Stuff kept from hot-reload attempt

Need to rewrite the ps1 lua metaprogram sometime soonish. Getting too bloated... need to consolidate code paths.

In this push codebase structure is starting to get a bit more realized. Decided todo now to match Pikuma's linking module files vods beginning to reorganize its codebase as well.
Atoms & atom components are not in their on *.atom.c files. (Not calling it tape.c as I don't really bake tapes like that outside of the unity c file so far...)

The lua metaprogram has had additional features added to it yet again to avoid hardcoding module handling and supporting multiple atom files per-module.
Either after the camera or cd-rom section I'll be most likely pausing to fully refactor the metaprogram. Possibly as a full re-write to get the loc minimal.

.gitignore

@@ -1,8 +1,9 @@
 build
 toolchain/armips
+toolchain/luajit-2.1
 toolchain/pcsx-redux
-# toolchain/psyq_iwyu
+toolchain/psyq_iwyu
-# toolchain/PSn00bSDK
+toolchain/PSn00bSDK
 
 *.exe
 *.elf
@@ -14,3 +15,9 @@ toolchain/pcsx-redux
 *.a
 .sentry-native
 .vscode/settings.json
+toolchain/lfs
+toolchain/lpeg
+
+scratch
+toolchain/libpsn00b
+scripts/pcsx_debug_helper.zip

.vscode/launch.json

@@ -4,7 +4,7 @@
 	// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
 	"version": "0.2.0",
 	"configurations": [
-		{
+{
 			"name": "Debug: Hello Psy-Q!",
 			"type": "gdb",
 			"request": "attach",
@@ -12,6 +12,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -20,10 +24,17 @@
 			"osx": {
 				"gdbpath": "gdb"
 			},
-			"executable": "${workspaceRoot}/build/hello_psyq.elf",
+			"executable": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
-				"load hello_psyq.elf",
+				"load hello_gte.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
 				"tbreak main",
 				"continue"
 			]
@@ -36,30 +47,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
-			"stopAtConnect": true,
+			"registerLimit": "1-32",
-			"gdbpath": "gdb-multiarch",
+			"frameFilters": false,
-			"windows": {
+			"showDevDebugOutput": false,
-				"gdbpath": "gdb-multiarch.exe"
+			"printCalls": false,
-			},
-			"osx": {
-				"gdbpath": "gdb"
-			},
-			"executable": "${workspaceRoot}/build/hello_gpu.elf",
-			"autorun": [
-				"monitor reset shellhalt",
-				"load hello_gpu.elf",
-				"tbreak main",
-				"continue"
-			]
-		},
-		{
-			"name": "Debug: Hello GTE Psy-Q!",
-			"type": "gdb",
-			"request": "attach",
-			"target": "localhost:3333",
-			"remote": true,
-			"cwd": "${workspaceRoot}/build",
-			"valuesFormatting": "parseText",
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -69,12 +60,124 @@
 				"gdbpath": "gdb"
 			},
 			"executable": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
 				"load hello_gte.elf",
 				"tbreak main",
 				"continue"
 			]
+		},
+		{
+			"name": "Debug: Hello GTE!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${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": "${workspaceRoot}/build/hello_gte.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_gte.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"tbreak main",
+				"continue"
+			]
+		},
+		{
+			"name": "Debug: Hello Joypad!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${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"
+			]
 		}
 	]
 }
+

