF32 headers from wire bytes.
+--
+-- This module contains the format constants and the byte-level walker.
+--- The metaprogram side keeps `read_u32_le` / `read_u16_le` as local forwarders; the helper side calls `E.*` directly.
+--
+-- **Adapter contract (explicit pass style):**
+-- The helper VM's `Support.File` exposes byte-read methods that require `self` (fileffi.lua:225-227),
+-- so callers wrap once in a 1-line adapter that strips `self`.
+-- The parsers here operate on the unwrapped form.
+-- Reads are flat function calls — `E.read_u8(adapter, off)`, `E.read_u32(adapter, off)`, `E.size(adapter)`.
+--   read_u8(adapter, off)   -> integer | nil
+--   read_u16(adapter, off)  -> integer | nil
+--   read_u32(adapter, off)  -> integer | nil
+--   size(adapter)           -> integer
+--
+-- **Convention:** every offset in the constants tables is a zero-based wire offset.
+-- The `+ 1` conversion happens only at the `string.byte` boundary inside the readers.
+--
+-- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
+-- spec: System V ABI gABI v1.2 §"Symbol Table" (Elf32_Sym layout)
+
+local M = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Little-endian readers (bit-weighted accumulator, math.floor only)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read a 4-byte little-endian unsigned integer from `adapter` at zero-based wire offset `off`.
+---
+--- Bit weights are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
+--- byte 0 contributes its value directly;
+--- byte 1 is shifted left by 8; byte 2 by 16; byte 3 by 24.
+---
+--- math.floor (not LuaJIT's `>>`) keeps the body portable across LuaJIT 2.0/2.1 and plain Lua 5.x. `string.byte` receives `+ 1` at the boundary.
+---
+--- **Call form:** explicit-pass. The reader receives `adapter` as the first positional argument and the offset as the second; no `self` is passed.
+--- Test fixtures declare `function(offset) ... end` and the parsers call them via dot syntax `adapter.read_u8_at(off)`.
+--- The colon form `adapter:read_u8_at(off)` would prepend the adapter table as `offset` and break the contract.
+--- @param adapter table
+--- @param off     integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u32(adapter, off)
+	return adapter.read_u8_at(off)
+		+ adapter.read_u8_at(off + 0x01) * 0x00000100
+		+ adapter.read_u8_at(off + 0x02) * 0x00010000
+		+ adapter.read_u8_at(off + 0x03) * 0x01000000
+end
+
+--- Read a 2-byte little-endian unsigned integer from `adapter` at zero-based wire offset `off`.
+--- @param adapter table
+--- @param off integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u16(adapter, off)
+	return adapter.read_u8_at(off)
+		+ adapter.read_u8_at(off + 0x01) * 0x00000100
+end
+
+--- Read a 1-byte unsigned integer from `adapter` at zero-based wire offset `off`.
+--- @param adapter table
+--- @param off integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u8(adapter, off)
+	return adapter.read_u8_at(off)
+end
+
+--- Total adapter byte length.
+--- @param adapter table
+--- @return integer
+function M.size(adapter)
+	return adapter.read_size()
+end
+
+--- Forwarders kept for backward compat with scripts/elf_dwarf.lua.
+--- The metaprogram side keeps `read_u32_le` / `read_u16_le`;
+--- both layers now use the same byte-level helpers under the hood.
+function M.read_u32_le(buf, off)
+	local byte_off = off + 1
+	return buf:byte(byte_off)
+		+ buf:byte(byte_off + 0x01) * 0x00000100
+		+ buf:byte(byte_off + 0x02) * 0x00010000
+		+ buf:byte(byte_off + 0x03) * 0x01000000
+end
+
+--- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
+--- @param buf string
+--- @param off integer  -- zero-based wire offset
+--- @return integer
+function M.read_u16_le(buf, off)
+	local byte_off = off + 1
+	return buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Format constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- ELF format constants (System V ABI gABI v1.2).
+M.ELFCLASS32  = 1    -- spec: gABI v1.2 §"ELF Header" — EI_CLASS byte
+M.ELFDATA2LSB = 1    -- spec: gABI v1.2 §"ELF Header" — EI_DATA byte
+M.EM_MIPS     = 8    -- spec: gABI v1.2 §"Machine Information" — MIPS architecture
+
+-- Section type constants (System V ABI gABI v1.2 §"Section Header Table").
+M.SHT_SYMTAB = 2     -- spec: gABI v1.2 §"Section Types" — symbol table
+M.SHT_STRTAB = 3     -- spec: gABI v1.2 §"Section Types" — string table
+M.SHT_NOBITS = 8     -- spec: gABI v1.2 §"Section Types" — no space in file
+
+-- Section flag constants (System V ABI gABI v1.2 §"Section Header Table").
+M.SHF_WRITE     = 0x1 -- spec: gABI v1.2 §"Section Attributes" — writable
+M.SHF_ALLOC     = 0x2 -- spec: gABI v1.2 §"Section Attributes" — occupies memory
+M.SHF_EXECINSTR = 0x4 -- spec: gABI v1.2 §"Section Attributes" — executable
+
+-- ---------------------------------------------------------------------------
+-- ELF32 header layout (System V ABI gABI v1.2 §"ELF Header" Table 1)
+-- ---------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets. The header is 52 bytes total (header_bytes = 0x34 = 52).
+M.ELF32_HEADER = {
+	magic_offset         = 0x00,  -- 4 bytes; expected "\127ELF"
+	magic                = "\127ELF",
+	class_offset         = 0x04,  -- 1 byte;  1 = ELF32, 2 = ELF64
+	endian_offset        = 0x05,  -- 1 byte;  1 = little-endian, 2 = big-endian
+	header_bytes         = 0x34,  -- ELF32 header is 52 bytes total
+	e_entry_offset       = 0x18,  -- 4-byte LE; entry-point virtual address
+	e_shoff_offset       = 0x20,  -- 4-byte LE; section-header table file offset
+	e_shentsize_offset   = 0x2E,  -- 2-byte LE; section-header entry size in bytes
+	e_shnum_offset       = 0x30,  -- 2-byte LE; number of section headers
+	e_shstrndx_offset    = 0x32,  -- 2-byte LE; index of section-name string table
+}
+
+-- ---------------------------------------------------------------------------
+-- ELF32 section-header layout (System V ABI gABI v1.2 §"Section Header Table")
+-- ---------------------------------------------------------------------------
+-- Each entry is 40 bytes (sh_entsize_bytes = 0x28 = 40);
+-- zero-based, field offsets relative to the start of the entry.
+M.ELF32_SECTION = {
+	sh_name_offset    = 0x00,  -- 4-byte LE; offset into .shstrtab
+	sh_type_offset    = 0x04,  -- 4-byte LE; section type (SHT_*)
+	sh_flags_offset   = 0x08,  -- 4-byte LE; section flags (SHF_*)
+	sh_addr_offset    = 0x0C,  -- 4-byte LE; virtual address at execution
+	sh_offset_offset  = 0x10,  -- 4-byte LE; section's file offset
+	sh_size_offset    = 0x14,  -- 4-byte LE; section's size in bytes
+	sh_link_offset    = 0x18,  -- 4-byte LE; link to a related section
+	sh_entsize_bytes  = 0x28,  -- spec: gABI v1.2 §"Section Header Table" — 40 bytes per entry
+}
+
+-- ---------------------------------------------------------------------------
+-- ELF32 symbol-table entry layout (System V ABI gABI v1.2 §"Symbol Table")
+-- ---------------------------------------------------------------------------
+-- Each entry is 16 bytes (sym_entry_bytes = 0x10 = 16);
+-- zero-based, field offsets relative to the start of the entry.
+M.ELF32_SYM = {
+	st_name          = 0x00,  -- 4-byte LE; offset into the linked string table
+	st_value         = 0x04,  -- 4-byte LE; symbol value (address / absolute)
+	st_size          = 0x08,  -- 4-byte LE; symbol size in bytes
+	st_info          = 0x0C,  -- 1 byte;  binding (high nibble) + type (low nibble)
+	sym_entry_bytes  = 0x10,  -- spec: gABI v1.2 §"Symbol Table" — 16 bytes per entry
+}
+
+-- DWARF32 initial-length terminator (DWARF4 §7.4) — kept here so the metaprogram's elf_dwarf.lua can drop its own copy of the same constant.
+M.dw_dwarf32_terminator = 0xFFFFFFFF
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Adapter validation
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Validate that `adapter` exposes the byte-read surface.
+--- Returns true on success, false + a stable error code on failure.
+--- The helper side calls this before parse_manifest to reject callers before any byte is read.
+--- @param adapter any
+--- @return boolean, string|nil
+function M.validate_adapter(adapter)
+	if type(adapter) ~= "table" then return false, "bad_file_adapter" end
+	if type(adapter.read_u8_at)  ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_u16_at) ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_u32_at) ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_size)   ~= "function" then return false, "bad_file_adapter" end
+	return true, nil
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- String-table reader
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Extract a NUL-terminated C string from `strtab` at zero-based offset `off`.
+--- Returns nil if `off` is out of range or the string is not NUL-terminated.
+--- @param strtab string
+--- @param off integer
+--- @return string|nil
+function M.get_str(strtab, off)
+	if off < 0 or off >= #strtab then return nil end
+	local  end_pos = strtab:find("\0", off + 1, true)
+	if not end_pos then return nil end
+	return strtab:sub(off + 1, end_pos - 1)
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Header / section / symbol walkers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read the ELF32 header through `adapter` and validate the magic, class, and data encoding.
+--- Returns a table on success:
+---   { e_entry, e_shoff, e_shentsize, e_shnum, e_shstrndx, error = nil }
+--- On failure returns nil + a stable error code:
+---   bad_magic, unsupported_elf_class, unsupported_elf_data, truncated_header
+--- The header's machine field is NOT validated here — callers (e.g. the helper's prime path) decide whether to require EM_MIPS before symbol reads.
+--- @param adapter table
+--- @return table|nil, string|nil
+function M.parse_elf32_headers(adapter)
+	local  ok, err = M.validate_adapter(adapter)
+	if not ok then return nil, err end
+
+	-- 4-byte magic: 0x7F 'E' 'L' 'F'.
+	-- The byte readers take the adapter explicitly.
+	-- The production `Support.File` adapter is wrapped by the caller to drop its implicit `self` so the parser shape is flat pass-style.
+	local b1 = M.read_u8(adapter, 0)
+	local b2 = M.read_u8(adapter, 1)
+	local b3 = M.read_u8(adapter, 2)
+	local b4 = M.read_u8(adapter, 3)
+	if not (b1 and b2 and b3 and b4)
+		or not (b1 == 0x7f and b2 == 0x45 and b3 == 0x4c and b4 == 0x46) then
+		return nil, "bad_magic"
+	end
+
+	local class = M.read_u8(adapter, M.ELF32_HEADER.class_offset)
+	if    class ~= M.ELFCLASS32 then
+		return nil, "unsupported_elf_class"
+	end
+
+	local data = M.read_u8(adapter, M.ELF32_HEADER.endian_offset)
+	if    data ~= M.ELFDATA2LSB then
+		return nil, "unsupported_elf_data"
+	end
+
+	local e_entry     = M.read_u32(adapter, M.ELF32_HEADER.e_entry_offset)
+	local e_shoff     = M.read_u32(adapter, M.ELF32_HEADER.e_shoff_offset)
+	local e_shentsize = M.read_u16(adapter, M.ELF32_HEADER.e_shentsize_offset)
+	local e_shnum     = M.read_u16(adapter, M.ELF32_HEADER.e_shnum_offset)
+	local e_shstrndx  = M.read_u16(adapter, M.ELF32_HEADER.e_shstrndx_offset)
+	if not (e_entry and e_shoff and e_shentsize and e_shnum and e_shstrndx) then
+		return nil, "truncated_header"
+	end
+
+	return {
+		e_entry     = e_entry,
+		e_shoff     = e_shoff,
+		e_shentsize = e_shentsize,
+		e_shnum     = e_shnum,
+		e_shstrndx  = e_shstrndx,
+		error       = nil,
+	}
+end
+
+--- Read one section-header entry from `adapter` at `sh_off`.
+--- Returns a table with the wire fields plus a (yet-unresolved) `name` field.
+--- @param adapter table
+--- @param sh_off  integer
+--- @return table|nil, string|nil  -- entry, error
+local function read_section_entry(adapter, sh_off)
+	local entry = {
+		sh_name   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_name_offset),
+		sh_type   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_type_offset),
+		sh_flags  = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_flags_offset),
+		sh_addr   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_addr_offset),
+		sh_offset = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_offset_offset),
+		sh_size   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_size_offset),
+		sh_link   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_link_offset),
+		name      = "",
+	}
+	if not (entry.sh_name and entry.sh_type and entry.sh_flags and entry.sh_addr
+		and entry.sh_offset and entry.sh_size and entry.sh_link) then
+		return nil, "truncated_section_headers"
+	end
+	return entry, nil
+end
+
+--- Walk every section header in `hdr` and return a 1-based array of entries
+--- (the section at logical index 0 is at array position 1, etc.).
+--- Each entry has the wire fields plus a resolved `name` derived from `.shstrtab`.
+--- Returns nil + a stable error code on failure: truncated_section_headers, missing_shstrtab, truncated_strtab
+--- @param adapter table
+--- @param hdr     table  -- the table returned by parse_elf32_headers
+--- @return table|nil, string|nil
+function M.walk_sections(adapter, hdr)
+	if not hdr or hdr.error then return nil, hdr and hdr.error or "truncated_section_headers" end
+
+	local file_size = M.size(adapter)
+	if hdr.e_shoff + hdr.e_shnum * hdr.e_shentsize > file_size then
+		return nil, "truncated_section_headers"
+	end
+
+	-- Read every section header first; we need .shstrtab to resolve names.
+	local sections = {}
+	for i = 0, hdr.e_shnum - 1 do
+		local sh_off = hdr.e_shoff + i * hdr.e_shentsize
+		local entry, err = read_section_entry(adapter, sh_off)
+		if not entry then return nil, err end
+		sections[i + 1] = entry
+	end
+
+	if hdr.e_shstrndx >= hdr.e_shnum then
+		return nil, "missing_shstrtab"
+	end
+
+	local  shstrtab = sections[hdr.e_shstrndx + 1]
+	if not shstrtab or shstrtab.sh_type ~= M.SHT_STRTAB then
+		return nil, "missing_shstrtab"
+	end
+	if shstrtab.sh_offset + shstrtab.sh_size > file_size then
+		return nil, "truncated_section_headers"
+	end
+	local  shstrtab_bytes = M.read_section_bytes(adapter, shstrtab)
+	if not shstrtab_bytes then return nil, "truncated_section_headers" end
+
+	for _, s in ipairs(sections) do
+		s.name = M.get_str(shstrtab_bytes, s.sh_name) or ""
+	end
+
+	return sections, nil
+end
+
+--- Read the bytes of one section. Returns a string, or nil if the adapter returns nil for any byte (out-of-bounds).
+--- The caller is responsible fors sizing the buffer (the section's sh_offset + sh_size must fit in adapter.size).
+--- @param adapter table
+--- @param section table  -- one entry from walk_sections
+--- @return string|nil
+function M.read_section_bytes(adapter, section)
+	local size = section.sh_size
+	if    size == 0 then return "" end
+	local out = {}
+	for i = 0, size - 1 do
+		local b = M.read_u8(adapter, section.sh_offset + i)
+		if    b == nil then return nil end
+		out[#out + 1] = string.char(b)
+	end
+	return table.concat(out)
+end
+
+--- Convenience: walk sections, then look up the named section, then read its bytes.
+--- Returns nil + a stable error code if the section is absent or out-of-bounds.
+--- @param adapter  table
+--- @param sections table  -- 1-based array from walk_sections
+--- @param name     string
+--- @return string|nil, string|nil
+function M.read_named_section(adapter, sections, name)
+	if not sections then return nil, "missing_section" end
+	for _, s in ipairs(sections) do
+		if   s.name == name then
+			local bytes = M.read_section_bytes(adapter, s)
+			if not bytes then return nil, "truncated_section_data" end
+			return bytes, nil
+		end
+	end
+	return nil, "missing_section"
+end
+
+--- Walk every SHT_SYMTAB section in `sections` and accumulate symbols by name.
+--- Each stored entry is `{ value = st_value, size = st_size, info = st_info, shndx = st_shndx }`.
+--- Both STB_LOCAL and STB_GLOBAL symbols are included; the live ELF stores `smem` as a local symbol.
+--- Returns nil + a stable error code on failure: missing_symtab_strtab, truncated_section_headers
+--- @param adapter  table
+--- @param sections table
+--- @return table|nil, string|nil
+function M.collect_symbols(adapter, sections)
+	if not sections then return nil, "missing_sections" end
+	local symbols = {}
+	local file_size = M.size(adapter)
+	for _, s in ipairs(sections) do
+		if s.sh_type == M.SHT_SYMTAB then
+			local  strtab = sections[s.sh_link + 1]
+			if not strtab or strtab.sh_type ~= M.SHT_STRTAB then
+				return nil, "missing_symtab_strtab"
+			end
+			if strtab.sh_offset + strtab.sh_size > file_size then
+				return nil, "truncated_section_headers"
+			end
+			local  strtab_bytes = M.read_section_bytes(adapter, strtab)
+			if not strtab_bytes then return nil, "truncated_section_headers" end
+			if s.sh_offset + s.sh_size > file_size then
+				return nil, "truncated_section_headers"
+			end
+			local  symtab_bytes = M.read_section_bytes(adapter, s)
+			if not symtab_bytes then return nil, "truncated_section_headers" end
+			local n = #symtab_bytes / M.ELF32_SYM.sym_entry_bytes
+			for j = 0, n - 1 do
+				local e = s.sh_offset + j * M.ELF32_SYM.sym_entry_bytes
+				local st_name  = M.read_u32(adapter, e + M.ELF32_SYM.st_name)
+				if    st_name then
+					local st_value = M.read_u32(adapter, e + M.ELF32_SYM.st_value)
+					local st_size  = M.read_u32(adapter, e + M.ELF32_SYM.st_size)
+					local st_info  = M.read_u8(adapter, e + M.ELF32_SYM.st_info)
+					-- st_shndx is at offset 14 (2 bytes) — derived from the layout
+					-- the metaprogram reads too. Inline the read to keep the
+					-- adapter as the only I/O surface.
+					local b1 = M.read_u8(adapter, e + 14)
+					local b2 = M.read_u8(adapter, e + 15)
+					if not (b1 and b2) then
+						return nil, "truncated_section_headers"
+					end
+					local st_shndx = b1 + b2 * 0x100
+					local name = M.get_str(strtab_bytes, st_name) or ""
+					if    name ~= "" then
+						symbols[name] = {
+							value = st_value,
+							size  = st_size,
+							info  = st_info,
+							shndx = st_shndx,
+						}
+					end
+				end
+			end
+		end
+	end
+	return symbols, nil
+end
+
+return M

scripts/elf_dwarf.lua

@@ -1,13 +1,8 @@
 --- elf_dwarf.lua — ELF32 + DWARF + atoms source-map utilities.
---- All ELF32 + DWARF-specific code lives here.
----
 --- **What this module contains:**
 ---   - **Format-constant tables** (the byte-offset / opcode / size encyclopedias for ELF32, DWARF4 aranges, DWARF5 rnglists, DWARF line-program, MIPS).
 ---     Every constant carries a spec:` comment naming the spec section that defines it.
 ---   - **I/O helpers**: little-endian byte read/write, ELF32 section walker, nm symbol reader, source-map parser, native directory glob.
----
---- **Conventions:** tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Native dependencies
@@ -16,6 +11,11 @@
 -- lfs is wired into package.cpath by `duffle_paths.lua` (vendored under `toolchain/lfs/lfs.dll`). 
 local lfs = require("lfs")
 
+-- scripts/elf32.lua contains format-constant tables + the byte-level walker.
+-- The this file re-exports `read_u32_le` / `read_u16_le` (and the DWARF32 terminator).
+-- TODO(Ed): Remove re-export.
+local E = require("elf32")
+
 local M = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -60,19 +60,19 @@ M.DW_AT = {
 }
 
 M.DW_FORM = {
-	addr        = 0x01,
-	data1       = 0x0B,
-	data2       = 0x05,
-	data4       = 0x06,
-	string      = 0x08,
-	strp        = 0x0E,
-	exprloc     = 0x18,
-	ref4        = 0x13,
-	udata       = 0x0F,
-	ref_sig8    = 0x20,
+	addr           = 0x01,
+	data1          = 0x0B,
+	data2          = 0x05,
+	data4          = 0x06,
+	string         = 0x08,
+	strp           = 0x0E,
+	exprloc        = 0x18,
+	ref4           = 0x13,
+	udata          = 0x0F,
+	ref_sig8       = 0x20,
 	implicit_const = 0x21,
-	flag_present = 0x19,
-	sec_offset  = 0x17,
+	flag_present   = 0x19,
+	sec_offset     = 0x17,
 }
 
 M.DW_ATE = {
@@ -104,34 +104,16 @@ M.MIPS_BYTES_PER_WORD = 0x04
 -- ----------------------------------------------------------------------------
 -- ELF32 (System V ABI gABI v1.2)
 -- ----------------------------------------------------------------------------
---- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors,
---- and section-relative values are zero-based wire offsets. Only Lua string APIs receive
---- a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors, and section-relative values are zero-based wire offsets. 
+--- Only Lua string APIs receive a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- ELF/DWARF field offsets are expressed in hex so they map directly to the zero-based byte positions in the binary file.
 ---
---- ELF/DWARF field offsets are expressed in hex so they map directly to the
---- zero-based byte positions in the binary file.
+--- The ELF32 header / section / sym layout tables are within scripts/elf32.lua.
+--- The metaprogram re-exports the DWARF32 initial-length terminator.
 
-
---- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
-M.ELF32 = {
-	magic_offset          = 0x00,         -- 4-byte magic "\127ELF" at file offset 0x00
-	magic                 = "\127ELF",
-	class_offset          = 0x04,         -- 1-byte; 1 = ELF32, 2 = ELF64
-	class_elf32           = 1,
-	endian_offset         = 0x05,         -- 1-byte; 1 = little-endian, 2 = big-endian
-	endian_little         = 1,
-	header_bytes          = 0x34,         -- spec: gABI v1.2 §"ELF Header" — ELF32 header is 52 bytes total
-	e_shoff_offset        = 0x20,         -- 4-byte LE; section-header table file offset
-	e_shentsize_offset    = 0x2E,         -- 2-byte LE; section-header entry size in bytes
-	e_shnum_offset        = 0x30,         -- 2-byte LE; number of section headers
-	e_shstrndx_offset     = 0x32,         -- 2-byte LE; index of section-name string table
-	sh_size_bytes         = 0x28,         -- spec: gABI v1.2 §"Section Header Table" — each entry is 40 bytes
-	sh_name_offset        = 0x00,         -- 4-byte LE; offset into .shstrtab
-	sh_type_offset        = 0x04,         -- 4-byte LE; section type (SHT_*)
-	sh_offset_offset      = 0x10,         -- 4-byte LE; section's file offset
-	sh_size_offset        = 0x14,         -- 4-byte LE; section's size in bytes
-	dw_dwarf32_terminator = 0xFFFFFFFF,   -- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
-}
+--- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
+M.dw_dwarf32_terminator = E.dw_dwarf32_terminator
+-- TODO(Ed): Remove re-export.
 
 -- ----------------------------------------------------------------------------
 -- DWARF4 .debug_aranges (per DWARF5 spec §7.4 — Address Range Table)
@@ -192,8 +174,7 @@ M.DWARF_LINE_OPS = {
 	DW_LNE_end_sequence = 1,    -- spec: §6.2.5.3
 	DW_LNE_set_address  = 2,    -- spec: §6.2.5.3
 	-- Standard opcode header (§6.2.5.1)
-	-- opcode_base + line_range are 1-byte header fields; hex so they map
-	-- directly to their position in the line-program header byte sequence.
+	-- opcode_base + line_range are 1-byte header fields; hex so they map directly to the line-program header byte sequence.
 	-- line_base stays signed decimal (=-5) since 0xFB obscures the spec semantics.
 	opcode_base              = 0x0D,
 	line_base                = -5,
@@ -204,6 +185,44 @@ M.DWARF_LINE_OPS = {
 	set_address_payload_size  = 0x05,   -- size = sub_opcode(1) + addr(4)
 }
 
+-- ----------------------------------------------------------------------------
+-- DWARF5 .debug_line (per DWARF5 spec §6.2.4 — Line Number Program Header)
+-- ----------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets from the start of the unit body
+-- (i.e. AFTER unit_length has been read and unit_length bytes skipped past unit_length's 4 bytes).
+--
+-- The DWARF3/4 line-program format differs:
+--   - It omits `address_size` (DWARF3 §6.2.4) + `segment_selector_size` (DWARF5 §6.2.4).
+--   - It uses null-terminated string lists for `include_directories` + `file_names`
+--     (vs. DWARF5's format_count + fields-list shape).
+-- These are documented inline at each parse site in read_line_unit_file_table below.
+
+--- spec: DWARF5 spec §6.2.4 (Line Number Program Header — version >= 5)
+M.DWARF5_DEBUG_LINE = {
+	-- Header fields (zero-based, AFTER unit_length has been read).
+	version_offset_post_il = 0x00,    -- 2-byte LE; expected = 5
+	addr_size_offset       = 0x02,    -- 1 byte;  expected = 4
+	seg_size_offset        = 0x03,    -- 1 byte;  expected = 0
+	header_length_offset   = 0x04,    -- 4-byte LE; length of program-header content that follows
+	program_header_start   = 0x08,    -- first byte of program-header content (after the 8 fixed bytes)
+
+	-- Per-form byte widths (used when reading directory / file-name entries).
+	form_addr_bytes        = 0x04,    -- DW_FORM_addr (32-bit) | DW_FORM_data4
+	form_strp_bytes        = 0x04,    -- DW_FORM_line_strp / DW_FORM_strp / DW_FORM_strp_sup
+	form_data16_bytes      = 0x10,    -- DW_FORM_data16 (MD5)
+
+	-- DWARF5 form codes (subset used in line-program directory + file tables).
+	form_line_strp         = 0x1A,    -- DWARF5 §7.5.6 — DW_FORM_line_strp (4-byte offset into .debug_line_str)
+	form_string            = 0x08,    -- DWARF4-compatible fallback (inline null-terminated; not in .debug_line_str)
+	form_udata             = 0x0F,    -- DW_FORM_udata (ULEB)
+	form_data16            = 0x18,    -- DW_FORM_data16 (16-byte MD5; gcc emits this for split debug info)
+
+	-- DWARF5 content-tag codes (DW_LNCT_* from §6.2.4.1 + §6.2.4.2).
+	lnct_path              = 0x01,
+	lnct_directory_index   = 0x02,
+	lnct_md5               = 0x05,    -- gcc with MD5 in file name table (rare)
+}
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- I/O helpers: little-endian byte read/write
 -- ════════════════════════════════════════════════════════════════════════════
@@ -211,41 +230,35 @@ M.DWARF_LINE_OPS = {
 --- Read a 4-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- Equivalent to `string.unpack("<I4", buf, off + 1)` but avoids the table-return shape + works under LuaJIT 2.1
 --- (which has partial `string.unpack` coverage).
----
 --- **Convention:** `off` is a zero-based wire offset; `+ 1` is applied only at the `string.byte` boundary.
 ---
---- **Byte weights** are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
---- byte 0 contributes its value directly; byte 1 is shifted left by 8
---- (= 0x100); byte 2 by 16 (= 0x10000); byte 3 by 24 (= 0x1000000).
+--- Thin forwarder: the canonical implementation lives in scripts/elf32.lua.
+--- The "second caller lifts" pattern keeps the metaprogram side fluent
+--- (`M.read_u32_le(buf, off)`) while the body is deduped.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u32_le(buf, off)
-	local byte_off = off + 1
-	return buf:byte(byte_off)
-		+ buf:byte(byte_off + 0x01) * 0x00000100
-		+ buf:byte(byte_off + 0x02) * 0x00010000
-		+ buf:byte(byte_off + 0x03) * 0x01000000
+	return E.read_u32_le(buf, off)
 end
 
 --- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- (`off` is zero-based; `+ 1` is applied only at the `string.byte` boundary.)
+--- Thin forwarder — see `M.read_u32_le` for the rationale.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u16_le(buf, off)
-	local byte_off = off + 1
-	return buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
+	return E.read_u16_le(buf, off)
 end
 
 -- Pure-Lua 5.3 LEB128 readers (no `bit` library). `2^shift` arithmetic matches the existing parser.
 -- Offsets are 0-based; returns (value, next_pos).
--- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua
--- can import them as file-scope locals per the 2nd-caller lift precedent
+-- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua can import them as file-scope locals per the 2nd-caller lift precedent
 -- (the uleb128 + sleb128 encoders were promoted the same way).
 function M.read_uleb128_at(buf, pos)
 	local value, shift = 0, 0
-	local len = #buf
+	local len          = #buf
 	while pos < len do
 		local b = buf:byte(pos + 1)
 		value = value + (b % 0x80) * (2 ^ shift)
@@ -389,13 +402,13 @@ local function read_form_value(buf, str_buf, pos, form)
 		-- The constant is declared in the abbrev; no value bytes in the DIE.
 		return nil, pos
 	elseif form == M.DW_FORM.ref_sig8 then
-		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): an 8-byte value identifying a type
-		-- by signature. The low 4 bytes (LE) are the type signature (content hash);
-		-- the high 4 bytes (LE) are a CU-relative offset into the matching type unit.
-		-- Consumers use the low 4 to look up the type unit (see M.find_type_unit_by_signature)
-		-- then the high 4 to resolve the specific type within it.
-		-- Return the low 4 as the primary value to preserve the (value, next_pos) shape;
-		-- the high 4 is exposed via M.read_ref_sig8 (which returns both halves).
+		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): An 8-byte value identifying a type by signature.
+		-- The low  4 bytes (LE) are the type signature (content hash);
+		-- The high 4 bytes (LE) are a CU-relative offset into the matching type unit.
+		-- Consumers use the low  4 to look up the type unit (see M.find_type_unit_by_signature)
+		--          then the high 4 to resolve the specific type within it.
+		-- Return the low  4 as the primary value to preserve the (value, next_pos) shape;
+		--        the high 4 is exposed via M.read_ref_sig8 (which returns both halves).
 		local _, _, next_pos = M.read_ref_sig8(buf, pos)
 		return M.read_u32_le(buf, pos), next_pos
 	else
@@ -409,34 +422,34 @@ end
 --- should use this directly rather than going through `read_form_value`,
 --- which only exposes the low 4 bytes to preserve its existing (value, next_pos) return shape.
 --- @param buf string
---- @param pos integer  -- zero-based wire offset
---- @return integer  -- low 4 bytes (LE), the type signature
---- @return integer  -- high 4 bytes (LE), the offset within the matching type unit
---- @return integer  -- cursor after the 8-byte value
+--- @param pos integer -- zero-based wire offset
+--- @return integer -- low 4 bytes (LE), the type signature
+--- @return integer -- high 4 bytes (LE), the offset within the matching type unit
+--- @return integer -- cursor after the 8-byte value
 function M.read_ref_sig8(buf, pos)
 	return M.read_u32_le(buf, pos), M.read_u32_le(buf, pos + 4), pos + 8
 end
 