code/duffle/atom_dsl.h

@@ -1,467 +0,0 @@
-/*
- * atom_dsl.h
- * ============================================================================
- *
- * ATOM DSL — annotation layer for tape atoms (lottes_tape.h).
- *
- * This header turns `__attribute__((annotate(...)))` and `_Pragma(...)` into
- * a small named DSL that the metaprogram can validate against.
- *
- *   The C compiler treats every macro below as a no-op:
- *     - atom_init / atom_terminate / atom_bind / atom_setup / atom_commit /
- *       atom_annot all expand to `__attribute__((annotate("..."))) MipsAtom_(name)`
- *       — accepted by GCC (with -Wno-attributes), absent at runtime.
- *     - atom_resource / atom_region / atom_group / atom_cadence / atom_async
- *       expand to `_Pragma("...")` — accepted by any C11 preprocessor.
- *
- *   The metaprogram (tape_atom_annotation_pass.lua) reads the source-as-written
- *   and validates:
- *     - every MipsAtom_ has one atom_*() annotation (no orphans)
- *     - phase is recognized (init/bind/setup/work/commit/terminate)
- *     - reads/writes reference canonical wave-context registers
- *     - rbind atoms reference a real Binds_* struct declaration
- *     - word-counts in tapre metadata agree with the body's actual .word count
- *     - resource/region/group/cadence/async pragmas are spelled correctly and
- *       reference known enum values
- *
- * ============================================================================
- *
- *  PUTTING IT ON AN ATOM — the canonical pattern
- *
- *      _tape_resources_
- *      atom_resource(cube_tri, "model_ship_cube")
- *      atom_region  (cube_tri, PRIM_ARENA)
- *      atom_group   (cube_tri, GROUP_RENDER_PRIMS)
- *      atom_cadence (cube_tri, CADENCE_FRAME)
- *
- *      atom_annot(cube_tri, phase_work,
- *                 tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
- *                 tape_regs(R_PrimCursor, R_FaceCursor))
- *      internal MipsAtom_(cube_tri) {
- *          atom_label(culling),
- *          // ... atom body ...
- *          atom_label(bounds_chk),
- *      };
- *
- *      atom_offset(culling, bounds_chk)         // ← branch target, validated
- *
- *  RBIND pattern — `Binds_*` is the contract
- *
- *      // Wave-context register layout (declarative):
- *      typedef struct Binds_TrackFaceBatch {
- *          U4 R_PrimCursor,  R_FaceCursor,
- *            R_VertBase,    R_OtBase;
- *      } Binds_TrackFaceBatch;
- *
- *      atom_resource(rbind_track_face_batch, "track_face_batch_42")
- *      atom_region  (rbind_track_face_batch, HEAP_3D)
- *      atom_group   (rbind_track_face_batch, GROUP_LOAD_FACES)
- *      atom_cadence (rbind_track_face_batch, CADENCE_ONDEMAND)
- *      atom_async   (rbind_track_face_batch, true)
- *
- *      atom_bind(rbind_track_face_batch, Binds_TrackFaceBatch,
- *                tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
- *      internal MipsAtom_(rbind_track_face_batch) { ... };
- *
- *  Annotation rules
- *  ----------------
- *      1. Each MipsAtom_(name) needs EXACTLY ONE atom_*() macro on the line
- *         immediately above. No annotation = orphan (warning). Two annotations
- *         on the same name = duplicate (error).
- *
- *      2. atom_init and atom_terminate take only the name.
- *
- *      3. atom_setup and atom_commit take name + reads.
- *
- *      4. atom_bind takes name + Binds_* type + writes.
- *
- *      5. atom_annot takes name + phase token + reads + writes.
- *         Phase tokens: phase_init / phase_bind / phase_setup / phase_work /
- *         phase_commit / phase_terminate.
- *
- *      6. Optional pragmas (atom_resource / atom_region / atom_group /
- *         atom_cadence / atom_async) attach metadata to the atom. They can
- *         appear in any order, with one per atom. They're independent of the
- *         atom_*() macro — multiple pragmatics are fine.
- *
- * ============================================================================
- *
- *  WHY A SEPARATE LAYER (not just put everything in source comments)?
- *
- *      Source comments are invisible to the compiler. Annotations live in the
- *      source as actual C tokens, so:
- *        - they can never silently get out of sync with the code (the build
- *          fails at preprocessing if the metaprogram disagrees)
- *        - they can be cross-validated against metadata (build fails if a
- *          WORD_COUNT entry drifts away from the .word count in source)
- *        - they make the C compiler a witness ("there's a marker here, and
- *          it's labelled, and it has arguments") without making the C compile
- *          itself do any work
- *
- * ============================================================================
- */
-
- #ifdef INTELLISENSE_DIRECTIVES
-#pragma once
-// #include <stdint.h>
-#endif
-
-/* ============================================================================
- * PHASE TOKENS — strings, used as the second arg to atom_annot(...)
- *
- * Why strings? They preserve the metaprogram's ability to read phase directly
- * from the source-as-written, even when the macro isn't expanded. The Lua
- * tool also has a MACRO_EXPANSION table for resolving phase_* source-level
- * references.
- *
- *   atom_annot(cube_tri, phase_work, ...)    ← legal
- *   atom_annot(cube_tri, "work", ...)        ← legal (and equivalent)
- *   atom_annot(cube_tri, phase_setup, ...)   ← legal
- *
- * ============================================================================*/
-
-#define phase_init       "init"
-#define phase_bind       "bind"
-#define phase_setup      "setup"
-#define phase_work       "work"
-#define phase_commit     "commit"
-#define phase_terminate  "terminate"
-
-/* ============================================================================
- * WAVE-CONTEXT REGISTERS — canonical register set for the tape wave model.
- *
- * The tape-atom runtime carries four registers across a wave:
- *
- *   R_PrimCursor  output pointer into the prim arena (next OT entry to write)
- *   R_FaceCursor  input pointer into the face array (next face to consume)
- *   R_VertBase    base pointer into the vertex arena (this wave's vertices)
- *   R_OtBase      base pointer into the ordering table (this wave's OT slot)
- *
- * Each atom declares its reads/writes against this canonical set. The Lua
- * tool rejects wave-context positions that reference any other register
- * (warning today — the C compiler's R_T4..R_T7 / R_RA / etc. aliases are
- * implementation details and not part of the typed surface).
- *
- * If your atom needs to touch GTE / SP / DMA / other side state, declare it
- * at the source level as you normally would — but DO NOT put those registers
- * in tape_regs(...). Wave-context is a closed set.
- *
- * ============================================================================*/
-
-/* ============================================================================
- * REGION TOKENS — memory regions atoms may allocate from or write into.
- *
- * Use atom_region(name, REGION) to declare. The Lua tool validates that the
- * region is in this set, AND that:
- *   - rbind atoms declare the source region (usually HEAP_3D or CDROM_STREAM)
- *   - work atoms declare the destination region (the arena they push to)
- *   - commit atoms must declare a region equal to what setup wrote, so the
- *     C-side mirror is consistent
- *
- * Add new regions by extending this list and the metaprogram's KNOWN_REGIONS.
- * Don't add regions ad-hoc — every new region becomes part of the contract.
- *
- * ============================================================================*/
-#define REGION_PRIM_ARENA     prim_arena       /* OT/prim packet arena */
-#define REGION_FACE_ARENA     face_arena       /* face index array */
-#define REGION_VERTEX_ARENA   vertex_arena     /* vertex pool */
-#define REGION_OT_ARENA       ot_arena         /* ordering-table array */
-#define REGION_HEAP_3D        heap_3d_models   /* loaded model heap */
-#define REGION_CDROM_STREAM   cdrom_stream     /* CDROM read buffer */
-#define REGION_VRAM           vram_heap        /* VRAM texture/GPU buffer */
-
-/* ============================================================================
- * CADENCE TOKENS — how often the atom runs.
- *
- *   frame    runs every vsync (rendering, input poll)
- *   once     runs exactly once per process lifetime (init, terminate)
- *   ondemand runs when triggered by event (CDROM load, async DMA complete)
- *
- * Used as a hint for the metaprogram to flag:
- *   - frame-cadence atoms that have side effects (they'll be hit many times,
- *     so avoid global state mutation unless it's idempotent)
- *   - once-cadence atoms inside "if (frame_count == 0)" guards (the guard
- *     is then provably one-shot, the metaprogram can lift initialization)
- *   - ondemand atoms that are missed by the wave scheduler (forces async
- *     and discards yield results without further processing)
- *
- * ============================================================================*/
-#define CADENCE_FRAME     frame
-#define CADENCE_ONCE      once
-#define CADENCE_ONDEMAND  ondemand
-
-/* ============================================================================
- * tape_regs(...) — wave-context register list
- *
- *   tape_regs(R_PrimCursor, R_FaceCursor)  →  (R_PrimCursor, R_FaceCursor)
- *
- * The macro produces a comma-evaluated expression that the C compiler
- * silently discards (it's wrapped in parentheses in the call argument
- * position — the result is never bound). The Lua tool pattern-matches the
- * "tape_regs(...)" token to extract the list.
- *
- * You can have at most one tape_regs(...) in the reads slot and one in the
- * writes slot of atom_annot. To declare multiple disjoint sets (rare), just
- * declare the union — the metaprogram doesn't track which reads need which
- * writes at this granularity.
- *
- * ============================================================================*/
-#define atom_reads(...)  (__VA_ARGS__)
-#define atom_writes(...) (__VA_ARGS__)
-
-/* ============================================================================
- * ATOM ANNOTATION MACROS
- *
- *   Each expands to `__attribute__((annotate("kind"))) MipsAtom_(name)` —
- *   the GCC attribute is accepted under -Wno-attributes (already in your
- *   build flags) and stripped at runtime. The annotation string is just the
- *   macro kind ("atom_annot", "atom_bind", etc.) — the metaprogram reads
- *   the macro call's full args list from the source-as-written.
- *
- * ============================================================================*/
-
-/* ----------------------------------------------------------------------------
- * atom_init — entry into tape_runtime_main
- *
- *   atom_init(tape_main)
- *   internal MipsAtom_(tape_main) { ... };
- *
- * Implies: no reads, no writes (wave-context not established yet).
- * ----------------------------------------------------------------------------*/
-#define atom_init(name) __attribute__((annotate("atom_init")))
-
-/* ----------------------------------------------------------------------------
- * atom_terminate — exit from tape_runtime_main
- *
- *   atom_terminate(tape_exit)
- *   internal MipsAtom_(tape_exit) { ... };
- *
- * Implies: no reads, no writes (wave-context destroyed at this point).
- * ----------------------------------------------------------------------------*/
-#define atom_terminate(name) __attribute__((annotate("atom_terminate")))
-
-/* ----------------------------------------------------------------------------
- * atom_setup — pre-work atom: prepares engine state (e.g., set_gte_world)
- *
- *   atom_setup(set_gte_world, tape_regs(R_TapePtr))
- *   internal MipsAtom_(set_gte_world) { ... };
- *
- * Reads:   anything (the engine state you're reading)
- * Writes:  engine state (GTE / DMA / etc. — declared in source, not part of
- *          wave-context, so doesn't go in tape_regs)
- *
- * The metaprogram checks that setup is followed (in atomic order) by a work
- * atom in the same wave — there's no point in setting up state if no one
- * reads it.
- * ----------------------------------------------------------------------------*/
-#define atom_setup(name, reads) __attribute__((annotate("atom_setup")))
-
-/* ----------------------------------------------------------------------------
- * atom_commit — post-work atom: flushes wave-context back to C-side state
- *
- *   atom_commit(sync_prim_cursor, tape_regs(R_PrimCursor))
- *   internal MipsAtom_(sync_prim_cursor) { ... };
- *
- * Reads:   wave-context registers (the ones you sync back to C)
- * Writes:  C-side mirror (declared in source — not part of wave-context)
- *
- * The metaprogram checks that commit is preceded (in atomic order) by a
- * work atom that wrote the registers this commit is reading.
- * ----------------------------------------------------------------------------*/
-#define atom_commit(name, reads) __attribute__((annotate("atom_commit")))
-
-/* ----------------------------------------------------------------------------
- * atom_bind — rbind atom: read wave-context registers from tape pointer
- *
- *   atom_bind(rbind_cube_tri, Binds_CubeTri,
- *             tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
- *   internal MipsAtom_(rbind_cube_tri) { ... };
- *
- * The binds_struct MUST be a typedef'd type (declared via
- * `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
- * The Lua tool cross-references this. Missing struct = error.
- *
- * Implicit: reads R_TapePtr, writes the four wave-context registers.
- * ----------------------------------------------------------------------------*/
-#define atom_bind(name, binds_struct, writes) __attribute__((annotate("atom_bind")))
-
-/* ----------------------------------------------------------------------------
- * atom_annot — generic work atom with explicit phase
- *
- *   atom_annot(cube_tri, phase_work,
- *              tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
- *              tape_regs(R_PrimCursor, R_FaceCursor))
- *   internal MipsAtom_(cube_tri) { ... };
- *
- * Use this for the bulk of your atoms. For init/setup/commit/bind, prefer
- * the convenience macros above — they pin the phase for you.
- *
- * The phase arg is one of: phase_init / phase_bind / phase_setup /
- * phase_work / phase_commit / phase_terminate. Spelling mistakes are errors.
- * ----------------------------------------------------------------------------*/
-#define atom_annot(name, phase, reads, writes) __attribute__((annotate("atom_annot")))
-
-/* ============================================================================
- * RESOURCE / GROUP / CADENCE / REGION / ASYNC — optional atom metadata
- *
- *   These don't annotate the atom semantically (phase/reads/writes do that).
- *   They attach extra context that the metaprogram uses to catch:
- *     - same resource loaded twice in different ways
- *     - atoms that span multiple regions (likely bug — pick one)
- *     - frame-cadence atoms that should be once-cadence (perf / correctness)
- *     - ondemand atoms that aren't async (CDROM races)
- *
- *   You can use as many as apply to a given atom, in any order, immediately
- *   above the atom_*() macro.
- *
- * ============================================================================*/
-
-/* ----------------------------------------------------------------------------
- * atom_resource — name the logical resource the atom references
- *
- *   atom_resource(cube_tri, "model_ship_cube")
- *   atom_resource(load_track_faces, "track_lavender_field_0x42")
- *   atom_resource(play_engine_sfx, "sfx_engine_loop")
- *
- * Use any human-readable string. The metaprogram:
- *   - validates resource strings are non-empty and don't contain control chars
- *   - flags duplicates across atoms with the same name (two atoms claiming
- *     ownership of a resource is usually a refactor artifact or bug)
- *   - flags references to resources that no atom actually defines
- *
- * The arg is a STRING LITERAL, so it can't accidentally alias a variable.
- * ----------------------------------------------------------------------------*/
-#define atom_resource(name, res_id) //_Pragma("atom " #name " resource=" res_id)
-
-/* ----------------------------------------------------------------------------
- * atom_region — name the memory region the atom touches
- *
- *   atom_region(cube_tri,   REGION_PRIM_ARENA)
- *   atom_region(load_faces, REGION_HEAP_3D)
- *   atom_region(load_tex,   REGION_VRAM)
- *
- * Use REGION_* tokens above. The metaprogram enforces the closed set.
- *
- * Edge cases the metaprogram catches:
- *   - rbind atom that doesn't declare a SOURCE region (where is it loading from?)
- *   - work atom with no destination region (where is it pushing to?)
- *   - region that disagrees with the Binds_* struct layout (you said it's a
- *     prim_arena rbind but the struct has 4 faces in it — wait, that's wrong)
- * ----------------------------------------------------------------------------*/
-#define atom_region(name, region) //_Pragma("atom " #name " region=" #region)
-
-/* ----------------------------------------------------------------------------
- * atom_group — bundle atoms into a logical batch (track-load, sound-load, etc.)
- *
- *   atom_group(load_track_face_42,  GROUP_LOAD_FACES)
- *   atom_group(load_track_face_43,  GROUP_LOAD_FACES)
- *   atom_group(swap_face_42_43,     GROUP_VISIBILITY_SWAP)
- *
- * Use any token as the group id. The metaprogram:
- *   - validates all atoms in a group emit their waves in the same tb_group
- *     (no spawning other waves inside a group)
- *   - flags groups with only one member (probably a typo — meant to be a group?)
- *   - validates cross-group edges (no atom reads what another group writes,
- *     unless explicitly grouped together)
- *
- * Useful when:
- *   - subdivisible work (track-face batches, polygon subdivision) needs to
- *     confirm that all batches of one logical visible scene are emitted
- *     together
- *   - async loads (CDROM -> VRAM) need to be grouped so all batches complete
- *     before the swap
- *
- * Use GROUPS for sound effects to track which sound plays during which atom,
- * which is needed if the sound tool ever has to validate "this atom is the
- * trigger for an audio play".
- * ----------------------------------------------------------------------------*/
-#define atom_group(name, group_id) //_Pragma("atom " #name " group=" #group_id)
-
-/* ----------------------------------------------------------------------------
- * atom_cadence — declare execution frequency
- *
- *   atom_cadence(render_frame,     CADENCE_FRAME)       // every vsync
- *   atom_cadence(load_track_faces, CADENCE_ONDEMAND)    // on demand
- *   atom_cadence(init_heap,        CADENCE_ONCE)        // process lifetime
- *
- * Default (no atom_cadence call) is CADENCE_FRAME — most atoms run every
- * frame. Override explicitly when not.
- *
- * The metaprogram's checks:
- *   - CADENCE_ONCE atoms inside `if (frame == 0)` or `if (!initialized)` are
- *     tagged, validating that guards are required (or warning if missing)
- *   - CADENCE_FRAME atoms that mutate state outside the wave context get
- *     flagged (likely a bug — state should persist through commits)
- *   - CADENCE_ONDEMAND atoms must have atom_async — otherwise the trigger
- *     mechanism is undefined
- * ----------------------------------------------------------------------------*/
-#define atom_cadence(name, cadence) //_Pragma("atom " #name " cadence=" #cadence)
-
-/* ----------------------------------------------------------------------------
- * atom_async — declare whether the atom yields / interacts with CDROM DMA
- *
- *   atom_async(load_track_tex,  true)   // CDROM read yield
- *   atom_async(load_vram,       true)   // VRAM upload DMA
- *   atom_async(render_frame,    false)  // pure compute, no async
- *
- * The metaprogram requires this for CADENCE_ONDEMAND atoms. For
- * CADENCE_FRAME, it's optional but documents intent.
- *
- * Note: CDROM ATOMS in Psy-Q are typically implemented as a chain of
- * "async-init" atom followed by a "wait-for-completion" atom. Both atoms
- * should be marked async=true, and both should have the same resource/group
- * tag (so the metaprogram can verify they're paired).
- * ----------------------------------------------------------------------------*/
-#define atom_async(name, is_async) //_Pragma("atom " #name " async=" #is_async)
-
-/* ============================================================================
- * WORD-COUNT ANNOTATION FOR A #define MAC
- *
- *   tape_words(mac_yield, 1)
- *   #define mac_yield()  \
- *       load_word(R_AtomJmp, R_TapePtr, 0), \
- *       add_ui_self(R_TapePtr, 4), \
- *       jump_reg(R_AtomJmp), \
- *       nop
- *
- * The compiler accepts the unknown _Pragma. The Lua tool reads it and
- * cross-checks against WORD_COUNT(mac_yield, 1) in tape_atom.metadata.h.
- * If they disagree, build fails.
- *
- * Use sparingly — only on multi-word macros (single-word ones don't need
- * drift tracking; they're checked by the .word-count pass anyway).
- *
- * ============================================================================*/
-#define tape_words(name, n) //_Pragma(#name " tape_atom words=" #n)
-
-/* ============================================================================
- * atom_label / atom_offset — branch target machinery
- *
- *   atom_label(culling)                      ← nothing in C; anchor only
- *   ... body ...
- *   atom_label(bounds_chk)                   ← another anchor
- *
- *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
- *
- * The metaprogram generates gen/atom_offsets.h with one
- *   #define atom_offset__culling__bounds_chk  ((target - branch_pos - 1))
- * per atom_offset(F, T) call. The preprocessor then expands your 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.
- * This catches:
- *   - typo in atom_label (no anchor → metaprogram doesn't emit the macro)
- *   - .offsets.h not regenerated after body edits
- *   - body edit that broke the offset math (recompile + retest picks it up
- *     in CPU emulator)
- *
- * ============================================================================*/
-
-#define atom_offset(F, T) atom_offset_ ## F ## _ ## T
-/* atom_label is a pure annotation for the metaprogram's offset calculations.
- * The macro expands to a C comment, so the C preprocessor strips it to
- * whitespace — NO instruction word is emitted in the asm. The metaprogram
- * still recognises the literal `atom_label(name)` token in source and
- * records the marker at the current pos. */
-#define atom_label(name) /* atom_label anchor: name */

code/duffle/dsl.atom.h

@@ -0,0 +1,159 @@
+/*
+ * dsl.atom.h
+ * ============================================================================
+ *
+ * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
+ * The metaprogram (scripts/passes/annotation.lua) reads source-as-written and validates:
+ *   - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X), atom_reads(...), atom_writes(...)) in any order and are optional.
+ *   - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a real Binds_* struct declaration.
+ *   - atom word-counts in word_counts.metadata.h match the body's actual .word count.
+ *
+ * Pure macro anntation.
+ * ---------------
+ * Don't want to constraint the macro usage to some attribute placment constraint, etc, don't want ot dela with the compiler.
+ * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip each expand to a C comment or to nothing
+ * (C preprocessor strips them to whitespace).
+ *
+ * ============================================================================
+ *  Usage:
+ *      MipsAtom_(cube_tri) atom_info(
+ *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor)
+ *      ){
+ *          atom_label(culling),
+ *          // ... atom body ...
+ *          atom_offset(culling, bounds_chk)         // branch target, validated
+ *          // ... atom body ...
+ *          atom_label(bounds_chk),
+ *      };
+ *
+ *
+ *  Data Binding pattern -- atom_bind as a sub-call of atom_info
+ *
+ *      // Wave-context register layout (declarative):
+ *      typedef Struct_(Binds_TrackFaceBatch) { 
+ * 					U4 PrimCursor; 
+ * 					U4 FaceCursor;
+ *          U4 VertBase;
+ * 					U4 OtBase;
+ *      };
+ *      MipsAtom_(rbind_track_face_batch) atom_info(
+ *         atom_bind(Binds_TrackFaceBatch)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *      ){ ... };
+ *
+ *  Annotation rules
+ *  ----------------
+ *      1. atom_info(...) is OPTIONAL. Atoms without atom_info are silently skipped by the metaprogram.
+ *      2. If present, atom_info takes up to three sub-calls, all order-independent within the arg list:
+ *           - atom_bind(Binds_X)
+ *           - atom_reads(...)
+ *           - atom_writes(...)
+ *      3. atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *      4. atom_reads(...) and atom_writes(...): Used to to check if registers are used correctly in macros: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
+ *      5. atom_label(name: Utilize with atom_offset as a target location.
+ *      6. atom_offset(F, T): Resolved by gen/atom_offsets.h, generated from the atom_label markers. Calculated during the offset pass of the lua  metaprogram.
+ */
+
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+
+/* ============================================================================
+ * atom_reads(...) / atom_writes(...)
+ *
+ * Used during the static analysis pass of the metaprogram to do
+ * ============================================================================*/
+#define atom_reads(...)  (__VA_ARGS__)
+#define atom_writes(...) (__VA_ARGS__)
+
+/* ----------------------------------------------------------------------------
+ * 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.
+ * ----------------------------------------------------------------------------*/
+#define atom_reg /* atom_reg: opt the preceding enum entry into the DWARF registry */
+
+/* ============================================================================
+ *   atom_info :
+ *      MipsAtom_(cube_tri) atom_info(
+ *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor)
+ *      ){ ... };
+ * 
+ *  - atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *  - atom_reads(...):    comma-list of registers
+ *  - atom_writes(...):   comma-list of registers
+ * ============================================================================*/
+#define atom_info(...)  /* atom_info(__VA_ARGS__) */
+
+/* ----------------------------------------------------------------------------
+ * DEBUG SOURCE-STEP MARKER
+ *
+ * Place `atom_dbg_skip` (BARE) before a MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_.
+ * The following declaration kind determines whether the marker selects a whole atom or a component inline view.
+ * The source scanner associates the marker with that declaration; placement diagnostics are handled by the annotation pass.
+ *
+ * Example:
+ *     atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+ *     atom_dbg_skip MipsAtomComp_(ac_yield) { ... };
+ *     atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { ... });
+ * ----------------------------------------------------------------------------*/
+#define atom_dbg_skip /* atom_dbg_skip: skip the following atom or component source view */
+
+/* ----------------------------------------------------------------------------
+ * Typed-view annotations (Registry for DWARF RR_<R_X> chain resolution)
+ *   atom_type(<T>) -- overloaded:
+ *      (a) enum-site default: `R_Foo = R_Tn, atom_reg atom_type(T)` 
+ *      Sets the per-alias default typed view in the register_alias_registry.
+ *      Consumed by the DWARF chain step (e) when no per-atom atom_ctx / atom_phase / atom_type callsite provides a stronger resolution.
+ *      (b) callsite override: `atom_reads(R_Foo atom_type(T), ...)` Overrides the per-alias default for THIS atom only.
+ *      Last-write-wins per R_Name; conflict -> error.
+ *   atom_ctx(<atom_name>) -- atom-info sub-call:
+ *       Propagate another atom's atom.rbind.fields (its Binds_* typed fields) into THIS atom's typed-view resolution.
+ *       The named atom must be an rbind atom (have `atom_bind(Binds_X)` in its `atom_info`).
+ *       Used as the escape hatch when atom_phase is not the natural correlation.
+ *   atom_phase(<label>) -- atom-info sub-call: 
+ *       Free-form C-identifier label for grouping atoms.
+ *       Within a phase, the FIRST atom in source-order that owns its own atom.rbind provides 
+ *       the Binds_* field types used by all other atoms in the same phase.
+ *       The preferred correlation mechanism; atom_ctx is the escape hatch for non-natural cases.
+ *
+ * All three expand to C comments
+ * (the bare-token convention matching `atom_reg` and `atom_dbg_skip`).
+ * The Lua scanner reads the bare tokens in source-as-written; the C preprocessor strips them.
+ * ----------------------------------------------------------------------------*/
+#define atom_type(T)        /* atom_type: associate <T> with the preceding enum entry (enum site) or this register (atom-info site) */
+#define atom_ctx(atom_name) /* atom_ctx: propagate <atom_name>'s Binds_* field types into this atom's typed views */
+#define atom_phase(label)   /* atom_phase: tag this atom with <label> for grouped typed-view resolution */
+
+/* ----------------------------------------------------------------------------
+ * atom_bind(Binds_X) -- rbind sub-call of atom_info
+ *
+ *   MipsAtom_(rbind_cube_tri) atom_info(
+ *         atom_bind(Binds_CubeTri)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *   ){ ... };
+ *
+ * The Binds_X MUST be a typedef'd type (declared via `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
+ * ----------------------------------------------------------------------------*/
+#define atom_bind(binds_struct)  /* atom_bind(binds_struct) */
+
+/* ============================================================================
+ * atom_label / atom_offset — branch target machinery
+ *
+ *   atom_label(culling)                      ← nothing in C; anchor only
+ *   ... body ...
+ *   atom_label(bounds_chk)                   ← another anchor
+ *
+ *   atom_offset(culling, bounds_chk)         ← resolved by gen/offsets.h
+ *
+ * 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/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.
+#define atom_label(name) /* atom_label anchor: name */

code/duffle/dsl.h

@@ -43,7 +43,9 @@
 
 #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 +69,8 @@
 #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.
@@ -135,16 +138,17 @@ enum { false = 0, true  = 1, true_overflow, };
 
 typedef void Proc_(VoidFn) (void);
 
-#define kilo(n)         (C_(U4, n) << 10)
+#define kilo(n)                (C_(U4, n) << 10)
-#define mega(n)         (C_(U4, n) << 20)
+#define mega(n)                (C_(U4, n) << 20)
-#define giga(n)         (C_(U4, n) << 30)
+#define giga(n)                (C_(U4, n) << 30)
-#define tera(n)         (C_(U4, n) << 40)
+#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 null                    C_(U4,    0)
-#define S_(data)         C_(U4, sizeof(data))
+#define nullptr                 C_(void*, 0)
+#define O_(type, field)         C_(U4, & C_(type*,0)->field)
+#define OA_(type, member, idx)  C_(U4, & C_(type*,0)->member[idx])
+#define OT_(field)              O_(typeof_ptr(& field), filed))
+#define S_(data)                C_(U4, sizeof(data))
 
 #define sop_1(op,a,b) C_(U1, s1_(a) op s1_(b))
 #define sop_2(op,a,b) C_(U2, s2_(a) op s2_(b))
@@ -218,3 +222,6 @@ IA_ void assert(U8 cond) { if(cond){return;} else{debug_trap(); ms_exit_process(
 #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
+/* reg_str(n) — Stringify an integer register id into the GCC asm string form (e.g. 12 → "$12").
- * string form (e.g. 12 → "$12"). Use this anywhere GCC's parser
+ * Use this anywhere GCC's parser expects a literal string identifying a register: clobber lists,
- * 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 —
- * asm templates, etc. The two-level macro is the standard preprocessor
+ * without it, `#n` would stringify the macro name `R_T4` to `"R_T4"` instead of expanding `R_T4` to its value first.
- * 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
+ * For declaring a register variable bound to a specific GPR, use the `rgcc(n)` bundle from gcc_asm.h instead —
- * `rgcc(n)` bundle from gcc_asm.h instead — it adds the `__asm__()`
+ * it adds the `__asm__()` qualifier around the string.
- * 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
+/* rgcc_ref(n) — GCC operand-reference form "%N". Not currently used by the placeholder-pun macros
- * by the placeholder-pun macros (the .word bodies are fully baked
+ * (the .word bodies are fully baked at compile time and have no runtime operand references),
- * at compile time and have no runtime operand references), but kept
+ * but kept here for completeness in case a future asm template needs to refer to a runtime input by position.
- * here for completeness in case a future asm template needs to refer
+ * Mirror of rgcc but produces "%N" instead of "$N". */
- * to a runtime input by position. Mirror of rgcc but produces "%N"
- * instead of "$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
+ * NOTE: we use `%0`, `%1`, ... not `%c0`, `%c1`, ... because GCC's asm-parser rejects `%cN` in this position with "invalid use of '%c'".
- * 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 `%cN` form is for printing *character* constants; for arbitrary
+ * the plain `%N` form is the right one. Both expand to the bare immediate.
- * 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/lottes_tape.offsets.h

@@ -1,9 +0,0 @@
-// Auto-generated by tape_atom_offset_gen.meta.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 +1,184 @@
+#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\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.c
+//   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).
+//   - mac_yield() is the safe default for atom-endings: 4 words, BD-slot of jr is mandatory nop.
+//   - mac_yield_load() + mac_yield_tail():
+//     - unconditional branch: mac_yield_load fills the branch's BD-slot (replaces a nop);
+//     - mac_yield_tail runs at the branch target (does NOT re-load R_AtomJmp).
+#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 */
+#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) \
+,	nop
+WORD_COUNT(mac_yield_tail, 3)
+
+/* atom_dbg_skip */
+#define mac_load_v2s2(rs_x, rs_y, r_base, offset) \
+	load_half( rs_x, r_base, O_(V3_S2,x)) \
+,	load_half( rs_y, r_base, 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_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_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)
+
+/* 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)) \
+,	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)) \
+,	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)) \
+,	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)
+
+#define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
+	load_upper_i(reg_transfer, cmd >> 16) \
+,	or_i_self(   reg_transfer, cmd & 0xFFFF) \
+,	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)
+
+#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_f3(r_ot_base, r_prim_cursor) \
+	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, (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_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_f3, 11)
+
+#define mac_insert_ot_tag_g4(r_ot_base, r_prim_cursor) \
+	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, (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_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_g4, 11)
+

code/duffle/gen/offsets.h

@@ -0,0 +1,53 @@
+// 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\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.c
+//   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: 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,
+};
+
+#pragma endregion duffle
+

code/duffle/gp.atom.c

@@ -0,0 +1,82 @@
+#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(U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_gcmd_push, {
+	load_upper_i(reg_transfer, cmd >> 16),
+	or_i_self(   reg_transfer, cmd & 0xFFFF),
+	store_word(  reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_store_rgb8(U1 rr, U1 rg, U1 rb, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rgb8, {
+	store_byte(rr, base, offset + O_(RGB8,r)),
+	store_byte(rg, base, offset + O_(RGB8,g)),
+	store_byte(rb, base, offset + O_(RGB8,b)),	
+})
+
+/* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
+ * byte offset. Internal helper used by the *_format_*_color macros. */
+FI_ Slice_MipsCode ac_pack_color_word(U4 r_base, 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_base, (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_ Slice_MipsCode ac_format_f3_color(U4 r_base, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+
+/* 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_ Slice_MipsCode ac_format_g4_color(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)
+MipsAtomComp_Proc_(ac_format_g4_color, {
+	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.
+ * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
+I_ Slice_MipsCode ac_insert_ot_tag_f3(U4 r_ot_base, U4 r_prim_cursor) MipsAtomComp_Proc_(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_ot_base),                             // 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_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
+})
+
+/* 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. */
+I_ Slice_MipsCode ac_insert_ot_tag_g4(U4 r_ot_base, U4 r_prim_cursor) MipsAtomComp_Proc_(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_ot_base),                              // 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_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

@@ -1,7 +1,6 @@
 /* ============================================================================
  *  duffle DSL Suffix Conventions
  *  ============================================================================
- *
  *  Every mnemonic in this header follows the same suffix grammar:
  *
  *  Primitive commands: gp0_cmd_poly_f3 = 0x20    (byte opcode)
@@ -26,8 +25,7 @@
  *    0. Opcode IDs             gp0_cmd_poly_f3 = 0x20
  *
  *  Vendor mnemonics (gte_mtc2, gte_mfc2, etc.) are NOT in this header.
- *  They live in the opt-in `gp_vendor_sym.h` for users who prefer the
+ *  They live in the opt-in `gp_vendor_sym.h` for users who prefer the PSYQ-style names.
- *  PSYQ-style names.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -41,33 +39,28 @@
 /* ============================================================================
  *  Hardware MMIO Addresses
  *  ============================================================================
- *
+ *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2 0x1F800000+.
- *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2
+ *  GP0 (offset 0x10) is the data port (commands + params).
- *  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).
- * The 16-bit upper half `IO_BASE_ADDR_HI16` is the form used by
+ * The 16-bit upper half `IO_BASE_ADDR_HI16` is the form used by tape-side macros that pin a register
- * tape-side macros that pin a register to hold the IO base and access
+ * to hold the IO base and access ports via offsets:
- * ports via offsets — `lui $reg, 0x1F80` (1 word) then `sw $data,
+ *   `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
- * GPIO_PORT*_OFFSET($reg)` (1 word). Mirrors the `IO_BASE_ADDR equ
+ * Mirrors the `IO_BASE_ADDR equ 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s. */
- * 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s.
- *
- * See lottes_tape.h `R_GpIoBase` + `mac_gp0_send_imm` for the
- * wave-context form that composes these primitives. */
 enum {
-    IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
+	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_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
 