--- DWARF5 §7.5.6 (Type Entries).
--- Walk all units in `info` and return the 0-based offset of the first unit
--- whose `DW_AT_type_signature` (8-byte value at the end of the unit header) equals `target_sig`.
--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
---
--- Unit header layout (from pos 0):
---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
--- @param info string  -- the .debug_info section bytes
--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
+--- DWARF5 §7.5.6 (Type Entries).
+--- Walk all units in `info` and return the 0-based offset of the first unit whose `DW_AT_type_signature`
+--- (8-byte value at the end of the unit header) equals `target_sig`.
+--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
+--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
+---
+--- Unit header layout (from pos 0):
+---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
+---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
+--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
+--- @param info          string      -- the .debug_info section bytes
+--- @param target_sig_lo integer     -- low 4 bytes (LE) of the desired signature
+--- @param target_sig_hi integer     -- high 4 bytes (LE) of the desired signature
+--- @return integer|nil, integer|nil -- unit offset, type_offset within the unit
 function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
-	local pos = 0
+	local pos         = 0
 	local section_len = #info
 	while pos + 4 < section_len do
 		local unit_length = M.read_u32_le(info, pos)
-		if unit_length == 0xFFFFFFFF then
+		if    unit_length == 0xFFFFFFFF then
 			return nil, nil  -- DWARF64 not supported
 		end
 		-- unit_length is the body size, NOT including the 4-byte unit_length field itself.
@@ -446,28 +459,27 @@ function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
 			return nil, nil  -- malformed
 		end
 		-- Per DWARF5 §7.5.6, the type_unit (DW_UT_type = 0x02) body layout is:
-		--   0:  version (2)
-		--   2:  unit_type (1) -- DW_UT_type = 0x02
-		--   3:  address_size (1)
-		--   4:  debug_abbrev_offset (4)
-		--   8:  type_signature (8)
-		--   16: type_offset (4)
-		--   20: <children>
+		--  0:  version (2)
+		--  2:  unit_type (1) -- DW_UT_type = 0x02
+		--  3:  address_size (1)
+		--  4:  debug_abbrev_offset (4)
+		--  8:  type_signature (8)
+		--  16: type_offset (4)
+		--  20: <children>
 		if body_end - body_start >= 20 then
 			-- read_ref_sig8 / write_u32_le / etc. are 1-indexed (string:byte);
-			-- pos / body_start / body_end are 0-based wire offsets, so the
-			-- 1-indexed byte at 0-based wire offset X is string:byte(X + 1).
+			-- pos / body_start / body_end are 0-based wire offsets, so the 1-indexed byte at 0-based wire offset X is string:byte(X + 1).
 			-- Per DWARF5 §7.5.6, the type_unit body is laid out as:
-			--   byte 0-1: version (2)
-			--   byte 2:   unit_type (1) -- DW_UT_type = 0x02
-			--   byte 3:   address_size (1)
-			--   byte 4-7: debug_abbrev_offset (4)
-			--   byte 8-15: type_signature (8)
-			--   byte 16-19: type_offset (4)
+			--  byte 0-1:   version (2)
+			--  byte 2:     unit_type (1) -- DW_UT_type = 0x02
+			--  byte 3:     address_size (1)
+			--  byte 4-7:   debug_abbrev_offset (4)
+			--  byte 8-15:  type_signature (8)
+			--  byte 16-19: type_offset (4)
 			local unit_type = info:byte(body_start + 2 + 1)  -- 0-based +2 = unit_type in 1-indexed
-			if unit_type == 0x02 then  -- DW_UT_type
+			if    unit_type == 0x02 then  -- DW_UT_type
 				local sig_lo, sig_hi, _ = M.read_ref_sig8(info, body_start + 8)  -- 0-based +8 = type_signature in 1-indexed
-				if sig_lo == target_sig_lo and sig_hi == target_sig_hi then
+				if    sig_lo == target_sig_lo and sig_hi == target_sig_hi then
 					local type_offset = M.read_u32_le(info, body_start + 16)  -- 0-based +16 = type_offset in 1-indexed
 					return pos, type_offset
 				end
@@ -514,7 +526,7 @@ end
 --- (we walk all `e_shnum` headers regardless of how many names are requested, to find the .shstrtab first).
 --- For frequent callers, pass the union of all needed sections in one call.
 --  Can add `.debug_info` + `.debug_loc` + `.debug_str_offsets` to the list without writing a 2nd ELF walker.
---- @param elf_path Path
+--- @param elf_path      Path
 --- @param section_names string[]  -- list of section names to read
 --- @return table<string, string>
 function M.read_elf_sections(elf_path, section_names)
@@ -533,75 +545,64 @@ function M.read_elf_sections(elf_path, section_names)
 		return result
 	end
 
-	local f = io.open(elf_path, "rb")
+	local  f = io.open(elf_path, "rb")
 	if not f then
 		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] io.open failed: %s\n", elf_path))
 		return result
 	end
 
-	-- Read the ELF32 header.
-	local header = f:read(M.ELF32.header_bytes)
-	if not header or #header < M.ELF32.header_bytes then
-		io.stderr:write("[elf_dwarf.read_elf_sections] ELF too small for ELF32 header\n")
+	local file_size
+	do
+		f:seek("end", 0)
+		file_size = f:seek("cur", 0)
+	end
+	local adapter = {
+		read_u8_at = function(offset)
+			f:seek("set", offset)
+			local b = f:read(1)
+			if not b then return nil end
+			return b:byte()
+		end,
+		read_u16_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			if not b1 or not b2 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+		end,
+		read_u32_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			local b3 = f:read(1)
+			local b4 = f:read(1)
+			if not b1 or not b2 or not b3 or not b4 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+				+ b3:byte() * 0x10000 + b4:byte() * 0x1000000
+		end,
+		read_size = function() return file_size end,
+	}
+
+	-- Delegate the header parse + section walk to E.*.
+	local hdr, hdr_err = E.parse_elf32_headers(adapter)
+	if not hdr then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] header parse failed: %s\n", tostring(hdr_err)))
 		f:close()
 		return result
 	end
 
-	-- Sanity-check magic + class + endianness.
-	if header:sub(M.ELF32.magic_offset + 1, M.ELF32.magic_offset + 0x04) ~= M.ELF32.magic then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not an ELF file\n")
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.class_offset + 1) ~= M.ELF32.class_elf32 then
-		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] not ELF32 (class=%d)\n", header:byte(M.ELF32.class_offset + 1)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.endian_offset + 1) ~= M.ELF32.endian_little then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not little-endian; unsupported\n")
+	local sections, walk_err = E.walk_sections(adapter, hdr)
+	if not sections then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] section walk failed: %s\n", tostring(walk_err)))
 		f:close()
 		return result
 	end
 
-	-- Parse section-header table location + dimensions from the header.
-	local e_shoff     = M.read_u32_le(header, M.ELF32.e_shoff_offset)
-	local e_shentsize = M.read_u16_le(header, M.ELF32.e_shentsize_offset)
-	local e_shnum     = M.read_u16_le(header, M.ELF32.e_shnum_offset)
-	local e_shstrndx  = M.read_u16_le(header, M.ELF32.e_shstrndx_offset)
-
-	-- Read the section-header string table (.shstrtab) so we can resolve section names from their `sh_name` offsets.
-	f:seek("set", e_shoff + e_shstrndx * e_shentsize)
-	local strtab_hdr = f:read(e_shentsize)
-	if not strtab_hdr or #strtab_hdr < e_shentsize then
-		io.stderr:write("[elf_dwarf.read_elf_sections] could not read .shstrtab header\n")
-		f:close()
-		return result
-	end
-	local strtab_offset = M.read_u32_le(strtab_hdr, M.ELF32.sh_offset_offset)
-	local strtab_size   = M.read_u32_le(strtab_hdr, M.ELF32.sh_size_offset)
-	f:seek("set", strtab_offset)
-	local strtab = f:read(strtab_size) or ""
-
-	-- Walk all section headers; collect (offset, size) for the wanted names.
-	local function read_section_bytes(sh_offset, sh_size)
-		f:seek("set", sh_offset)
-		return f:read(sh_size) or ""
-	end
-
-	for sh_idx = 0, e_shnum - 1 do
-		f:seek("set", e_shoff + sh_idx * e_shentsize)
-		local sh = f:read(e_shentsize)
-		if not sh or #sh < e_shentsize then break end
-		local sh_name   = M.read_u32_le(sh, M.ELF32.sh_name_offset)
-		local sh_offset = M.read_u32_le(sh, M.ELF32.sh_offset_offset)
-		local sh_size   = M.read_u32_le(sh, M.ELF32.sh_size_offset)
-
-		-- Extract the name (null-terminated C string in strtab).
-		local name_end = strtab:find("\0", sh_name + 1, true) or (sh_name + 1)
-		local name     = strtab:sub(sh_name + 1, name_end - 1)
-		if wanted[name] then
-			result[name] = read_section_bytes(sh_offset, sh_size)
+	-- Resolve the requested sections.
+	for _, s in ipairs(sections) do
+		if wanted[s.name] then
+			local bytes = E.read_section_bytes(adapter, s)
+			if bytes then result[s.name] = bytes end
 		end
 	end
 
@@ -610,57 +611,95 @@ function M.read_elf_sections(elf_path, section_names)
 end
 
 --- Read ELF symbol addresses by walking the `.symtab` + `.strtab` sections directly (no `nm` subprocess). 
---- Returns a map `{name -> {addr, size_bytes}}` for every `code_<name>` symbol.
+--- Returns a map `{name -> {addr, size_bytes}}` for every defined symbol.
 ---
 --- **Conventions:**
 --- - ELF32 symtab entry = 16 bytes (`st_name:4 + st_value:4 + st_size:4 + st_info:1 + st_other:1 + st_shndx:2`); offsets within each entry are zero-based wire offsets.
 --- - Direct Lua `string.byte`/`string.sub`/`string.find` boundaries receive `+ 1`.
 --- - We filter on STB_GLOBAL (high nibble of st_info = 1) to match `nm`'s default (external symbols only). STB_WEAK excluded.
---- - The `code_` prefix is stripped (MipsAtom_ macros emit bare atom names, no `code_` prefix).
+--- - Keys are the ELF symbol names as written (the C ident).
 --- - `st_size > 0` filter excludes undefined/imported symbols.
+---
 --- @param elf_path Path
 --- @return table<string, {integer, integer}>
 function M.read_nm(elf_path)
 	local addrs = {}
 
-	-- Read .symtab + .strtab via the existing ELF walker (no subprocess).
-	local sections = M.read_elf_sections(elf_path, {".symtab", ".strtab"})
-	local symtab   = sections[".symtab"]
-	local strtab   = sections[".strtab"]
-	if not symtab or not strtab or #symtab == 0 or #strtab == 0 then
-		-- No symbol table (e.g. stripped ELF). Return empty.
+	-- Existence check first; an empty or missing ELF returns an empty map.
+	if lfs.attributes(elf_path, "mode") ~= "file" then
 		return addrs
 	end
 
-	-- Iterate the 16-byte ELF32 symtab entries.
-	-- Each entry (zero-based): st_name at 0, st_value at 4, st_size at 8, st_info at 12, st_other at 13, st_shndx at 14.
-	local SYM_ENTRY_BYTES = 0x10
-	local SYM_ST_NAME     = 0x00
-	local SYM_ST_VALUE    = 0x04
-	local SYM_ST_SIZE     = 0x08
-	local SYM_ST_INFO     = 0x0C
-	local n_syms = #symtab / SYM_ENTRY_BYTES
-	for i = 0, n_syms - 1 do
-		local entry_off  = i * SYM_ENTRY_BYTES
-		local st_info    = symtab:byte(entry_off + SYM_ST_INFO + 1)
-		-- High nibble = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
-		-- Use math.floor(/16) instead of bit.rshift for LuaJIT 2.1 compat
-		-- (LuaJIT's `>>` is 5.3+, but math.floor(x/16) works on all versions).
-		local binding    = math.floor(st_info / 16)
-		if binding == 0 or binding == 1 then            -- STB_LOCAL or STB_GLOBAL
-			local st_size   = M.read_u32_le(symtab, entry_off + SYM_ST_SIZE)
-			if st_size > 0 then
-				local st_name_off = M.read_u32_le(symtab, entry_off + SYM_ST_NAME)
-				-- Extract the name from .strtab (null-terminated C string).
-				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
-				local name     = strtab:sub(st_name_off + 1, name_end - 1)
-				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — MipsAtom_ macros strip the `code_` prefix).
-				-- The atoms_source_map pass already filters out non-atom symbols via the source-map.txt cross-ref.
-				if name and #name > 0 then
-					local st_value = M.read_u32_le(symtab, entry_off + SYM_ST_VALUE)
-					addrs[name] = { st_value, st_size }
-				end
-			end
+	local f = io.open(elf_path, "rb")
+	if not f then
+		return addrs
+	end
+
+	-- Build the file adapter for E.*.
+	local file_size
+	do
+		f:seek("end", 0)
+		file_size = f:seek("cur", 0)
+	end
+	local adapter = {
+		read_u8_at = function(offset)
+			f:seek("set", offset)
+			local b = f:read(1)
+			if not b then return nil end
+			return b:byte()
+		end,
+		read_u16_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			if not b1 or not b2 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+		end,
+		read_u32_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			local b3 = f:read(1)
+			local b4 = f:read(1)
+			if not b1 or not b2 or not b3 or not b4 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+				+ b3:byte() * 0x10000 + b4:byte() * 0x1000000
+		end,
+		read_size = function() return file_size end,
+	}
+
+	-- Delegate the header + section walk to E.*.
+	local hdr, hdr_err = E.parse_elf32_headers(adapter)
+	if not hdr then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] header parse failed: %s\n", tostring(hdr_err)))
+		f:close()
+		return addrs
+	end
+
+	local sections, walk_err = E.walk_sections(adapter, hdr)
+	if not sections then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] section walk failed: %s\n", tostring(walk_err)))
+		f:close()
+		return addrs
+	end
+
+	-- E.collect_symbols returns every defined symbol (no binding filter).
+	-- The metaprogram then applies its STB_LOCAL / STB_GLOBAL + size>0 filter, matching `nm`'s default (external symbols only).
+	local symbols, sym_err = E.collect_symbols(adapter, sections)
+	if not symbols then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] symbol collection failed: %s\n", tostring(sym_err)))
+		f:close()
+		return addrs
+	end
+
+	f:close()
+
+	for name, entry in pairs(symbols) do
+		-- High nibble of st_info = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
+		-- math.floor(/16) is portable across LuaJIT 2.0/2.1 and plain Lua 5.x.
+		local binding = math.floor(entry.info / 16)
+		if (binding == 0 or binding == 1) and entry.size > 0 then
+			addrs[name] = { entry.value, entry.size }
 		end
 	end
 
@@ -741,8 +780,8 @@ function M.sleb128(n)
 		local b = n % (LEB_DATA_MASK + 1) -- extract low 7 bits
 		n = (n - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
 		-- Termination: remaining value bits fit in the sign bit of the last byte.
-		if     n == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
-		if     n == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if n == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if n == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
 		if more then b = b + LEB_CONT_BIT end
 		bytes[#bytes + 1] = string.char(b)
 	end
@@ -773,20 +812,238 @@ end
 --- @param n integer  -- any integer (negative allowed)
 --- @return integer
 function M.sleb128_size(n)
-	local more = true
+	local more  = true
 	local bytes = 0
-	local v = n
+	local v     = n
 	while more do
 		local b = v % (LEB_DATA_MASK + 1) -- extract low 7 bits
 		v = (v - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
-		if     v == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
-		if     v == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if v == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if v == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
 		if more then b = b + LEB_CONT_BIT end
 		bytes = bytes + 1
 	end
 	return bytes
 end
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- DWARF5 line-program file-table reader
+-- ════════════════════════════════════════════
+
+--- Read every line-program unit in `.debug_line` and produce one entry per file across all units.
+--- Returns three parallel maps keyed by 1-based file index.
+---
+--- Wire format notes:
+---   * The `.debug_line` section may contain MULTIPLE line-program units
+---     File indices are 1-based, **per unit**; we concatenate all units and the index ranges from 1..N₁ in unit 1, N₁+1..N₁+N₂ in unit 2, etc.
+---     Per-unit indices (the way gcc emits them, and the way `DW_LNS_set_file` references them in the line program)
+---     are returned via the `basename_to_index` map only when the unit boundary happens to align with the metaprogram's per-atom `inv.call_file`
+---   * Per spec, the `.debug_line_str` section (DWARF5 §7.5.6) holds the strings referenced by `DW_FORM_line_strp`.
+---     The legacy DWARF3 format embeds strings directly with null terminators. This helper handles BOTH.
+---   * File entries may have multiple forms (gcc -gdwarf-5 with `DW_LNCT_directory_index` emits 2 forms: path + dir_index). 
+---     The helper supports:
+---     - DW_FORM_line_strp (DWARF5; offset into .debug_line_str)
+---     - DW_FORM_string (DWARF4-compat; inline null-terminated in .debug_line)
+---     - DW_FORM_udata (ULEB128)
+---     - DW_FORM_data16 (16-byte MD5; ignored — skip the form's bytes)
+---   * Symlink-canonicalisation: each path's `paths[i]` is stored verbatim from the wire
+---     (mixed `/` and `\` accepted; the basename is taken via the last path separator). Caller normalises as needed.
+---
+--- Behavior on failure: writes to stderr and returns nil.
+--- Helpers consumed by `passes/dwarf_injection.lua::init_file_index_lookup(elf_path)` calls this once at pass start to populate the module-level `basename_to_index` map;
+--- downstream `resolve_provenance_file_index(path)` consumers consult the map directly.
+---
+--- @param elf_path string  -- absolute path to the post-link ELF (typically the gcc-emitted `.elf` BEFORE dwarf_injector's splice; both shapes work since the splice preserves `.debug_line`)
+--- @return table|nil, table|nil, table|nil
+---   basename_to_index: { [basename] = 1-based-per-unit-file-index, ... }
+---   basenames:         { [1-based-per-unit-file-index] = basename, ... }
+---   paths:             { [1-based-per-unit-file-index] = full path (mixed slashes), ... }
+function M.read_line_unit_file_table(elf_path)
+	local sections = M.read_elf_sections(elf_path, { ".debug_line", ".debug_line_str" })
+	local line     = sections[".debug_line"]
+	local lstr     = sections[".debug_line_str"] or ""
+	if not line or line == "" then
+		io.stderr:write("[elf_dwarf.read_line_unit_file_table] no .debug_line section in: " .. tostring(elf_path) .. "\n")
+		return nil
+	end
+
+	local basenames         = {}
+	local basename_to_index = {}
+	local paths             = {}
+
+	--- Read one form-code's bytes from `buf` at position `p` according to `form`.
+	--- Returns (value, after) where `value` is:
+	---   * the resolved string (DW_FORM_line_strp / DW_FORM_string)
+	---   * the ULEB128 number  (DW_FORM_udata)
+	---   * nil + skip-bytes    (DW_FORM_data16; we don't surface the MD5)
+	local function read_form(buf, lstr_buf, p, form)
+		if form == M.DWARF5_DEBUG_LINE.form_line_strp then
+			local strp    = M.read_u32_le(buf, p)
+			local end_pos = lstr_buf:find("\0", strp + 1, true) or (#lstr_buf + 1)
+			return lstr_buf:sub(strp + 1, end_pos - 1), p + M.DWARF5_DEBUG_LINE.form_strp_bytes
+		elseif form == M.DWARF5_DEBUG_LINE.form_string then
+			local nul = buf:find("\0", p + 1, true) or (#buf + 1)
+			return buf:sub(p + 1, nul - 1), nul
+		elseif form == M.DWARF5_DEBUG_LINE.form_udata then
+			local  v, after = M.read_uleb128_at(buf, p)
+			return v, after
+		elseif form == M.DWARF5_DEBUG_LINE.form_data16 then
+			return nil, p + M.DWARF5_DEBUG_LINE.form_data16_bytes
+		else
+			-- Unsupported form in a directory/file-table entry: best-effort skip.
+			-- We do NOT stderr-write because the crt0.s DWARF5 line unit (gcc-as emitted) uses DW_FORM_addr (0x01) for what is effectively a path entry, which is non-standard.
+			-- The C-unit's DWARF3 paths are read via the parallel DWARF3 path and never see this error.
+			-- Callers should consult `basename_to_index` for the paths they care about and ignore this unit if it produced none.
+			return nil, p
+		end
+	end
+
+	--- Parse one DWARF-version-3-style unit (DWARF3/4 line program; gcc default in the PS1 toolchain still emits DWARF3 for line programs in `-g` mode).
+	--- Layout: null-terminated directory list, then path(null) + dir_idx(ULEB) + time(ULEB) + size(ULEB) file entries terminated by an empty null.
+	--- `content_start` = zero-based wire offset of the first byte of program-header content (after version + header_length fields).
+	--- @return unit_basenames { [idx_in_unit_1_based] = basename }
+	--- @return unit_paths     { [idx_in_unit_1_based] = full path }
+	local function parse_dwarf3_unit(buf, content_start, body_end)
+		local up = content_start
+		-- 5 fixed bytes: min_insn, default_is, line_base (signed), line_range, opcode_base
+		up = up + 5
+		local opcode_base = buf:byte(content_start + 5)
+		up = up + (opcode_base - 1)  -- std_opcode_lengths
+		local dirs = {}
+		while up < body_end do
+			local nul = buf:find("\0", up + 1, true) or (body_end + 1)
+			if    nul > body_end then break end
+			local len = nul - up - 1
+			if    len == 0 then up = nul break end
+			dirs[#dirs + 1] = buf:sub(up + 1, nul - 1)
+			up = nul
+		end
+		local unit_basenames = {}
+		local unit_paths = {}
+		while up < body_end do
+			local nul = buf:find("\0", up + 1, true) or (body_end + 1)
+			if    nul > body_end or nul == up + 1 then up = nul break end
+			local path = buf:sub(up + 1, nul - 1)
+			up = nul
+			local didx,  up_next  = M.read_uleb128_at(buf, up); up = up_next
+			local _time, up_next2 = M.read_uleb128_at(buf, up); up = up_next2
+			local _size, up_next3 = M.read_uleb128_at(buf, up); up = up_next3
+			local idx = #unit_basenames + 1
+			local bs  = path:match("[^/\\]+$") or path
+			unit_paths[idx]     = path
+			unit_basenames[idx] = bs
+			dirs[1] = dirs[1] or ""  -- safety: gcc emits "" sentinel dir at 0
+			if didx > 0 and dirs[didx] then
+				unit_paths[idx] = dirs[didx] .. "/" .. path
+			end
+		end
+		return unit_basenames, unit_paths
+	end
+
+	--- Parse one DWARF-version-5-style unit (DWARF5 line program; used by modern gcc with `-gdwarf-5`).
+	--- `content_start` is the first byte of program-header content (after the 8 fixed bytes version+addr_size+seg_size+header_length).
+	--- @return same shape as parse_dwarf3_unit
+	local function parse_dwarf5_unit(buf, lstr_buf, content_start, body_end)
+		local up = content_start
+		-- 6 fixed bytes: min_insn, max_ops_per_insn, default_is, line_base, line_range, opcode_base
+		up = up + 6
+		local opcode_base = buf:byte(content_start + 6)
+		up = up + (opcode_base - 1)  -- std_opcode_lengths
+		-- directories
+		local dir_format_count, after = M.read_uleb128_at(buf, up); up = after
+		local dir_formats = {}
+		for i = 1, dir_format_count do
+			local f, a2 = M.read_uleb128_at(buf, up); up = a2
+			dir_formats[i] = f
+		end
+		local dir_count, a3 = M.read_uleb128_at(buf, up); up = a3
+		local dirs = {}
+		for i = 1, dir_count do
+			local combined = ""
+			for j = 1, dir_format_count do
+				local v, a4 = read_form(buf, lstr_buf, up, dir_formats[j])
+				up = a4
+				if j == 1 and type(v) == "string" then combined = v end
+			end
+			dirs[i] = combined
+		end
+		-- file names
+		local file_format_count, after2 = M.read_uleb128_at(buf, up); up = after2
+		local file_formats = {}
+		for i = 1, file_format_count do
+			local f, a2 = M.read_uleb128_at(buf, up); up = a2
+			file_formats[i] = f
+		end
+		local file_count, a3 = M.read_uleb128_at(buf, up); up = a3
+		local unit_basenames = {}
+		local unit_paths     = {}
+		for i = 1, file_count do
+			local combined = ""
+			local didx = 0
+			for j = 1, file_format_count do
+				local v, a4 = read_form(buf, lstr_buf, up, file_formats[j])
+				up = a4
+				if j == 1 and type(v) == "string" then combined = v end
+				if j == 2 and type(v) == "number"  then didx  = v end
+			end
+			local idx = #unit_basenames + 1
+			local bs  = combined:match("[^/\\]+$") or combined
+			unit_paths[idx]     = combined
+			unit_basenames[idx] = bs
+			if didx > 0 and dirs[didx] then
+				unit_paths[idx] = dirs[didx] .. "/" .. combined
+			end
+		end
+		return unit_basenames, unit_paths
+	end
+
+	--- Walk every line-program unit in the section.
+	local p = 0
+	local section_end = #line
+	while p + 4 <= section_end do
+		local unit_length = M.read_u32_le(line, p)
+		if    unit_length == 0xFFFFFFFF then
+			io.stderr:write("[elf_dwarf.read_line_unit_file_table] 64-bit DWARF (initial-length 0xFFFFFFFF); not supported\n")
+			return nil
+		end
+		local body_start = p + 4
+		local body_end   = p + 4 + unit_length
+		if body_end > section_end then break end
+		local version = M.read_u16_le(line, body_start)
+		local unit_basenames, unit_paths
+		if version >= 5 then
+			-- DWARF5 header: version(2) + addr_size(1) + seg_size(1) + header_length(4) + content
+			local header_length_offset = body_start + 6  -- past version(2) + addr_size(1) + seg_size(1) - wait that's wrong; past hdr len is at +6
+			local content_start = body_start + 8  -- past version(2) + addr_size(1) + seg_size(1) + header_length(4)
+			unit_basenames, unit_paths = parse_dwarf5_unit(line, lstr, content_start, body_end)
+		elseif version >= 2 then
+			-- DWARF2/3/4 header: version(2) + header_length(4) + content
+			local content_start = body_start + 6  -- past version(2) + header_length(4)
+			unit_basenames, unit_paths = parse_dwarf3_unit(line, content_start, body_end)
+		else
+			io.stderr:write(string.format("[elf_dwarf.read_line_unit_file_table] unsupported DWARF version %d (offset 0x%x)\n", version, p))
+			p = body_end
+			goto continue
+		end
+		-- Per-unit 1-based file indices are aligned with `inv.call_file` values because the metaprogram emits `DW_LNS_set_file` with the per-unit index.
+		-- When multiple units are present (crt0.s + C unit), the per-unit index in each unit matches the metaprogram's intent (gcc always sets file in unit-local terms).
+		-- We therefore store directly without global re-indexing; the caller is responsible for knowing which unit the file-index applies to.
+		-- For DWARF3 (C unit is the unit that matters for atom line tables), this matches.
+		-- For DWARF5 (crt0.s + C unit), each carries its own per-unit file-table map;
+		-- the atom-side DW_LNS_set_file(N) refers to the C unit's indices, NOT crt0.s's.
+		-- Since the C unit is the one with full include_directories + 12 entries, we can use it directly.
+		for idx, bs in pairs(unit_basenames) do
+			basenames[idx] = bs
+			paths[idx] = unit_paths[idx]
+			basename_to_index[bs] = idx
+		end
+		p = body_end
+		::continue::
+	end
+
+	return basename_to_index, basenames, paths
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- I/O helpers: atoms source-map + native directory glob
 -- ════════════════════════════════════════════════════════════════════════════

scripts/gdb/gdb_tape_atoms.gdb

@@ -3,53 +3,53 @@
 # Wrapper for the tape-atom step-debug helpers.
 # The 9 user commands are defined here as STUBS (degraded-state messages). 
 # The real implementations + the per-atom data tables are emitted by `passes/atoms_source_map.lua` 
-# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/gen/gdb_tape_atoms_runtime.gdb`.
+# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/gdb_tape_atoms_runtime.gdb`.
 # Sourcing that file RE-DEFINES the commands with real implementations.
 #
-# If `build/gen/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
+# If `build/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
 # The user just needs to re-run `build_psyq.ps1` to regenerate.
 
-# ── Stub commands (defined here so they're always present, even if the runtime file is missing). The runtime file overrides these if sourced. ──
+# ?? Stub commands (defined here so they're always present, even if the runtime file is missing). The runtime file overrides these if sourced. ??
 
 define tape_atoms
-  echo "[gdb_tape_atoms] STUB: runtime file build/gen/gdb_tape_atoms_runtime.gdb not found."
+  echo "[gdb_tape_atoms] STUB: runtime file build/gdb_tape_atoms_runtime.gdb not found."
   echo "[gdb_tape_atoms] STUB: run .\\build_psyq.ps1 to regenerate, then re-source this file."
 end
 document tape_atoms
-  List every tape atom symbol in the loaded ELF (code_<name>) with its .rodata address and word count.
+  List every tape atom symbol in the loaded ELF with its .rodata address and word count.
   STUB state: runtime file not sourced. Run build_psyq.ps1 to regenerate.
 end
 
 define break_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document break_atom
   Set a breakpoint at the start of tape atom <name>. STUB state.
 end
 
 define step_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document step_atom
   Resume execution until the next atom boundary. STUB state.
 end
 
 define next_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document next_atom
   Alias for step_atom. STUB state.
 end
 
 define where_in_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document where_in_atom
   Report current atom name, .rodata addr, word offset, and source line (if known). STUB state.
 end
 
 define stepi_inside_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document stepi_inside_atom
   One MIPS-instruction step, then where_in_atom. STUB state.
@@ -89,17 +89,17 @@ document wave_ctx
 end
 
 
-# ── Source the runtime file (re-defines commands with real impls + data). ──
+# ?? Source the runtime file (re-defines commands with real impls + data). ??
 