-    /* Offsets from IO_BASE_ADDR to each port. Used by tape-side macros
+	/* 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:
+		* that pin a register to IO_BASE_ADDR and access ports via offsets:
-     *   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
+		*   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
-     *   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
+		*   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
-    GPIO_PORT0_OFFSET = 0x1810,
+	GPIO_PORT0_OFFSET = 0x1810,
-    GPIO_PORT1_OFFSET = 0x1814,
+	GPIO_PORT1_OFFSET = 0x1814,
 
-    HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
+	HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
-    HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
+	HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
 };
 
 #define HW_GP0 C_(U4 V_*, HW_GP0_ADDR)
@@ -79,93 +72,83 @@ enum {
 /* ============================================================================
  *  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.
- *  8-bit GP0 opcodes (the upper byte of a primitive's first word).
+ *  NO macro body past this point uses a raw shift or raw mask.
- *  These are the BYTE only; pre-baked 32-bit words are in §10.4.
+ *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
- *  The layer-1 bitfield-layout constants live in the same enum block
- *  so the encoder in §10.4 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.
  * ============================================================================ */
 enum {
-    gp0_cmd_Nop             = 0x00,
+	gp0_cmd_Nop             = 0x00,
 
-    /* Cache management */
+	/* Cache management */
-    gp0_cmd_ClearCache      = 0x01,
+	gp0_cmd_ClearCache      = 0x01,
-    gp0_cmd_FillVram        = 0x02,
+	gp0_cmd_FillVram        = 0x02,
-    gp0_cmd_CopyVram        = 0x80,
+	gp0_cmd_CopyVram        = 0x80,
-    gp0_cmd_CopyVramChained = 0x81,
+	gp0_cmd_CopyVramChained = 0x81,
-    gp0_cmd_ReadVram        = 0xC0,
+	gp0_cmd_ReadVram        = 0xC0,
 
-    /* Polygons */
+	/* Polygons */
-    gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
+	gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
-    gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
+	gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
-    gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
+	gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
-    gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
+	gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
-    gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
+	gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
-    gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
+	gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
-    gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
+	gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
-    gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
+	gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
 
-    /* Lines */
+	/* Lines */
-    gp0_cmd_line_f2         = 0x40,
+	gp0_cmd_line_f2         = 0x40,
-    gp0_cmd_line_g2         = 0x50,
+	gp0_cmd_line_g2         = 0x50,
 
-    /* Sprites + Tiles + Rects */
+	/* Sprites + Tiles + Rects */
-    gp0_cmd_sprt_1          = 0x64,
+	gp0_cmd_sprt_1          = 0x64,
-    gp0_cmd_sprt_8          = 0x74,
+	gp0_cmd_sprt_8          = 0x74,
-    gp0_cmd_sprt_16         = 0x7C,
+	gp0_cmd_sprt_16         = 0x7C,
-    gp0_cmd_tile_1          = 0x60,
+	gp0_cmd_tile_1          = 0x60,
-    gp0_cmd_tile_8          = 0x68,
+	gp0_cmd_tile_8          = 0x68,
-    gp0_cmd_tile_16         = 0x70,
+	gp0_cmd_tile_16         = 0x70,
 
-    /* State setters (not drawing primitives; set render context).
+	/* State setters (not drawing primitives; set render context). */
-     * Per PSX-SPX graphicsprocessingunitgpu.md §"GP0 Other Commands". */
+	gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
-    gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
+	gp0_cmd_SetTextureWindow     = 0xE2,
-    gp0_cmd_SetTextureWindow     = 0xE2,
+	gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
-    gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
+	gp0_cmd_SetDrawArea_BotRight = 0xE4,
-    gp0_cmd_SetDrawArea_BotRight = 0xE4,
+	gp0_cmd_SetDrawOffset        = 0xE5,
-    gp0_cmd_SetDrawOffset        = 0xE5,
+	gp0_cmd_SetMaskBit           = 0xE6,
-    gp0_cmd_SetMaskBit           = 0xE6,
 
-    /* bitfield shifts / widths / masks ----
+	/* bitfield shifts / widths / masks ----
-     *
+		* Generic GP0/GP1 command byte (upper 8 bits of every word sent to either port). */
-     * Generic GP0/GP1 command byte (upper 8 bits of every word sent
+	gp0_cmd_shift = 24,
-     * to either port). Used by `enc_gp0_cmd(cmd)` and friends below. */
+	gp0_cmd_width =  8,
-    gp0_cmd_shift = 24,
+	gp0_cmd_mask  = 0xFF,
-    gp0_cmd_width =  8,
-    gp0_cmd_mask  = 0xFF,
 
-    /* Color word layout (lives in Poly_F3.color, Poly_G4.c0..c3, etc.):
+	/* Color word layout (lives in Poly_F3.color, Poly_G4.c0..c3, etc.):
-     *   bits 31..24 = command byte
+		*   bits 31..24 = command byte
-     *   bits 23..16 = BLUE
+		*   bits 23..16 = BLUE
-     *   bits 15..08 = GREEN
+		*   bits 15..08 = GREEN
-     *   bits 07..00 = RED     (PSX GPU is BGR, NOT RGB) */
+		*   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_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_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_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,
+	gp0_color_red_shift   =  0, gp0_color_red_width   = 8, gp0_color_red_mask   = 0xFF,
 };
 
 /* ============================================================================
  *  Layer 1.5 (per-field encoders) + Layer 2 (composite) + Layer 3 (semantic GP0 word builders)
  *  ============================================================================
- *
+ *  Layer 1.5 encoders take one field's value, mask it to its own width, and shift it to its own position. 
- *  Layer 1.5 encoders take one field's value, mask it to its own width,
+ *  Mirrors `enc_op` / `enc_rs` / `enc_rt` in mips.h and `enc_gte_sf` / `enc_gte_mx` in gte.h. 
- *  and shift it to its own position. Mirrors `enc_op` / `enc_rs` /
+ *  Layer-2 composite encoders OR the per-field encoders together; layer-3 semantic macros delegate to the composites.
- *  `enc_rt` in mips.h and `enc_gte_sf` / `enc_gte_mx` in gte.h. Layer-2 composite encoders 
- *  OR the per-field encoders together; layer-3 semantic macros delegate to the composites.
  *  No raw shifts or magic numbers in any macro body below this point.
  * ============================================================================ */
 
 /* ---- 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_mask)         << gp0_cmd_shift)
 
-#define enc_gp0_color_cmd(cmd) (((cmd) & gp0_color_cmd_mask)     << gp0_color_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_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_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_b(b)     (((b)   & gp0_color_blue_mask)  << 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))
@@ -186,73 +169,73 @@ 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_clear_cache()       enc_gp0_cmd_word(gp0_cmd_ClearCache)
-#define gp0_word_fill_vram()   enc_gp0_cmd_word(gp0_cmd_FillVram)
+#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_copy_vram()         enc_gp0_cmd_word(gp0_cmd_CopyVram)
-#define gp0_word_read_vram()   enc_gp0_cmd_word(gp0_cmd_ReadVram)
+#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)
  *  ============================================================================
- *
+ *  GP1 status bits are read from HW_GP1;
- *  GP1 status bits are read from HW_GP1; ctrl writes use GP1 commands
+ *  ctrl writes use GP1 commands packed into 32-bit words 
- *  packed into 32-bit words (cmd byte in the upper 8 bits via
+ *  (cmd byte in the upper 8 bits via `enc_gp0_cmd(cmd)`).
- *  `enc_gp0_cmd(cmd)` — never a raw shift).
  * ============================================================================ */
 enum {
-    gp1_cmd_Reset                  = 0x00,
+	gp1_cmd_Reset                  = 0x00,
-    gp1_cmd_ResetCmdBuffer         = 0x01,
+	gp1_cmd_ResetCmdBuffer         = 0x01,
-    gp1_cmd_AcknowledgeIRQ         = 0x02,
+	gp1_cmd_AcknowledgeIRQ         = 0x02,
-    gp1_cmd_DisplayEnable          = 0x03,
+	gp1_cmd_DisplayEnable          = 0x03,
-    gp1_cmd_DMADirection           = 0x04,
+	gp1_cmd_DMADirection           = 0x04,
-    gp1_cmd_StartDisplayArea       = 0x05,
+	gp1_cmd_StartDisplayArea       = 0x05,
-    gp1_cmd_HorizontalDisplayRange = 0x06,
+	gp1_cmd_HorizontalDisplayRange = 0x06,
-    gp1_cmd_VerticalDisplayRange   = 0x07,
+	gp1_cmd_VerticalDisplayRange   = 0x07,
-    gp1_cmd_DisplayMode            = 0x08,
+	gp1_cmd_DisplayMode            = 0x08,
-    /* Note: GP1 only has commands 0x00..0x08. 
+	/* Note: GP1 only has commands 0x00..0x08. 
-		 * The state-setter commands (SetTextureWindow, * SetDrawArea*, 
+		* The state-setter commands (SetTextureWindow, * SetDrawArea*, SetDrawOffset, SetMaskBit) 
-		 * SetDrawOffset, SetMaskBit) live in the GP0 enum as * 0xE1..0xE6. 
+		* live in the GP0 enum as * 0xE1..0xE6.
-		 * DrawArea word builders are below as GP0s * macros 
+		* DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
-		 * (since they emit GP0 commands). */
 
-    /* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
+	/* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
-     * Bit positions match the encoder shifts below; values are the
+		* 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). */
-     * *payload* bits only (the cmd byte is OR'd in by enc_gp1_disp_mode_word). */
+	gp1_disp_HRes_256   = 0x0,
-    gp1_disp_HRes_256   = 0x0,
+	gp1_disp_HRes_320   = 0x1,
-    gp1_disp_HRes_320   = 0x1,
+	gp1_disp_HRes_512   = 0x2,
-    gp1_disp_HRes_512   = 0x2,
+	gp1_disp_HRes_640   = 0x3,
-    gp1_disp_HRes_640   = 0x3,
+	gp1_disp_VRes_240   = 0x0,
-    gp1_disp_VRes_240   = 0x0,
+	gp1_disp_VRes_480   = 0x1,
-    gp1_disp_VRes_480   = 0x1,
+	gp1_disp_Color15    = 0x0,
-    gp1_disp_Color15    = 0x0,
+	gp1_disp_Color24    = 0x1,
-    gp1_disp_Color24    = 0x1,
+	gp1_disp_VInterlace = 0x1,
-    gp1_disp_VInterlace = 0x1,
 
-    /* ---- Layer 1: GP1 display-mode + range + draw-area shifts/masks ---- */
+	/* ---- 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_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_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_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,
+	gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1, gp1_disp_interlace_mask = 0x1,
 
-    /* GP1 horizontal display range: bits 0..11 = X2, bits 12..23 = X1 */
+	/* 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_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_hrange_x2_shift =  0, gp1_hrange_x2_width = 12, gp1_hrange_x2_mask = 0xFFF,
 
-    /* GP1 vertical display range: bits 0..9 = Y2, bits 10..19 = Y1 */
+	/* 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_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_vrange_y2_shift =  0, gp1_vrange_y2_width = 10, gp1_vrange_y2_mask = 0x3FF,
 
-    /* GP1 draw area (top-left or bottom-right): bits 0..9 = X, bits 10..19 = Y
+	/* 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) */
+		* (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_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_y_shift = 10, gp1_draw_y_width = 10, gp1_draw_y_mask = 0x3FF,
 };
 
 /* ---- 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_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_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_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_interlace(i) (((i) & gp1_disp_interlace_mask) << 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)
@@ -267,12 +250,16 @@ enum {
 #define enc_gp1_vrange_word(y1, y2)        (enc_gp0_cmd(gp1_cmd_VerticalDisplayRange)   | enc_gp1_vrange_y1(y1) | enc_gp1_vrange_y2(y2))
 
 /* ---- Layer 2: GP0 state-setter composite encoders ----
- * GP0(0xE3) SetDrawArea top-left and GP0(0xE4) SetDrawArea bottom-right
+ * GP0(0xE3) SetDrawArea top-left and GP0(0xE4) SetDrawArea bottom-right both use the same X/Y 10-bit signed payload as GP1 DisplayRange. */
- * both use the same X/Y 10-bit signed payload as GP1 DisplayRange. */
 #define enc_gp0_draw_area_tl_word(x, y)    (enc_gp0_cmd(gp0_cmd_SetDrawArea_TopLeft)  | enc_gp1_draw_x(x) | enc_gp1_draw_y(y))
 #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)
@@ -290,47 +277,43 @@ enum {
 /* ============================================================================
  *  Pre-baked GPU state words
  *  ============================================================================
- *
+ *  Common command words for boot-time GPU init and standard display configurations.
- *  Common command words for boot-time GPU init and standard
- *  display configurations. Each one is a pure compile-time integer
- *  constant ready to drop into a `.word` directive.
- *
- *  These are the equivalents of the `gp_HorizontalDisplayRange_3168_608`,
- *  `gp_VerticalDisplayRange_264_24`, `gp_DisplayMode_320x240_15bit_NTSC`,
- *  `gp_SetDrawMode_DrawAllowed`, `gp_DMA_*` `.equ`s from the pre-rewrite
- *  gp.h / graphics_hello/gp.s, rebuilt using the layer-cake encoders so
- *  no magic numbers appear in any body.
  * ============================================================================ */
 
 /* ---- Display enable (1-bit payload on DisplayEnable cmd) ---- */
 #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_Off            = 0,
-    gp1_dma_dir_FIFO           = 1,
+	gp1_dma_dir_FIFO           = 1,
-    gp1_dma_dir_CPU_to_GPU     = 2,
+	gp1_dma_dir_CPU_to_GPU     = 2,
-    gp1_dma_dir_GPUREAD_to_CPU = 3,
+	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 */
+	/* NTSC horizontal range: X1=608, X2=3168 */
-    gp1_hrange_NTSC_x1 = 0x260,
+	gp1_hrange_NTSC_x1 = 0x260,
-    gp1_hrange_NTSC_x2 = 0xC60,
+	gp1_hrange_NTSC_x2 = 0xC60,
-    /* PAL horizontal range (same as NTSC for most CRTs) */
+	/* PAL horizontal range (same as NTSC for most CRTs) */
-    gp1_hrange_PAL_x1  = 0x260,
+	gp1_hrange_PAL_x1  = 0x260,
-    gp1_hrange_PAL_x2  = 0xC60,
+	gp1_hrange_PAL_x2  = 0xC60,
 
-    /* NTSC vertical range: Y1=24, Y2=264 */
+	/* NTSC vertical range: Y1=24, Y2=264 */
-    gp1_vrange_NTSC_y1 = 24,
+	gp1_vrange_NTSC_y1 = 24,
-    gp1_vrange_NTSC_y2 = 264,
+	gp1_vrange_NTSC_y2 = 264,
-    /* PAL vertical range: Y1=24, Y2=504 */
+	/* PAL vertical range: Y1=24, Y2=504 */
-    gp1_vrange_PAL_y1  = 24,
+	gp1_vrange_PAL_y1  = 24,
-    gp1_vrange_PAL_y2  = 504,
+	gp1_vrange_PAL_y2  = 504,
 };
 
 #define gp1_word_horizontal_range_ntsc enc_gp1_hrange_word(gp1_hrange_NTSC_x1, gp1_hrange_NTSC_x2)
@@ -339,16 +322,51 @@ enum {
 #define gp1_word_vertical_range_pal    enc_gp1_vrange_word(gp1_vrange_PAL_y1,  gp1_vrange_PAL_y2)
 
 /* ---- Draw-mode setting (TPage / draw-area allowance) ---- */
-/* The pre-baked "drawing enabled" word is the standard post-init state. */
+/* 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_320x240 enc_gp0_draw_area_br_word(319, 239)
-#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(640, 480)
+#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(639, 479)
 
 #pragma endregion GPU Ports & Commands
 
@@ -359,9 +377,9 @@ enum {
  *  Read from HW_GP1; the lower bits are DMA-block-size (variable-width).
  * ============================================================================ */
 enum {
-    gp1_Status_BitReady          = 31,
+	gp1_Status_BitReady          = 31,
-    gp1_Status_BitSendingDMA     = 25,
+	gp1_Status_BitSendingDMA     = 25,
-    gp1_Status_DMABlockSizeShift =  0,
+	gp1_Status_DMABlockSizeShift =  0,
 };
 
 #define gp1_status_is_ready()       ((HW_GP1[0] >> gp1_Status_BitReady)      & 1)
@@ -372,17 +390,14 @@ enum {
 /* ============================================================================
  *  Primitive structs (8 polygon variants + tag)
  *  ============================================================================
+ *  Each struct follows the GPU-documented memory layout for the corresponding primitive command. 
+ *  The PolyTag is the OT-link header; the rest of the struct is the primitive's body.
  *
- *  Each struct follows the GPU-documented memory layout for the corresponding
+ *  The current working layouts match the existing demo 
- *  primitive command. The PolyTag is the OT-link header; the rest of the
+ *  (floor_tri uses Poly_F3; cube_tri uses Poly_G4). 
- *  struct is the primitive's body.
+ *  They are NOT necessarily byte-identical to the PSX-SPX reference layout. 
- *
+ *  The demo layout uses color+vertex interleaving that doesn't match the standard PSX SDK file format. 
- *  The current working layouts match the existing demo (floor_tri uses
+ *  For PSX-SDK file compatibility, the textured variants (FT*, GT*) would need layout adjustments.
- *  Poly_F3; cube_tri uses Poly_G4). They are NOT necessarily byte-identical
- *  to the PSX-SPX reference layout — the demo layout uses color+vertex
- *  interleaving that doesn't match the standard PSX SDK file format. For
- *  PSX-SDK file compatibility, the textured variants (FT*, GT*) would need
- *  layout adjustments; out of scope for this track.
  * ============================================================================ */
 
 /* ---------- RGB8 (3-byte packed color) ---------- */
@@ -390,13 +405,13 @@ 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 {
+// enum {
-    polytag_len_bits  =  8,
+// 	PolyTag_len_bits  =  8,
-    polytag_addr_bits = 24,
+// 	PolyTag_addr_bits = 24,
-};
+// };
 typedef Struct_(PolyTag) {
 	union {
-		U4 bf_addr_len;
+		U4 code;
 		struct {
 			U4 addr: 24;
 			U4 len:  8;
@@ -404,112 +419,111 @@ typedef Struct_(PolyTag) {
 	};
 };
 
-/* DSL cast convention: every cast uses `C_()`, every pointer qualifier
+/* DSL cast convention: every cast uses `C_()`, every pointer qualifier is `R_` (restrict) or `V_` (volatile).
- * is `R_` (restrict) or `V_` (volatile). No raw C-style casts. RHS values
+ * No raw C-style casts. RHS values are assumed to be `U4` — caller passes a `U4` directly. */
- * 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
+/* `set_code` is no longer in the new PolyTag design — the code byte lives in the primitive body
- *  in the primitive body (e.g. `((Poly_F3*)(p))->code`), not in the tag.
+ * (e.g. `((Poly_F3*)(p))->code`), not in the tag.
- *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the
+ *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters,
- *  `set_poly_*` setters, which set both the tag's length and the code. */
+ *  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;
+	U4   tag;
-    RGB8 color;
+	RGB8 color;
-    B1   code;
+	B1   code;
-    union {
+	union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
-        A3_V2_S2 points;
+		A3_V2_S2 points;
-    };
+	};
 };
 
 /* ---------- Poly_F4 (Flat Quad; 6 words) ---------- */
 typedef Struct_(Poly_F4) {
-    U4   tag;
+	U4   tag;
-    RGB8 color;
+	RGB8 color;
-    B1   code;
+	B1   code;
-    union {
+	union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
-        A4_V2_S2 points;
+		A4_V2_S2 points;
-    };
+	};
 };
 
-/* ---------- Poly_G3 (Gouraud Triangle; 6 words) ---------- */
+/* ---------- Poly_G3 (Gouraud Triangle; 7 words) ---------- */
 typedef Struct_(Poly_G3) {
-    U4    tag; RGB8 c0; B1 code;
+	U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
+	V2_S2 p2;
 };
 
-/* ---------- Poly_G4 (Gouraud Quad; 5 words in the demo's interleaved layout) ---------- */
+/* ---------- Poly_G4 (Gouraud Quad; 9 words) ---------- */
 typedef Struct_(Poly_G4) {
-    U4    tag; RGB8 c0; B1 code;
+	U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
+	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;
+	U4   tag;
-    RGB8 color;
+	RGB8 color;
-    B1   code;
+	B1   code;
-    U4   tpage;
+	U4   tpage;
-    U4   clut;
+	U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
+	V2_S2 p2; U1 u2; U1 v2;
 };
 
-/* ---------- Poly_FT4 (Flat Textured Quad; placeholder layout) ---------- */
+/* ---------- Poly_FT4 (Flat Textured Quad) ---------- */
 typedef Struct_(Poly_FT4) {
-    U4   tag;
+	U4   tag;
-    RGB8 color;
+	RGB8 color;
-    B1   code;
+	B1   code;
-    U4   tpage;
+	U4   tpage;
-    U4   clut;
+	U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
+	V2_S2 p2; U1 u2; U1 v2;
-    V2_S2 p3; U1 u3; U1 v3;
+	V2_S2 p3; U1 u3; U1 v3;
 };
 
-/* ---------- Poly_GT3 (Gouraud Textured Triangle; placeholder layout) ---------- */
+/* ---------- Poly_GT3 (Gouraud Textured Triangle) ---------- */
 typedef Struct_(Poly_GT3) {
-    U4    tag; RGB8 c0; B1 code;
+	U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
+	V2_S2 p2;
-    U4    tpage;
+	U4    tpage;
-    U4    clut;
+	U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
+	V2_S2 tp2; U1 u2; U1 v2;
 };
 
-/* ---------- Poly_GT4 (Gouraud Textured Quad; placeholder layout) ---------- */
+/* ---------- Poly_GT4 (Gouraud Textured Quad) ---------- */
 typedef Struct_(Poly_GT4) {
-    U4    tag; RGB8 c0; B1 code;
+	U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
+	V2_S2 p3;
-    U4    tpage;
+	U4    tpage;
-    U4    clut;
+	U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
+	V2_S2 tp2; U1 u2; U1 v2;
-    V2_S2 tp3; U1 u3; U1 v3;
+	V2_S2 tp3; U1 u3; U1 v3;
 };
 
-/* ---------- Primitive setters (C-level, no emitted words) ----------
+/* ---------- Primitive setters (C-level) ----------
  * DSL cast convention: every cast via C_(), every pointer via R_/V_. */
 #define set_poly_f3(p)  set_len(p, 4),  C_(Poly_F3_R, p)->code = gp0_cmd_poly_f3
 #define set_poly_ft3(p) set_len(p, 7),  C_(Poly_FT3_R,p)->code = gp0_cmd_poly_ft3
@@ -530,7 +544,6 @@ typedef Struct_(Poly_GT4) {
 /* ============================================================================
  *  Texture Page (TPage) bit layout
  *  ============================================================================
- *
  *  The TPage data word sent via GP0(0x2X) has:
  *    bits  0..3   = texture page X    (4 bits, 64-px units, 0..16)
  *    bit   4      = texture page Y    (1 bit,  64-px units, 0/1)
@@ -542,26 +555,29 @@ typedef Struct_(Poly_GT4) {
  *    bits 12..31  = reserved (zero)
  * ============================================================================ */
 enum {
-    /* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
+	/* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
-    gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4, gp0_tpage_x_mask            = 0xF,
+	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_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_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_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_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_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,
+	gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1, gp0_tpage_tex_disable_mask  = 0x1,
 
-    /* TPage color-depth payload values (NOT bit positions — these go in
+	/* TPage color-depth payload values (NOT bit positions — these go in
-     * the 2-bit field at gp0_tpage_color_depth_shift). */
+	 * the 2-bit field at gp0_tpage_color_depth_shift). */
-    gp0_tpage_color_4bpp  = 0x0,
+	gp0_tpage_color_4bpp  = 0x0,
-    gp0_tpage_color_8bpp  = 0x1,
+	gp0_tpage_color_8bpp  = 0x1,
-    gp0_tpage_color_16bpp = 0x2,
+	gp0_tpage_color_16bpp = 0x2,
 
-    /* TPage semi-transparency mode payload values (NOT bit positions). */
+	/* Default TPage value libpsyx's SetDefDrawEnv writes (matches the `li v1, 10; sh v1, 20(v0)` sequence at C11_only.elf:0x8001273C). */
-    gp0_tpage_semi_trans_none  = 0x0,
+	gp0_tpage_default    = 10,
-    gp0_tpage_semi_trans_alpha = 0x1,
+
-    gp0_tpage_semi_trans_add   = 0x2,
+	/* TPage semi-transparency mode payload values (NOT bit positions). */
-    gp0_tpage_semi_trans_sub   = 0x3,
+	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. ---- */
@@ -575,26 +591,25 @@ enum {
 
 /* ---- 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_x(x)                       \
-   | enc_gp0_tpage_y(y)                       \
+	| enc_gp0_tpage_y(y)                       \
-   | enc_gp0_tpage_semi_trans(semi_trans)     \
+	| enc_gp0_tpage_semi_trans(semi_trans)     \
-   | enc_gp0_tpage_color_depth(color_depth)   \
+	| enc_gp0_tpage_color_depth(color_depth)   \
-   | enc_gp0_tpage_dither(dither)             \
+	| enc_gp0_tpage_dither(dither)             \
-   | enc_gp0_tpage_draw_to_disp(draw_to_disp) \
+	| enc_gp0_tpage_draw_to_disp(draw_to_disp) \
-   | enc_gp0_tpage_tex_disable(tex_disable))
+	| 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
 /* ============================================================================
  *  CLUT (Color Look-Up Table) semantics
  *  ============================================================================
- * 
  *  CLUT is loaded into VRAM by sending a GP0 command whose payload is:
  *    bits  0..5   = Y in 16-px units (palette row)
  *    bits  6..14  = X in 16-px units (palette column)
@@ -602,12 +617,12 @@ 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 ---- */
+	/* ---- Layer 1: CLUT bitfield shifts / widths / masks ---- */
-    gp0_clut_y_shift = 0, gp0_clut_y_width = 6, gp0_clut_y_mask = 0x3F,
+	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,
+	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. */
+	/* CLUT-load cmd-byte variants — the upper byte of the GP0 word. */
-    gp0_clut_cmd_Load4bpp = 0x20,
+	gp0_clut_cmd_Load4bpp = 0x20,
-    gp0_clut_cmd_Load8bpp = 0x25,
+	gp0_clut_cmd_Load8bpp = 0x25,
 };
 