 # Try to source from project-root-relative path first (the typical case).
 # If the user is in a different CWD, the source will fail and stubs remain.
-# The runtime file path is computed relative to the ELF's source map convention (build/gen/gdb_tape_atoms_runtime.gdb).
+# The runtime file path is computed relative to the ELF's source map convention (build/gdb_tape_atoms_runtime.gdb).
 echo [gdb_tape_atoms] Wrapper loaded. Sourcing runtime file...
 # Suppress the "Redefine command" prompts that would otherwise appear when the runtime file overrides the 9 stub commands defined above.
-# The runtime's `define` blocks are intended to overwrite — there's no ambiguity to confirm.
+# The runtime's `define` blocks are intended to overwrite ? there's no ambiguity to confirm.
 set confirm off
 
 # Source the runtime file (re-defines commands with real impls + data).
-source build/gen/gdb_tape_atoms_runtime.gdb
+source build/gdb_tape_atoms_runtime.gdb
 set confirm on
 echo [gdb_tape_atoms] Runtime sourced successfully (9 commands now have real implementations).

scripts/launch_pcsx_debug.ps1

@@ -1,10 +1,9 @@
 # scripts/launch_pcsx_debug.ps1
 #
-# One-shot launcher for debug sessions: starts pcsx-redux with the .ps-exe
-# loaded, the gdb stub enabled, AND the pcsx_debug_helper Lua plugin loaded
-# so external CLI tools (gdb's `shell` command, etc.)
-# can read GTE state via http://localhost:8080/api/v1/lua/gte
-# (the gdb stub doesn't expose COP2 at all).
+# One-shot launcher for debug sessions:
+# Starts pcsx-redux with the .ps-exe loaded, the gdb stub enabled, 
+# AND the pcsx_debug_helper Lua plugin loaded so external CLI tools (gdb's `shell` command, etc.)
+# can read GTE state via http://localhost:8080/api/v1/lua/gte (the gdb stub doesn't expose COP2 at all).
 #
 # usage:
 #   .\scripts\launch_pcsx_debug.ps1
@@ -19,11 +18,11 @@
 
 [CmdletBinding()]
 param(
-    [string]$PcsxPath = (Join-Path $PSScriptRoot '..\toolchain\pcsx-redux\vsprojects\x64\Release\pcsx-redux.exe'),
-    [string]$ExePath  = (Join-Path $PSScriptRoot '..\build\hello_gte.ps-exe'),
+    [string]$PcsxPath  = (Join-Path $PSScriptRoot '..\toolchain\pcsx-redux\vsprojects\x64\Release\pcsx-redux.exe'),
+    [string]$ExePath   = (Join-Path $PSScriptRoot '..\build\hello_gte.ps-exe'),
     [string]$HelperZip = (Join-Path $PSScriptRoot 'pcsx_debug_helper.zip'),
-    [int]   $GdbPort  = 3333,
-    [int]   $WebPort  = 8080
+    [int]   $GdbPort   = 3333,
+    [int]   $WebPort   = 8080
 )
 
 $ErrorActionPreference = 'Stop'
@@ -84,7 +83,8 @@ try {
     $r = Invoke-WebRequest -Uri "http://localhost:$WebPort/api/v1/lua/gte" -UseBasicParsing -TimeoutSec 5
     $firstLine = ([System.Text.Encoding]::UTF8.GetString($r.Content) -split "`n")[0]
     Write-Host "GTE handler OK: $firstLine" -ForegroundColor Green
-} catch {
+}
+catch {
     Write-Warning "GTE handler NOT responding: $_"
     Write-Host "Check the pcsx-redux Lua Console for debug cli messages." -ForegroundColor Yellow
 }

scripts/passes/annotation.lua

@@ -7,18 +7,11 @@
 ---
 --- Ownership: the canonical `ctx.shared.corpus` supplies cross-source registries, while each `src.scan` supplies its source's declarations and bodies.
 --- A context without `ctx.shared.corpus` is rejected with an explicit canonical-corpus message.
----
---- Writes `<ctx.out_root>/<dir_basename>.errors.h` once per module, with `#error` directives for findings that the C compile surfaces.
---- `passes/report.lua` renders annotations.txt from `corpus.sources_by_dir`, re-validating each source through `M.validate()`.
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
 
 -- Bootstrap follows the entry scripts; `scripts/duffle_paths.lua` sets package.path and package.cpath. See `ps1_meta.lua` for the rationale.
 -- `debug.getinfo(1, "S").source` locates this file for standalone and orchestrated runs, then `duffle_paths.lua` returns the loaded `duffle` module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local write_file     = duffle.write_file
-local ensure_dir     = duffle.ensure_dir
 
 -- The annotation pass reads the source-derived registries from scan_source:
 --   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
@@ -29,11 +22,11 @@ local ensure_dir     = duffle.ensure_dir
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -52,28 +45,28 @@ local ensure_dir     = duffle.ensure_dir
 --- @field warnings table[]
 
 --- @class AtomAnnotation
---- @field line    integer       -- source line of the atom_info call
---- @field macro   string        -- the macro name (always "atom_info" in the new shape)
---- @field name    string        -- the atom name
---- @field kind    string        -- always "info"
+--- @field line    integer       -- Source line of the atom_info call
+--- @field macro   string        -- Macro name (always "atom_info" in the new shape)
+--- @field name    string        -- Atom name
+--- @field kind    string        -- Always "info"
 --- @field binds   string|nil    -- Binds_X name if any
 --- @field reads   string[]      -- R_* names (read targets)
 --- @field writes  string[]      -- R_* names (write targets)
---- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
+--- @field errors  string[]|nil  -- Parse-time errors from scan_source (atom_info body malformed)
 
---- @class DebugSkipMarker  -- sub-shape of scan_source.lua's @class DebugSkipMarker
---- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
+--- @class DebugSkipMarker  -- Sub-shape of scan_source.lua's @class DebugSkipMarker
+--- @field marker_kind    string                 -- Exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
 --- @field marker_line    integer
---- @field args           string|nil             -- trimmed text inside the parens (nil when has_parens is false)
+--- @field args           string|nil             -- Trimmed text inside the parens (nil when has_parens is false)
 --- @field has_parens     boolean
 --- @field is_bare        boolean                -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
 --- @field pending        boolean                -- true while awaiting the following declaration
---- @field superseded_by_marker_line integer|nil -- set on a marker that was bumped out of the pending slot
+--- @field superseded_by_marker_line integer|nil -- Set on a marker that was bumped out of the pending slot
 --- @field target_kind    string|nil             -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
 
 --- @class Finding
---- @field line integer -- source line (or 0 for pass-level)
---- @field msg  string  -- finding message
+--- @field line integer -- Source line (or 0 for pass-level)
+--- @field msg  string  -- Finding message
 
 --- @class Findings
 --- @field errors   Finding[]
@@ -81,14 +74,14 @@ local ensure_dir     = duffle.ensure_dir
 --- @field info     Finding[]
 
 --- @class PipeCtx
---- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
---- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
---- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
---- @field types          table<string, RegTypeDefault> -- from scan_source
---- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
---- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field atom_index     table<string, AtomAnnotation> -- Name -> AtomAnnotation (only kind=="atom")
+--- @field binds_index    table<string, BindsStruct>    -- Name -> BindsStruct
+--- @field annot_counts   table<string, integer>        -- Name -> annotation count (for unique_annotation check)
+--- @field types          table<string, RegTypeDefault> -- From scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- From scan_source
+--- @field seen_defaults  table<string, integer>        -- Duplicate atom_dbg_reg_default detection
 --- @field seen_field     table<string, integer>        -- Binds_* -> count of fields (set/checked by check_binds_no_duplicate_fields)
---- @field _scan          SourceScan                    -- full scan payload (typed-view sub-calls live here) 
+--- @field _scan          SourceScan                    -- Full scan payload (typed-view sub-calls live here) 
 
 --- @class AnnotatedResult
 --- @field atoms    AtomEntry[]
@@ -102,11 +95,10 @@ local ensure_dir     = duffle.ensure_dir
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-check functions (the CHECK_RULES table's payload)
 -- ════════════════════════════════════════════════════════════════════════════
---
 --- The dispatcher in `validate()` routes each result by convention: existence checks write errors[] and shape checks write warnings[].
 --- `macro_word_drift` writes errors[] for missing or mismatched metadata and info[] for a match.
 
---- Check: every annotated atom must have a matching MipsAtom_(name) declaration.
+--- Check: Every annotated atom must have a matching MipsAtom_(name) declaration.
 --- @param a        AtomAnnotation
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
@@ -119,11 +111,11 @@ local function check_atom_decl_exists(a, pipe_ctx, findings)
 	end
 end
 
---- Check: every atom may have AT MOST ONE annotation.
---- Post-loop: needs full-corpus `annot_counts` from pipe_ctx.
+--- Check:     Every atom may have AT MOST ONE annotation.
+--- Post-loop: Needs full-corpus `annot_counts` from pipe_ctx.
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
-local function check_unique_annotation(pipe_ctx, findings)
+local function check_unique_annotation(_item, pipe_ctx, findings)
 	for name, n in pairs(pipe_ctx.annot_counts) do
 		if n > 1 then
 			findings.errors[#findings.errors + 1] = {
@@ -153,9 +145,10 @@ end
 --- Check: TAPE_WORDS(mac_X, N) ↔ WORD_COUNT(mac_X, N) drift.
 --- Three outcomes: missing (error), mismatch (error), match (info).
 --- @param m  MacroEntry
---- @param wc table<string, integer> -- the shared word-count table (from ctx.shared.word_counts)
+--- @param wc table<string, integer> -- Shared word-count table (from ctx.shared.word_counts)
 --- @param findings Findings
-local function check_macro_word_drift(m, wc, findings)
+local function check_macro_word_drift(m, pipe_ctx, findings)
+	local wc = (pipe_ctx and pipe_ctx.word_counts) or {}
 	local  declared = wc[m.name]
 	if not declared then
 		findings.errors[#findings.errors + 1] = {
@@ -311,7 +304,7 @@ local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
 	end
 end
 
--- Check: debug-skip markers must satisfy shape + placement constraints.
+-- Check: Debug-skip markers must satisfy shape + placement constraints.
 --- Walks the priority list once; each marker produces at most one error, so one source defect yields one finding.
 --- Priority order (first defect wins):
 ---   1. marker_kind ~= "atom_dbg_skip" -> legacy/renamed spelling (use `atom_dbg_skip`)
@@ -322,12 +315,11 @@ end
 ---   6. unsupported target_kind     -> marker precedes an unrelated declaration
 --- Valid markers stamp `debug_skip` on whole-atom, bare-component, and proc-component declaration records in scan_source.lua.
 --- @param marker DebugSkipMarker
---- @param _pipe_ctx PipeCtx     -- unused today; kept for plex-shape consistency with per_annot
+--- @param _pipe_ctx PipeCtx     -- Unused; kept for consistency with per_annot // TODO(Ed): Remove?
 --- @param findings Findings
 local function check_skip_marker(marker, _pipe_ctx, findings)
 	local kind = marker.marker_kind
 	local line = marker.marker_line
-
 	-- Left `scan.debug_skip_markers` with production records for `atom_dbg_skip` only; other identifiers take the walker's unrelated branch.
 
 	if marker.has_parens then
@@ -378,8 +370,6 @@ local function check_skip_marker(marker, _pipe_ctx, findings)
 end
 
 --- Warn when a source references an unregistered alias.
----
---- R_TapePtr, R_AtomJmp, R_PrimCursor, R_FaceCursor, R_VertBase, and R_OtBase opt in through `#define atom_reg` in lottes_tape.h.
 --- When a source uses an unregistered R_X, this check emits one pass-level info entry for that source and directs C-ABI register names to explicit alias registration.
 --- @param _src     SourceFile
 --- @param pipe_ctx PipeCtx
@@ -433,54 +423,30 @@ local CHECK_RULES = {
 -- ════════════════════════════════════════════════════════════════════════════
 -- Validation
 -- ════════════════════════════════════════════════════════════════════════════
---
--- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
+-- Pure check: Read from src.scan, run validations, emit findings. The scan was done once upstream.
 
 --- Builds one pass-wide pipe_ctx from the merged `corpus.*` registries and source-ordered `corpus.atom_infos`; per-source declarations and bodies remain in `src.scan`.
 --- The module ownership contract above requires callers to construct `ctx.shared.corpus` through `build_ctx`; the error message below enforces that gate.
 --- @param ctx PassCtx
 --- @return PipeCtx
 local function build_corpus_pipe_ctx(ctx)
-	local corpus = ctx.shared and ctx.shared.corpus
-	if not corpus then
-		error("annotation requires ctx.shared.corpus "
-			.. "(the canonical corpus is the source of truth; "
-			.. "no per-source fallback is supported)", 0)
-	end
-
-	-- `corpus.atom_infos` preserves source order and duplicates; I precompute counts here for `check_unique_annotation` and the per-source checks.
+	local view = duffle.corpus_view(ctx)
 	local annot_counts = {}
-	for _, info in ipairs(corpus.atom_infos or {}) do
+	for _, info in ipairs(view.atom_infos) do
 		if info and info.atom_name then
 			annot_counts[info.atom_name] = (annot_counts[info.atom_name] or 0) + 1
 		end
 	end
-
-	-- Every consumer of these fields observes mutations via the canonical corpus without independently mutable registry construction.
-	return {
-		-- Cross-source lookup tables from corpus.
-		register_alias_registry  = corpus.register_alias_registry or {},
-		type_name_registry       = corpus.type_name_registry      or {},
-		atom_views               = corpus.atom_views              or {},
-		atom_ctxs                = corpus.atom_ctxs               or {},
-		atom_phases              = corpus.atom_phases             or {},
-		binds_by_name            = corpus.binds_by_name           or {},
-		atoms_by_name            = corpus.atoms_by_name           or {},
-		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
-		atom_infos_list          = corpus.atom_infos              or {},
-		-- Corpus-wide annotation count aggregation (post-rule consumes this).
-		annot_counts             = annot_counts,
-		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
-		collisions               = corpus.collisions              or {},
-		-- `check_macro_word_drift` reads `corpus.word_counts`, populated by word_count_eval.run.
-		word_counts              = corpus.word_counts or {},
-	}
+	view.annot_counts    = annot_counts
+	view.atom_infos_list = view.atom_infos
+	view.word_counts     = ctx.shared.corpus.word_counts or {}
+	return view
 end
 
 --- Validate one source against its pre-scanned SourceScan payload + the corpus-wide pipe_ctx.
 --- @param ctx             PassCtx
 --- @param src             SourceFile
---- @param corpus_pipe_ctx PipeCtx|nil  -- built once per pass from corpus registries; nil builds the same projection here.
+--- @param corpus_pipe_ctx PipeCtx|nil  -- Built once per pass from corpus registries; nil builds the same projection here.
 --- @return AnnotatedResult
 local function validate(ctx, src, corpus_pipe_ctx)
 	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
@@ -489,8 +455,8 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	-- Project the pre-scanned atoms to the AtomEntry shape this pass needs.
 	local atoms = {}
 	for _, a in ipairs(scan.atoms) do
-		if a.kind == "atom" then
-			atoms[#atoms + 1] = { line = a.line, name = a.raw_name }
+		if a.kind == "atom" or a.kind == "atom_proc" then
+			atoms[#atoms + 1] = { line = a.line, name = a.raw_name or a.name }
 		end
 	end
 
@@ -510,14 +476,8 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	end
 
 	-- Build a per-source pipe_ctx: shared lookups come from `corpus_pipe_ctx`, while declarations, bodies, types, views, defaults, and occurrences come from `src.scan`.
-	local seen_defaults = {}
-	for reg, _ in pairs(scan.types or {}) do
-		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
-	end
-	local atom_infos_list = {}
-	for _, ai in ipairs(scan.atom_infos or {}) do
-		atom_infos_list[#atom_infos_list + 1] = ai
-	end
+	local seen_defaults   = {}; for reg, _ in pairs (scan.types      or {}) do seen_defaults[reg] = (seen_defaults[reg] or 0) + 1 end
+	local atom_infos_list = {}; for _, ai  in ipairs(scan.atom_infos or {}) do atom_infos_list[#atom_infos_list + 1] = ai         end
 
 	local pipe_ctx = {
 		atom_index               = {},
@@ -554,38 +514,28 @@ local function validate(ctx, src, corpus_pipe_ctx)
 
 	-- THE per-annotation pipeline. ONE loop. CHECK_RULES dispatches per_annot rules.
 	for _, a in ipairs(annots) do
-		for _, rule in ipairs(CHECK_RULES) do
-			if rule.per_annot then rule.per_annot(a, pipe_ctx, findings) end
-		end
+		duffle.run_check_rules(CHECK_RULES, "per_annot", a, pipe_ctx, findings)
 	end
 
 	-- Post-loop rules (one-shot checks that need full-corpus aggregation in pipe_ctx).
-	for _, rule in ipairs(CHECK_RULES) do
-		if rule.post then rule.post(pipe_ctx, findings) end
-	end
+	duffle.run_check_rules(CHECK_RULES, "post", nil, pipe_ctx, findings)
 
 	-- scan_source records each marker in scan.debug_skip_markers; this loop validates each record independently and emits at most one error per marker.
 	-- Valid markers stamp `debug_skip = true` on the following atom or component declaration, which downstream consumers read directly.
 	local skip_markers = scan.debug_skip_markers or {}
 	for _, marker in ipairs(skip_markers) do
-		for _, rule in ipairs(CHECK_RULES) do
-			if rule.per_skip_marker then rule.per_skip_marker(marker, pipe_ctx, findings) end
-		end
+		duffle.run_check_rules(CHECK_RULES, "per_skip_marker", marker, pipe_ctx, findings)
 	end
 
 	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
-	local wc = corpus_pipe_ctx.word_counts
+	pipe_ctx.word_counts = corpus_pipe_ctx.word_counts
 	for _, m in ipairs(scan.macros) do
-		for _, rule in ipairs(CHECK_RULES) do
-			if rule.per_macro then rule.per_macro(m, wc, findings) end
-		end
+		duffle.run_check_rules(CHECK_RULES, "per_macro", m, pipe_ctx, findings)
 	end
 
 	-- Per-source rules (reg defaults, atom_view layout, compute-register type overrides, Binds_* field uniqueness).
 	-- Each per_source rule sees the full scan payload via pipe_ctx.
-	for _, rule in ipairs(CHECK_RULES) do
-		if rule.per_source then rule.per_source(src, pipe_ctx, findings) end
-	end
+	duffle.run_check_rules(CHECK_RULES, "per_source", src, pipe_ctx, findings)
 
 	-- Information summary (always emitted).
 	findings.info[#findings.info + 1] = {
@@ -605,40 +555,6 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	}
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Per-DIRECTORY (per-module) output: errors.h + annotations.txt
--- ════════════════════════════════════════════════════════════════════════════
-
---- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
---- Empty directories (no errors, no atoms) produce no file.
-local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
-	if atoms_count == 0 and #errors == 0 then
-		return nil
-	end
-	local out_path = ctx.out_root .. "/" .. dir_basename .. ".errors.h"
-	local lines    = {
-		"// Auto-generated by ps1_meta.lua (passes/annotation.lua) — DO NOT EDIT",
-		string.format("// Module: %s   Sources: %d", dir_basename, #sources),
-		"#pragma once",
-		"",
-	}
-	if #errors == 0 then
-		lines[#lines + 1] = "// annotation pass OK"
-	else
-		for _, e in ipairs(errors) do
-			local src_tag = ""
-			if e.source then
-				local src_name = e.source:match("([^/\\]+)$") or e.source
-				src_tag = src_name .. ": "
-			end
-			lines[#lines + 1] = string.format('#error "%s%s (line %d)"', src_tag, e.msg, e.line)
-		end
-	end
-	ensure_dir(ctx.out_root)
-	write_file(out_path, table.concat(lines, "\n") .. "\n")
-	return out_path
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- M.run — orchestrator entry
 -- ════════════════════════════════════════════════════════════════════════════
@@ -683,11 +599,6 @@ function M.run(ctx)
 				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
 			end
 		end
-
-		local err_path = emit_module_errors_h(ctx, dir_basename, dir_atoms, dir_errors, dir_sources)
-		if err_path then
-			table.insert(outputs, { errors_h = err_path })
-		end
 	end
 
 	return { outputs = outputs, errors = errors, warnings = warnings }

scripts/passes/atoms_source_map.lua

@@ -1,23 +1,21 @@
 --- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
 ---
 --- Writer: this pass, given `atom.paths` (the per-atom mutable surface owned by `emission_model`). Readers:
---- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and the gdb-runtime
---- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and 
+--- the gdb-runtime wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
 ---
---- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: one
---- `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
+--- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: 
+---  one `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
 ---
---- **Two output forms** (per the workspace's per-emission-form pattern from `guide_metaprogram_ssdl.md`):
----   1. **Sourcemap.txt form** — `<out_root>/<basename>.atoms.sourcemap.txt`. Format-version-tagged for forward-compat.
----      Lives in `<out_root>/` (build/gen). Mirrors the convention used by `annotation.lua`
----      (`<out_root>/<basename>.errors.h`) and `static_analysis.lua` (`<out_root>/<basename>.static_analysis.txt`).
+--- Two output forms:
+---   1. Markdown form: Handled by `passes/report.lua` (writes `<module>.atoms.md`).
+---      The render functions `render_source_map` + `render_provenance` are exported for `report.lua` to call directly.
 ---      Compile artifacts (`*.macs.h`, `*.offsets.h`) stay in `<source_dir>/gen/`.
----   2. **gdb-runtime form** — `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`. A pure gdb command script — addresses come
----      from `nm`, the 9 user commands are static `define ... end` blocks. Emitted when `ctx.flags.gdb_runtime` is true
----      AND `ctx.flags.elf_path` points to an existing ELF. Useful for `gdb-multiarch --without-python` users
----      (the common case on Windows MinGW builds) — `source <path>` loads it with no Python / Tcl / Guile required.
+---   2. `gdb_tape_atoms_runtime.gdb`: Post-link opt-in (`ctx.flags.gdb_runtime`),
+---      so the gdb wrapper script + the generated runtime script share the same canonical location.
+---      Triggered by `--post-link` or `--gdb-runtime`.
 ---
---- **Output format** (sourcemap.txt form):
+--- Output forma (sourcemap.txt form):
 ---   ```
 ---   # FORMAT_VERSION 1
 ---   # auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT
@@ -30,10 +28,7 @@
 ---   ...
 ---   ENDATOM
 ---   ```
----
 --- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
----
---- **Conventions:** tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -43,8 +38,8 @@
 -- (works both standalone + when require'd). `duffle_paths.lua` sets package.path then returns `require("duffle")` 
 -- at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle    = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local elf_dwarf = require("elf_dwarf")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local elf_dwarf      = require("elf_dwarf")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -73,8 +68,8 @@ local FORMAT_VERSION = 1
 --- @param atom table
 --- @return table[], integer
 local function canonical_word_entries(atom)
-	local paths = atom.paths or {}
-	local events = paths.word_events or {}
+	local paths      = atom.paths or {}
+	local events     = paths.word_events or {}
 	local word_items = {}
 	for _, item in ipairs(paths.items or {}) do
 		if item.kind == "word" then word_items[#word_items + 1] = item end
@@ -88,6 +83,7 @@ local function canonical_word_entries(atom)
 			line       = event.call_line or item.line or 0,
 			text       = event.call_text or item.call_text or "",
 			body_line  = event.body_line or item.body_line or item.line or 0,
+			gpr_keys   = event.gpr_keys,
 			invocation = (event.outermost_invocation_id
 				and paths.invocations
 				and paths.invocations[event.outermost_invocation_id]) or nil,
@@ -99,41 +95,38 @@ end
 --- Render one atom's provenance stanza. Format 1 line shapes:
 ---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>" BODY <line>` (component invocation)
 ---   `WORD N  CALL <src-path>:<src-line>  RAW` (raw `.word` outside any mac_* component)
---- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was
---- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
---- @param src table
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was declared in `corpus.word_counts` 
+--- (populated by word_count_eval + components passes).
+--- @param src  table
 --- @param atom table
---- @param wc table -- identity alias of corpus.word_counts
+--- @param wc   table -- identity alias of corpus.word_counts
 --- @return string[], integer
 local function emit_provenance_stanza(src, atom, wc)
-	local lines    = {}
-	local rel_path = src.path:gsub("\\\\", "/")
+	local lines          = {}
+	local rel_path       = src.path:gsub("\\\\", "/")
 	local entries, total = canonical_word_entries(atom)
-
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
 
 	for _, entry in ipairs(entries) do
-		local inv = entry.invocation
+		local inv         = entry.invocation
 		local macro_count = inv and wc["mac_" .. inv.component_name]
 		if inv and macro_count ~= nil then
-			lines[#lines + 1] = string.format(
-				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
-				entry.pos, rel_path, entry.line, inv.component_name,
-				inv.def_path or "", inv.def_line or 0, entry.body_line)
+			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d'
+				, entry.pos, rel_path, entry.line, inv.component_name
+				, inv.def_path or "", inv.def_line or 0, entry.body_line)
 		else
-			lines[#lines + 1] = string.format(
-				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
 		end
 	end
 
-	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
+	lines[1]          = lines[1]:gsub(" 0$", " " .. tostring(total))
 	lines[#lines + 1] = "ENDATOM"
 	return lines, total
 end
 
 --- Render the full provenance file content for one source.
 --- @param src table
---- @param wc table
+--- @param wc  table
 --- @return string
 local function render_provenance(src, wc)
 	local lines = {}
@@ -166,8 +159,8 @@ end
 --- @param wc     table
 --- @return string[], integer
 local function emit_atom_stanza(src, atom)
-	local lines    = {}
-	local rel_path = src.path:gsub("\\\\", "/")
+	local lines          = {}
+	local rel_path       = src.path:gsub("\\\\", "/")
 	local entries, total = canonical_word_entries(atom)
 
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
@@ -183,10 +176,10 @@ end
 --- Render the full source map file content for one source (one .atoms.sourcemap.txt per source). Mirrors offsets.lua's
 --- `project_atoms` shape: scan.atoms + scan.raw_atoms, no kind filter.
 --- @param src table
---- @param wc table
+--- @param wc  table
 --- @return string
 local function render_source_map(src)
-	local lines = {}
+	local lines       = {}
 	lines[#lines + 1] = "# FORMAT_VERSION " .. FORMAT_VERSION
 	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
 
@@ -220,16 +213,16 @@ end
 --- @param ctx PassCtx
 --- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
 local function build_atom_table(ctx)
-	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
-	local corpus = ctx.shared and ctx.shared.corpus
+	local addrs   = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local corpus  = ctx.shared and ctx.shared.corpus
 	local matched = {}
 
 	for _, src in ipairs(corpus.source_order or {}) do
 		local file_base = src.path:match("([^/\\\\]+)$") or src.path
 		local function append(atom)
 			if not atom.paths then return end
-			local name = atom.raw_name or atom.name
-			local info = addrs[name]
+			local  name = atom.raw_name or atom.name
+			local  info = addrs[name]
 			if not info then return end
 			local entries, total = canonical_word_entries(atom)
 			matched[#matched + 1] = {
@@ -252,9 +245,10 @@ local function build_atom_table(ctx)
 	return matched
 end
 
---- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`, the convenience
---- vars set in `emit_gdb_runtime` provide printf args, and each command is a static sequence of `printf` / `tbreak` /
---- `if ... end` blocks. The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`,
+--- the convenience vars set in `emit_gdb_runtime` provide printf args, and 
+--- each command is a static sequence of `printf` / `tbreak` / `if ... end` blocks.
+--- The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
 ---
 --- Why hardcoded per-atom: gdb's `$` substitution doesn't concat inside var names — `$__atom_name_$__i` in a `while`
 --- loop resolves to one literal identifier, not `name_i`. Compile-time emission is the only path.
@@ -267,12 +261,12 @@ local function append_gdb_commands(lines, matched)
 	for _, a in ipairs(matched) do
 		-- gdb 12.1 quirk: literals in printf args require an attached target.
 		-- Use the per-atom convenience vars set above as printf args.
-		lines[#lines + 1] = string.format('  printf "  code_%%-32s @ 0x%%08x  %%4d words\\n", $__atom_name_%d, $__atom_addr_%d, $__atom_words_%d',
+		lines[#lines + 1] = string.format('  printf "  %%-32s @ 0x%%08x  %%4d words\\n", $__atom_name_%d, $__atom_addr_%d, $__atom_words_%d',
 			a.idx, a.idx, a.idx)
 	end
 	lines[#lines + 1] = "end"
 	lines[#lines + 1] = "document tape_atoms"
-	lines[#lines + 1] = "  List every tape atom symbol in the loaded ELF (code_<name>) with .rodata addr + word count."
+	lines[#lines + 1] = "  List every tape atom symbol in the loaded ELF with .rodata addr + word count."
 	lines[#lines + 1] = "end"
 	lines[#lines + 1] = ""
 
@@ -291,10 +285,10 @@ local function append_gdb_commands(lines, matched)
 	for _, a in ipairs(matched) do
 		lines[#lines + 1] = string.format("define break_atom_%s", a.name)
 		lines[#lines + 1] = string.format("  break *$__atom_addr_%d", a.idx)
-		lines[#lines + 1] = string.format('  printf "  Breakpoint set at code_%s (0x%%08x)\\n", $__atom_addr_%d', a.name, a.idx)
+		lines[#lines + 1] = string.format('  printf "  Breakpoint set at %s (0x%%08x)\\n", $__atom_addr_%d', a.name, a.idx)
 		lines[#lines + 1] = "end"
 		lines[#lines + 1] = string.format("document break_atom_%s", a.name)
-		lines[#lines + 1] = string.format("  Set a breakpoint at code_%s.", a.name)
+		lines[#lines + 1] = string.format("  Set a breakpoint at %s.", a.name)
 		lines[#lines + 1] = "end"
 		lines[#lines + 1] = ""
 	end
@@ -329,7 +323,7 @@ local function append_gdb_commands(lines, matched)
 		-- Precompute end_addr (gdb 12.1's expression evaluator chokes on `addr + words*4`).
 		lines[#lines + 1] = string.format("  set $__end_%d = $__atom_addr_%d + $__atom_words_%d * 4", a.idx, a.idx, a.idx)
 		lines[#lines + 1] = string.format("  if $__pc >= $__atom_addr_%d && $__pc < $__end_%d", a.idx, a.idx)
-		lines[#lines + 1] = string.format('    printf "atom:    code_%%s\\n", $__atom_name_%d', a.idx)
+		lines[#lines + 1] = string.format('    printf "atom:    %%s\\n", $__atom_name_%d', a.idx)
 		lines[#lines + 1] = '    printf "addr:    0x%08x\\n", $__pc'
 		lines[#lines + 1] = string.format("    set $__word = ($__pc - $__atom_addr_%d) / 4", a.idx)
 		lines[#lines + 1] = string.format('    printf "word:    %%d/%%d\\n", $__word, $__atom_words_%d', a.idx)
@@ -399,8 +393,7 @@ local function append_gdb_commands(lines, matched)
 end
 
 --- Emit the gdb-runtime file (post-link). Pure gdb scripting — addresses come from `mipsel-none-elf-nm -S`, get embedded
---- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb
---- source-time.
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb source-time.
 --- @param ctx PassCtx
 local function emit_gdb_runtime(ctx)
 	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
@@ -458,7 +451,22 @@ local function emit_gdb_runtime(ctx)
 	-- Confirmation line for the source operator.
 	lines[#lines + 1] = 'printf "[gdb_tape_atoms] runtime loaded %d atoms from %s\\n", $__atom_count, $__elf_path'
 
-	local out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	local out_path
+	-- Move out of `<out_root>/gdb_tape_atoms_runtime.gdb` to `<out_root>/../gdb_tape_atoms_runtime.gdb` when the conventional `<out_root>` is `<build>/gen`
+	-- (any equivalent spelling — relative, absolute backslash, absolute forward-slash, trailing-separator variants).
+	-- This puts the gdb runtime alongside the ELF at `build/` rather than under the report subdir.
+	local function ends_with_gen_dir(p)
+		if type(p) ~= "string" then return false end
+		return p:match("[/\\]gen[/\\]?$") ~= nil or p == "build/gen" or p == "build\\gen"
+	end
+	if ends_with_gen_dir(ctx.out_root) then
+		-- Strip the trailing `/gen` segment, then write the runtime script under `build/`.
+		-- e.g. "C:/projects/Pikuma/ps1/build/gen" -> "C:/projects/Pikuma/ps1/build".
+		local parent = ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		out_path = parent .. "/gdb_tape_atoms_runtime.gdb"
+	else
+		out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	end
 	duffle.ensure_dir(duffle.dirname(out_path))
 	duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
 	-- io.stderr:write(string.format("[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
@@ -470,10 +478,73 @@ end
 
 local M = {}
 
---- Pass entry. For each source that declares at least one `MipsAtom_(name)` / `MipsCode code_<name>`, emit two files
---- in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
---- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation). When `ctx.flags.gdb_runtime`
---- is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
+-- Expose the pure render functions so `report.lua` and the focused tests can call them directly without triggering the file-emit path.
+M.render_source_map = render_source_map
+M.render_provenance = render_provenance
+
+--- Render ONE atom's sourcemap stanza.
+--- @param atom table   -- atom record (must have `atom.paths` populated)
+--- @return string
+function M.render_atom_source_map(atom)
+	assert(type(atom) == "table", "render_atom_source_map: atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_source_map: atom.paths must be a table")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		local word_line = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+		local keys = {}
+		for pos = 1, 16 do
+			local k = entry.gpr_keys and entry.gpr_keys[pos]
+			if type(k) == "string" and k:sub(1, 7) == "reguse:" then
+				keys[#keys + 1] = k
+			end
+		end
+		if #keys > 0 then
+			word_line = word_line .. "  KEYS " .. table.concat(keys, ",")
+		end
+		lines[#lines + 1] = word_line
+	end
+	lines[#lines + 1] = "ENDATOM"
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render ONE atom's provenance stanza — no per-file format header, no enumeration of other atoms.
+---
+--- `rel_path` is the source path (forward-slashes) embedded in every `CALL` line.
+--- The .md caller (report.lua) is expected to derive this once per `## <source>` heading and pass it down for each atom in that source.
+--- @param atom      table   -- atom record (must have `atom.paths` populated)
+--- @param wc        table   -- identity alias of `corpus.word_counts`
+--- @param rel_path  string  -- source path (forward-slashes) for `CALL` fields
+--- @return string
+function M.render_atom_provenance(atom, wc, rel_path)
+	assert(type(atom)       == "table", "render_atom_provenance:  atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_provenance:  atom.paths must be a table")
+	assert(type(rel_path)   == "string", "render_atom_provenance: rel_path must be a string")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		local inv = entry.invocation
+		local macro_count = inv and wc and wc["mac_" .. inv.component_name]
+		if inv and macro_count ~= nil then
+			lines[#lines + 1] = string.format(
+				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				entry.pos, rel_path, entry.line, inv.component_name,
+				inv.def_path or "", inv.def_line or 0, entry.body_line)
+		else
+			lines[#lines + 1] = string.format(
+				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Pass entry. For each source that declares at least one tape atom,
+--- emit two files in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
+--- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation).
+--- When `ctx.flags.gdb_runtime` is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -495,34 +566,8 @@ function M.run(ctx)
 		}
 	end
 
-	-- Always emit the text form (per-source).
-	for _, src in ipairs(corpus.source_order) do
-		local has_projection = false
-		for _, atom in ipairs((src.scan or {}).atoms or {}) do
-			if (atom.kind == "atom" or atom.kind == "raw_atom") and atom.paths then
-				has_projection = true; break
-			end
-		end
-		if not has_projection then
-			for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
-				if atom.paths then has_projection = true; break end
-			end
-		end
-		if has_projection then
-			local basename = duffle.basename_no_ext(src.path)
-			-- (1) atoms.sourcemap.txt — format-1 per-word call-site map.
-			local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
-			local sourcemap_body = render_source_map(src)
-			-- (2) atoms.provenance.txt — format-1 per-word definition/body map.
-			local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
-			local prov_body = render_provenance(src, wc)
-			duffle.ensure_dir(duffle.dirname(sourcemap_path))
-			duffle.write_file_lf(sourcemap_path, sourcemap_body)
-			duffle.write_file_lf(prov_path,    prov_body)
-			outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
-			outputs[#outputs + 1] = { kind = "report", path = prov_path }
-		end
-	end
+	-- atoms.sourcemap.txt + atoms.provenance.txt content moved to report.lua via `<module>.atoms.md` markdown file.
+	-- This pass emits only the post-link gdb_runtime artifact (see emit_gdb_runtime below).
 