 /* ---- Layer 1.5: CLUT per-field encoders ---- */
@@ -627,7 +642,6 @@ enum {
 /* ============================================================================
  *  TIM file format constants and headers
  *  ============================================================================
- *
  *  TIM (Sony .TIM texture image) file structure:
  *    +0x00  U4 file_id (always 0x10 = TIM magic)
  *    +0x04  U4 version (always 0x00 for v1)
@@ -645,31 +659,30 @@ enum {
  *         +0x06 U2 px_height
  *         +0x08 ... pixel data
  *
- *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that
+ *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that emits the necessary GP0 commands.
- *  emits the necessary GP0 commands. Stoppped for now at the
+ *  Stoppped for now at the struct + enum level.
- *  struct + enum level for this track.
  * ============================================================================ */
 enum {
-    tim_file_id_magic = 0x10,
+	tim_file_id_magic = 0x10,
-    tim_type_4bpp     = 0x00,
+	tim_type_4bpp     = 0x00,
-    tim_type_8bpp     = 0x01,
+	tim_type_8bpp     = 0x01,
-    tim_type_16bpp    = 0x02,
+	tim_type_16bpp    = 0x02,
-    tim_type_32bpp    = 0x03,
+	tim_type_32bpp    = 0x03,
-    tim_type_mixed    = 0x04,
+	tim_type_mixed    = 0x04,
-    tim_flag_has_clut = 0x08,
+	tim_flag_has_clut = 0x08,
 };
 
 typedef Struct_(TIM_Header) {
-    U4 file_id;        /* always 0x10 = "TIM" magic */
+	U4 file_id;       /* always 0x10 = "TIM" magic */
-    U4 version;       /* ignored; always 0 */
+	U4 version;       /* ignored; always 0 */
-    U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
+	U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
 };
 typedef Struct_(TIM_SectionHeader) {
-    U4 section_length;  /* bytes in this section including this header */
+	U4 section_length;  /* bytes in this section including this header */
-    U2 org_x;           /* origin in VRAM */
+	U2 org_x;           /* origin in VRAM */
-    U2 org_y;
+	U2 org_y;
-    U2 width;           /* width in pixels */
+	U2 width;           /* width in pixels */
-    U2 height;          /* height in pixels */
+	U2 height;          /* height in pixels */
 };
 #pragma endregion TIM File Format
 
@@ -678,35 +691,24 @@ typedef Struct_(TIM_SectionHeader) {
  *  Tape-side GPU operations (NOT in this header)
  *  ============================================================================
  *
- *  No `mac_gp0_send` or related macros live in gp.h. Rationale: the
+ *  No `mac_gp0_send` or related macros live in gp.h.
- *  Lottes tape model uses OT-DMA for primitive submission, so atom bodies
+ *  Rationale: the Lottes tape model uses OT-DMA for primitive submission, so atom bodies write to main RAM (the OT/primitive buffer)
- *  write to main RAM (the OT/primitive buffer) and to GTE state — never
+ *  and to GTE state — never directly to the GPU ports at 0x1F801810 / 0x1F801814.
- *  directly to the GPU ports at 0x1F801810 / 0x1F801814. See
+ *  See `mac_format_f3_color`, `mac_insert_ot_tag`, `mac_gte_store_f3` in lottes_tape.h for the patterns atom bodies actually use.
- *  `mac_format_f3_color`, `mac_insert_ot_tag`, `mac_gte_store_f3` in
- *  lottes_tape.h for the patterns atom bodies actually use.
  *
- *  If a feature need arises requires tape-side GPU port writes (e.g. DMA-kick to
+ *  If a feature need arises requires tape-side GPU port writes 
- *  start GPU consumption of the OT, VBlank sync via GP1 status poll),
+ *  (e.g. DMA-kick to start GPU consumption of the OT, VBlank sync via GP1 status poll),
- *  the right home is `lottes_tape.h` alongside the rest of the `mac_*`
+ *  the right home is `lottes_tape.h` alongside the rest of the `mac_*` family:
- *  family — the encoder infrastructure is already in place:
+ *    1. The caller pins a register to hold the IO base, e.g. register U4 r_io rgcc(R_T4) = IO_BASE_ADDR;
+ *       The compiler emits `lui R_T4, IO_BASE_ADDR_HI16` outside the atom body (in the C prologue before tape_run).
+ *    2. The atom body uses `store_word(R_data, R_T4, GPIO_PORT0_OFFSET)` to write to GP0, and `store_word(R_data, R_T4, GPIO_PORT1_OFFSET)`
+ *       to write to GP1. Both are preprocessor-encodable because R_T4 is a fixed register and the GPIO_PORT*_OFFSET constants
+ *       fit in the `sw`'s 16-bit signed offset field. No placeholder-pun, no asm constraints, no hidden register choice.
+ *       Same pattern as the old graphics_hello/hello_gp_routines.s `reg_io_offset`/`gcmd_push` convention.
  *
- *    1. The caller pins a register to hold the IO base, e.g.
+ *  This mirrors the existing tape-side wave-context discipline: 
- *         register U4 r_io rgcc(R_T4) = IO_BASE_ADDR;
+ *  the caller binds the IO-base register via `rgcc()`, the macro assumes the binding is in effect,
- *       The compiler emits `lui R_T4, IO_BASE_ADDR_HI16` outside the
+ *  and the encoding falls out at preprocessor time. 
- *       atom body (in the C prologue before tape_run).
+ *  No additional GPU-domain macro layer required.
- *
- *    2. The atom body uses `store_word(R_data, R_T4, GPIO_PORT0_OFFSET)`
- *       to write to GP0, and `store_word(R_data, R_T4, GPIO_PORT1_OFFSET)`
- *       to write to GP1. Both are preprocessor-encodable because R_T4 is
- *       a fixed register and the GPIO_PORT*_OFFSET constants fit in the
- *       `sw`'s 16-bit signed offset field. No placeholder-pun, no asm
- *       constraints, no hidden register choice. Same pattern as the
- *       old graphics_hello/hello_gp_routines.s `reg_io_offset`/`gcmd_push`
- *       convention.
- *
- *  This mirrors the existing tape-side wave-context discipline: the
- *  caller binds the IO-base register via `rgcc()`, the macro assumes
- *  the binding is in effect, and the encoding falls out at preprocessor
- *  time. No additional GPU-domain macro layer required.
  * ============================================================================ */
 #pragma endregion Tape-Side Macros

code/duffle/gp_vendor_sym.h

@@ -2,10 +2,8 @@
  *  duffle DSL — GPU Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the PSYQ-style CamelCase aliases for the canonical duffle GPU
+ *  Provides the PSYQ-style CamelCase aliases for the duffle GPU primitive setters and OT operations.
- *  primitive setters and OT operations. The duffle snake_case names are
+ *  The duffle snake_case names are primary; this header is for users who prefer the PSYQ SDK function names from the legacy C API.
- *  primary; this header is for users who prefer the PSYQ SDK function
- *  names from the legacy C API.
  *
  *  USAGE:   #include "duffle/gp_vendor_sym.h"   // after gp.h
  *
@@ -23,15 +21,11 @@
  *    OT operations:
  *      AddPrim(ot, p)  -> orderingtbl_add_primitive(ot, p)
  *
- *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and
+ *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and descriptive; no vendor alias is provided for them.
- *  descriptive; no vendor alias is provided for them.
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (tape_atom.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries (the ones emitted by
- *  mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13
- *  (objdump byte-identical) holds.
  *
+ *  The vendor mnemonics are NOT registered with the duffle word-count metadata (word_counts.metadata.h).
+ *  They expand to the duffle macros which DO have word-count entries
+ *  (the ones emitted by mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13 (objdump byte-identical) holds.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/gte.atom.c

@@ -0,0 +1,76 @@
+#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(U4 r_face_cusor, U4 r_i0, U4 r_i1, U4 r_i2) atom_dbg_skip MipsAtomComp_Proc_(ac_load_tri_indices, {
+	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)),
+})
+
+/* 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(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_f3, {
+	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(U4 r_vert_base, U4 r_v0, U4 r_v1, U4 r_v2) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_load_tri_verts, {
+	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)), 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)), 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)), 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(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_g4_p012, {
+	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(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_g4_p3, { gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3)) })
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Bsked Atoms
+
+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()
+};
+
+#pragma endregion Baked Atoms

code/duffle/gte.h

@@ -17,9 +17,8 @@
  *            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
+ *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are NOT in this header.
- *  NOT in this header. They live in the opt-in `gte_vendor_sym.h` for
+ *  They are in the opt-in `gte_vendor_sym.h` for users who prefer the textbook MIPS assembly mnemonics.
- *  users who prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -34,76 +33,27 @@
  *  gte.h — Geometry Transformation Engine (COP2) for the PS1
  * ============================================================================
  *
- *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word`
+ *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word` constants from C.
- *  constants from C. No GCC inline-assembly string syntax in the code body.
+ *  No GCC inline-assembly string syntax in the code body.
- *
- *  PHILOSOPHY
- *  ----------
- *  1. A 32-bit instruction word is composed from per-field encoders. Each
- *     encoder knows only its own bit range; the composite ORs them together.
- *     No magic numbers inside any encoder body. Every shift and mask is a
- *     named constant from the bitfield-layout enum below.
- *
- *  2. Pure (compile-time) instructions. Every GTE *command* (RTPS, RTPT,
- *     NCLIP, MVMVA, …) and every COP2 *transfer* (ctc2/cfc2) with a constant
- *     rs/rt/rd — are emitted as a single integer constant via
- *     `asm_inline(...)` from gcc_asm.h. The C compiler constant-folds
- *     these into `.word` directives in .rodata.
- *
- *  3. Runtime-base-register instructions (lwc2, swc2, lw, sw, …) cannot be
- *     a pure compile-time word because the `rs` field is chosen by the
- *     compiler at codegen. For these we use a "placeholder-pun" pattern:
- *     a fixed register number (R_T4 = $12) is baked into the rs field of
- *     the `.word` constant, and the macro declares a `"r"(arg)` input
- *     constraint plus a clobber on the same register. The compiler is
- *     therefore *forced* to bind `arg` to that exact register, and the
- *     constant is correct.
- *
- *  USAGE
- *  -----
- *      // Pure command sequence — all bits compile-time:
- *      asm volatile(
- *          asm_inline( gte_cmd_rtpt , gte_cmd_nclip , gte_cmd_avsz3 )
- *          asm_clobber( clbr_volatile_gprs )
- *      );
- *
- *      // Runtime-base-register load — caller picks the base GPR:
- *      register V3_S2* p_in_12 __asm__("$12") = verts[0].ptr;
- *      gte_load_v0(p_in_12, R_T4);   // R_T4 = 12 = $t4 = $12
- *
- *      // Three independent bases for an RTPT pipeline:
- *      register V3_S2* p0 gcc_reg(R_T4) = verts[0].ptr;
- *      register V3_S2* p1 gcc_reg(R_T5) = verts[1].ptr;
- *      register V3_S2* p2 gcc_reg(R_T6) = verts[2].ptr;
- *      gte_load_v0(p0, R_T4);
- *      gte_load_v1(p1, R_T5);
- *      gte_load_v2(p2, R_T6);
- *      gte_rtpt();
  *
  *  STYLE NOTES
  *  -----------
- *  - Per-field encoders are named `enc_gte_<field>(value)` and each one
+ *  - Per-field encoders are named `enc_gte_<field>(value)` and each one self-masks its argument before shifting.
- *    self-masks its argument before shifting. Mirrors the `enc_op / enc_rs
+ *    Mirrors the `enc_op / enc_rs / enc_rt / ...` family in mips.h.
- *    / 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.
- *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of
+ *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined for the common cases so call sites read like assembly source.
- *    the per-field encoders, plus the COP2/CO base.
+ *  - All register/field values are enums (not `#define`s) so they show up in debugger symbol tables and IDE autocomplete.
- *  - 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
  *  --------
- *  - gcc_asm.h: the `.word` emitter (`asm_inline`, `asm_clobber`, clobbers)
+ *  - mips.h: The MIPS encoder layer this builds on.
- *  - mips.h:    the MIPS encoder layer this builds on
  */
 
 /* C2 data registers */
 
 /* --- 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
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
- * 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
@@ -222,8 +172,8 @@ 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. Mirrors the
+ * configurable field, used by the ENC_GTE_CMD encoder. 
- * OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
+ * Mirrors the OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
  */
 
 	gte_shift_sf  = 19,  gte_width_sf  = 1,  gte_mask_sf  = 0x1,
@@ -236,10 +186,8 @@ enum {
 