 	-- Optionally emit the gdb-runtime form (post-link, one file per build).
 	if ctx.flags and ctx.flags.gdb_runtime then

scripts/passes/auto_reg.lua

@@ -0,0 +1,343 @@
+--- passes/auto_reg.lua — Per-phase automatic GPR allocator + gen/auto_reg.h emitter.
+---
+--- Reads the per-source + corpus-level `atom_auto_regs` + `phase_auto_regs` registries populated by `passes/scan_source.lua`.
+--- Runs a deterministic first-fit allocator in the `R_T0..R_T7 + R_V0..R_V1` pool (10 physical GPRs).
+--- Emits one `#define R_<Sym>_Code R_Tn_Code` per marker into per-directory `gen/auto_reg.h`.
+---
+--- User-pinned GPRs : The corpus's `register_alias_registry` is consulted to exclude GPRs the user has pinned via 
+--- `atom_reg` + `_Code` defs (e.g. carriers like `R_ResolveScratch = R_T4 atom_reg`). 
+--- These GPRs are unavailable to EVERY atom's source pool.
+--- Carriers are preserved across atoms by context discipline and must never be reallocated.
+--- Per-atom body parsing also catches alias references (R_<Alias>) and hardcoded R_Tn references,
+--- so the user can write either `R_T4` or `R_ResolveScratch` in an atom body and the pass will
+--- exclude R_T4 from that atom's pool.
+---
+--- Conflict detection: If the user hardcodes `R_Tn` in an atom body that shares a phase with an auto-reg that picked `R_Tn`,
+--- emit `phase_register_clash` as an info finding (no build stop).
+--- Should be unreachable after the user-pinning + body-parsing fix above; kept as a defensive safety net.
+---
+--- Pool exhaustion: If a phase declares more `R_<Sym>` mappings than the 10-register pool can hold,
+--- emit `phase_register_pool_exhausted` as a build-stopping error.
+
+--- @class AutoRegResult
+--- @field outputs  table[]  -- {kind=, path=} entries
+--- @field errors   table[]  -- {line=, msg=}  entries (build-stops)
+--- @field warnings table[]  -- {line=, msg=}  entries (build-continues)
+
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- THE GPR ALLOCATION POOL — what's allocatable, and (more importantly) WHY
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- The auto-reg pass picks physical GPRs for `atom_auto_reg(...)` / `phase_auto_reg(...)` markers.
+-- The 24-register pool covers R2-R25 (the user/atom allocatable surface):
+--   R_T0..R_T7, R_V0..R_V1, R_A0..A3, R_S0..S7, R_T8..T9.
+-- Excluded (and never added to the pool):
+--   R_0          (code 0)  — hardwired zero. Cannot be written.
+--   R_AT         (code 1)  — assembler temporary. Reserved by the MIPS O32 ABI.
+--   R_A0..A3                — explicitly omitted above even though their integer codes
+--                              map to POOL entries; the pool-construction loop below
+--                              only references the POOL string literals, never the
+--                              integer codes, so they are NOT auto-allocated by default.
+--                              (A0-A3 become available when the user adds them to
+--                              POOL or hardcodes an R_A0 reference in the atom body.)
+--   R_K0/K1      (codes 26-27) — kernel / interrupt handler reserves. Never touched by user code.
+--   R_GP/SP/FP/RA (codes 28-31) — R_SP/R_FP/R_RA are tape-runtime carriers between
+--                              tape_enter and tape_exit; R_GP stays the host global pointer.
+---
+local POOL = {
+	"R_T0", "R_T1", "R_T2", "R_T3",
+	"R_T4", "R_T5", "R_T6", "R_T7",
+	"R_V0", "R_V1",
+	"R_A0", "R_A1", "R_A2", "R_A3",
+	"R_S0", "R_S1", "R_S2", "R_S3",
+	"R_S4", "R_S5", "R_S6", "R_S7",
+	"R_T8", "R_T9",
+}
+
+-- Map from integer MIPS GPR code (the `code` field on AliasEntry) to the physical GPR ident in POOL.
+-- The standard MIPS O32 ABI register numbering matches mips.h's R_*_Code #defines (mips.h).
+-- Only the POOL entries matter for auto_reg — non-pool aliases
+-- (R_AT=1, R_A0..A3=4..7, R_T8=24, R_T9=25, R_K0/K1=26..27, R_GP/SP/FP/RA=28..31) 
+-- are deliberately omitted — see the comment block above for the WHY of each exclusion.
+local INT_CODE_TO_POOL_GPR = {
+	[2]  = "R_V0", [3]  = "R_V1",
+	[4]  = "R_A0", [5]  = "R_A1", [6]  = "R_A2", [7]  = "R_A3",
+	[8]  = "R_T0", [9]  = "R_T1", [10] = "R_T2", [11] = "R_T3",
+	[12] = "R_T4", [13] = "R_T5", [14] = "R_T6", [15] = "R_T7",
+	[16] = "R_S0", [17] = "R_S1", [18] = "R_S2", [19] = "R_S3",
+	[20] = "R_S4", [21] = "R_S5", [22] = "R_S6", [23] = "R_S7",
+	[24] = "R_T8", [25] = "R_T9",
+}
+
+-- Stable sort for deterministic allocation order.
+local function stable_sort_keys(tbl)
+	local keys = {}
+	for k in pairs(tbl) do keys[#keys + 1] = k end
+	table.sort(keys)
+	return keys
+end
+
+-- Allocate one phase's auto-reg mappings.
+-- Returns (allocated_map, errors). On pool exhaustion, errors is populated and the function halts.
+local function allocate_phase(phase_label, decls)
+	-- Deep-copy POOL into a fresh sequence table. The original `table.unpack and table.unpack(POOL) or { unpack(POOL) }`
+	-- idiom wraps the unpacked values in a single inner table under LuaJIT 5.1 (`table.unpack` is nil; the `or` returns one value),
+	-- which corrupts the pool into `{ {R_T0, R_T1, ...} }` — making `table.remove(pool, 1)` return the inner table on iteration.
+	local pool   = {}
+	for i = 1, #POOL do pool[i] = POOL[i] end
+	local result = {}
+	local errors = {}
+	for _, sym in ipairs(stable_sort_keys(decls)) do
+		local next_gpr = table.remove(pool, 1)
+		if not next_gpr then
+			errors[#errors + 1] = {
+				line = 0,
+				msg  = string.format("phase_register_pool_exhausted: "
+					.. "phase '%s' requested symbol '%s' but the pool has no remaining registers "
+					.. "(max 24 per phase: R_T0..R_T7 + R_V0..R_V1 + R_A0..R_A3 + R_S0..R_S7 + R_T8..R_T9). Split the phase or use hardcoded GPRs."
+					, phase_label, sym),
+			}
+			return result, errors
+		end
+		result[sym] = next_gpr
+	end
+	return result, errors
+end
+
+-- Build two projections from corpus.register_alias_registry:
+--   user_pinned   -- { [physical_gpr_ident] = true } -- GPRs unavailable to auto_reg globally (wave-context carriers, file-scope pinned aliases)
+--   alias_to_gpr  -- { [alias_ident]        = physical_gpr_ident } -- for body parsing
+-- Both projections are derived from the same set of entries: every AliasEntry in register_alias_registry has `has_atom_reg = true` 
+-- (only those entries are added to the registry; see passes/scan_source.lua parse_enum_entry).
+-- Each entry's `code` is the integer MIPS GPR number (0..31); INT_CODE_TO_POOL_GPR translates it back to the physical GPR ident.
+-- Aliases whose `code` points to a non-POOL GPR (e.g. R_S0, R_T8, R_K1) are ignored — 
+-- they don't affect the auto_reg pool, and they're already excluded from POOL above.
+local function build_user_pins(corpus)
+	local user_pinned  = {}
+	local alias_to_gpr = {}
+	if not corpus.register_alias_registry then return user_pinned, alias_to_gpr end
+	for alias_name, alias_entry in pairs(corpus.register_alias_registry) do
+		if alias_entry.has_atom_reg and alias_entry.code then
+			local gpr = INT_CODE_TO_POOL_GPR[alias_entry.code]
+			if gpr then
+				user_pinned[gpr]         = true
+				alias_to_gpr[alias_name] = gpr
+			end
+		end
+	end
+	return user_pinned, alias_to_gpr
+end
+
+-- Find every physical GPR referenced in the atom body, via EITHER:
+--   (a) A hardcoded physical GPR ident (R_T\d+|R_V\d+|R_A\d+|R_S\d+) — the existing regex;
+--   (b) An alias ident (R_<Alias>) resolved via alias_to_gpr back to its physical GPR ident.
+-- Returns { [physical_gpr_ident] = count }. Clash-detection and source-pool-exclusion logic
+-- only needs the presence of each GPR (boolean test), but keeping count preserves the
+-- original find_hardcoded_rn shape so callers can switch without churn.
+-- The alias pattern is sorted lexicographically to keep the regex deterministic.
+local function find_used_gprs(body_text, alias_to_gpr)
+	local found = {}
+	-- (a) Hardcoded physical GPRs (R_T0..R_T7, R_V0..R_V1, R_A0..R_A3, R_S0..R_S7).
+	for gpr in body_text:gmatch("(R_T%d+|R_V%d+|R_A%d+|R_S%d+)") do
+		found[gpr] = (found[gpr] or 0) + 1
+	end
+	-- (b) Alias references (R_<Alias>) resolved to physical GPRs via the registry.
+	--      Sorted by name so the regex is byte-stable across runs.
+	if alias_to_gpr and next(alias_to_gpr) then
+		local aliases = {}
+		for alias_name in pairs(alias_to_gpr) do
+			aliases[#aliases + 1] = alias_name
+		end
+		table.sort(aliases)
+		local pattern = "(" .. table.concat(aliases, "|") .. ")"
+		for alias_name in body_text:gmatch(pattern) do
+			local gpr = alias_to_gpr[alias_name]
+			if gpr and not found[gpr] then
+				found[gpr] = 1
+			end
+		end
+	end
+	return found
+end
+
+-- Emit one gen/auto_reg.h header per directory.
+local function emit_auto_reg_h(out_dir, dir, sources, mappings)
+	if not mappings or next(mappings) == nil then return end
+	local out_path = out_dir .. "/" .. "auto_reg.h"
+	duffle.ensure_dir(out_dir)
+	local lines = {
+		"#ifdef INTELLISENSE_DIRECTIVES",
+		"#pragma once",
+		"#endif",
+		"// Auto-generated by ps1_meta.lua (passes/auto_reg.lua) — DO NOT EDIT",
+		"// Directory: " .. dir:gsub("/", "\\"),
+	}
+	for _, src in ipairs(sources) do
+		lines[#lines + 1] = "//   source: " .. src.path
+	end
+	lines[#lines + 1] = "// Per-phase register allocations resolved by the lua pass."
+	lines[#lines + 1] = "// R_<Sym>_Code = <chosen GPR's _Code constant> for every marker in this directory."
+	lines[#lines + 1] = ""
+	for _, sym in ipairs(stable_sort_keys(mappings)) do
+		local gpr         = mappings[sym]
+		local gpr_code    = gpr .. "_Code"
+		lines[#lines + 1] = "#define " .. sym .. "_Code " .. gpr_code
+	end
+	lines[#lines + 1] = ""
+	duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
+	print("  -> " .. out_path)
+	return out_path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Pass entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+--- @param ctx PassCtx
+--- @return AutoRegResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("auto_reg.run requires ctx.shared.corpus", 0)
+	end
+
+	-- 0. Build the user-pinned GPR exclusion set + alias-to-GPR resolution map.
+	--  Wave-context carriers (e.g. `R_ResolveScratch = R_T4 atom_reg` in hello_camera.atom.c)
+	--  MUST NOT be allocated to any auto-reg marker — they're preserved across atoms by the wave-context discipline.
+	--  The corpus's register_alias_registry is the source of truth for these opt-in pins.
+	--  Body references to those aliases (via alias_to_gpr) are also excluded on a per-atom basis in step 2 below.
+	local user_pinned, alias_to_gpr = build_user_pins(corpus)
+
+	-- 1. Allocate phase pools first (phase declarations take precedence over per-atom declarations).
+	local phase_allocations = {}
+	for phase_label, decls in pairs(corpus.phase_auto_regs or {}) do
+		local mapping, errs = allocate_phase(phase_label, decls)
+		for sym, gpr in pairs(mapping) do
+			phase_allocations[phase_label]      = phase_allocations[phase_label] or {}
+			phase_allocations[phase_label][sym] = gpr
+		end
+		for _, e in ipairs(errs) do
+			errors[#errors + 1] = e
+		end
+	end
+
+	-- 2. Allocate per-atom auto-regs. If the atom scope matches a phase, reuse the phase pool.
+	-- Otherwise, allocate a private pool for the atom.
+	-- The phase membership is in `corpus.atom_phases[phase_label].atoms` (an array of atom names declared via `atom_phase(<phase>)`
+	-- in the atom's `atom_info` line). Build a reverse map `atom_name -> phase_label` so the lookup is O(1) per atom scope.
+	local atom_name_to_phase = {}
+	for phase_label, entry in pairs(corpus.atom_phases or {}) do
+		for _, atom_name in ipairs(entry.atoms or {}) do
+			atom_name_to_phase[atom_name] = phase_label
+		end
+	end
+
+	local atom_allocations = {}
+	for atom_scope, decls in pairs(corpus.atom_auto_regs or {}) do
+		local phase_label = atom_name_to_phase[atom_scope]
+		-- Build the atom's source pool: start with the full POOL, subtract:
+		--   (a) every GPR already committed (phase allocations + prior atom allocations)
+		--   (b) every USER-PINNED GPR (wave-context carriers + file-scope pinned aliases)
+		--   (c) every GPR referenced in the atom's body — either hardcoded R_X or alias R_Xxx
+		--       (the latter resolved via alias_to_gpr; this catches cases where the user wrote R_ResolveScratch instead of R_T4 directly)
+		-- Atoms whose scope matches a phase share the global pool with the phase allocations;
+		-- the original `source_pool = phase_allocations[phase_label]` form used the phase
+		-- allocation MAP as a pool, but that map has no array part, so `table.remove(source_pool, 1)`
+		-- returned nil and every atom-with-phase marker errored with `phase_register_pool_exhausted`.
+		local used = {}
+		for _, m in pairs(phase_allocations) do for _, gpr in pairs(m) do used[gpr] = true end end
+		for _, m in pairs(atom_allocations)  do for _, gpr in pairs(m) do used[gpr] = true end end
+		-- (c) Body references — scan the atom body for hardcoded + alias-resolved GPRs.
+		--     Folded into `used` so the source_pool exclusion is a single check.
+		local atom = corpus.atoms_by_name and corpus.atoms_by_name[atom_scope]
+		if atom and atom.body then
+			local body_used = find_used_gprs(atom.body, alias_to_gpr)
+			for gpr in pairs(body_used) do used[gpr] = true end
+		end
+		local source_pool = {}
+		for _, gpr in ipairs(POOL) do
+			-- Exclude (a) prior commitments, (b) USER-PINNED GPRs (wave-context carriers
+			-- declared via atom_reg + _Code defs, preserved across atoms globally).
+			if not used[gpr] and not user_pinned[gpr] then
+				source_pool[#source_pool + 1] = gpr
+			end
+		end
+		local result = {}
+		for _, sym in ipairs(stable_sort_keys(decls)) do
+			local next_gpr = table.remove(source_pool, 1)
+			if not next_gpr then
+				errors[#errors + 1] = {
+					line = 0,
+					msg  = string.format("phase_register_pool_exhausted: atom '%s' requested symbol '%s' "
+						.. "but no free registers remain in its scope pool."
+						, atom_scope, sym),
+				}
+			else
+				result[sym] = next_gpr
+			end
+		end
+		atom_allocations[atom_scope] = result
+	end
+
+	-- 3. Conflict-with-hardcoded detection (defensive — should be unreachable now).
+	-- The source_pool exclusion in step 2 (b) + (c) already accounts for both user-pinned GPRs
+	-- and body-referenced GPRs (hardcoded R_Tn OR alias R_<Alias>).
+	-- An auto-reg allocation that matched an existing body reference would be impossible by construction.
+	-- This warning is kept as a defensive safety net for cases the body scanner might miss
+	-- (e.g. macros that expand to register references the scanner cannot resolve).
+	-- For each resolved (scope, sym) -> R_Tn mapping, scan the atom body source for used GPRs.
+	for atom_scope, decls in pairs(atom_allocations) do
+		local atom = corpus.atoms_by_name and corpus.atoms_by_name[atom_scope]
+		if atom and atom.body then
+			local used_in_body = find_used_gprs(atom.body, alias_to_gpr)
+			for sym, allocated_gpr in pairs(decls) do
+				if used_in_body[allocated_gpr] and used_in_body[allocated_gpr] > 0 then
+					warnings[#warnings + 1] = {
+						line  = atom.line or 0,
+						msg   = string.format("phase_register_clash: atom '%s' has hardcoded '%s' in its body AND an auto-reg marker '%s' "
+							.. "that was allocated to '%s' (same phase). Resolve by removing the hardcoded reference or renaming the auto-reg."
+							, atom_scope, allocated_gpr, sym, allocated_gpr),
+					}
+				end
+			end
+		end
+	end
+
+	-- 4. Emit per-directory gen/auto_reg.h.
+	-- For each source directory that has atom_auto_regs or phase_auto_regs entries, emit one header.
+	local sources_by_dir = corpus.sources_by_dir or {}
+	for dir, sources in pairs(sources_by_dir) do
+		local per_dir_mappings = {}
+		for _, src in ipairs(sources) do
+			-- Collect every (sym -> gpr) entry that originated from a source in this directory.
+			-- `src.scan.atom_auto_regs` is keyed by ATOM SCOPE NAME; `pairs(t)` iterates KEYS so `scope_name` here is the scope ident (e.g. "cube_g4_face").
+			-- The previous `for _, scan_atom_auto` form silently assigned the VALUE (a `{sym = sym}` table) to the variable,
+			-- which made `atom_allocations[scan_atom_auto]` a table-indexed lookup that never resolved.
+			for scope_name in pairs(src.scan and src.scan.atom_auto_regs or {}) do
+				for sym, gpr in pairs(atom_allocations[scope_name] or {}) do
+					per_dir_mappings[sym] = gpr
+				end
+			end
+			for scope_name in pairs(src.scan and src.scan.phase_auto_regs or {}) do
+				for sym, gpr in pairs(phase_allocations[scope_name] or {}) do
+					per_dir_mappings[sym] = gpr
+				end
+			end
+		end
+		local out_dir  = dir .. "/gen"
+		local out_path = emit_auto_reg_h(out_dir, dir, sources, per_dir_mappings)
+		if out_path then outputs[#outputs + 1] = { auto_reg_h = out_path } end
+	end
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/components.lua

@@ -3,13 +3,15 @@
 --- Ownership: `corpus.word_counts`, `corpus.components`, and `corpus.component_body_index`.
 --- Scanner owns `declaration_comment` and `debug_skip` on each declaration record; this pass projects both forward.
 ---
---- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations,
---- then resolves the function-args string from the preceding `FI_ MipsAtom ac_X(...)` declaration via a backward walk.
+--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations (kind="comp_bare" / "comp_proc"),
+--- then resolves the function-args string from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration via a backward walk.
 ---
---- Emits one `<dir_basename>.macs.h` per source with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
+--- `MipsAtom_Proc_(X, ab, { body })` declarations (kind="atom_proc") are ATOMS, not components, and are deliberately excluded —
+--- the ELF symbol is the C ident. Raw `MipsCode code_*` is leftover, not the atom rule.
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
+--- Emits one `gen/macs.h` per *immediate source directory* with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
+--- All sources inside the same directory contribute to the same file (per-directory aggregation).
+--- The directory itself is the namespace, so the filename does not repeat the module name.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -29,7 +31,7 @@ local duffle          = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- Atom component declaration identifiers.
 local ATOM_COMP_PROC  = "MipsAtomComp_Proc_"
-local MIPS_ATOM       = "MipsAtom"  -- prefix on the function declaration that wraps an AtomComp_Proc_
+local MIPS_ATOM       = "Slice_MipsCode"  -- prefix on the function declaration that wraps an AtomComp_Proc_
 
 -- Component-name prefixes.
 local AC_PREFIX       = "ac_"  -- arg to MipsAtomComp_(ac_X); the X is the atom name
@@ -41,43 +43,44 @@ local MAC_PREFIX_LEN  = 4
 local BYTE_NEWLINE    = 10
 local BYTE_SLASH      = 47
 
--- Source dir basename used as the output `.macs.h` filename.
+-- Output gen subdirectory + filename (per-directory aggregation; the directory name is the namespace).
 local GEN_SUBDIR      = "gen"
+local MACS_FILENAME   = "macs.h"
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
---- @field out_root           string                 -- output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
+--- @field sources            SourceFile[]           -- All source files in the build
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
+--- @field shared             table                  -- Cross-pass shared state
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
+--- @field project_root       string                 -- Project root (e.g. "code/")
+--- @field upstream           table<string, table>   -- Per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @field verbose            boolean                -- log diagnostic info
+--- @field verbose            boolean                -- Log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
---- @field errors   table[]  -- {line=, msg=} entries; build-stops
---- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+--- @field errors   table[]  -- {line=, msg=}  entries; build-stops
+--- @field warnings table[]  -- {line=, msg=}  entries; build-succeeds
 
 --- @class Component
---- @field name       string     -- atom name (without `ac_` prefix)
---- @field body       string     -- brace-delimited body (without the braces)
---- @field args       string|nil -- function-args string (function form only)
---- @field line       integer    -- source line of the declaration
---- @field comment    string|nil -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
---- @field kind       string     -- "comp_bare" | "comp_proc"
---- @field debug_skip boolean    -- mirror of `a.debug_skip` (scanner-owned); true iff a bare `atom_dbg_skip` marker immediately preceded the declaration
+--- @field name       string     -- Atom name (without `ac_` prefix)
+--- @field body       string     -- Brace-delimited body (without the braces)
+--- @field args       string|nil -- Function-args string (function form only)
+--- @field line       integer    -- Source line of the declaration
+--- @field comment    string|nil -- Scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @field kind       string     -- "comp_bare" | "comp_proc"  (atom_proc is NOT a component — see `project_components`)
+--- @field debug_skip boolean    -- Mirror of `a.debug_skip` (scanner-owned); true iff a bare `atom_dbg_skip` marker immediately preceded the declaration
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Local helpers (file I/O + path normalization)
@@ -93,48 +96,21 @@ local M = {}
 -- so this file reads it forward rather than re-walking the source.
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation of the given name.
+--- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation.
 --- Returns the args string (e.g., `"U4 off, U4 code, U1 r, U1 g, U1 b"`) or nil if no function declaration is found.
 ---
---- Convention: function form is
----   `FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
---- We find the LAST occurrence of `"ac_X("` before `before_pos` and extract the args from inside the parens.
---- We then verify the preceding context ends with `MipsAtom`
---- (the function-decl keyword with possible qualifiers between).
+--- After the `sym` arg was dropped from MipsAtomComp_Proc_, the component name
+--- and the args both come from the preceding `FI_ Slice_MipsCode ac_X(args)`
+--- declaration. The shared `duffle.find_function_decl_for` helper does the
+--- backward walk; this function returns just the args.
 ---
 --- @param source     string
---- @param name       string
+--- @param name       string  (retained for signature stability; unused — the walk derives the name)
 --- @param before_pos integer
 --- @return string|nil
 local function find_function_args_for(source, name, before_pos)
-	-- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
-	local name_open  = name .. "("
-	local last_idx   = nil
-	local scan_pos   = 1
-	while true do
-		-- Pass `before_pos + 1` so string.find only returns positions < before_pos + 1
-		-- (string.find's 4th arg `plain` is true; we use the 3rd arg `init` for the upper bound).
-		local found = source:find(name_open, scan_pos, true)
-		if not found or found >= before_pos then break end
-		last_idx = found
-		scan_pos = found + #name_open
-	end
-	if not last_idx then return nil end
-
-	-- Verify the preceding context ends with "MipsAtom" (with possible qualifiers between).
-	local before   = source:sub(1, last_idx - 1)
-	local trimmed  = duffle.trim(before)
-	if trimmed:sub(-#MIPS_ATOM) ~= MIPS_ATOM then
-		-- Preceding context is not a function declaration.
-		return nil
-	end
-
-	local open_paren = last_idx + #name  -- position of "("
-	-- scan: MipsAtom ac_X(
-	local inner      = duffle.read_parens(source, open_paren)
-	-- scan: MipsAtom ac_X(<args>)
-	if not inner then return nil end
-	return inner
+	local _, args_inner = duffle.find_function_decl_for(source, before_pos, #MIPS_ATOM)
+	return args_inner
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -154,6 +130,60 @@ local function extract_arg_names(args_str)
 	for _, tok in ipairs(tokens) do
 		local trimmed = duffle.trim(tok)
 		if trimmed ~= "" then
+			-- Strip trailing block comment (/* ... */) from the token, if present.
+			-- split_top_level_commas only skips block comments at TOP LEVEL (between commas),
+			-- not block comments embedded WITHIN a token between a parameter and a trailing comma.
+			-- Without this strip, the identifier-walk below stops at the `/` of `*/` and returns
+			-- the wrong name (or nothing). See `test_extract_arg_names_handles_trailing_block_comments`.
+			local trimmed_end = #trimmed
+			if trimmed_end >= 2 and trimmed:sub(trimmed_end - 1, trimmed_end) == "*/" then
+				-- Find the matching `/*` that opens the trailing comment.
+				-- Walk back from the `*/` looking for `/*` (whitespace + `/*`).
+				local close_pos = trimmed_end - 1   -- position of the second-to-last char
+				-- Walk back: skip trailing whitespace, then look for the `/*` opener.
+				while close_pos > 1 do
+					local ch = trimmed:sub(close_pos, close_pos)
+					if ch == " " or ch == "\t" or ch == "\n" or ch == "\r" then
+						close_pos = close_pos - 1
+					else
+						break
+					end
+				end
+				-- Now scan back from close_pos for the `/*` opener (slashes are at close_pos-1 and close_pos-2).
+				local opener_pos = nil
+				local scan = close_pos - 3
+				while scan >= 1 do
+					if trimmed:sub(scan, scan + 1) == "/*" then
+						opener_pos = scan
+						break
+					end
+					scan = scan - 1
+				end
+				if opener_pos then
+					-- Truncate everything from opener_pos onwards.
+					trimmed = duffle.trim(trimmed:sub(1, opener_pos - 1))
+				end
+			end
+			if trimmed == "" then goto continue end
+			-- Strip trailing array suffix `[N]` if present.
+			-- Example: `Reg r_data[4]` → identifier is `r_data`, not `4`.
+			trimmed_end = #trimmed
+			if trimmed_end >= 4 and trimmed:sub(trimmed_end, trimmed_end) == "]" then
+				-- Walk back: skip digits, expect `[`.
+				local bracket_pos = trimmed_end - 1
+				while bracket_pos > 1 do
+					local ch = trimmed:sub(bracket_pos, bracket_pos)
+					if ch >= "0" and ch <= "9" then
+						bracket_pos = bracket_pos - 1
+					else
+						break
+					end
+				end
+				if bracket_pos >= 1 and trimmed:sub(bracket_pos, bracket_pos) == "[" then
+					trimmed = duffle.trim(trimmed:sub(1, bracket_pos - 1))
+				end
+			end
+			if trimmed == "" then goto continue end
 			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
 			-- then walk back over the identifier chars (alnum + `_`).
 			local ident_end = #trimmed
@@ -177,12 +207,21 @@ local function extract_arg_names(args_str)
 			ident_start = ident_start + 1
 			local name = trimmed:sub(ident_start, ident_end)
 			if name ~= "" then names[#names + 1] = name end
+			::continue::
 		end
 	end
 	if #names == 0 then return nil end
 	return names
 end
 
+local function formal_arg_names(args_str)
+	local names = extract_arg_names(args_str)
+	if not names then return nil end
+	if names[1] == "ab" then table.remove(names, 1) end
+	if #names == 0 then return nil end
+	return names
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Component projection (read from pre-scanned SourceScan)
 -- ════════════════════════════════════════════════════════════════════════════
@@ -200,7 +239,16 @@ end
 local function project_components(source, scan)
 	local out = {}
 	for _, a in ipairs(scan.atoms) do
+		-- Only `MipsAtomComp_(ac_X)` (kind="comp_bare") and `MipsAtomComp_Proc_(ac_X, ...)` (kind="comp_proc")
+		-- are COMPONENTS — they get inlined via `mac_<name>` aliases inside atom bodies.
+		-- `MipsAtom_Proc_` (kind="atom_proc") is an ATOM (ends with `mac_yield()`); it gets emitted via
+		-- `tb_emit` of the C ident, NOT inlined as a macro. Including `atom_proc` here
+		-- would incorrectly emit `mac_<name>` aliases for atoms, polluting `gen/macs.h`.
+		-- See `docs/duffle_dsl_primer.md` §"mac_* aliases" for the contract.
 		if a.kind == "comp_bare" or a.kind == "comp_proc" then
+			-- Function-args lookup is meaningful for `MipsAtomComp_Proc_` components
+			-- (the macro sits inside `FI_ Slice_MipsCode ac_X(...)`); the alias expansion
+			-- discards the `ab` (atom-builder) arg the same way both forms do.
 			local args = find_function_args_for(source, a.raw_name, a.ident_pos)
 			-- Comment ownership: scan_source.lua stamps `declaration_comment` on the record by walking backward past any associated bare marker.
 			-- The pass reads `declaration_comment` directly.
@@ -212,6 +260,7 @@ local function project_components(source, scan)
 				body_off    = a.body_off,
 				body_tokens = a.body_tokens,
 				args        = args,
+				arg_names   = formal_arg_names(args),
 				comment     = comment,
 				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
 				debug_skip  = a.debug_skip == true,
@@ -231,7 +280,6 @@ end
 --
 -- Skips `//` sequences that are inside string or character literals
 -- (a rough heuristic — sufficient for component bodies which don't have those constructs).
---
 --- @param s string
 --- @return string
 local function convert_line_comments_to_block(s)
@@ -281,6 +329,24 @@ local function strip_mac_prefix(ident)
 	return ident
 end
 
+--- Strip a leading delay marker (`LdSlot_` / `BdSlot_` / `GteDelay_` / `DmaSlot_`)
+--- plus following whitespace and block comments. Returns the remainder, or ""
+--- when the token is only the marker.
+--- `BdSlot_ nop` becomes `nop`. Bare `LdSlot_` becomes "".
+--- @param tok string
+--- @return string
+local function strip_leading_delay_marker(tok)
+	local ident = duffle.read_ident(tok, 1)
+	if not ident or not duffle.DELAY_MARKERS[ident] then return tok end
+	local rest = tok:sub(#ident + 1):match("^%s*(.*)$") or ""
+	while rest:sub(1, 2) == "/*" do
+		local close = rest:find("*/", 3, true)
+		if not close then return "" end
+		rest = rest:sub(close + 2):match("^%s*(.*)$") or ""
+	end
+	return rest
+end
+
 --- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
 --- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
 --- @param name         string -- the component name (without `mac_`)
@@ -299,16 +365,30 @@ local function word_count_rec(name, comp_by_name, wc, cache)
 		for _, t in ipairs(tokens) do
 			local trimmed = t.tok
 			if trimmed ~= "" then
-				local lookup = strip_mac_prefix(duffle.read_ident(trimmed, 1))
-				if lookup and comp_by_name[lookup] then
-					-- It's a `mac_X(...)` call. Recurse.
-					n = n + word_count_rec(lookup, comp_by_name, wc, cache)
-				elseif lookup and wc and wc[lookup] then
-					-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2, nop2 = 2).
-					n = n + wc[lookup]
-				else
-					-- Unrecognized token. Fall back to 1 word.
-					n = n + 1
+				local work = trimmed
+				while true do
+					local marker = duffle.read_ident(work, 1)
+					if marker and duffle.DELAY_MARKERS[marker] then
+						work = strip_leading_delay_marker(work)
+						if work == "" then break end
+					else
+						break
+					end
+				end
+				if work ~= "" then
+					local lookup = strip_mac_prefix(duffle.read_ident(work, 1))
+					if lookup == "atom_label" or lookup == "atom_offset" then
+						-- Pure metaprogram anchors; emit zero words.
+					elseif lookup and comp_by_name[lookup] then
+						-- It's a `mac_X(...)` call. Recurse.
+						n = n + word_count_rec(lookup, comp_by_name, wc, cache)
+					elseif lookup and wc and wc[lookup] then
+						-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2, nop2 = 2).
+						n = n + wc[lookup]
+					else
+						-- Unrecognized token. Fall back to 1 word.
+						n = n + 1
+					end
 				end
 			end
 		end
@@ -339,6 +419,128 @@ local function count_all_components(components, wc)
 	return counts
 end
 
+-- ═══════════════════════════════════════════
+-- Per-component metadata derivation (replaces the hardcoded `M.GP0_MACRO_CONTRIB` + `M.INSTRUCTION_LATENCY[mac_*]` tables that previously lived in `duffle.lua`).
+--
+-- Each `MipsAtomComp_(ac_X) { body }` definition in `code/duffle/lottes_tape.h` is the canonical source.
+-- The `mac_X(...)` macros are GENERATED from these definitions by `emit_component_macros_h` for tape-side composition;
+-- the metaprogram must NEVER walk the generated variants to derive metadata.
+-- Always walk the original `MipsAtomComp_` body via `cc.body_tokens`.
+-- ═══════════════════════════════════════════
+
+--- (internal) Recursive cycle-cost derivation. Sum `latency[ident]` per emitted instruction in the component body,
+--- recursing through nested `mac_*` calls (so `mac_format_g4_color`'s cost = 4 × `mac_pack_color_word`'s cost).
+--- Special rule: `mac_yield`'s cost = 0 (per `lottes_tape.h:125-130` "the runtime cost lands in the next atom's prologue").
+--- @param name         string  -- component bare name (e.g. "yield", "pack_color_word")
+--- @param comp_by_name table<string, Component>
+--- @param latency      table<string, integer>
+--- @param cache        table<string, integer>  -- shared memoization; `-1` sentinel detects cycles
+--- @return integer
+local function cycle_cost_rec(name, comp_by_name, latency, cache)
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1
+	local cc = comp_by_name[name]
+	local n
+	if cc then
+		if name == "yield" then
+			-- mac_yield's cost is 0 by convention (the runtime cost lands in the next atom's prologue).
+			n = 0
+		else
+			n = 0
+			local tokens = cc.body_tokens
+			for _, t in ipairs(tokens) do
+				local trimmed = t.tok
+				if trimmed ~= "" then
+					local ident = duffle.read_ident(trimmed, 1)
+					if ident and ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+						-- Nested `mac_X(...)` call: recurse.
+						local nested = ident:sub(MAC_PREFIX_LEN + 1)
+						n = n + cycle_cost_rec(nested, comp_by_name, latency, cache)
+					else
+						-- Leaf instruction or pseudo-macro.
+						local isa = duffle.instr(ident)
+						local gte = duffle.gte(ident)
+						n = n + ((isa and isa.cycles) or (gte and gte.cycles) or latency[ident] or 1)
+					end
+				end
+			end
+		end
+	else
+		n = 1
+	end
+	cache[name] = n
+	return n
+end
+
+--- (internal) Recursive GP0 prim-buffer contribution. Count `store_word` / `store_half` / `store_byte` 
+--- calls in the component body that target `R_PrimCursor` (these are the RAM-side prim-buffer words the macro contributes), recursing through nested `mac_*` calls.
+--- Only `R_PrimCursor`-targeting stores count. Stores targeting other registers (e.g. `R_OtBase`, heap pointers) are not prim-buffer contributions.
+--- @param name         string
+--- @param comp_by_name table<string, Component>
+--- @param cache        table<string, integer>
+--- @return integer
+local function gp0_contrib_rec(name, comp_by_name, cache)
+	if name:match("^insert_ot_tag") then
+		cache[name] = 0
+		return 0
+	end
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1
+	local cc = comp_by_name[name]
+	local n
+	if cc then
+		n = 0
+		local tokens = cc.body_tokens
+		for _, t in ipairs(tokens) do
+			local trimmed = t.tok
+			if    trimmed ~= "" then
+				local ident = duffle.read_ident(trimmed, 1)
+				if    ident and ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+					-- Nested `mac_X(...)` call: recurse.
+					local nested = ident:sub(MAC_PREFIX_LEN + 1)
+					n = n + gp0_contrib_rec(nested, comp_by_name, cache)
+				elseif ident == "gte_sw" then
+					n = n + 1
+				elseif ident == "store_word" or ident == "store_half" or ident == "store_byte" then
+					if trimmed:find("R_PrimCursor", 1, true)
+						or trimmed:find("O_(Poly_", 1, true)
+						or trimmed:find("r_prim_cursor", 1, true)
+						or trimmed:find("r_primitive_cursor", 1, true)
+						or trimmed:find("r_base", 1, true)
+					then
+						n = n + 1
+					end
+				end
+			end
+		end
+	else
+		n = 0
+	end
+	cache[name] = n
+	return n
+end
+
+--- Compute `cycle_cost` + `gp0_contrib` for every component in `components` in a single pass.
+--- Memoization cache is built ONCE (per source) and shared across both helpers so that
+--- a nested `mac_Y` reference inside a `mac_X` body computes its values once.
+--- @param components Component[]
+--- @param latency    table<string, integer>
+--- @return table<string, {cycle_cost=integer, gp0_contrib=integer}>
+local function compute_components_metadata(components, latency)
+	local comp_by_name = {}
+	for _, cc in ipairs(components) do comp_by_name[cc.name] = cc end
+	local cc_cache  = {}
+	local gc_cache  = {}
+	local out       = {}
+	for _, c in ipairs(components) do
+		out[c.name] = {
+			cycle_cost = cycle_cost_rec(c.name, comp_by_name, latency, cc_cache),
+			gp0_contrib = gp0_contrib_rec(c.name, comp_by_name, gc_cache),
+		}
+	end
+	return out
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-component emit logic
 -- ════════════════════════════════════════════════════════════════════════════
@@ -348,8 +550,8 @@ end
 --- @param s string
 --- @return string[]
 local function split_comment_lines(s)
-	local out  = {}
-	local pos  = 1
+	local out   = {}
+	local pos   = 1
 	local s_len = #s
 	while pos <= s_len do
 		local nl = s:find("\n", pos, true)
@@ -364,12 +566,16 @@ local function split_comment_lines(s)
 end
 
 --- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
+--- For `MipsAtomComp_Proc_` components, the leading `ab` (atom-builder) arg is dropped:
+--- the generated `mac_<name>` macros are inline-expansion aliases for baked atoms; their bodies don't reference `ab`
+--- (the builder is only consumed by the procedural `atombuilder_unroll` line that `MipsAtomComp_Proc_` appends after the body).
+--- Inline callers therefore don't need to thread a builder context.
 --- @param args_str string|nil
 --- @return string
 local function signature_from_args(args_str)
-	local arg_names = extract_arg_names(args_str)
-	if    arg_names and #arg_names > 0 then
-		return table.concat(arg_names, ", ")
+	local names = formal_arg_names(args_str)
+	if names then
+		return table.concat(names, ", ")
 	end
 	return "..."
 end
@@ -383,16 +589,103 @@ local function strip_trailing_continuation(lines)
 	end
 end
 
+--- Classify a token as a "pure delay marker token" (a delay-marker identifier
+--- with no following instruction — only whitespace and/or block comments).
+--- Examples that match:
+---   * `GteDelay_`                               → marker alone
+---   * `GteDelay_  /* RT diagonal: D1 = a.x... */` → marker + block comment
+---   * `GteDelay_  /* RT diagonal: ... */\n\t`    → marker + comment + trailing whitespace
+--- Examples that DO NOT match (these contain a real instruction after the marker
+--- and must be preserved verbatim so the instruction still gets emitted):
+---   * `GteDelay_ nop2`
+---   * `GteDelay_ add_si(r.dst_ptr, r.scratch, dst_offset)`
+---
+--- Why this classification matters: the metaprogram emits tokens separated by `,`
+--- and joins them with `\<newline>` line continuations. After C preprocessor
+--- phase 2 (line splicing), the macro body collapses to a single logical line.
+--- Each delay-marker identifier expands to empty (its definition
+--- `#define GteDelay_ // ...` consumes the `//` line comment during preprocessing
+--- of the definition itself, leaving an empty replacement list). When a token
+--- is purely a delay marker with only a trailing comment, the `,` the metaprogram
+--- normally adds before each token-after-the-first brackets empty content and
+--- produces the syntax error `,,` (`expected expression before ',' token`) at
+--- C compile. The metaprogram therefore emits such tokens WITHOUT the leading
+--- `,` (see `token_skips_leading_comma`) — but the marker + trailing comment
+--- are still emitted verbatim so the annotation is preserved in `gen/macs.h`.
+--- @param tok string  -- a single token from split_top_level_commas (already trimmed at the start, may contain trailing whitespace + block comment)
+--- @return boolean
+local function is_pure_delay_marker_token(tok)
+	local markers = duffle.DELAY_MARKERS
+	if type(markers) ~= "table" then return false end
+
+	-- Identify a leading delay-marker identifier (e.g. `GteDelay_`).
+	local ident_end = 1
+	while ident_end <= #tok do
+		local ch = tok:sub(ident_end, ident_end)
+		if ch:match("[%w_]") then
+			ident_end = ident_end + 1
+		else
+			break
+		end
+	end
+	local ident = tok:sub(1, ident_end - 1)
+	if not markers[ident] then return false end
+
+	-- Walk the remainder: only whitespace and block comments are allowed.
+	local scan = ident_end
+	while scan <= #tok do
+		local ch = tok:sub(scan, scan)
+		if    ch:match("%s") then
+			scan = scan + 1
+		elseif ch == "/" and tok:sub(scan + 1, scan + 1) == "*" then
+			local close = tok:find("*/", scan + 2, true)
+			if not close then return false end
+			scan = close + 2
+		else
+			-- Non-whitespace, non-block-comment content: a real instruction
+			-- follows the marker (e.g. `GteDelay_ nop2`); keep this token intact.
+			return false
+		end
+	end
+	return true
+end
+
+--- Classify a token's "leading comma requirement".
+--- Pure delay-marker tokens (`GteDelay_` / `LdSlot_` / `BdSlot_` / `DmaSlot_`
+--- followed by whitespace + optional block comment and NOTHING ELSE) expand
+--- to empty at C preprocessor time. Emitting them WITHOUT the leading `,`
+--- separator that the metaprogram normally adds before each token after the
+--- first keeps exactly one `,` between the surrounding real expressions in
+--- the spliced macro body:
+---
+---   * before this rule:  `<tok1> ,\t<gdelay> ,\t<tok3>` → after expansion
+---                        `<tok1> ,  /* comment */ , <tok3>` → `,,` syntax error.
+---   * after  this rule:  `<tok1> \t<gdelay> ,\t<tok3>` → after expansion
+---                        `<tok1>  /* comment */ , <tok3>` → `<tok1>, <tok3>` — valid.
+---
+--- Tokens like `GteDelay_ nop2` keep the leading `,` (the marker is followed
+--- by a real instruction, so the marker + instruction together need the
+--- separator on the LEFT to land between two real expressions).
+--- @param tok string
+--- @return boolean  -- true if the token needs NO leading `,` separator.
+local function token_skips_leading_comma(tok)
+	return is_pure_delay_marker_token(tok)
+end
+
 --- Emit the `#define mac_X(sig) \<newline>\t<tok1> \<newline>,\t<tok2> ...` block.
 --- Converts `//` line comments to `/* */` block comments in each token so they don't break the C macro `\` line continuations.
+---
+--- Pure delay-marker tokens (`GteDelay_` / `LdSlot_` / `BdSlot_` / `DmaSlot_` with only a trailing block comment, no real instruction) are emitted WITHOUT a leading `,` separator; the annotation IS preserved in the generated header (so the comment + marker remain visible to anyone reading `gen/macs.h`), but the C preprocessor expands the marker to empty, so leaving the `,` separator out is what stops the `,,` syntax error. See `token_skips_leading_comma` for the contract.
 local function emit_macro_body(lines, c, sig, tokens)
 	for tok_idx = 1, #tokens do
 		tokens[tok_idx] = convert_line_comments_to_block(tokens[tok_idx])
 	end
+	if #tokens == 0 then return end
 	lines[#lines + 1] = "#define mac_" .. c.name .. "(" .. sig .. ") \\"
 	lines[#lines + 1] = "\t" .. tokens[1] .. " \\"
 	for tok_idx = 2, #tokens do
-		lines[#lines + 1] = ",\t" .. tokens[tok_idx] .. " \\"
+		local sep = token_skips_leading_comma(tokens[tok_idx]) and "\t" or ",\t"
+		lines[#lines + 1] = sep .. tokens[tok_idx] .. " \\"
 	end
 	strip_trailing_continuation(lines)
 end
@@ -411,7 +704,7 @@ local function build_component_lines(c, counts)
 
 	-- Marker comment: emitted once for every skipped component.
 	-- The marker is scanner-owned (declared by `atom_dbg_skip` immediately before the declaration in the source);
-	-- the components pass projects `c.debug_skip` and emits the marker as a generated comment.
+	-- This pass projects `c.debug_skip` and emits the marker as a generated comment.
 	if c.debug_skip then
 		lines[#lines + 1] = "/* atom_dbg_skip */"
 	end
@@ -424,9 +717,9 @@ local function build_component_lines(c, counts)
 
 	local tokens = duffle.split_top_level_commas(c.body)
 	for i = 1, #tokens do tokens[i] = duffle.trim(tokens[i]) end
-	local sig    = signature_from_args(c.args)
+	local sig = signature_from_args(c.args)
 	-- Direct lookup against the per-source precomputed `counts` table (built once by count_all_components).
-	local n      = counts[c.name]
+	local n   = counts[c.name]
 
 	if n > 0 then
 		emit_macro_body(lines, c, sig, tokens)
@@ -445,9 +738,15 @@ end
 
 --- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
 --- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
---- @param src SourceFile
+--- @param dir     string       -- Absolute source directory
+--- @param sources SourceFile[] -- Sources contributing to this directory (for the header comment)
 --- @return string[]
-local function header_boilerplate(src)
+local function header_boilerplate(dir, sources)
+	local source_lines = { "// Directory: " .. duffle.to_absolute_path(dir) .. "/" }
+	for _, src in ipairs(sources) do
+		source_lines[#source_lines + 1] = "//   source: " .. duffle.to_absolute_path(src.path)
+	end
+	local source_blob = table.concat(source_lines, "\n")
 	return {
 		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
 		-- The build does manual unity includes (the user controls include order), so the pragma is only active for IDE/tooling.
@@ -455,7 +754,7 @@ local function header_boilerplate(src)
 		"#pragma once",
 		"#endif",
 		"// Auto-generated by ps1_meta.lua — DO NOT EDIT",
-		"// Source: " .. duffle.to_absolute_path(src.path),
+		source_blob,
 		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
 		"",
 		-- Self-contained: define WORD_COUNT if not already defined.
@@ -468,30 +767,30 @@ local function header_boilerplate(src)
 	}
 end
 
---- Compute the output path for one source's `.macs.h` file.
---- The pre-rework convention uses the *directory* basename (not the source file basename)
----  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
---- This matches what the C codebase #includes.
---- @param src SourceFile
---- @return string  -- the output directory
---- @return string  -- the full output path
-local function compute_macs_h_path(src)
-	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
-	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+--- Compute the per-directory output path for `.macs.h`.
+---  e.g. any source in `code/duffle/` produces `code/duffle/gen/macs.h` regardless of source filename.
+--- The directory name is the namespace; the filename does not repeat it.
+--- @param dir string  -- Absolute source directory
+--- @return string  -- Output directory
+--- @return string  -- Full output path
+local function compute_macs_h_path(dir)
+	local  out_dir  = dir .. "/" .. GEN_SUBDIR
+	local  out_path = out_dir .. "/" .. MACS_FILENAME
 	return out_dir, out_path
 end
 
---- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
+--- Emit a per-directory `.macs.h` header with the aggregated `mac_X` macros + `WORD_COUNT` entries.
 --- Writes in BINARY mode so LF line endings are preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
 --- @param ctx        PassCtx
---- @param src        SourceFile
---- @param components Component[]
---- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
---- @return string|nil  -- path to the written file (nil if no components)
-local function emit_component_macros_h(ctx, src, components, counts)
+--- @param dir        string                  -- Absolute source directory
+--- @param sources    SourceFile[]            -- Sources contributing to this directory (for the header comment)
+--- @param components Component[]             -- Aggregated components from all sources in this directory
+--- @param counts     table<string, integer>  -- Precomputed word counts (from count_all_components)
+--- @return string|nil -- Path to the written file (nil if no components)
+local function emit_component_macros_h(ctx, dir, sources, components, counts)
 	if #components == 0 then return nil end
-	local out_dir, out_path = compute_macs_h_path(src)
-	local lines             = header_boilerplate(src)
+	local out_dir, out_path = compute_macs_h_path(dir)
+	local lines             = header_boilerplate(dir, sources)
 
 	for _, c in ipairs(components) do
 		for _, l in ipairs(build_component_lines(c, counts)) do
@@ -526,40 +825,44 @@ local function update_canonical_word_counts(corpus, components, counts)
 end
 
 --- @class ComponentDef
---- @field name       string  -- bare name (without ac_/mac_ prefix)
---- @field line       integer -- definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
---- @field path       string  -- absolute source path of the definition
---- @field kind       string  -- "comp_bare" | "comp_proc"
---- @field debug_skip boolean -- mirror of the scanner-owned `a.debug_skip`; consumers read this directly
+--- @field name       string  -- Bare name (without ac_/mac_ prefix)
+--- @field line       integer -- Definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
+--- @field path       string  -- Absolute source path of the definition
+--- @field kind       string  -- "comp_bare" | "comp_proc"  (atom_proc is NOT a component)
+--- @field debug_skip boolean -- Mirror of the scanner-owned `a.debug_skip`; consumers read this directly
 
 --- (internal) Populate `corpus.components` with this source's components-by-name map.
 --- First declaration wins; later declarations of the same bare name are dropped and recorded as a collision via `corpus.collisions` (kind = "component").
---- The pass does NOT write to `ctx.shared.components` (ownership follows the canonical contract).
---- The `debug_skip` field mirrors the scanner-owned declaration record (`c.debug_skip`).
+--- The pass does NOT write to `ctx.shared.components`.
 --- No parallel skip map is built here; consumers that need the per-component skip state read `corpus.components[name].debug_skip` directly.
---- @param corpus     table  -- the corpus
+--- The `cycle_cost` + `gp0_contrib` fields are populated from `metadata[c.name]` (computed by `compute_components_metadata` against the original `MipsAtomComp_` body).
+--- @param corpus     table   -- the corpus
 --- @param src        SourceFile
 --- @param components Component[]
-local function update_canonical_components(corpus, src, components)
+--- @param metadata    table<string, {cycle_cost=integer, gp0_contrib=integer}>
+local function update_canonical_components(corpus, src, components, metadata)
 	local rel_path = src.path:gsub("\\", "/")
 	for _, c in ipairs(components) do
 		-- Keyed by bare name (e.g. `yield`, `load_tri_indices`).
 		-- The atoms_source_map pass looks up components by bare name from the corpus;
 		-- `mac_` prefix lives at the call-site identifier and is stripped before lookup.
+		local m = metadata and metadata[c.name] or nil
 		if corpus.components[c.name] == nil then
 			corpus.components[c.name] = {
-				name       = c.name,
-				line       = c.line,
-				path       = rel_path,
-				kind       = c.kind or "comp_bare",
-				debug_skip = c.debug_skip == true,
+				name        = c.name,
+				line        = c.line,
+				path        = rel_path,
+				kind        = c.kind or "comp_bare",
+				debug_skip  = c.debug_skip == true,
+				cycle_cost  = m and m.cycle_cost  or nil,
+				gp0_contrib = m and m.gp0_contrib or nil,
 			}
 		else
 			-- A second declaration of the same bare name: record a typed collision so static-analysis + the report can surface it.
 			-- Identical-shape declarations (same path + line) reuse the first-wins entry without a collision record.
 			local existing = corpus.components[c.name]
 			if existing.path ~= rel_path or existing.line ~= c.line then
-				local kind     = c.kind or "comp_bare"
+				local kind       = c.kind or "comp_bare"
 				local first_kind = existing.kind or "comp_bare"
 				corpus.collisions[#corpus.collisions + 1] = {
 					kind              = "component",
@@ -592,6 +895,7 @@ local function update_canonical_component_body_index(corpus, src, components, sc
 				source      = src.path,
 				declaration = c.line,
 				kind        = c.kind,
+				arg_names   = c.arg_names,
 			}
 		end
 	end
@@ -624,21 +928,39 @@ function M.run(ctx)
 	--   * `corpus.component_body_index[name]` — body / line_of / source index
 	-- The pass writes to the corpus only; consumers read from the corpus directly.
 
-	for _, src in ipairs(corpus.source_order) do
-		-- project_components reads from src.scan + does backward lookups on src.text
-		local components = project_components(src.text, src.scan)
-		if #components > 0 then
-			-- Compute all component word counts once per source.
-			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
-			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
-			local counts = count_all_components(components, corpus.word_counts)
-			local macs_path = emit_component_macros_h(ctx, src, components, counts)
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/macs.h`.
+	-- The directory itself is the namespace. `corpus.sources_by_dir` preserves source-order within each bucket (matches `corpus.source_order`).
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
+	for dir, sources in pairs(sources_by_dir) do
+		-- Aggregate components from every source in this directory.
+		-- `project_components` returns nil for sources with no `MipsAtomComp_` declarations; we skip those.
+		local aggregated_components = {}
+		local metadata_per_source   = {}
+		for _, src in ipairs(sources) do
+			local per_source = project_components(src.text, src.scan) or {}
+			for _, c in ipairs(per_source) do
+				aggregated_components[#aggregated_components + 1] = c
+			end
+			if #per_source > 0 then
+				metadata_per_source[src] = compute_components_metadata(per_source, {})
+			end
+		end
+		if #aggregated_components > 0 then
+			-- Compute word counts across the aggregated set. `corpus.word_counts` carries the
+			-- same-source + prior-directory entries so the recursive lookup sees both.
+			local counts    = count_all_components(aggregated_components, corpus.word_counts)
+			local macs_path = emit_component_macros_h(ctx, dir, sources, aggregated_components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				-- Populate the projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
-				update_canonical_word_counts(corpus, components, counts)
-				update_canonical_components(corpus, src, components)
-				update_canonical_component_body_index(corpus, src, components, src.scan)
+				-- Populate the projections AFTER disk emission (byte-identical `.macs.h` contract).
+				update_canonical_word_counts(corpus, aggregated_components, counts)
+				for _, src in ipairs(sources) do
+					local per_source = project_components(src.text, src.scan) or {}
+					if #per_source > 0 then
+						update_canonical_components(corpus, src, per_source, metadata_per_source[src])
+						update_canonical_component_body_index(corpus, src, per_source, src.scan)
+					end
+				end
 			end
 		end
 	end

scripts/passes/dwarf_injection.lua

@@ -2,7 +2,7 @@
 ---
 --- Reads the post-link ELF directly (io.open; walks the ELF32 section header table to find
 --- `.debug_info` + `.debug_abbrev` + `.debug_str` + `.debug_line` + `.debug_aranges` + `.debug_rnglists`),
---- APPENDS synthetic DWARF line-program sequences for every `code_<name>` atom, EXTENDS the `.debug_aranges`
+--- APPENDS synthetic DWARF line-program sequences for every tape atom, EXTENDS the `.debug_aranges`
 --- and main-CU range tables with the atom ranges, and INSERTS synthetic atom/component DIE children into the
 --- existing main compilation unit in `.debug_info` (no second compilation unit).
 --- Per-atom `DW_TAG_subprogram` + per-register `DW_TAG_variable` entries make
@@ -71,9 +71,11 @@ local DW_LNE_set_address  = DWARF_LINE_OPS.DW_LNE_set_address
 local DW_RLE_end_of_list   = DWARF5_RNGLISTS.end_of_list
 local DW_RLE_start_length  = DWARF5_RNGLISTS.start_length
 
--- File index 11 in the existing main line unit is hello_gte_tape.c.
--- The injector extends that unit rather than appending an unreferenced unit.
-local ATOM_SOURCE_FILE_INDEX = 11
+-- File-index lookup for the existing main line unit (Unit 2).
+-- Populated at pass start by `init_file_index_lookup(elf_path)` from the runtime ELF (see `elf_dwarf.read_line_unit_file_table`).
+local _file_index_by_basename    = nil  -- [basename] = 1-based line-table file index
+local _file_path_by_index        = nil  -- [1-based index] = full source path (diagnostics / future consumers)
+local _default_atom_source_index = nil  -- any valid index used in opaque-row fallbacks
 
 -- RR_<R_Name> debug-visible variables come from the merged register_alias_registry filtered to aliases whose code is a valid MIPS GPR 0..31
 -- (see collect_per_source_registries + by_alias in build_inserted_children).
@@ -98,16 +100,76 @@ local ABBREV_INLINED_SUBROUTINE  = 0x6C   -- 108: DW_TAG_inlined_subroutine with
 -- (each field transitions from tape memory to GPR at load_pc + 8 = MIPS I load-delay slot boundary).
 local ABBREV_BIND_VAR_LOCLIST = 0x6D   -- 109: DW_TAG_variable no children + DW_AT_type = ref4 + DW_AT_location = sec_offset
 -- Typed-view pointer_type (for the synthetic V4_S2* / V3_S2* / U4* / void* chains).
--- MUST be a fresh abbrev code in the appended table — emitting uleb128(9) collides with GCC's
--- existing abbrev 9 (a pointer_type that carries DW_AT_byte_size + DW_AT_type), so gdb misparses
--- our 4-byte ref4 as (byte_size, type[0..2]) and lands the cursor mid-attribute.
+-- MUST be a fresh abbrev code in the appended table — emitting uleb128(9) collides with GCC's existing abbrev 9
+-- (a pointer_type that carries DW_AT_byte_size + DW_AT_type), so gdb misparses our 4-byte ref4 as (byte_size, type[0..2]) and lands the cursor mid-attribute.
 local ABBREV_TYPED_VIEW_POINTER = 0x6E   -- 110: DW_TAG_pointer_type no children + DW_AT_type = ref4 (typed-view / U4 / void chain)
 
+-- One row per DIE kind build_inserted_children emits.
+-- attrs list form + the value key filled from the atom / registry / local table.
+local DIE_SCHEMA = {
+	base_type = {
+		abbrev = ABBREV_BASE_TYPE,
+		attrs  = {
+			{ form = "string", key = "name" },
+			{ form = "data1",  key = "byte_size" },
+			{ form = "data1",  key = "encoding" },
+		},
+	},
+	abstract_subprogram = {
+		abbrev = ABBREV_ABSTRACT_SUBPROGRAM,
+		attrs  = {
+			{ form = "string", key = "name" },
+			{ form = "data1",  key = "inline" },
+			{ form = "data1",  key = "external" },
+			{ form = "udata",  key = "decl_file" },
+			{ form = "udata",  key = "decl_line" },
+		},
+	},
+	subprogram = {
+		abbrev = ABBREV_SUBPROGRAM,
+		attrs  = {
+			{ form = "string", key = "name" },
+			{ form = "addr",   key = "low_pc" },
+			{ form = "addr",   key = "high_pc" },
+			{ form = "string", key = "linkage_name" },
+		},
+	},
+	variable = {
+		abbrev = ABBREV_VARIABLE,
+		attrs  = {
+			{ form = "string",     key = "name" },
+			{ form = "ref4",       key = "type" },
+			{ form = "exprloc",    key = "location" },
+		},
+	},
+	structure_type = {
+		abbrev = ABBREV_STRUCT_TYPE,
+		attrs  = {
+			{ form = "string", key = "name" },
+			{ form = "data1",  key = "byte_size" },
+		},
+	},
+	member = {
+		abbrev = ABBREV_MEMBER,
+		attrs  = {
+			{ form = "string", key = "name" },
+			{ form = "data2",  key = "data_member_location" },
+			{ form = "ref4",   key = "type" },
+		},
+	},
+	pointer_type = {
+		abbrev = ABBREV_TYPED_VIEW_POINTER,
+		attrs  = {
+			{ form = "ref4", key = "type" },
+		},
+	},
+}
+
 -- DWARF5 §7.7.3 loclist opcodes.
 local DW_LLE_end_of_list    = 0x00
 local DW_LLE_start_length   = 0x08
-local DW_OP_reg0            = 0x50   -- base reg op; regN = 0x50 + N
-local DW_OP_breg0           = 0x70   -- base breg op; bregN = 0x70 + N (SLEB offset)
+local DW_OP_reg0            = 0x50 -- base reg op; regN = 0x50 + N
+local DW_OP_breg0           = 0x70 -- base breg op; bregN = 0x70 + N (SLEB offset)
 local DW_OP_piece           = 0x93
 local MIPS_LOAD_DELAY_BYTES = 0x08 -- 1 load word + 1 BD-slot word
 