 /* --- 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
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
- * 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. */
- * 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 */
@@ -267,8 +215,9 @@ 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_OFX_Code  24
-#define gte_cr_OFY_Code  31
+#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,
@@ -290,21 +239,16 @@ 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
+ * Spans the 2x2 {From, To} × {Data, Control} register classes that the GTE exposes:
- * 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
+ * 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
- * numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
+ * (which target the data register file on any coprocessor).
- * (which target the data register file on any coprocessor). They are
+ * They are re-aliased here so the four-way table reads like the spec mnemonics (MFC2 / CFC2 / MTC2 / CTC2)
- * re-aliased here so the four-way table reads like the spec mnemonics
+ * and so the encoding lives next to its only consumer (this header).
- * (MFC2 / CFC2 / MTC2 / CTC2) and so the encoding lives next to its
- * only consumer (this header).
  *
- * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2)
+ * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2) live in gte_vendor_sym.h. */
- * 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 */
@@ -314,11 +258,11 @@ 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
+ *  - sub: one of sub_mfc2 / sub_cfc2 / sub_mtc2 / sub_ctc2
- *   - rt:  GPR source/dest
+ *  - rt:  GPR source/dest
- *   - rd:  COP2 register index (0..31):
+ *  - rd:  COP2 register index (0..31):
- *            data class → C2_VXY0_Code..C2_LZCR_Code (gte_in_v0_xy..gte_math_accum2 aliases)
+ *    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 */
+ *    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))
 
 
@@ -358,8 +302,8 @@ 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
+ * For the typical user-facing vector-level load (xy + z as two instructions),
- * instructions), use the higher-level `gte_load_vN` macros below. */
+ * 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)
 
@@ -367,13 +311,12 @@ enum { _C2_TX_SUBS_ = 0
  * Opcode is always MIPS_OP_COP2, RS is always 1 (CO).
  * The lower 25 bits are the GTE-specific command payload.
  *
- * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs`
+ * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs` pattern in mips.h:
- * pattern in mips.h: each one self-masks and shifts its own field, so a
+ * Each one self-masks and shifts its own field, so a caller can build up a GTE command piece by piece
- * caller can build up a GTE command piece by piece (handy for state-driven
+ * (handy for state-driven MVMVA emitters that vary one field at a time).
- * MVMVA emitters that vary one field at a time).
  *
- * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command
+ * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command word in one go.
- * word in one go. It just ORs the per-field encoders together. */
+ * 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. */
@@ -397,41 +340,35 @@ enum { _C2_TX_SUBS_ = 0
 
 /* GTE command words for the common cases.
  *
- * These are pure compile-time integer constants — the C compiler
+ * These are pure compile-time integer constants — the C compiler constant-folds them into `.word` directives in .rodata.
- * constant-folds them into `.word` directives in .rodata. Use them
+ * Use them inside `asm_inline(...)` blocks (see `gte_rtpt` below for the idiom).
- * inside `asm_inline(...)` blocks (see `gte_rtpt` below for the
- * canonical idiom).
  *
  * 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 (standard
+ * The SF / MX / V / CV / LM fields are all zero in the common cases 
- * rotation-matrix, no scaling factor, V0 vector, translation vector,
+ * (standard rotation-matrix, no scaling factor, V0 vector, translation vector, no clamp),
- * no clamp), so the only varying bits are the `cmd` field.
+ * so the only varying bits are the `cmd` field.
  *
- * Naming follows the file's convention: `gte_cmd_*` is the raw
+ * Naming convention:
- * 6-bit `cmd` field id, `gte_cmdw_*` is the fully-encoded 32-bit
+ * - `gte_cmd_*` : Raw 6-bit `cmd` field id 
- * instruction word ready to drop into a `.word` directive.
+ * - `gte_cmdw_* : 32-bit instruction word ready to drop into a `.word` directive.
  *
  * --------------------------------------------------------------------------
  *  PsyQ-compatibility note (RTPS/RTPT):
- *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and
+ *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and RTPS as `cop2 0x0180001`.
- *  RTPS as `cop2 0x0180001`. Both have `0x20` set in the upper-reserved
+ *  Both have `0x20` set in the upper-reserved region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
- *  region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
+ *  Per psx-spec these bits are reserved/must-be-zero,
- *  Per psx-spec these bits are reserved/must-be-zero, but the real GTE
+ *  but the real GTE hardware and PCSX-Redux's GTE model both IGNORE them on these two commands
- *  hardware and PCSX-Redux's GTE model both IGNORE them on these two
+ *  (the perspective divide happens regardless of `sf`).
- *  commands (the perspective divide happens regardless of `sf`).
  *
- *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits
+ *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits clear),
- *  clear), PCSX-Redux's GTE checks those bits more strictly than the
+ *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops — 
- *  silicon does and RTPT silently no-ops — the floor's screen
+ *  the floor's screen coordinates come out as raw projection-of-rotation (Z never divided),
- *  coordinates come out as raw projection-of-rotation (Z never
+ *  `nclip` ends up wrong, and the triangle is culled.
- *  divided), `nclip` ends up wrong, and the triangle is culled.
  *
- *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to
+ *  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.
- *  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.
- *  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))
@@ -440,6 +377,8 @@ 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_mvmva  (gte_cmd_base | enc_gte_cmd(gte_cmd_mvmva))
 
 #define gte_cmdw_rotate_translate_perspective_single gte_cmdw_rtps
@@ -462,20 +401,16 @@ 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
+ * Uses string-style GCC inline asm with `%0` substitution because the base register `r0` is a runtime GPR chosen by the compiler.
- * base register `r0` is a runtime GPR chosen by the compiler — it cannot
+ * It cannot be encoded into a static `.word` constant.
- * be encoded into a static `.word` constant.
  *
- * Usage:
+ * Usage: asm_gte_load_v0(svector_ptr);
- *   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
@@ -484,8 +419,7 @@ 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
+ * These are pure compile-time integers; the C compiler constant-folds them into .word directives. */
- * them into .word directives. */
 
 enum {
 	GTE_Z_Offset = 4
@@ -500,26 +434,21 @@ enum {
 
 /* gte_load_vN(r_ptr, base) — placeholder-punned lwc2 loaders
  *
- * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen
+ * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen GTE vector register, where `<base>` is the GPR number you pass in
- * 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).
  *
  * The caller MUST bind `r_ptr` to that same GPR via a register variable:
  *     register V3_S2* p_in_12 __asm__("$12") = my_ptr;
  *     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
+ * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register `p_in_12` can live in),
- * `p_in_12` can live in), which is exactly the register the .word
+ * which is exactly the register the .word constants expect.
- * constants expect. A `"$12"` clobber would conflict with the
+ * A `"$12"` clobber would conflict with the register-variable binding ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
- * register-variable binding ("asm specifier for variable conflicts
+ * 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.
- * 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
  * ----------------------
- * Any caller-saved GPR is safe. Recommended default for an RTPT-style
+ * Any caller-saved GPR is safe. Recommended default for an RTPT-style 3-pointer pipeline:
- * 3-pointer pipeline:
  *   gte_load_v0(p0, R_T4);  // $12
  *   gte_load_v1(p1, R_T5);  // $13
  *   gte_load_v2(p2, R_T6);  // $14
@@ -532,8 +461,7 @@ enum {
  *   clobbers section   : "$2", "$8", ..., "memory"      (from asm_clobber)
  *   3 colons total, GCC-legal. No string-syntax mnemonics in the .word body.
  *
- * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that
+ * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that starts the clobbers section. */
- * starts the clobbers section. */
 #define gte_load_v0(r_ptr, base) asm volatile(       \
 	asm_words( gte_lw_v0_xy(base), gte_lw_v0_z(base) ) \
 	asm_rpins,   r_use(r_ptr)                          \
@@ -552,12 +480,10 @@ enum {
 	asm_clobber: rlit(R_V0), rlit(R_T0), rlit(R_T1), rlit(R_RA), clb_mem_drain \
 )
 
-/* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — the canonical prelude to gte_cmd_rtpt.
+/* 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. Caller
- * must bind each `pN` to `bN` via a register variable.
  *
+ * 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")
@@ -576,29 +502,20 @@ enum {
 
 /**
  * @brief Rotate, Translate and Perspective Triple (23 cycles)
- *
+ * @details Performs rotation, translation and perspective calculation of three vertices at once.
- * @details Performs rotation, translation and perspective calculation of three
+ * The equation performed is the same as gte_rtps() only repeated three times for each vertex.
- * vertices at once. The equation performed is the same as gte_rtps() only
+ * The result of the first vertex is stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then C2_SXY2.
- * 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
+ *   1. Two `nop` words fill the COP2 pipeline latency — the GTE takes ~8 cycles per perspective divide, 
- *      takes ~8 cycles per perspective divide, and the nops let any
+ *      and the nops let any preceding lwc2/swc2 retire before RTPT starts reading its inputs from V0/V1/V2.
- *      preceding lwc2/swc2 retire before RTPT starts reading its
+ *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the pre-baked encoders above) — 
- *      inputs from V0/V1/V2.
+ *      `0x0280030` decoded as `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields zero
- *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the
+ *      (standard rotation, no scaling, V0 vector, translation vector, no clamp).
- *      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
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI)
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
+ * plus the standard "memory" barrier. Does not clobber any COP2 data/control register — 
- * data/control register — those have to be saved by the caller if
+ * 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.).
- * they need to survive across the call (RTPT writes SXY0..2, SZ0..3,
- * OTZ, MAC0..3, IR0..3, etc.).
  */
 #define gte_rtpt()                        \
 	asm volatile(                           \
@@ -614,32 +531,24 @@ enum {
 
 /**
  * @brief Normal clipping (8 cycles)
- *
+ * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for backface culling.
- * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for
+ * If the value of C2_MAC0 is negative, the coordinates are inverted and thus the triangle is back facing.
- * 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
+ *   1. Two `nop` words fill the COP2 pipeline latency
- *      pipeline takes a few cycles per op, and the nops let any
+ *      - the GTE pipeline takes a few cycles per op, and the nops let any preceding
- *      preceding lwc2/swc2/RTPT retire before NCLIP starts reading
+ *      lwc2/swc2/RTPT retire before NCLIP starts reading its inputs from SXY0/SXY1/SXY2.
- *      its inputs from SXY0/SXY1/SXY2.
+ *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the pre-baked encoders above)
- *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the
+ *      - `0x01400006` decoded as  `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields zero.
- *      pre-baked encoders above) - `0x01400006` decoded as
+ *      NCLIP is spec-clean in the original PsyQ source (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat` quirk),
- *      `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields
+ *      so `gte_cmdw_nclip` does NOT OR in any reserved bits.
- *      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
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI) plus the standard "memory" barrier.
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
+ * Does not clobber any COP2 data/control register.
- * data/control register - those have to be saved by the caller if
+ * Those have to be saved by the caller if they need to survive across the call (NCLIP writes MAC0 only;
- * they need to survive across the call (NCLIP writes MAC0 only; it
+ * it is purely a sign-of-double-product computation on SXY0..2).
- * is purely a sign-of-double-product computation on SXY0..2).
  */
 #define gte_nclip()                       \
 	asm volatile(                           \
@@ -665,13 +574,10 @@ 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
+ * Memory layout at r0: five contiguous 32-bit words (offsets 0..16), each holding two packed 16-bit matrix elements.
- * control registers (RT11..RT22, indices 0..4) via ctc2.
+ * The first 1.5 rows of a standard PSX SDK MATRIX struct (where each row is laid out as
- *
- * 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):
@@ -686,41 +592,36 @@ 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
+ * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
- * 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.
- * instructions to read from the right base. The `"r"(r0)` constraint
+ * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field of each lw, so the lw instructions will 
- * alone doesn't force a specific GPR — it just lets GCC pick one.
+ * only do the right thing if $12 / $13 / $14 hold the matrix base at runtime.
- * 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
+ * We clobber $12/$13/$14 (the ones we use as scratch inside the inline asm)
- * inline asm) plus the system clobbers; we don't clobber `r0` because
+ * plus the system clobbers; we don't clobber `r0` because the `rgcc` binding already says "this variable lives in $12".
- * the `rgcc` binding already says "this variable lives in $12".
  *
- * WARNING: Incomplete by design. The source macro only writes RT11..RT22
+ * WARNING: Incomplete by design. The source macro only writes RT11..RT22 (5 of 9 rotation elements);
- * (5 of 9 rotation elements); RT23 and the entire RT3x row are left
+ * RT23 and the entire RT3x row are left untouched.
- * untouched. Real libpsn00b SetRotMatrix writes all 9. Use only when the
+ * Real libpsn00b SetRotMatrix writes all 9. Use only when the GTE's remaining rotation entries are already correct,
- * GTE's remaining rotation entries are already correct, or you will
+ * or you will get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
- * get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
  */
 #define asm_gte_matrix_set_rotation(r0) \
 	asm volatile(                         \
 		asm_words(                          \
 			  load_word(R_T5, R_T4,  0)       \
 			, load_word(R_T6, R_T4,  4)       \
-			, gte_mt(   R_T5,        0)       \
+			, gte_mv_to_data_r(   R_T5,  0)   \
-			, gte_mt(   R_T6,        1)       \
+			, gte_mv_to_data_r(   R_T6,  1)   \
 			, load_word(R_T5, R_T4,  8)       \
 			, load_word(R_T6, R_T4, 12)       \
 			, load_word(R_T4, R_T4, 16)       \
-			, gte_mt(   R_T5,  2)             \
+			, gte_mv_to_data_r(   R_T5,  2)   \
-			, gte_mt(   R_T6,  3)             \
+			, gte_mv_to_data_r(   R_T6,  3)   \
-			, gte_mt(   R_T4,  4)             \
+			, gte_mv_to_data_r(   R_T4,  4)   \
 		)                                   \
 		, r_use(r0)                         \
 		asm_clobber: clbr_volatile_gprs, rlit(R_T4), rlit(R_T5), rlit(R_T6) \

code/duffle/gte_vendor_sym.h

@@ -2,10 +2,8 @@
  *  duffle DSL — GTE Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2
+ *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2 instructions as thin aliases to the duffle macros in gte.h.
- *  instructions as thin aliases to the canonical duffle macros in gte.h.
+ *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
- *  The duffle names are primary; this header is for users who prefer
- *  the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/gte_vendor_sym.h"   // after gte.h
  *
@@ -21,12 +19,6 @@
  *      gte_swc2(rt, base, off)  -> gte_sw(rt, base, off)
  *      (the lower-level vector variants gte_lw_v0_xy etc. don't have
  *      vendor mnemonics; they're already gte_-prefixed and short)
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (tape_atom.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries. Verification: V3 (objdump
- *  byte-identical) holds.
- *
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/lottes_tape.h