@@ -153,44 +215,66 @@ local DW_AT_inline           = 0x20   -- DWARF5 §7.7.1: DW_AT_inline (used by a
 local DW_AT_decl_file        = 0x3A   -- DWARF5 §7.7.1: DW_AT_decl_file (1-based file index into the CU's file table)
 local DW_AT_decl_line        = 0x3B   -- DWARF5 §7.7.1: DW_AT_decl_line
 
--- File index lookup table for the existing main line unit (Unit 2).
--- Provenance paths come back with mixed slashes; we normalize to basename and look up against the line unit's actual file table.
--- Current scope has two provenance basenames: hello_gte_tape.c (the atom's call site) and lottes_tape.h (the component definition).
--- Both live in the existing gcc-generated line unit; their 1-based indices are stable across rebuilds because the include order
--- in code/gte_hello/hello_gte.c determines the unit's file table.
--- gcc only adds a file to the line table when it has actual line-number entries;
--- headers that are pure macros/typedefs (dsl.h, memory.h, math.h, mips.h, gp.h, gte.h, etc.) never appear.
--- lottes_tape.h is the FIRST include that emits line entries (MipsAtomComp_ declarations), so it is the FIRST entry after the primary file.
-local PROVENANCE_BASENAME_TO_FILE_INDEX = {
-	["hello_gte_tape.c"] = ATOM_SOURCE_FILE_INDEX,  -- = 11
-	["lottes_tape.h"]    = 2,
-}
+-- Replaced the hardcoded `ATOM_SOURCE_FILE_INDEX = 11` and the `PROVENANCE_BASENAME_TO_FILE_INDEX` table below with a runtime lookup
+-- (`init_file_index_lookup` + `resolve_provenance_file_index`) that reads the actual `.debug_line` file table from the post-link ELF.
+
+--- Populate the module-level file-index lookup table from the `.debug_line` section of the post-link ELF pointed at by `elf_path`.
+--- This MUST be called exactly once at pass start (from `M.run`) before any `resolve_provenance_file_index` invocation;
+--- downstream callers handle a nil table as "no file info available; fall back to errors".
+---
+--- The lookup uses `elf_dwarf.read_line_unit_file_table` (which parses both DWARF3 and DWARF5 line-program units — 
+--- the crt0.s assembler-side DWARF5 unit may emit non-standard form codes for paths and is intentionally skipped).
+--- @param elf_path string|nil
+local function init_file_index_lookup(elf_path)
+	if not elf_path or elf_path == "" then return end
+	local b2i, _basenames, paths = elf_dwarf.read_line_unit_file_table(elf_path)
+	if type(b2i) ~= "table" or type(paths) ~= "table" then
+		io.stderr:write("[dwarf_injection] read_line_unit_file_table returned no file table for: " .. tostring(elf_path) .. "\n")
+		return
+	end
+	_file_index_by_basename = b2i
+	_file_path_by_index    = paths
+	-- Pick any valid index for the opaque-row fallbacks at lines 466 + 570
+	-- (both sites legitimately want "any file index"; gdb resolves whatever index we emit to whatever that file's line happens to be).
+	for idx in pairs(paths) do
+		_default_atom_source_index = idx
+		break
+	end
+end
 
 --- Resolve an absolute provenance path to the line-unit file index used by the emitting line program.
---- Normalizes mixed `/` and `\` separators to a basename and looks it up against the known file table.
+--- Normalizes mixed `/` and `\` separators to a basename and looks it up against the runtime-computed file table populated by `init_file_index_lookup`.
 ---
---- Fails loudly on an unknown provenance basename: adding a new component source file requires extending
---- `PROVENANCE_BASENAME_TO_FILE_INDEX` so the line-program emission contract stays explicit.
---- Silent fallback to ATOM_SOURCE_FILE_INDEX would mask the new-file case by misattributing component rows to the atom's source file.
---- @param path string  -- absolute provenance path (e.g. "C:/.../lottes_thttps://www.youtube.com/watch?v=ORM4yLkdKx8ape.h" or "C:\\...\\lottes_tape.h")
---- @return integer  -- 1-based line-unit file index
+--- Returns 0 (the DWARF `set_file(0)` "no file change" sentinel) when the basename is not in the file table.
+--- This is a normal occurrence: the compiler only adds a file to the `.debug_line` file table when the file has line-numbered content (i.e., code).
+--- Files containing only static-array data (e.g. `MipsAtomComp_` declarations in `gp.atom.c`, `psyq.atom.c`, `pad.atom.c` — the OT-tag inserts, etc.) produce no line numbers,
+--- so gcc omits them from the file table.
+--- The DWARF emitter then keeps the previous line-program file state instead of pointing at a file that has no entries to walk.
+--- A stderr warning is emitted per-miss so the user can audit which files the compiler dropped.
+--- @param path string  -- absolute provenance path (mixed slashes accepted)
+--- @return integer  -- 1-based line-unit file index, or 0 on miss (DWARF no-change sentinel)
 local function resolve_provenance_file_index(path)
+	if _file_index_by_basename == nil then
+		error("[dwarf_injection] resolve_provenance_file_index called before init_file_index_lookup. Is M.run being entered correctly (with --elf)?")
+	end
 	if path == nil or path == "" then
 		error("[dwarf_injection] resolve_provenance_file_index: empty path")
 	end
-	-- Normalize backslashes → forward slashes (paths arrive with mixed separators from the provenance file: forward slashes from Lua's io.lines;
-	-- backslashes if the input ever round-trips through Windows shell expansion).
+	-- Normalize backslashes → forward slashes (paths arrive with mixed separators from the provenance file).
 	local normalized = path:gsub("\\", "/")
 	-- Take the last path component (the basename).
 	local basename = normalized:match("([^/]+)$") or normalized
-	local idx      = PROVENANCE_BASENAME_TO_FILE_INDEX[basename]
-	if idx == nil then
-		error(string.format(
-			"[dwarf_injection] resolve_provenance_file_index: unknown provenance basename '%s' (from '%s'). "
-			.. "Extend PROVENANCE_BASENAME_TO_FILE_INDEX in passes/dwarf_injection.lua.",
-			basename, path))
+	local idx      = _file_index_by_basename[basename]
+	if idx ~= nil then return idx end
+	-- Last-resort exact-path match (handles paths that don't reduce to a known basename).
+	for i, p in pairs(_file_path_by_index) do
+		if p and p:gsub("\\", "/") == normalized then return i end
 	end
-	return idx
+	-- File is in the corpus but gcc omitted it from the .debug_line file table (data-only content).
+	-- Return 0 = DWARF `set_file(0)` no-change sentinel so the line program keeps its prior file state.
+	io.stderr:write(string.format("[dwarf_injection] line-table miss: '%s' (basename '%s') not in .debug_line file table; "
+		.. "falling back to set_file(0)\n", path, basename))
+	return 0
 end
 
 local DW_FORM_addr           = 0x01
@@ -205,7 +289,6 @@ local DW_FORM_sec_offset     = 0x17   -- 4-byte section-relative offset (into .d
 
 -- DW_OP_reg0 + DW_OP_piece are declared above (lines 114-116) alongside the other DWARF5 §7.7.3 loclist opcodes.
 
-
 local DW_ATE_unsigned        = 0x07   -- DWARF5 §7.8.1: DW_ATE_unsigned (used for U4 base type)
 
 -- (DW_LANG_Mips_Assembler = 0x8001 was used in the, but we want this CU to look like a C TU so VSCode's Variables pane treats it as code.)
@@ -425,9 +508,9 @@ end
 --- Statement-state rules:
 ---   * A marked whole atom emits one opaque is_stmt=false range row and no nested component rows; its subprogram symbol/range remains available.
 ---   * Per-row policy at every other PC:
----       - Call-site row of any invocation's first word:  is_stmt = true   (unconditional; `want_call = true`).
+---       - Call-site row of any invocation's first word:     is_stmt = true   (unconditional; `want_call = true`).
 ---       - Body row of any invocation (first or subsequent): is_stmt = not inv.debug_skip  (`want_body = not inv.debug_skip`).
----       - RAW word (no containing invocation):            is_stmt = true   (unconditional).
+---       - RAW word (no containing invocation):              is_stmt = true   (unconditional).
 ---   * The previous per-word `marked_idx` ancestor walk and the GDB 12 zero-instruction-prologue duplicate row at atom entry are DELETED; the new
 ---     first-word emission IS the entry statement.
 ---   * Whole-atom suppression wins over component markers; no nested inversion.
@@ -461,7 +544,7 @@ local function build_atom_sequence(atom)
 	if atom.debug_skip then
 		return table.concat({
 			set_address(atom.addr),
-			set_file(ATOM_SOURCE_FILE_INDEX),
+			set_file(resolve_provenance_file_index(atom.src_path)),
 			advance_line(atom.entries[1].line - 1),
 			negate_stmt(),
 			copy_op(),
@@ -508,10 +591,9 @@ local function build_atom_sequence(atom)
 	--   * If the invocation's body has any NESTED invocations (parent_id == top_inv.id), the body's
 	--     first content is the call_line of the earliest nested invocation (by start_pos).
 	--   * Otherwise (only RAW words in the body), it's the line of the first raw word = body_lines[1].
-	-- This is the value the multi-row PC's body_lines[1] row must reference for source-order display:
-	-- `anc.body_lines[1]` is the line of the FIRST WORD (which for an outer whose body starts with a
-	-- nested expansion is inside the inner's expansion = wrong for display purposes); `anc.body_first_line`
-	-- is the body's first content line in the parent's source (= correct for display).
+	-- This is the value the multi-row PC's body_lines[1] row must reference for source-order display: `anc.body_lines[1]` is the line of the FIRST WORD 
+	-- (which for an outer whose body starts with a nested expansion is inside the inner's expansion = wrong for display purposes);
+	-- `anc.body_first_line` is the body's first content line in the parent's source (= correct for display).
 	local body_first_line_of = {}
 	for _, top_inv in ipairs(invs) do
 		local earliest_nested_call_line = nil
@@ -565,7 +647,7 @@ local function build_atom_sequence(atom)
 		parts[#parts + 1] = copy_op()
 	end
 
-	local call_file_idx = ATOM_SOURCE_FILE_INDEX
+	local call_file_idx = resolve_provenance_file_index(atom.src_path)
 
 	-- --- Atom entry (idx 1) -------------------------------------------------
 	local entry_1         = atom.entries[1]
@@ -574,13 +656,11 @@ local function build_atom_sequence(atom)
 	-- If atom entry 1 starts inside an invocation, walk the ancestry and emit a call-site row + (when applicable)
 	-- a body_lines[1] row for every active ancestor. For a non-nested invocation this is just the one pair;
 	-- for nested invocations this emits the outer call-site + body_lines[1] rows BEFORE the inner pair so the debugger displays
-	-- the outer body line at the inner's first word
-	-- (PROBLEM B fix).
+	-- the outer body line at the inner's first word (PROBLEM B fix).
 	--
-	-- A marked OUTERMOST ancestor's body_lines[1] row is suppressed at this PC (the existing full-skip
-	-- contract is preserved for the marked outer range); its call-site row IS still emitted as a
-	-- statement. Marked INNER ancestors always emit their body_lines[1] row with is_stmt=false
-	-- (the per-invocation `want_body = not inv.debug_skip` predicate).
+	-- A marked OUTERMOST ancestor's body_lines[1] row is suppressed at this PC (the existing full-skip contract is preserved for the marked outer range);
+	-- Its call-site row IS still emitted as a statement.
+	-- Marked INNER ancestors always emit their body_lines[1] row with is_stmt=false (the per-invocation `want_body = not inv.debug_skip` predicate).
 	if #entry_1_ancestry == 0 then
 		-- RAW word at atom entry: single call-site row, always a statement target.
 		emit_row(call_file_idx, entry_1.line, true)
@@ -588,15 +668,12 @@ local function build_atom_sequence(atom)
 		-- Atom starts in an invocation. Walk the ancestry outermost-first.
 		-- Each ancestor emits one call-site row (statement) and one body_lines[1] row
 		-- (statement iff unmarked; suppressed for marked outermost).
-		-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content
-		-- line in the parent's source), NOT anc.body_lines[1] (= the line of the first WORD,
-		-- which is wrong when the outer's body starts with a nested call).
+		-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content line in the parent's source),
+		-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call).
 		for ai, anc in ipairs(entry_1_ancestry) do
 			assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-			assert(anc.body_lines[1] ~= nil
-				, "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
-			assert(anc.call_path and anc.call_path ~= ""
-				, "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
+			assert(anc.body_lines[1] ~= nil,              "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
+			assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation "            .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
 			emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line,    true)
 			local is_outermost = (ai == 1)
 			if not (is_outermost and anc.debug_skip) then
@@ -615,29 +692,23 @@ local function build_atom_sequence(atom)
 
 		if inv and idx == inv.start_pos + 1 then
 			-- First word of the innermost active invocation (PROBLEM B fix — nested-display rule).
-			-- Walk the active ancestry outermost-first; for each ancestor emit a call-site row
-			-- (statement) + a body_lines[1] row. The inner-most invocation's call-site + body pair
-			-- become the LAST two rows in the sequence. Marked outermost ancestors suppress their
-			-- body_lines[1] row at this PC (the existing full-skip contract is preserved for the
-			-- marked outer range); all OTHER ancestors emit body_lines[1] with is_stmt = not debug_skip.
+			-- Walk the active ancestry outermost-first; for each ancestor emit a call-site row (statement) + a body_lines[1] row.
+			-- The inner-most invocation's call-site + body pair become the LAST two rows in the sequence.
+			-- Marked outermost ancestors suppress their body_lines[1] row at this PC (the existing full-skip contract is preserved for the marked outer range);
+			-- all OTHER ancestors emit body_lines[1] with is_stmt = not debug_skip.
 			--
 			-- This re-emits the outer ancestor's call-site + body rows at the inner's first word PC
-			-- for debugger context: source-level stepping now shows the outer body line (not the
-			-- inner body line) when stepping into the inner. PROBLEM B fix.
-			-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content
-			-- line in the parent's source), NOT anc.body_lines[1] (= the line of the first WORD,
-			-- which is wrong when the outer's body starts with a nested call: gdb 12.1 picks the
-			-- displayed line as the LAST row at the same PC in byte-stream order, so the disc=1 row's
-			-- value matters for what's shown when stepping into the nested case).
+			-- for debugger context: source-level stepping now shows the outer body line (not the inner body line) when stepping into the inner. PROBLEM B fix.
+			-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content line in the parent's source), 
+			-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call: 
+			-- gdb 12.1 picks the displayed line as the LAST row at the same PC in byte-stream order,
+			-- so the disc=1 row's value matters for what's shown when stepping into the nested case).
 			local ancestry = ancestry_idx[idx]
 			for ai, anc in ipairs(ancestry) do
 				assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-				assert(anc.body_lines[1] ~= nil
-					, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d",
-						anc.component_name, anc.start_pos, #(anc.body_lines or {})))
-				assert(anc.call_path and anc.call_path ~= ""
-					, "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
-				emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line,    true)
+				assert(anc.body_lines[1] ~= nil, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d", anc.component_name, anc.start_pos, #(anc.body_lines or {})))
+				assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
+				emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line, true)
 				local is_outermost = (ai == 1)
 				if not (is_outermost and anc.debug_skip) then
 					emit_row(resolve_provenance_file_index(anc.def_path),  body_first_line_of[anc.id] or anc.body_lines[1], not anc.debug_skip)
@@ -647,16 +718,13 @@ local function build_atom_sequence(atom)
 			-- Subsequent body word of the innermost active invocation: `body_lines[k]` is indexed by the 1-based offset of this word inside the invocation.
 			-- Both `idx` (1-based DWARF entry index) and `inv.start_pos` (0-based emitted-word position stamped at `emit_invoke_begin`) come from the same
 			-- monotonic counter, so `idx - inv.start_pos` is exactly the 1-based k (the first word of the invocation has `idx == inv.start_pos + 1`, hence `k == 1`).
-			-- atom_dbg_step_ux_20260725: `want_body = not inv.debug_skip`. The previous `want = not marked_idx[idx]`
-			-- (which suppressed ALL body rows when any ancestor was marked) is replaced by the per-invocation
-			-- predicate. Marked invocations emit non-statement body rows at every body word; unmarked
-			-- invocations emit statement body rows.
+			-- atom_dbg_step_ux_20260725: `want_body = not inv.debug_skip`.
+			-- The previous `want = not marked_idx[idx]` (which suppressed ALL body rows when any ancestor was marked) is replaced by the per-invocation predicate.
+			-- Marked invocations emit non-statement body rows at every body word; unmarked invocations emit statement body rows.
 			assert(inv.body_lines, "missing body_lines: emitter did not run emission-model")
 			local words_into = idx - inv.start_pos
-			assert(inv.body_lines[words_into] ~= nil
-				, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d",
-					words_into, inv.component_name, inv.start_pos, #(inv.body_lines or {}), idx))
-			emit_row(resolve_provenance_file_index(inv.def_path),       inv.body_lines[words_into], not inv.debug_skip)
+			assert(inv.body_lines[words_into] ~= nil, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d", words_into, inv.component_name, inv.start_pos, #(inv.body_lines or {}), idx))
+			emit_row(resolve_provenance_file_index(inv.def_path), inv.body_lines[words_into], not inv.debug_skip)
 		else
 			-- RAW word: single call-site row, always a statement target (the word itself is unmarked).
 			emit_row(call_file_idx, entry.line, true)
@@ -696,15 +764,17 @@ end
 ---     `{comp_name, call_file, call_line, comp_file, comp_line, start_pos, end_pos, body_lines, debug_skip}`. `body_lines[k]`
 ---     is the k-th word's source line within the component body.
 ---
---- @param corpus    table -- the corpus from `ctx.shared.corpus`
---- @param addrs     table -- ELF symbols keyed by atom name from `elf_dwarf.read_nm`
---- @return table[]  -- list of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
+--- @param corpus table -- From `ctx.shared.corpus`
+--- @param addrs  table -- ELF symbols keyed by atom name from `elf_dwarf.read_nm`
+--- @return table[]  -- List of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
 local function build_atom_table(corpus, addrs)
 	-- Cross-ref: keep only atoms present in BOTH the nm symbol table AND `corpus.atoms_by_name`. Output is sorted by ascending addr.
 	local atoms_by_name = corpus.atoms_by_name or {}
 
 	-- Per-atom ingest. Returns nil if the atom is absent from the corpus; the caller skips it via the `if atom then ...` guard.
-	local function ingest_atom(name, info)
+	-- `src_path` is the absolute source path that declared this atom; the build_atom_table iteration below threads `src.path` through.
+	-- This is consumed by `build_atom_sequence::set_file(...)` for opaque-row fallbacks + raw-word rows (atoms where no invocation ancestry exists).
+	local function ingest_atom(name, info, src_path)
 		local atom_record = atoms_by_name[name]
 		if not atom_record then return nil end
 
@@ -712,8 +782,8 @@ local function build_atom_table(corpus, addrs)
 		local word_events      = paths.word_events or {}
 		local invocations_proj = paths.invocations or {}
 		-- Build the dense entries list from `word_events`.
-		--   `word_events[i].i`     = the 0-based `.word` position
-		--   `call_line`            = the root atom's physical source line for that word (stamped by emission_model)
+		--   `word_events[i].i`  = the 0-based `.word` position
+		--   `call_line`         = the root atom's physical source line for that word (stamped by emission_model)
 		local entries = {}
 		for idx, ev in ipairs(word_events) do
 			entries[#entries + 1] = {
@@ -730,6 +800,7 @@ local function build_atom_table(corpus, addrs)
 			words      = #word_events,
 			entries    = entries,
 			debug_skip = atom_record.debug_skip == true,
+			src_path   = src_path or "",
 		}
 
 		-- Consume invocation records from `atom.paths.invocations`. It is the single producer of per-invocation body_lines, per-invocation debug_skip,
@@ -753,9 +824,25 @@ local function build_atom_table(corpus, addrs)
 	end
 
 	local out = {}
-	for name, info in pairs(addrs) do
-		local atom = ingest_atom(name, info)
-		if atom then out[#out + 1] = atom end
+	-- Walk every source's atom list (which preserves source order + per-source src_path).
+	-- Cross-ref with the nm symbol table; atoms absent from `addrs` are skipped
+	-- (an atom declared in source but not emitted as a symbol is a metaprogram or atom-info bug, not a source-correlation bug — emit_no_emit would catch it upstream).
+	for _, src in ipairs((corpus and corpus.source_order) or {}) do
+		local src_path = src.path or ""
+		for _, atom_rec in ipairs(((src.scan or {}).atoms) or {}) do
+			local info = addrs[atom_rec.name or atom_rec.raw_name]
+			if info then
+				local atom = ingest_atom(atom_rec.name or atom_rec.raw_name, info, src_path)
+				if atom then out[#out + 1] = atom end
+			end
+		end
+		for _, atom_rec in ipairs(((src.scan or {}).raw_atoms) or {}) do
+			local info = addrs[atom_rec.name or atom_rec.raw_name]
+			if info then
+				local atom = ingest_atom(atom_rec.name or atom_rec.raw_name, info, src_path)
+				if atom then out[#out + 1] = atom end
+			end
+		end
 	end
 	table.sort(out, function(a, b) return a.addr < b.addr end)
 	return out
@@ -800,27 +887,33 @@ end
 ---   load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
 ---   ...
 ---
+--- Also matches `load_half` / `load_half_u` / `load_byte` / `load_byte_u` (any MIPS load instruction with `(R_<reg>, R_<base>, O_(<Binds_X>, FieldName))` shape).
+--- Every field's `byte_size` + `offset` determine which load to emit; this function only records the (reg, field) pair.
+---
 --- The GPR for each `R_<reg>` is looked up in the merged register_alias_registry; aliases absent from the registry
 --- (no `atom_reg` opt-in) are silently skipped — the resulting rbind record will be incomplete and the atom will fail to bind a usable piece chain.
 --- This is intentional: silently falling back to a hardcoded GPR would mask the missing opt-in.
 ---
---- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level
---- statements (each entry is a single `load_word(...)` call or other statement).
---- @param body_tokens table[]  -- the atom's pre-tokenized body statements (from atom.body_tokens)
---- @param binds_name  string   -- expected Binds_X name (skip pairs with mismatching binds)
---- @param registries  table    -- merged registries from collect_per_source_registries
---- @return table[]             -- list of {reg = <MIPS index>, field = <field name>}
+--- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level statements (each entry is a single `load_*` call or other statement).
+--- @param body_tokens table[]  -- The atom's pre-tokenized body statements (from atom.body_tokens)
+--- @param binds_name  string   -- Expected Binds_X name (skip pairs with mismatching binds)
+--- @param registries  table    -- Merged registries from collect_per_source_registries
+--- @return table[]             -- List of {reg = <MIPS index>, field = <field name>}
 local function parse_body_load_pairs(body_tokens, binds_name, registries)
 	local pairs = {}
 	local reg_index_by_name = (registries and registries.register_alias_registry) or {}
+	-- One regex that matches any of: load_word, load_half, load_half_u, load_byte, load_byte_u, gte_lw, gte_lwc2.
+	-- The captured ident is `kind`; `inner` holds the parens body for arg parsing.
+	local load_pattern = "^(load_word|load_half|load_half_u|load_byte|load_byte_u|gte_lw|gte_lwc2)%s*%((.*)%)$"
 	for _, t in ipairs(body_tokens or {}) do
-		local tok = duffle.trim(t.tok or "")
-		-- Match "load_word(...)" — the entire call is one body_tokens entry.
-		local inner = tok:match("^load_word%s*%((.*)%)$")
-		if inner then
+		local tok         = duffle.trim(t.tok or "")
+		local kind, inner = tok:match(load_pattern)
+		if kind then
 			local args = duffle.split_top_level_commas(inner)
-			-- Expected shape: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
-			if #args >= 3 then
+			-- Expected shape for an rbind piece-chain load: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
+			-- The second arg MUST be R_TapePtr — loads from other bases (e.g. `load_byte_u(R_RawStatus, R_PadRaw, 0)`)
+			-- are field-derivative loads that read already-bound tape values; they're NOT a new piece-chain.
+			if #args >= 3 and duffle.trim(args[2]) == "R_TapePtr" then
 				local reg_name  = duffle.trim(args[1])
 				local third_arg = duffle.trim(args[3])
 				-- Match O_(Binds_<X>, FieldName)
@@ -839,9 +932,7 @@ local function parse_body_load_pairs(body_tokens, binds_name, registries)
 end
 
 --- Collect every rbind atom + the matching Binds_X struct + (reg, field) pairs.
----
 --- Inputs come from the dep-closed `scan-source` pass (the per-source `src.scan` payload is preserved on each `corpus.source_order` entry).
----
 --- Returns:
 ---   rbind_atoms     = {[atom_name] = {binds, fields, regs, byte_size, info_line}}
 ---   rbind_structs   = {[binds_name] = {byte_size, fields, atom_names}}
@@ -850,9 +941,9 @@ end
 --- The piece chain uses (DW_OP_regN, DW_OP_piece, ULEB128(field_size)).
 ---
 --- Binds fields come from `scan.binds`; the per-source `scan.binds[i].fields` already carries the typed-field record after the scan-source generalization.
---- @param corpus     table   -- the corpus from `ctx.shared.corpus`
---- @param atom_table table[] -- the cross-ref'd atom table from build_atom_table
---- @param registries table   -- merged registries from collect_per_source_registries
+--- @param corpus     table   -- From `ctx.shared.corpus`
+--- @param atom_table table[] -- Cross-ref'd atom table from build_atom_table
+--- @param registries table   -- Merged registries from collect_per_source_registries
 --- @return table, table      -- (rbind_atoms, rbind_structs)
 local function parse_rbind_atoms(corpus, atom_table, registries)
 	registries = registries or {}
@@ -886,7 +977,7 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 	local body_tokens_by_atom = {}
 	for _, src in ipairs((corpus and corpus.source_order) or {}) do
 		local scan = src.scan
-		if scan then
+		if    scan then
 			for _, atom in ipairs(scan.atoms or {}) do
 				body_tokens_by_atom[atom.name] = atom.body_tokens
 			end
@@ -905,8 +996,8 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 
 	for atom_name, ai in pairs(ai_by_atom) do
 		if ai.binds then
-			local struct     = rbind_structs[ai.binds]
-			local body_toks  = body_tokens_by_atom[atom_name]
+			local struct    = rbind_structs[ai.binds]
+			local body_toks = body_tokens_by_atom[atom_name]
 			if struct and body_toks then
 				local pairs = parse_body_load_pairs(body_toks, ai.binds, registries)
 				if #pairs > 0 then
@@ -914,7 +1005,7 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 						binds     = ai.binds,
 						fields    = struct.fields, -- {name, offset} from scan.binds
 						bytes     = struct.bytes,
-						regs      = pairs,         -- ordered list of {reg, field}
+						regs      = pairs,         -- Ordered list of {reg, field}
 						info_line = ai.info_line,
 					}
 					table.insert(struct.atom_names, atom_name)
@@ -941,7 +1032,8 @@ end
 --- (the final unit, referenced by the main CU's DW_AT_stmt_list).
 ---
 --- This builder extends the main compilation unit.
---- A detached synthetic line unit has no DW_AT_stmt_list referencing it, so gdb ignored it (a previous experiment); byte 13 is the first special opcode, not the extended-opcode marker.
+--- A detached synthetic line unit has no DW_AT_stmt_list referencing it, so gdb ignored it (a previous experiment);
+--- byte 13 is the first special opcode, not the extended-opcode marker.
 --- The existing final unit already contains hello_gte_tape.c as file index 11 and ends with a valid end_sequence.
 --- We preserve its bytes, append independent atom sequences, and increase only that unit's DWARF32 unit_length.
 --- @param existing   string -- existing section bytes, byte-for-byte
@@ -952,9 +1044,7 @@ local function build_dwarf_line_section(existing, atom_table)
 
 	-- Build the sequences.
 	local sequences = {}
-	for _, atom in ipairs(atom_table) do
-		sequences[#sequences + 1] = build_atom_sequence(atom)
-	end
+	for _, atom in ipairs(atom_table) do sequences[#sequences + 1] = build_atom_sequence(atom) end
 	local appended = table.concat(sequences)
 
 	-- Walk DWARF32 line units and retain the final unit's bounds.
@@ -962,8 +1052,8 @@ local function build_dwarf_line_section(existing, atom_table)
 	local unit_pos, last_pos, last_length, last_end = 0, nil, nil, nil
 	while unit_pos < #existing do
 		if unit_pos + 4 > #existing then return existing end
-		local unit_length = elf_dwarf.read_u32_le(existing, unit_pos)
-		if    unit_length == elf_dwarf.ELF32.dw_dwarf32_terminator then return existing end
+		local unit_length =  elf_dwarf.read_u32_le(existing, unit_pos)
+		if    unit_length == elf_dwarf.dw_dwarf32_terminator then return existing end
 		local unit_end_excl = unit_pos + 4 + unit_length
 		if unit_end_excl > #existing then return existing end
 		last_pos, last_length, last_end = unit_pos, unit_length, unit_end_excl
@@ -1007,13 +1097,13 @@ local function build_dwarf_aranges_section(existing, atom_table)
 	-- We bump the unit's length field accordingly.
 	--
 	-- Unit structure (DWARF4 §7.21):
-	--   unit_length (4)
-	--   version (2)
+	--   unit_length       (4)
+	--   version           (2)
 	--   debug_info_offset (4)  -- CU DIE offset in .debug_info
-	--   address_size (1)
-	--   segment_size (1)
-	--   entries... (4-byte addr + 4-byte length)
-	--   terminator (8 bytes: addr=0, length=0)
+	--   address_size      (1)
+	--   segment_size      (1)
+	--   entries...        (4-byte addr + 4-byte length)
+	--   terminator        (8 bytes: addr=0, length=0)
 
 	-- Walk all units and emit each one (preserving existing structure).
 	-- For the LAST unit, replace the terminator with my entries + new term.
@@ -1024,7 +1114,7 @@ local function build_dwarf_aranges_section(existing, atom_table)
 	while i < #existing do
 		-- Read this unit's length.
 		local ul = elf_dwarf.read_u32_le(existing, i)
-		if    ul == elf_dwarf.ELF32.dw_dwarf32_terminator then
+		if    ul == elf_dwarf.dw_dwarf32_terminator then
 			-- DWARF64 marker - not supported.
 			io.stderr:write("[dwarf_injection] WARN: .debug_aranges contains a DWARF64 marker (0xFFFFFFFF); the 64-bit extension is not supported by this metaprogram; passing through unchanged\n")
 			return existing
@@ -1315,7 +1405,7 @@ local function build_new_abbrev()
 		attr(   DW_AT_name,         DW_FORM_string)
 	  .. attr(DW_AT_low_pc,       DW_FORM_addr)
 	  .. attr(DW_AT_high_pc,      DW_FORM_addr)
-	  .. attr(DW_AT_linkage_name, DW_FORM_string))  -- equals DW_AT_name; lets gdb's symbol-table lookup resolve to our subprogram (not the gcc global `code_<name>` const U4 array)
+	  .. attr(DW_AT_linkage_name, DW_FORM_string))  -- equals DW_AT_name; gdb resolves the subprogram, not the gcc global array
 
 	local abbrev_variable = abbrev(ABBREV_VARIABLE, DW_TAG_variable, false,  -- DW_CHILDREN_no
 		attr(   DW_AT_name,     DW_FORM_string)
@@ -1488,13 +1578,12 @@ end
 ---        DW_AT_location  = piece-chain (DW_FORM_exprloc)
 ---        DW_AT_type      = ref4 → structure_type DIE
 ---
---- This function does NOT emit the final 0 byte (root terminator). build_debug_info_section splices bytes ahead of the root terminator
---- and preserves existing DIE bytes exactly.
+--- This function does NOT emit the final 0 byte (root terminator).
+--- build_debug_info_section splices bytes ahead of the root terminator and preserves existing DIE bytes exactly.
 ---
 --- ref4 basis: DW_FORM_ref4 is CU-relative (offset from the first byte of the CU header).
 --- Our inserted DIEs live in the main CU, so every ref4 = (target section offset) - main_cu_offset.
 --- Per-die section offsets are tracked via the running `next_offset` cursor (= section offset of the NEXT byte to emit).
----
 --- @param main_cu_offset integer    -- 0-based section offset of the main CU's unit_length field
 --- @param main_cu_end_excl integer  -- 0-based section offset of the first byte AFTER the main CU
 --- @param atom_table table[]        -- atoms (with atom.rbind set if rbind; atom.invocations set if mac_X(...) calls)
@@ -1553,6 +1642,26 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		S.bytes[#S.bytes + 1] = s
 		S.next_offset       = S.next_offset + #s
 	end
+	local function emit_die(schema_name, values)
+		local row = DIE_SCHEMA[schema_name]
+		emit(uleb128(row.abbrev))
+		for _, attr in ipairs(row.attrs) do
+			local v = values[attr.key]
+			if attr.form == "string" then
+				emit(v .. "\0")
+			elseif attr.form == "data1" then
+				emit(string.char(v))
+			elseif attr.form == "udata" then
+				emit(uleb128(v))
+			elseif attr.form == "addr" or attr.form == "ref4" then
+				emit(elf_dwarf.write_u32_le(v))
+			elseif attr.form == "data2" then
+				emit(elf_dwarf.write_u16_le(v))
+			elseif attr.form == "exprloc" then
+				emit(v)
+			end
+		end
+	end
 	local function ref4_of(section_offset)
 		if section_offset == nil then return 0 end
 		return section_offset - main_cu_offset
@@ -1561,10 +1670,11 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 
 	-- 1) Emit the base_type DIE first (member ref4s reference it).
 	local base_type_section_offset = S.next_offset
-	emit(uleb128(ABBREV_BASE_TYPE))
-	emit("unsigned int\0")                   -- DW_FORM_string (DW_AT_name)
-	emit(string.char(4))                     -- DW_FORM_data1  (DW_AT_byte_size)
-	emit(string.char(DW_ATE_unsigned))       -- DW_FORM_data1  (DW_AT_encoding)
+	emit_die("base_type", {
+		name      = "unsigned int",
+		byte_size = 4,
+		encoding  = DW_ATE_unsigned,
+	})
 	-- The function body below reads S.next_offset directly via the `next_offset` function;
 	-- this keeps offsets synchronized with emitted data.
 	local function next_offset() return S.next_offset end
@@ -1614,6 +1724,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	--
 	-- The table is small + explicit — the prototype principle treats the typed-view struct layout as data, not derived state.
 	local STRUCT_MEMBER_TABLE = {
+		-- TODO(Ed): This hardcoding is brittle...
+		-- TODO(Ed): Better to just have a table for the fundamental types in duffle/dsl.h, we can derive the rest via typedef parsing...
 		-- 2-element signed short vector (rare; placeholder for future use).
 		V2_S2 = { byte_size = 4, members = {
 			{ name = "x", offset = 0, byte_size = 2 },
@@ -1750,7 +1862,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	emit(uleb128(ABBREV_TYPED_VIEW_POINTER))  -- DW_TAG_pointer_type (abbrev 110; NOT 9; void chain target)
 	emit(elf_dwarf.write_u32_le(ref4_of(void_chain_offset)))  -- 4-byte ref4: points at the void base_type's tag byte
 	-- type_chain_offsets["void|1"] is what step (f) of the per-RR_<R_Name> chain looks up.
-	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type. The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
+	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type.
+	-- The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
 	-- Capture the pointer_type's offset (the last-thing-emitted DIE start) and overwrite the lookup.
 	-- The pointer_type was emitted as: uleb(9) (1 byte) + 4-byte ref4 = 5 bytes. Its tag byte is at void_chain_offset + 8 (the base_type's 8 bytes: 1 tag + 5 name + 1 byte_size + 1 encoding).
 	local ptr_void_offset = void_chain_offset + 8
@@ -1816,25 +1929,26 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	for _, comp_name in ipairs(sorted_comp_names) do
 		local def = component_defs[comp_name]
 		abstract_offsets[comp_name] = next_offset()
-		emit(uleb128(ABBREV_ABSTRACT_SUBPROGRAM))
-		emit("mac_" .. comp_name .. "\0") -- DW_FORM_string (DW_AT_name)
-		emit(string.char(DW_INL_inlined)) -- DW_FORM_data1 (DW_AT_inline)
-		emit(string.char(0x01))           -- DW_FORM_data1 (DW_AT_external=1)
-		-- decl_file + decl_line resolve the abstract origin back to its definition site even when no inlined_subroutine instance maps to it.
-		emit(uleb128(resolve_provenance_file_index(def.def_file)))  -- DW_FORM_udata (DW_AT_decl_file)
-		emit(uleb128(def.def_line))                                 -- DW_FORM_udata (DW_AT_decl_line)
+		emit_die("abstract_subprogram", {
+			name      = "mac_" .. comp_name,
+			inline    = DW_INL_inlined,
+			external  = 0x01,
+			decl_file = resolve_provenance_file_index(def.def_file),
+			decl_line = def.def_line,
+		})
 	end
 
 	-- 4) Emit per-atom DW_TAG_subprograms (children of main CU).
-	--    Subprogram names match nm symbols without a `code_` prefix.
+	--    Subprogram names match the written C ident (the ELF symbol).
 	--    The gcc global `<name>[]` is a DW_TAG_variable without children; our subprogram has the wave-context var children.
 	--    gdb's symbol resolution picks our subprogram (it has low_pc/high_pc + children) over the gcc global for function-context lookups.
 	for _, atom in ipairs(atom_table) do
-		emit(uleb128(ABBREV_SUBPROGRAM))
-		emit(atom.name .. "\0") -- DW_FORM_string (DW_AT_name)
-		emit(elf_dwarf.write_u32_le(atom.addr))
-		emit(elf_dwarf.write_u32_le(atom.addr + atom.size_bytes))
-		emit(atom.name .. "\0") -- DW_FORM_string (DW_AT_linkage_name; same as DW_AT_name for non-mangled C)
+		emit_die("subprogram", {
+			name          = atom.name,
+			low_pc        = atom.addr,
+			high_pc       = atom.addr + atom.size_bytes,
+			linkage_name  = atom.name,
+		})
 
 	-- Per debug-visible R_ alias (filtered to GPR 0..31 in `by_alias`): DW_TAG_variable.
 	-- Precedence chain (per atom, per RR_<R_Name>):
@@ -1866,12 +1980,13 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		end
 		local atom_view = (registries.atom_views or {})[atom.name]
 		-- Build the atom-name lookup table once (cheap; O(atom_table)) so step (b) and step (d) can resolve rbind_atom names.
+		-- TODO(Ed): Bad assignment?
 		local atom_by_name = atom_by_name or (function() local m = {}; for _, a in ipairs(atom_table) do if a.name then m[a.name] = a end end; return m end)()
 		-- step (b) inputs: this atom's `atom_ctx(<rbind_atom>)` (resolved from the registries' atom_ctxs)
 		local this_ctx = registries.atom_ctxs and registries.atom_ctxs[atom.name]
 		if this_ctx and this_ctx.rbind_atom then
 			local rbind = atom_by_name_global[this_ctx.rbind_atom]
-			if rbind and rbind.rbind and rbind.rbind.fields then
+			if    rbind and rbind.rbind and rbind.rbind.fields then
 				atom_view_ctx_fields = {}
 				for _, f in ipairs(rbind.rbind.fields) do atom_view_ctx_fields[f.name] = f end
 				if rbind.rbind.regs then
@@ -1891,11 +2006,11 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		end
 		if my_phase_label then
 			local group = (registries.atom_phases or {})[my_phase_label]
-			if group and group.atoms then
+			if    group and group.atoms then
 				for _, group_atom_name in ipairs(group.atoms) do
 					if group_atom_name ~= atom.name then
 						local cand = atom_by_name_global[group_atom_name]
-						if cand and cand.rbind and cand.rbind.fields then
+						if    cand and cand.rbind and cand.rbind.fields then
 							atom_view_phase_fields = {}
 							for _, f in ipairs(cand.rbind.fields) do atom_view_phase_fields[f.name] = f end
 							if cand.rbind.regs then
@@ -1916,7 +2031,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			-- (a) per-atom callsite atom_type(R_X, <T>): most specific; user explicit override for THIS atom only.
 			function(r_name, alias_code)
 				local override = atom_view and atom_view.reg_type_overrides and atom_view.reg_type_overrides[r_name]
-				if override and override.pointer_depth and override.pointer_depth > 0 then
+				if    override and override.pointer_depth and override.pointer_depth > 0 then
 					return type_chain_offsets[override.type_name .. "|" .. override.pointer_depth]
 				end
 			end,
@@ -1924,7 +2039,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local ctx_field_name = reg_to_field_ctx and reg_to_field_ctx[alias_code]
 				local ctx_f          = ctx_field_name and atom_view_ctx_fields and atom_view_ctx_fields[ctx_field_name]
-				if ctx_f and ctx_f.pointer_depth and ctx_f.pointer_depth > 0 then
+				if    ctx_f and ctx_f.pointer_depth and ctx_f.pointer_depth > 0 then
 					return type_chain_offsets[ctx_f.type_name .. "|" .. ctx_f.pointer_depth]
 				end
 			end,
@@ -1932,7 +2047,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local field_name = reg_to_field[alias_code]
 				local f         = field_name and field_type_by_name[field_name]
-				if f and f.pointer_depth and f.pointer_depth > 0 then
+				if    f and f.pointer_depth and f.pointer_depth > 0 then
 					return type_chain_offsets[f.type_name .. "|" .. f.pointer_depth]
 				end
 			end,
@@ -1940,12 +2055,13 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local phase_field_name = reg_to_field_phase and reg_to_field_phase[alias_code]
 				local phase_f          = phase_field_name and atom_view_phase_fields and atom_view_phase_fields[phase_field_name]
-				if phase_f and phase_f.pointer_depth and phase_f.pointer_depth > 0 then
+				if    phase_f and phase_f.pointer_depth and phase_f.pointer_depth > 0 then
 					return type_chain_offsets[phase_f.type_name .. "|" .. phase_f.pointer_depth]
 				end
 			end,
 			-- (e) enum-site atom_type(<T>) default on the registry entry.
 			function(r_name, alias_code)
+				-- TODO(Ed): Bad definition?
 				if alias and alias.default_type and alias.default_depth and alias.default_depth > 0 then
 					return type_chain_offsets[alias.default_type .. "|" .. alias.default_depth]
 				end
@@ -1954,8 +2070,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 
 		-- Iterate `by_alias` in sorted order; Lua's pairs() is non-deterministic, so sorting ensures byte-identical DWARF output across builds.
 		for _, r_name in ipairs(by_alias_order) do
-			local alias = by_alias[r_name]
-			local rr_name = "RR_" .. strip_r_prefix(r_name)
+			local alias      = by_alias[r_name]
+			local rr_name    = "RR_" .. strip_r_prefix(r_name)
 			local alias_code = alias.code
 			emit(uleb128(ABBREV_VARIABLE))
 			emit(rr_name .. "\0") -- DW_FORM_string (DW_AT_name)
@@ -1977,7 +2093,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		-- Two PC ranges cover every field: [atom.addr, last_load+8) describes each field as tape memory (DW_OP_bregN + offset) piece,
 		-- and [last_load+8, atom.end) describes each field as a GPR (DW_OP_regN) piece.
 		if atom.rbind then
-			local binds_name = atom.rbind.binds
+			local binds_name      = atom.rbind.binds
 			local loclists_offset = loclists_offsets[atom.name] or 0
 			emit(uleb128(ABBREV_BIND_VAR_LOCLIST))
 			emit("bind_args\0")                  -- DW_FORM_string (DW_AT_name)
@@ -2030,7 +2146,7 @@ end
 ---
 --- Fails safely by returning existing sections unchanged if the table walker can't find the table terminator (malformed input).
 ---
---- @param existing string  -- existing .debug_abbrev bytes, byte-for-byte
+--- @param existing           string   -- existing .debug_abbrev bytes, byte-for-byte
 --- @param main_abbrev_offset integer  -- 0-based offset into `existing` of the main CU's abbrev table
 --- @return string, integer  -- (new_abbrev_bytes, offset_where_duplicate_table_starts = #existing)
 local function build_debug_abbrev_section(existing, main_abbrev_offset)
@@ -2048,7 +2164,7 @@ local function build_debug_abbrev_section(existing, main_abbrev_offset)
 end
 
 --- Build the new .debug_str: existing strings + new strings appended.
---- @param existing string  -- existing .debug_str bytes, byte-for-byte
+--- @param existing   string  -- existing .debug_str bytes, byte-for-byte
 --- @param atom_table table[]
 --- @param registries table  -- merged registries from collect_per_source_registries
 --- @return string  -- existing bytes plus the deterministic appended strings
@@ -2058,7 +2174,6 @@ local function build_debug_str_section(existing, atom_table, registries)
 end
 
 --- Build the new .debug_info: SPLICE inserted DIEs into the MAIN CU as children.
----
 --- This implementation:
 ---   1. Builds the inserted-children bytes (base_type, struct_types, subprograms with their RR_* + bind_args children) via build_inserted_children.
 ---   2. Patches the main CU's `unit_length` field to account for the inserted bytes.
@@ -2068,14 +2183,14 @@ end
 ---
 --- The crt CU (everything before main_cu_start) is preserved.
 
---- @param existing string            -- existing .debug_info section bytes
---- @param main_cu_start integer      -- 0-based offset of the main CU's unit_length field
---- @param main_cu_end_excl integer   -- 0-based offset of the first byte AFTER the main CU
+--- @param existing          string   -- existing .debug_info section bytes
+--- @param main_cu_start     integer  -- 0-based offset of the main CU's unit_length field
+--- @param main_cu_end_excl  integer  -- 0-based offset of the first byte AFTER the main CU
 --- @param new_abbrev_offset integer  -- 0-based offset into the new .debug_abbrev of the duplicate main table
---- @param atom_table table[]
---- @param rbind_structs table     -- {[binds_name] = {bytes, fields, atom_names}}
---- @param loclists_offsets table  -- {[atom_name] = section-relative offset}
---- @param registries table        -- merged registries from collect_per_source_registries
+--- @param atom_table        table[]
+--- @param rbind_structs     table    -- {[binds_name] = {bytes, fields, atom_names}}
+--- @param loclists_offsets  table    -- {[atom_name] = section-relative offset}
+--- @param registries        table    -- merged registries from collect_per_source_registries
 --- @return string  -- the rebuilt .debug_info bytes
 local function build_debug_info_section(existing, main_cu_start, main_cu_end_excl, new_abbrev_offset, atom_table, rbind_structs, loclists_offsets, registries)
 	-- 1) Build the inserted children bytes (just before the main CU's root terminator).
@@ -2133,13 +2248,13 @@ local SECTION_WRITERS = {
 -- Write a list of `{name, data}` section records to disk via SECTION_WRITERS.
 -- @param results table[]  -- list of `{name=, data=}` records to write
 -- @param ctx PassCtx
--- @param basename string   -- output file basename (e.g. "hello_gte")
--- @return table             -- list of {name_bin = path} entries to append to M.run's outputs
+-- @param basename string  -- output file basename (e.g. "hello_gte")
+-- @return table           -- list of {name_bin = path} entries to append to M.run's outputs
 local function write_sections(results, ctx, basename)
 	local outputs = {}
 	for _, r in ipairs(results) do
 		local path = SECTION_WRITERS[r.name](ctx.out_root, basename)
-		local f = io.open(path, "wb")
+		local  f   = io.open(path, "wb")
 		if not f then
 			io.stderr:write(string.format("[dwarf_injection] failed to open %s for write\n", path))
 		else
@@ -2166,7 +2281,7 @@ function M.run(ctx)
 	end
 
 	-- Guard: --elf is required.
-	local elf_path = ctx.flags and ctx.flags.elf_path
+	local  elf_path = ctx.flags and ctx.flags.elf_path
 	if not elf_path or elf_path == "" then
 		io.stderr:write("[dwarf_injection] --elf flag missing\n")
 		return { outputs = {}, errors = {}, warnings = {} }
@@ -2177,7 +2292,8 @@ function M.run(ctx)
 
 	-- Read the existing DWARF sections directly (no subprocess; io.open + manual ELF32 section-header walk).
 	-- We need all 8 sections: .debug_line / .debug_aranges / .debug_rnglists get extended (additional rows appended to the existing unit),
-	-- and .debug_info / .debug_abbrev / .debug_str / .debug_loc / .debug_loclists get spliced (the main CU's unit_length is patched; no new compile unit is appended; .debug_loc/.debug_loclists may not exist in the source ELF so we add-section them on splice).
+	-- and .debug_info / .debug_abbrev / .debug_str / .debug_loc / .debug_loclists get spliced
+	-- (the main CU's unit_length is patched; no new compile unit is appended; .debug_loc/.debug_loclists may not exist in the source ELF so we add-section them on splice).
 	-- The per-section dispatch is inlined in the writers loop below.
 	local existing_sections = elf_dwarf.read_elf_sections(elf_path, {
 		".debug_line", ".debug_aranges", ".debug_rnglists",
@@ -2186,14 +2302,17 @@ function M.run(ctx)
 		-- reading them just returns "" which is the "missing" case the builder handles.
 		".debug_loc", ".debug_loclists",
 	})
+	-- Resolve the per-file line-table indices from the same .debug_line bytes;
+	-- this MUST run before any atom sequence is emitted (build_atom_sequence below
+	-- calls resolve_provenance_file_index when populating call-site / body rows).
+	init_file_index_lookup(elf_path)
 	-- Skip state lives in `corpus.atoms_by_name[*].debug_skip` (whole-atom) and `atom.paths.invocations[*].debug_skip` (per-invocation).
 	-- `corpus` is the sole canonical source projection.
-	local corpus = (ctx.shared and ctx.shared.corpus) or {}
+	local corpus     = (ctx.shared and ctx.shared.corpus) or {}
 	local registries = collect_per_source_registries(corpus)
-	-- Read nm symbols (the ONLY disk-side input to the atom table) and join
-	-- them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
+	-- Read nm symbols (the ONLY disk-side input to the atom table) and join them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
 	-- Disk source-map/provenance text is not consulted (those are diagnostic artifacts; semantic inputs are in memory).
-	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local addrs      = elf_dwarf.read_nm(ctx.flags.elf_path)
 	local atom_table = build_atom_table(corpus, addrs)
 
 	-- Detect rbind atoms + index Binds_* struct fields from the corpus.
@@ -2217,7 +2336,7 @@ function M.run(ctx)
 		duffle.ensure_dir(ctx.out_root)
 
 		-- Step 0: layout validation. Bail out safely if the .debug_info layout doesn't match what we expect (crt CU + DWARF5 main CU + final 0 byte).
-	-- A layout mismatch means the gcc emission changed; the safest response is to leave existing sections unchanged and emit no synthetic data, so the build's debug-info step never silently produces broken DWARF.
+		-- A layout mismatch means the gcc emission changed; the safest response is to leave existing sections unchanged and emit no synthetic data, so the build's debug-info step never silently produces broken DWARF.
 		local existing_info   = existing_sections[".debug_info"]   or ""
 		local existing_abbrev = existing_sections[".debug_abbrev"] or ""
 		local main_cu_start, main_cu_end_excl, main_abbrev_offset = find_main_cu_layout(existing_info)
@@ -2252,7 +2371,7 @@ function M.run(ctx)
 		local new_info  = build_debug_info_section(existing_info, main_cu_start, main_cu_end_excl, new_abbrev_offset, atom_table, rbind_structs, loclists_offsets, registries)
 
 		-- Step 2b: rebuild .debug_str now that we know which RR_<R_Name> entries get emitted.
-	-- This aligns with build_debug_info_section's by_alias loop.
+		-- This aligns with build_debug_info_section's by_alias loop.
 		local new_str = build_debug_str_section(existing_sections[".debug_str"] or "", atom_table, registries)
 
 		-- Step 3-5: independent sections.
@@ -2277,5 +2396,6 @@ end
 M.compute_loclists_offsets_for_test     = compute_loclists_offsets
 M.build_debug_loclists_section_for_test = build_debug_loclists_section
 M.tape_piece_size_for_test              = tape_piece_size
+M.build_atom_table_for_test             = build_atom_table
 
 return M

scripts/passes/emission_model.lua

@@ -7,7 +7,7 @@
 --- Public boundary:
 ---   * `M.run(ctx)` is the only entry point.
 ---   * The pass returns `{outputs = {}, errors = ..., warnings = ...}`.
----     Pass kind = `validation` → `PASS_KIND_STOP_ON_ERROR.validation` preserves the existing build-stopping policy.
+---     Pass kind = `validation`. Findings record on the result; the orchestrator does not exit non-zero.
 ---
 --- Source-order discipline:
 ---   * `corpus.source_order` sets the source-record order.
@@ -41,7 +41,6 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- Convert the recursive walk's body-relative line numbers into physical source lines once.
 -- The walker builds `line_of` from `body_text` and stamps body-relative line numbers (1..N) into `item.line` and `invocation.call_line`.
 -- This function converts those values to physical source lines at the close site with the forwarded source `line_of` closure.
---
 -- `call_line` discipline:
 --   * ROOT invocations (`inv.parent_id == 0`) receive body-relative `call_line` values directly from `M.LineIndex(body_text)` in the walker.
 --     The source `line_of` closure supplies physical lines at the close site, so this function converts each root value exactly once.
@@ -59,10 +58,9 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	-- `root_body_line` is the physical source line of the ATOM HEADER byte containing the opening `{`; that byte is one byte BEFORE `atom_record.body_off`.
 	-- The walker assigns line 2 to the body's first content line because line 1 is the trailing `\n` after `{`. Body-text line k therefore maps to `root_body_line + (k - 1)`.
 	-- `body_off - 1` points at the opening `{`, whose line index identifies the header line. `body_off` points after `{` and would shift every word row forward by one line.
-	local root_body_line = root_line_of(atom_record.body_off - 1)
-		or atom_record.line or 0
+	local root_body_line  = root_line_of(atom_record.body_off - 1) or atom_record.line or 0
 	local component_index = corpus.component_body_index or {}
-	local word_items = {}
+	local word_items      = {}
 
 	for _, item in ipairs(projection.items) do
 		if item.kind == "word" then word_items[#word_items + 1] = item end
@@ -102,7 +100,7 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	end
 
 	-- Normalize `inv.call_line` to a physical source line.
-	--   * ROOT invocations (`parent_id == 0`) carry body-relative `call_line` values from `M.LineIndex(body_text)`; convert them once with `root_body_line`.
+	--   * ROOT  invocations (`parent_id == 0`) carry body-relative `call_line` values from `M.LineIndex(body_text)`; convert them once with `root_body_line`.
 	--   * INNER invocations (`parent_id ~= 0`) carry physical `call_line` values from the component's `line_of`; retain them unchanged.
 	for _, inv in ipairs(projection.invocations) do
 		if inv.parent_id == 0 then
@@ -114,14 +112,14 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	-- `atoms_source_map` and `dwarf_injection` read `inv.body_lines[k]` directly from the invocation record created here.
 	-- Component words already carry physical `item.line` values from the walker's COMPONENT line index, so `body_line_for` returns them unchanged.
 	for _, inv in ipairs(projection.invocations) do
-		local sw = inv.start_word
-		local ew = inv.end_word
+		local sw  = inv.start_word
+		local ew  = inv.end_word
 		local bls = {}
 		for i = sw, ew do
 			local it = projection.items and projection.items[i]
-			if it and it.kind == "word" then
+			if    it and it.kind == "word" then
 				local fake_event = { invocation_ids = { inv.id } }
-				bls[#bls + 1] = body_line_for(fake_event, it) or 0
+				bls[#bls + 1]    = body_line_for(fake_event, it) or 0
 			end
 		end
 		inv.body_lines = bls
@@ -157,8 +155,28 @@ local function project_atom(atom_record, src, corpus)
 	local body  = atom_record.body or ""
 	local wc    = corpus.word_counts or {}
 	local cbi   = corpus.component_body_index or {}
+	local schema = nil
+	if atom_record.reg_use_schema_name then
+		schema = corpus.reg_use_schemas and corpus.reg_use_schemas[atom_record.reg_use_schema_name]
+	end
 	-- That construction site stamps `invocation.debug_skip` while appending each record to `proj.invocations`.
-	local proj  = duffle.project_emission(body, cbi, wc, corpus.components)
+	local proj  = duffle.project_emission(body, cbi, wc, corpus.components, {
+		reg_use_schema = schema,
+		reg_use_param  = atom_record.reg_use_param_name,
+		atom_name      = atom_record.name,
+		schema_name    = atom_record.reg_use_schema_name,
+	})
+	if atom_record.reg_use_schema_name and not schema then
+		proj.errors[#proj.errors + 1] = {
+			kind = "reguse_missing_schema",
+			msg  = string.format("RegUse schema %q is missing", atom_record.reg_use_schema_name),
+		}
+	end
+	for _, err in ipairs(corpus.reg_use_errors or {}) do
+		if err.schema_name == atom_record.reg_use_schema_name then
+			proj.errors[#proj.errors + 1] = err
+		end
+	end
 	local paths = {
 		tokens       = atom_record.body_tokens or {},
 		line_in_body = duffle.build_body_line_index(body),
@@ -190,11 +208,11 @@ function M.run(ctx)
 	if type(corpus.source_order) ~= "table" then error("emission_model: ctx.shared.corpus.source_order is required", 0) end
 
 	-- Project once, collect errors + warnings for one atom.
-	-- Kind must be one of: atom | raw_atom | comp_bare | comp_proc.
+	-- Kind must be one of: atom | atom_proc | raw_atom | comp_bare | comp_proc.
 	local function process_atom(atom, src)
 		if not (atom and atom.body) then return end
 		local kind = atom.kind
-		if kind ~= "atom" and kind ~= "raw_atom" and kind ~= "comp_bare" and kind ~= "comp_proc" then
+		if kind ~= "atom" and kind ~= "atom_proc" and kind ~= "raw_atom" and kind ~= "comp_bare" and kind ~= "comp_proc" then
 			return
 		end
 		local proj = project_atom(atom, src, corpus)
@@ -217,7 +235,7 @@ function M.run(ctx)
 	end
 
 	-- Walk `corpus.source_order`; within each source, visit atoms followed by raw_atoms.
-	-- Recognized kinds (atom | raw_atom | comp_bare | comp_proc) each receive the atom.paths projection via duffle.project_emission.
+	-- Recognized kinds (atom | atom_proc | raw_atom | comp_bare | comp_proc) each receive the atom.paths projection via duffle.project_emission.
 	-- Components are macros inlined into atom bodies; focused tests and isolated component analyses consume atom.paths directly.
 	for _, src in ipairs(corpus.source_order) do
 		local scan = src.scan or {}

scripts/passes/offsets.lua

@@ -1,14 +1,22 @@
 --- passes/offsets.lua — Branch-offset generator.
 ---
 --- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
---- for `MipsAtom_(name)` and `MipsCode code_<name>` declarations, computes the word offset
+--- for `MipsAtom_(name)` and leftover `MipsCode code_*` declarations, computes the word offset
+--- (ELF symbol is the C ident; raw `code_*` is leftover, not the atom rule)
 --- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
---- `<dir_basename>.offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+--- `gen/offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+---
+--- Per-directory aggregation: every source in the same directory contributes to the same `gen/offsets.h`.
+--- The directory itself is the namespace; the filename does not repeat the module name.
+---
+--- (Task 12.16 note: atom-namespaced enum names — e.g., `atom_offset__normalize_v3s4__srav_path__aligned_done` —
+--- were considered to prevent cross-atom label collisions, but the C-side `atom_offset(F, T)` macro in
+--- `code/duffle/dsl.atom.h` doesn't know the current atom_name at expansion time, so any namespacing
+--- on the metaprogram side breaks the C build. Reverted. The C-side would need a per-atom
+--- `CURRENT_ATOM` #define (set by `MipsAtom_`/`MipsAtom_Proc_` macros) plus an updated `atom_offset`
+--- macro that uses it. That's a coordinated refactor — deferred to a future track.)
 ---
 --- The offset is `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding: branch_offset = relative_pc_in_words - 1).
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -16,12 +24,11 @@
 
 -- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
--- both standalone and when require'd from the orchestrator.
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -39,40 +46,42 @@ local OFFSET_MACRO_COL = 44
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field shared             table                  -- cross-pass shared state
---- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared             table -- Cross-pass shared state
+--- @field shared.corpus      table -- Corpus projection
 --- @field shared.word_counts table
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string -- Output root (e.g. "build/gen")
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
---- @field errors   table[]  -- {line=, msg=} entries; build-stops
---- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+--- @field errors   table[]  -- {line=, msg=}  entries; build-stops
+--- @field warnings table[]  -- {line=, msg=}  entries; build-succeeds
 
 --- @class BranchOffset
---- @field tag         string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
---- @field target      string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field branch_word integer  -- branch word position within the atom body
---- @field offset      integer  -- computed `target_word - branch_word - 1`
+--- @field tag               string      -- Marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field target            string      -- Target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field branch_word       integer     -- Branch word position within the atom body
+--- @field offset            integer     -- Computed per consuming instruction (see `compute_offsets`)
+--- @field consuming_encoder string|nil  -- Instruction consuming the offset (e.g. "branch_le_zero", "jump", "call_addr")
+--- @field consuming_arg_pos integer|nil -- 1-based arg position within the consuming instruction's arg list
 
 --- @class AtomData
---- @field name        string         -- atom name
---- @field total_words integer        -- total word count of the atom body
---- @field offsets     BranchOffset[] -- per-branch offset list
+--- @field name        string         -- Atom name
+--- @field total_words integer        -- Total word count of the atom body
+--- @field offsets     BranchOffset[] -- Per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Canonical marker projection
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- MARKER_PROJECTORS is the marker-kind data table.
--- The emission-model pass already records marker word positions;
+-- The emission-model pass already records marker word positions + consuming-instruction context;
 -- this pass only projects those records into the label/branch lookup shape needed by offset computation.
 local MARKER_PROJECTORS = {
 	label = function(state, marker)
@@ -80,9 +89,11 @@ local MARKER_PROJECTORS = {
 	end,
 	offset = function(state, marker)
 		state.branches[#state.branches + 1] = {
-			tag         = marker.name,
-			target      = marker.target,
-			branch_word = marker.word_index,
+			tag                = marker.name,
+			target             = marker.target,
+			branch_word        = marker.word_index,
+			consuming_encoder  = marker.consuming_encoder,
+			consuming_arg_pos  = marker.consuming_arg_pos,
 		}
 	end,
 }
@@ -95,7 +106,7 @@ local function project_markers(markers)
 	local state = { labels = {}, branches = {} }
 	for _, marker in ipairs(markers or {}) do
 		local project = MARKER_PROJECTORS[marker.kind]
-		if project then project(state, marker) end
+		if    project then project(state, marker) end
 	end
 	return state.labels, state.branches
 end
@@ -104,23 +115,52 @@ end
 -- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
---- @param labels table<string, integer>
+--- Compute branch offsets per consuming instruction.
+--- Disposition table:
+---   `branch_*`           -> relative offset: `target_word - branch_word - 1` (MIPS branch-immediate encoding).
+---   `jump` / `call_addr` -> same value as `branch_*` (a relative word offset). 
+---     The duffle headers' `enc_i` macro truncates the value to the immediate-field width (16 bits for branches, 26 bits for jumps).
+---     For tape-atom bodies within a single module, this works for `j`/`jal` because the linker's symbol resolution produces the correct 26-bit absolute target via standard `j` relocations.
+---     For cross-module `j`/`jal` (atom body in one module, target in another), the linker emits a `R_MIPS_26` relocation against the lower 26 bits; the upper 4 bits come from the PC of the delay slot following the `j`.
+---     The metaprogram doesn't know either at compile time, so the emitted value is the relative word offset that the duffle `enc_i` macro places in the immediate field; the toolchain handles the rest.
+---   `jump_reg` / `call_reg` / `jump_link` -> ERROR. Register-form jumps have no offset field; `atom_offset` is invalid.
+---
+--- Top-level `atom_offset(F, T)` markers (where the marker is the entire token — `consuming_encoder` == nil) default to `branch_*` behavior (relative offset).
+--- This preserves backward compatibility for any top-level marker that may exist outside a control-transfer instruction.
+--- @param labels   table<string, integer>
 --- @param branches table[]
+--- @param errors   table[]
 --- @return BranchOffset[]
-local function compute_offsets(labels, branches)
+local function compute_offsets(labels, branches, errors)
 	local results = {}
 	for _, br in ipairs(branches) do
 		local target = labels[br.target]
 		if not target then
-			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")")
+			errors[#errors + 1] = {
+				line = br.line or 0,
+				msg  = "Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")",
+			}
+		else
+			local consuming = br.consuming_encoder
+			if consuming == "jump_reg" or consuming == "call_reg" or consuming == "jump_link" then
+				errors[#errors + 1] = {
+					line = br.line or 0,
+					msg  = "atom_offset cannot be used with " .. consuming
+						.. " (register-form jumps have no offset field); at word " .. br.branch_word,
+				}
+			else
+				-- All other consuming instructions (including `branch_*`, `jump`, `call_addr`, and nil for top-level markers) use the same relative offset value.
+				-- The MIPS encoding differs per opcode but the duffle `enc_i` macro handles the truncation to the immediate-field width.
+				results[#results + 1] = {
+					target             = br.target,
+					tag                = br.tag,
+					branch_word        = br.branch_word,
+					offset             = target - br.branch_word - 1,
+					consuming_encoder  = br.consuming_encoder,
+					consuming_arg_pos  = br.consuming_arg_pos,
+				}
+			end
 		end
-		results[#results + 1] = {
-			target      = br.target,
-			tag         = br.tag,
-			branch_word = br.branch_word,
-			offset      = target - br.branch_word - 1,
-		}
 	end
 	return results
 end
@@ -167,66 +207,74 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
---- Generate the per-source .offsets.h header.
---- @param source_path string
---- @param atoms_data  AtomData[]
+--- Generate the per-directory .offsets.h header.
+--- @param dir          string     -- the absolute source directory
+--- @param sources      table[]    -- sources contributing to this directory (for the header comment)
+--- @param atoms_data   AtomData[]
 --- @return string
-local function generate_header(source_path, atoms_data)
-	local basename = duffle.basename_no_ext(source_path)
+local function generate_header(dir, sources, atoms_data)
+	local dir_basename = duffle.basename_no_ext(dir)
 
 	local lines = {}
 	local function add(s) lines[#lines + 1] = s end
 
 	add("// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT")
-	add("// Source: " .. source_path)
+	add("// Directory: " .. dir:gsub("/", "\\") .. "\\")
+	for _, src in ipairs(sources) do
+		add("//   source: " .. src.path:gsub("/", "\\"))
+	end
 	add("#pragma once")
 	add("")
-	add("#pragma region " .. basename)
+	add("#pragma region " .. dir_basename)
 	add("")
 	add("")
 	for _, atom in ipairs(atoms_data) do
 		emit_atom_offsets(add, atom)
 	end
-	add("#pragma endregion " .. basename)
+	add("#pragma endregion " .. dir_basename)
 	add("")
 	return table.concat(lines, "\n") .. "\n"
 end
 
 local M = {}
 
---- (internal) Process one source: render offsets from canonical atom paths.
+--- (internal) Aggregate atoms from every source in one directory, render the per-directory `offsets.h`.
 --- Returns the offsets_h path if a header was written, or nil.
---- @param ctx PassCtx
---- @param src SourceFile
+--- @param ctx     PassCtx
+--- @param dir     string       -- the absolute source directory
+--- @param sources SourceFile[] -- sources in this directory
+--- @param errors  table[]
 --- @return string|nil  -- the offsets_h path
-local function process_source(ctx, src)
+local function process_directory(ctx, dir, sources, errors)
 	local atoms_data = {}
-	local scan = src.scan or {}
 
 	local function append_atom(atom)
-		local paths = atom and atom.paths
+		local  paths = atom and atom.paths
 		if not paths then return end
 		local labels, branches = project_markers(paths.markers)
 		atoms_data[#atoms_data + 1] = {
 			name        = atom.raw_name or atom.name,
 			total_words = #(paths.word_events or {}),
-			offsets     = compute_offsets(labels, branches),
+			offsets     = compute_offsets(labels, branches, errors),
 		}
 	end
 
-	for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
-	for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	for _, src in ipairs(sources) do
+		local scan = src.scan or {}
+		for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+		for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	end
 	if #atoms_data == 0 then return nil end
 
-	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
+	local out_path = dir .. "/gen/offsets.h"
 	duffle.ensure_dir(duffle.dirname(out_path))
-	duffle.write_file(out_path, generate_header(src.path:gsub("/", "\\"), atoms_data))
+	duffle.write_file(out_path, generate_header(dir, sources, atoms_data))
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
---- containing constants for every marker recorded in atom.paths.
+--- For each canonical source-directory, emits a per-directory `gen/offsets.h`
+--- containing constants for every marker recorded in atom.paths across every source in that directory.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -235,12 +283,17 @@ function M.run(ctx)
 	local warnings = {}
 
 	local corpus = ctx.shared and ctx.shared.corpus
-	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
-		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	if type(corpus) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order.", 0)
 	end
 
-	for _, src in ipairs(corpus.source_order) do
-		local out_path = process_source(ctx, src)
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`.
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
+	for dir, sources in pairs(sources_by_dir) do
+		local out_path = process_directory(ctx, dir, sources, errors)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }
 		end

scripts/passes/word_count_eval.lua

@@ -93,7 +93,7 @@ end
 --- Load the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts`.
 --- Generated `.macs.h` files are OUTPUT artifacts and are NOT scanned as inputs.
 --- Current component counts are computed and inserted by `passes/components.lua`
---- after the components pass iterates `corpus.source_order` and writes each source's `<dir_basename>.macs.h` file.
+--- after the components pass iterates `corpus.source_order` and writes each source-directory's `gen/macs.h` file.
 ---
 --- Contract:
 ---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).

scripts/ps1_meta.lua

@@ -2,24 +2,21 @@
 ---
 --- Dispatches to pass modules under `scripts/passes/`, resolving dependencies topologically (Kahn's algorithm + cycle detection).
 ---
---- **Architecture**:
----   - **PASSES table** — declarative dep graph (data, not code).
----   - **FLAG_HANDLERS table** — maps CLI flags to handlers.
----   - **parse_args** → **build_ctx** (resolves unity/direct includes or exact sources; no semantic scanning) → **topo_sort** → **dispatch_passes**.
+--- Architecture:
+---   - PASSES table: Declarative dep graph (data, not code).
+---   - FLAG_HANDLERS table: Maps CLI flags to handlers.
+---   - parse_args → build_ctx (resolves unity/direct includes or exact sources) → topo_sort → dispatch_passes.
 ---   - The first pass in the dep graph is `scan-source` (see `passes/scan_source.lua`).
 ---     It calls `duffle.scan_source` once per source to produce the fat `SourceScan` payload, which is attached to each `src.scan`. 
 ---     Every other pass that reads source structure depends on `scan-source` and consumes `src.scan` as a read-only. 
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
----
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Bootstrap: load `duffle_paths.lua` via this script's own path.
 -- Use `arg[0]` when this file is the entry script (`arg[0]` ends in "ps1_meta.lua");
--- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path, not ours).
+-- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path).
 -- That single statement: (a) sets `package.path` + `package.cpath`, (b) at the bottom returns `require("duffle")`.
 -- So the dofile's return value is the duffle module.
 local _is_entry_script = arg and arg[0] and arg[0]:match("ps1_meta%.lua$") ~= nil
@@ -57,45 +54,44 @@ local PASS_FLAG_DISPATCH_KEY   = "__pass__"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class PassDescriptor
---- @field module string       -- module name passed to require()
+--- @field module string       -- Module name passed to require()
 --- @field kind   string       -- "shared" | "header-output" | "validation" | "diagnostic" | "report"
 ---                            -- Report severity is independent from process exit policy (see PASS_KIND_STOP_ON_ERROR).
---- @field deps   string[]     -- names of upstream passes
---- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of
----                            -- (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
+--- @field deps   string[]     -- Names of upstream passes
+--- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field metadata_path      string                 -- path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
---- @field shared.corpus      table                  -- canonical authored-source/project projection
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
+--- @field shared             table                  -- Cross-pass shared state
+--- @field shared.corpus      table                  -- Authored-source/project projection
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
 --- @field project_root       string                 -- PS1 repository root
 --- @field flags              table                  -- CLI flags + per-pass stash
---- @field verbose            boolean                -- if true, log diagnostic info
+--- @field verbose            boolean                -- If true, log diagnostic info
 
 --- @class Finding
---- @field line integer  -- source line (or 0 for pass-level)
---- @field msg  string   -- finding message
+--- @field line integer  -- Source line (or 0 for pass-level)
+--- @field msg  string   -- Finding message
 
 --- @class PassResult
---- @field outputs  PassOutputEntry[] -- emitted file paths
---- @field errors   Finding[]         -- build-stops (per-pass kind policy)
---- @field warnings Finding[]         -- informational
+--- @field outputs  PassOutputEntry[] -- Emitted file paths
+--- @field errors   Finding[]         -- Build-stops (per-pass kind policy)
+--- @field warnings Finding[]         -- Informational
 
 --- @class ParsedArgs
---- @field requested_set string[]   -- pass names to run (explicit --all expanded)
---- @field sources       string[]   -- exact --source values, retained in CLI order
+--- @field requested_set string[]   -- Pass names to run (explicit --all expanded)
+--- @field sources       string[]   -- Exact --source values, retained in CLI order
 --- @field unity_root    string|nil -- --unity-root value; mutually exclusive with sources
 --- @field metadata      string     -- --metadata value
 --- @field out_root      string     -- --out-root value (default "build/gen")
 --- @field project_root  string     -- PS1 repository root (derived from metadata by default)
---- @field verbose       boolean    -- if true, log diagnostic info
+--- @field verbose       boolean    -- If true, log diagnostic info
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- PASSES Table
@@ -122,6 +118,12 @@ local PASSES = {
 		kind   = "header-output",
 		deps   = {"scan-source", "word-counts"},
 	},
+	auto_reg = {
+		module = "passes.auto_reg",
+		kind   = "header-output",
+		deps   = {"components"},
+		groups = { "pre-link" },
+	},
 	["emission-model"] = {
 		module = "passes.emission_model",
 		kind   = "validation",
@@ -140,8 +142,7 @@ local PASSES = {
 	},
 	["static-analysis"] = {
 		module = "passes.static_analysis",
-		-- "diagnostic" — every `error`/`warning` finding is written to the report file;
-		-- the orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
+		-- "diagnostic" — every `error`/`warning` finding is written to the report file.
 		-- Report severity is independent from process exit policy.
 		kind   = "diagnostic",
 		deps   = {"scan-source", "word-counts", "components", "emission-model"},
@@ -160,18 +161,18 @@ local PASSES = {
 	report = {
 		module = "passes.report",
 		kind   = "report",
-		deps   = {"annotation", "static-analysis"},
+		deps   = {"annotation", "static-analysis", "atoms-source-map"},  -- +atoms-source-map (consolidated-report-files refactor, 2026-07-26)
 		groups = { "pre-link" },
 	},
 }
 
 -- ────────────────────────────────────────────────────────────────────────────
--- Phase-root selection: derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
+-- Phase-root selection: Derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
 -- topo_sort closes the transitive deps from there; dispatch_passes runs every resolved pass without phase-filtering.
 -- ────────────────────────────────────────────────────────────────────────────
 
---- @param group_name string  -- the build-phase group ("pre-link" | "post-link")
---- @return string[]          -- sorted root pass names belonging to that group
+--- @param group_name string  -- Build-phase group ("pre-link" | "post-link")
+--- @return string[]          -- Sorted root pass names belonging to that group
 local function roots_for_group(group_name)
 	local names = {}
 	for name, pass in pairs(PASSES) do
@@ -204,22 +205,19 @@ local function request_roots_for_group(args, group_name)
 	end
 end
 
--- Pass-kind taxonomy: Which kinds stop the build on errors?
---
--- Report severity is independent from process exit policy. A "diagnostic" pass still writes every `error`/`warning` finding into its report file,
--- but `report_validation_errors` returns early for non-stopping kinds, so nothing is printed to stderr and the orchestrator does not exit non-zero.
+-- Pass-kind taxonomy: findings always print. No pass kind stops the build.
+-- Report severity is independent from process exit policy.
 -- Adding a new pass kind requires listing it here explicitly; an unknown kind must not silently fall back to "true".
 local PASS_KIND_STOP_ON_ERROR = {
 	["shared"]        = false,
-	["header-output"] = true,
-	["validation"]    = true,
+	["header-output"] = false,
+	["validation"]    = false,
 	["diagnostic"]    = false,
 	["report"]        = false,
 }
 
 -- Closed set of CLI flags -> pass names.
--- Per-pass flags (e.g. --word-counts) live here; phase flags (--pre-link, --post-link, --all)
--- live in FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
+-- Per-pass flags (e.g. --word-counts); phase flags (--pre-link, --post-link, --all) are within FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
 -- dwarf-injection is *also* a per-pass opt-in flag, but its selection + opt-in state are both owned by the explicit FLAG_HANDLERS entry below
 -- (it sets args.flags.dwarf_injection and appends "dwarf-injection" to requested_set), so it is intentionally absent from this table.
 local PASS_FLAG_TO_NAME = {
@@ -248,7 +246,6 @@ end
 
 -- Per-flag handlers. Each handler takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it). 
 -- Returning nil + os.exit() handles termination flags (--help). 
-
 local FLAG_HANDLERS = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -274,9 +271,9 @@ PASS_FLAGS:
   Or pick any subset:
     --scan-source         Scan sources into the fat SourceScan payload
     --word-counts         Load metadata.h + scan for existing .macs.h
-    --components          Generate <module>/gen/<basename>.macs.h
+    --components          Generate <srcdir>/gen/macs.h (per-directory aggregation)
     --validate            Run atom annotation DSL validation
-    --offsets             Generate <module>/gen/<basename>.offsets.h
+    --offsets             Generate <srcdir>/gen/offsets.h (per-directory aggregation)
     --atoms-source-map    Generate <basename>.atoms.sourcemap.txt per source
     --dwarf-injection     [opt-in] Select the post-link dwarf-injection pass + set the opt-in flag. Requires --elf.
     --static-analysis     Static analysis: GTE pipeline-fill, mac_yield, ABI handoff, cycle budget
@@ -319,8 +316,7 @@ local function require_flag_value(argv, arg_idx, flag)
 	local next_known = type(value) == "string"
 		and (FLAG_HANDLERS[value] ~= nil or PASS_FLAG_TO_NAME[value] ~= nil)
 	if value == nil or next_known then
-		io.stderr:write("ps1_meta: " .. flag .. " requires "
-			.. FLAG_VALUE_NAMES[flag] .. "\n")
+		io.stderr:write("ps1_meta: " .. flag .. " requires " .. FLAG_VALUE_NAMES[flag] .. "\n")
 		os.exit(EXIT_INTERNAL_ERROR)
 	end
 	return value, arg_idx + 1
@@ -330,12 +326,10 @@ end
 -- Termination flags like --help call os.exit() instead.
 -- Populated AFTER print_help so the --help handler can reference it as an upvalue (Lua resolves locals at closure-call time, 
 -- but if the closure is defined before the local, it falls back to _G). 
-FLAG_HANDLERS["--help"] = function(args)
-	print_help()
-	os.exit(0)
-end
 
-FLAG_HANDLERS["--verbose"] = function(args) args.verbose = true end
+FLAG_HANDLERS["--help"]    = function(args) print_help(); os.exit(0) end
+FLAG_HANDLERS["--verbose"] = function(args) args.verbose = true      end
+
 FLAG_HANDLERS["--source"] = function(args, argv, arg_idx)
 	local value, value_idx = require_flag_value(argv, arg_idx, "--source")
 	args.sources[#args.sources + 1] = value
@@ -492,14 +486,13 @@ end
 --- @param args ParsedArgs
 --- @return PassCtx
 local function build_ctx(args)
-	local normalized_project_root = duffle.normalize_path(args.project_root)
-	local project_root            = normalized_project_root
+	local normalized_project_root  = duffle.normalize_path(args.project_root)
+	local project_root             = normalized_project_root
 	local project_root_is_absolute = normalized_project_root:match("^%a:/")
 		or normalized_project_root:sub(1, 2) == "//"
 		or normalized_project_root:sub(1, 1) == "/"
 	if not project_root_is_absolute then
-		-- canonical_path_key validates ordinary relative paths and rejects
-		-- drive-relative paths before the absolute-path rewrite is performed.
+		-- canonical_path_key validates ordinary relative paths and rejects drive-relative paths before the absolute-path rewrite is performed.
 		duffle.canonical_path_key(normalized_project_root)
 		project_root = duffle.normalize_path(duffle.to_absolute_path(normalized_project_root))
 	else
@@ -513,8 +506,7 @@ local function build_ctx(args)
 			project_root = project_root,
 		})
 		if not ok_resolve then
-			io.stderr:write("ps1_meta: cannot resolve --unity-root "
-				.. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
+			io.stderr:write("ps1_meta: cannot resolve --unity-root " .. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
 			os.exit(EXIT_INTERNAL_ERROR)
 		end
 		resolution = resolved
@@ -530,8 +522,7 @@ local function build_ctx(args)
 			local path = duffle.normalize_path(input_path)
 			local key_ok, key_or_error = pcall(duffle.canonical_path_key, path)
 			if not key_ok then
-				error("ps1_meta: invalid --source " .. input_path .. ": "
-					.. tostring(key_or_error), 0)
+				error("ps1_meta: invalid --source " .. input_path .. ": " .. tostring(key_or_error), 0)
 			end
 			local file = io.open(path, "r")
 			if not file then
@@ -629,9 +620,7 @@ local function topo_sort(passes, requested_set)
 		changed = false
 		for name, _ in pairs(needed) do
 			local pass = passes[name]
-			if not pass then
-				error("unknown pass '" .. name .. "' requested")
-			end
+			if not pass then error("unknown pass '" .. name .. "' requested") end
 			for _, dep in ipairs(pass.deps) do
 				if not needed[dep] then
 					needed[dep] = true
@@ -700,19 +689,19 @@ end
 -- Main Orchestrator
 -- ════════════════════════════════════════════════════════════════════════════
 
---- (internal) If the pass's kind is in PASS_KIND_STOP_ON_ERROR and it reported errors, write each error to stderr.
---- Returns true if any validation errors were reported.
+--- (internal) Write every pass error to stderr.
+--- Returns true only when the pass kind still stops the build.
 --- @param pass_name string
 --- @param pass      PassDescriptor
 --- @param result    PassResult
 --- @return boolean
 local function report_validation_errors(pass_name, pass, result)
-	local   has_errors = result.errors and #result.errors > 0
-	if not (has_errors and PASS_KIND_STOP_ON_ERROR[pass.kind]) then return false end
+	local has_errors = result.errors and #result.errors > 0
+	if not has_errors then return false end
 	for _, e in ipairs(result.errors) do
 		io.stderr:write(string.format("[%s] line %d: %s\n", pass_name, e.line or 0, e.msg or ""))
 	end
-	return true
+	return PASS_KIND_STOP_ON_ERROR[pass.kind] == true
 end
 
 --- (internal) Run each pass in `order` in topological sequence.

scripts/update_deps.ps1

@@ -14,16 +14,21 @@ $url_armips     = 'https://github.com/Kingcom/armips.git'
 $url_pcsx_redux = 'https://github.com/grumpycoders/pcsx-redux.git'
 $url_psyq_iwyu  = 'https://github.com/johnbaumann/psyq_include_what_you_use.git'
 $url_lpeg       = 'https://github.com/roberto-ieru/LPeg.git'
+# $url_mkpsxiso   = 'https://github.com/Lameguy64/mkpsxiso.git'
+
+$url_mkpsxiso_win64 = 'https://github.com/Lameguy64/mkpsxiso/releases/download/v2.30/mkpsxiso-2.30-win64.zip'
 
 $path_armips     = join-path $path_toolchain 'armips'
 $path_pcsx_redux = join-path $path_toolchain 'pcsx-redux'
 $path_psyq_iwyu  = join-path $path_toolchain 'psyq_iwyu'
 $path_lpeg       = join-path $path_toolchain 'lpeg'
+$path_mkpsxiso   = join-path $path_toolchain 'mkpsxiso'
 
 clone-gitrepo $path_armips     $url_armips
 clone-gitrepo $path_lpeg       $url_lpeg
 clone-gitrepo $path_pcsx_redux $url_pcsx_redux
 clone-gitrepo $path_psyq_iwyu  $url_psyq_iwyu
+# clone-gitrepo $path_mkpsxiso   $url_mkpsxiso
 
 $path_armips_build = join-path $path_armips 'build'
 verify-path $path_armips_build
@@ -56,7 +61,111 @@ if (-not $msbuild_exe) {
 }
 
 $path_pcsx_sln = join-path $path_pcsx_redux 'vsprojects\pcsx-redux.sln'
-&	$msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
+
+# ════════════════════════════════════════════════════════════════════════════
+# NuGet restore — required before MSBuild.
+# pcsx-redux's .vcxproj files use the legacy packages.config style with
+# hardcoded `<Import Project="..\packages\{id}.{ver}\...">` directives.
+# MSBuild's `/t:Restore` won't fetch missing packages here (the local
+# packages\ dir is checked but no package-source lookup happens), and
+# `dotnet restore` errors on packages.config projects, so we walk every
+# packages.config, parse out the <package id version/> entries, and pull
+# any missing .nupkg directly from api.nuget.org's flat container.
+# ════════════════════════════════════════════════════════════════════════════
+$path_pcsx_packages = join-path $path_pcsx_redux 'vsprojects\packages'
+$nuget_flat_container = 'https://api.nuget.org/v3-flatcontainer'
+
+# Collect required (id, version) pairs from every packages.config.
+$required_packages = @{}
+Get-ChildItem -Path (join-path $path_pcsx_redux 'vsprojects') -Filter 'packages.config' -Recurse -ErrorAction SilentlyContinue |
+    ForEach-Object {
+        [xml]$xml = Get-Content -LiteralPath $_.FullName -Raw
+        foreach ($pkg in $xml.packages.package) {
+            $key = '{0}|{1}' -f $pkg.id, $pkg.version
+            $required_packages[$key] = @{ id = $pkg.id; version = $pkg.version }
+        }
+    }
+
+# Ensure the packages root exists.
+if (-not (Test-Path -LiteralPath $path_pcsx_packages)) {
+    New-Item -ItemType Directory -Path $path_pcsx_packages -Force | Out-Null
+}
+
+# Download anything missing.
+# Skip the package entirely if its dir already has any contents (the legacy packages.config style means the targets file location varies per package 
+# — `luajit.native` puts it at build/native/, `glfw` puts it elsewhere — so we can't probe a specific path; just check whether the dir is non-empty).
+Add-Type -AssemblyName System.IO.Compression.FileSystem
+foreach ($pkg in $required_packages.Values) {
+    $pkgDir = Join-Path $path_pcsx_packages ('{0}.{1}' -f $pkg.id, $pkg.version)
+    if ((Test-Path -LiteralPath $pkgDir) -and `
+        (@(Get-ChildItem -LiteralPath $pkgDir -Recurse -ErrorAction SilentlyContinue).Count -gt 0)) {
+        continue
+    }
+    $url = '{0}/{1}/{2}/{1}.{2}.nupkg' -f $nuget_flat_container, $pkg.id, $pkg.version
+    $nupkg = Join-Path $pkgDir ('{0}.{1}.nupkg' -f $pkg.id, $pkg.version)
+    New-Item -ItemType Directory -Path $pkgDir -Force | Out-Null
+    Write-Host "Fetching NuGet package: $($pkg.id) $($pkg.version)"
+    try {
+        Invoke-WebRequest -Uri $url -OutFile $nupkg -UseBasicParsing -ErrorAction Stop
+        [System.IO.Compression.ZipFile]::ExtractToDirectory($nupkg, $pkgDir)
+        Remove-Item -LiteralPath $nupkg -Force
+    } catch {
+        $msg = $_.Exception.Message
+        if ($msg -match '404') {
+            Write-Host "  Not on nuget.org (vendored?) — skipping $url"
+        } else {
+            Write-Warning "Failed to fetch $url — $msg"
+        }
+        if (Test-Path -LiteralPath $nupkg) { Remove-Item -LiteralPath $nupkg -Force }
+    }
+}
+
+# ════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════
+# isoffi.lua size guard — `core.vcxproj` #includes src/core/isoffi.lua into luaiso.cc via the `-- lualoader, R"EOF(...)EOF"` trick.
+# The raw string literal between R"EOF(-- and -- )EOF" must stay under ~16,379 bytes or MSVC (19.44) fails with C2026 (its actual raw-string limit is 16,384, minus 5 bytes for the `-- lualoader, ` prefix).
+# If the upstream file grows past that, trim it: remove license header, trailing whitespace, blank separators, inline comments, and shrink 4-space indent to 2-space.
+# Idempotent — only writes when the raw string exceeds the limit.
+# ════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════
+$path_isoffi = join-path $path_pcsx_redux 'src\core\isoffi.lua'
+if (Test-Path -LiteralPath $path_isoffi) {
+    $content = Get-Content -LiteralPath $path_isoffi -Raw -Encoding utf8
+    $startMarker = $content.IndexOf('R"EOF(--')
+    $endMarker   = $content.IndexOf('-- )EOF"')
+    $literalLen  = if ($startMarker -ge 0 -and $endMarker -gt $startMarker) { $endMarker - ($startMarker + 8) } else { -1 }
+    # Effective MSVC raw-string limit for the lualoader prefix is 16379 bytes.
+    if ($literalLen -gt 16379) {
+        Write-Host "isoffi.lua raw string is $literalLen bytes (>16379); trimming for MSVC C2026 limit."
+        $lines     = $content -split "`n"
+        $markerIdx = -1
+        for ($i = 0; $i -lt $lines.Length; $i++) {
+            if ($lines[$i] -match '^-- \)EOF"') { $markerIdx = $i; break }
+        }
+        $newLines = @()
+        for ($i = 0; $i -lt $lines.Length; $i++) {
+            $lineNum = $i + 1
+            $line    = $lines[$i]
+            # Keep the first line and the EOF-marker line untouched.
+            if ($i -eq 0 -or $i -eq $markerIdx) { $newLines += $line; continue }
+            # Drop the GPL license header (lines 2-17).
+            if ($lineNum -ge 2 -and $lineNum -le 17) { continue }
+            # Drop blank separator lines.
+            if ($line -match '^\s*$') { continue }
+            # Drop trailing whitespace.
+            $line = $line -replace '\s+$', ''
+            # Drop inline comments (anything from `--` to end of line).
+            $line = $line -replace '\s*--.*$', ''
+            # Shrink 4-space indent to 2-space.
+            $line = $line -replace '^(    )', '  '
+            if ($line -match '^\s*$') { continue }
+            $newLines += $line
+        }
+        ($newLines -join "`n") | Out-File -LiteralPath $path_isoffi -Encoding utf8 -NoNewline
+        $newLen = ((Get-Content -LiteralPath $path_isoffi -Raw -Encoding utf8) -replace '.*R"EOF\(--', '' -replace '-- \)EOF".*', '').Length
+        Write-Host "isoffi.lua trimmed: $literalLen -> $newLen bytes of raw string content."
+    }
+}
+
+& $msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
 
 # Locate luajit via scoop. `luajit.exe` is on PATH via scoop's shim;
 # we use `scoop prefix` to find the install root for the include dir (needed to compile lpeg against luajit's headers).
@@ -70,8 +179,8 @@ if (-not $luajit_prefix -or -not (Test-Path (Join-Path $luajit_prefix 'bin/luaji
 # Discover the luajit include dir by globbing `include/luajit-*`.
 # This avoids hardcoding a specific version (e.g. `luajit-2.1`).
 $luajit_include_root = Join-Path $luajit_prefix 'include'
-$lua_inc_dir = Get-ChildItem -Path $luajit_include_root -Directory -Filter 'luajit-*' -ErrorAction SilentlyContinue |
-	Select-Object -First 1 -ExpandProperty FullName
+$lua_inc_dir         = Get-ChildItem -Path $luajit_include_root -Directory -Filter 'luajit-*' -ErrorAction SilentlyContinue |
+Select-Object -First 1 -ExpandProperty FullName
 if (-not $lua_inc_dir) {
 	write-error "No 'luajit-*' include dir found under '$luajit_include_root'. The scoop luajit install may be broken."
 	exit 1
@@ -90,7 +199,7 @@ $lpeg_compile_args = @(
 	'-o', 'lpeg.dll'
 ) + $lpeg_sources + @('-lluajit-5.1')
 push-location $path_lpeg
-&	gcc @lpeg_compile_args
+& gcc @lpeg_compile_args
 pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
@@ -101,8 +210,8 @@ pop-location
 
 $path_lfs = join-path $path_toolchain 'lfs'
 verify-path $path_lfs
-$lfs_src = join-path $path_pcsx_redux 'third_party\luafilesystem\src\lfs.c'
-$lfs_dll = join-path $path_lfs 'lfs.dll'
+$lfs_src        = join-path $path_pcsx_redux 'third_party\luafilesystem\src\lfs.c'
+$lfs_dll        = join-path $path_lfs 'lfs.dll'
 $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 & gcc -O2 -shared "-I$lua_inc_dir" -o $lfs_dll $lfs_src $lfs_dll_import
 
@@ -112,6 +221,15 @@ $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 # ════════════════════════════════════════════════════════════════════════════
 
 $path_openbios = join-path $path_pcsx_redux 'src\mips\openbios'
+
+# Wipe stale *.dep files across src\mips.
+# These cache absolute paths to the GCC headers directory; if the toolchain was upgraded (e.g. v14.2.0 → v16.1.0)
+# Make reads the stale paths and aborts with "no rule to make target .../stddef.h".
+# `make clean` in openbios only clears its own dir — subdirs like common/crt0/, modplayer/, and shell/ keep their stale .dep files.
+# Easier to just delete the lot before each build than to teach every Makefile about deepclean recursion.
+Get-ChildItem -Path (join-path $path_pcsx_redux 'src\mips') -Recurse -Filter '*.dep' -ErrorAction SilentlyContinue |
+    ForEach-Object { Remove-Item -LiteralPath $_.FullName -Force }
+
 push-location $path_openbios
 &	make clean
 &	make