@@ -1,82 +1,210 @@
 #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 "dsl.atom.h"
 #endif
 
-typedef U4 const MipsCode;
-#define MipsAtom_(sym) MipsCode tmpl(code,sym) [] align_(4) =
-
 #pragma region Tape Drive
-/* ---------------------------------------------------------------------------
+/* -----------------------------------------------------------------------------
- *  TAPE DRIVE ABI & REGISTER ALIASES
+ *  TAPE DRIVE ABI
- * ---------------------------------------------------------------------------
+ * -----------------------------------------------------------------------------
- *  We map the MIPS temporary registers to a persistent global workspace.
+ * Note(Ed): One of the main purposes of this codebase is to help me
- *  The C compiler is completely unaware of these bindings.
+ * learn this, as such the information below may 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 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".
+ * 
+ * This 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. Stack usage is non-existent.
+ * Push/Pop, FIFO, or Arena/Bump data structures are used by atoms explicitly.
+ * In it's current form withe C11 macro dsl, the user also has to do 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 can't ignore
+ * excessive argument shuffle across workload or waves and thier phases.
+ * Crossing ABI boundaries to other runtimes that do has obviouss penalties.
+ * 
+ * Learning data-oreinted code becomes a natural progression. Your not fighting
+ * a stack-based procedural paradigm that wants to argument shuffle on by lack of 
+ * constraints on how the user may "call" a procedure. The user doesn't
+ * have to hammer down "rules" or patterns to know how to massage the compiler
+ * to dissolve those call frames 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.
+ * 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 is 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_AtomJmp    = R_T9,
+	R_AtomJmp  = R_T8 atom_reg,  /* debug-visible; tape yield handshake scratch */
-	R_TapePtr    = R_T8,  /* The Instruction Stream Pointer */
+	R_TapePtr  = R_T9 atom_reg,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,  /* Input data cursor */
+/* Stringification codes for the GCC inline assembler clobber lists. */
+#define R_AtomJmp_Code R_T8_Code
+#define R_TapePtr_Code R_T9_Code
 
-	R_PrimCursor = R_T7,  /* VRAM output cursor (primitive buffer) */
+// R_InCursor = R_T4,
-	R_FaceCursor = R_T4,  /* Input data cursor (indices/faces) */
+// #define R_InCursor_Code    R_T4_Code
-	R_VertBase   = R_T5,  /* Base address of the vertex array */
-	R_OtBase     = R_T6,  /* Base address of the Ordering Table */
 
-/* Stringification codes for the GCC inline assembler clobber lists */
+// Reserved Registers (Callee-saved):
-#define R_TapePtr_Code     R_T8_Code
+// - R_T9: Holds the Tape Ptr which we need to increment
-#define R_InCursor_Code    R_T4_Code
+//         If we hit a wall with register allocations we can clobber V0 & V1 (return values), defering as opt-in by user.
+// - R_RA: Not sure??
+// Needed by ac_yield but can be used as atom scratch:
+// - R_T8: Will be used as the atom jump register.
 
-#define R_PrimCursor_Code  R_T7_Code
+// All allocatable registers for mips atoms:
-#define R_FaceCursor_Code  R_T4_Code
+	R_TScratchVolatile = R_AT, // This one is reserved for psuedo instructions, but you can technically use it.
-#define R_VertBase_Code    R_T5_Code
+	R_TScratch0  = R_T0,
-#define R_OtBase_Code      R_T6_Code
+	R_TScratch1  = R_T1,
+	R_TScratch2  = R_T2,
+	R_TScratch3  = R_T3,
+	R_TScratch4  = R_T4,
+	R_TScratch5  = R_T5,
+	R_TScratch6  = R_T6,
+	R_TScratch7  = R_T7,
+	R_TScratch8 =  R_T8,
+	R_TScratch10 = R_V0, // Tend to be used with gte DMAs
+	R_TScratch11 = R_V1, // Tend to be used with gte DMAs
+	// Note(Ed): We can technically clobber these, but don't unless we hit a bottleneck.
+	// A 0-2
+	// S 0-7
 };
 
-/* The 'Exit' Atom */
+typedef U4 const MipsCode; // Underlying type to mips asm words.
-MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+typedef Slice_(MipsCode);
 
-/* Generalized Tape Engine Runner */
+typedef U4 const MipsAtom; // Underlying type to an array of mips asm words that must terminate with an ac_yield.
-FI_ void tape_run(Slice_U4 tape) { register U4* tp rgcc(R_TapePtr) = tape.ptr; asm volatile(
+#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+
+// 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_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+// expands to:
+//   FI_ Slice_MipsCode 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); }
+
+/* Line-table anchor: gcc only adds a file to the .debug_line file table when the
+   file contains line-numbered content. Files containing only:
+     - `MipsAtomComp_` static-array declarations, or
+     - `MipsAtomComp_Proc_` (force-inline) function bodies whose line info gets
+       attributed to the call site at the include point are otherwise omitted from the file table,
+			 which breaks the DWARF injection when it tries to resolve atom-component provenance paths.
+
+   Place `ATOM_FILE_LINE_MARKER();` once at file scope in any `.atom.c` that defines atoms.
+	 The macro expands to a file-scope `internal U4 const` declaration keeps the file in the line table.
+	 The constant is in `.rodata` and unreferenced; the linker may eliminate it.
+	 The two-level concat + `__LINE__` suffix makes the identifier unique per call site
+	 (the identifier embeds the source line, so duplicates across `#include`d files don't collide). */
+#define ATOM_FILE_DEBUGGER_LINE_MARKER(file_name) internal U4 const tmpl(atom_file_debugger_line_marker,file_name) = 0
+
+ typedef Slice_(MipsAtom); typedef Slice_MipsAtom Tape;
+
+/* The 'Exit' Atom */
+atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+
+// TODO(Ed): When we have a substantial workload/throughput, profile each of these to see impact at ABI boundaries.
+
+/* Tape Runner (Default) */
+FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
-			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
+		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
-		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
+		, add_ui_self(R_TapePtr, S_(MipsAtom)) /* Advance tape */
-		, load_word(  R_AtomJmp, R_TapePtr, 0)         /* Bootstrap the first jump */
+		, call_reg(   R_AtomJmp)               /* jalr $t9 */
-		, add_ui_self(R_TapePtr, S_(MipsCode))         /* Advance tape */
+		, nop                                  /* Branch delay slot */
-		, 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_rpins, r_use(tape_ptr)
 	asm_clobber: 
-			rlit(R_AT)
+		rlit(R_AT),
-		, rlit(R_V0), rlit(R_V1)
+		rlit(R_V0), rlit(R_V1), // We clobber these for GTE ACs (that don't expose register selection, might expose them in the future...)
-		, rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3)
+		rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
-		/* Tell GCC the tape engine owns and destroys the workspace registers */
+		rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8),
-		, rlit(R_PrimCursor), rlit(R_FaceCursor), rlit(R_VertBase), rlit(R_OtBase)
+		clb_mem_drain 
-		, rlit(R_T9)
-		, clb_mem_drain 
 ); }
 
+/* Tape Runner (Static and Arg Clobbers) */
+FI_ void tape_run_a02_s07(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
+	asm_words(
+		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
+		, add_ui_self(R_TapePtr, S_(MipsAtom)) /* Advance tape */
+		, call_reg(   R_AtomJmp)               /* jalr $t9 */
+		, nop                                  /* Branch delay slot */
+	)
+	asm_rpins, r_use(tape_ptr)
+	asm_clobber: 
+		rlit(R_AT),
+		rlit(R_V0), rlit(R_V1), rlit(R_A0), rlit(R_A1), rlit(R_A2),
+		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_S0), rlit(R_S1), rlit(R_S2), rlit(R_S3), rlit(R_S4),
+		rlit(R_S5), rlit(R_S6), rlit(R_S7),
+		clb_mem_drain 
+); }
+
+// 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 }; }
 
-#define tb_emit_(tb, atom) tb_emit(tb, tmpl(code,atom))
 FI_ void tb_emit(TapeBuilder* tb, MipsCode* 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_U4 tb_end  (TapeBuilder* tb) { tb_emit(tb,code_tape_exit); return (Slice_U4){ C_(U4*,tb->ptr), tb->used }; }
+FI_ Tape tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Tape){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Slice_U4){ 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,code_tape_exit))
+#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
@@ -85,108 +213,34 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
  *  These do NOT yield. They are expanded inline inside Tape Atoms.
  * ---------------------------------------------------------------------------*/
 
-/* The 'Yield' sequence for Tape Atoms.
+// The 'Yield' sequence for Tape Atoms (mac_yield).
- * Loads the next pointer from the tape, advances the tape, and jumps. 
+//   - mac_yield() is the safe default for atom-endings: 4 words, BD-slot of jr is mandatory nop.
- * Cost: ~ 4 cycles */
+//   - mac_yield_load() + mac_yield_tail():
-#define mac_yield()                               \
+//     - unconditional branch: mac_yield_load fills the branch's BD-slot (replaces a nop);
-	  load_word(R_AtomJmp, R_TapePtr, 0)            \
+//     - mac_yield_tail runs at the branch target (does NOT re-load R_AtomJmp).
-	, add_ui_self(         R_TapePtr, S_(MipsCode)) \
-	, jump_reg( R_AtomJmp)                          \
-	, nop
 
-/* Words: 3; Loads 3 S2 indices from the face array */
+atom_dbg_skip MipsAtomComp_(ac_yield) {
-#define mac_load_tri_indices(rId_0, rId_1, rId_2) \
+	load_word(R_AtomJmp, R_TapePtr, 0),
-	  load_half_u(rId_0, R_FaceCursor, 0 * S_(S2)) \
+	add_ui_self(         R_TapePtr, S_(MipsCode)),
-	, load_half_u(rId_1, R_FaceCursor, 1 * S_(S2)) \
+	jump_reg( R_AtomJmp), nop,
-	, load_half_u(rId_2, R_FaceCursor, 2 * S_(S2))
+};
 
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
+atom_dbg_skip MipsAtomComp_(ac_yield_load) {
-#define mac_load_tri_verts(rId_0, rId_1, rId_2) \
+	load_word(R_AtomJmp, R_TapePtr, 0),
-	  shift_lleft(R_AT, rId_0, 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, rId_1, 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, rId_2, 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)
 
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list */
+atom_dbg_skip MipsAtomComp_(ac_yield_tail) {
-#define mac_insert_ot_tag(r_otz, prim_type) \
+	add_ui_self(R_TapePtr, S_(MipsCode)),
-	  shift_lleft( R_T1, r_otz, S_(U4)/2)                       /* T1 = r_otz * S_(U4) */ \
+	jump_reg(  R_AtomJmp), nop,
-	, add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
+};
-	, load_word(   R_AT, R_T1,         O_(PolyTag,bf_addr_len)) /* AT = old_ot_head */ \
-	, load_upper_i(R_V0, (S_(prim_type)/S_(U4) - S_(PolyTag)/S_(U4)) << polytag_len_bits) /* V0 = S_(prim_type without tag field) */ \
-	, 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,bf_addr_len)) /* 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,bf_addr_len)) /* 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.
- * Args: off = U4 byte offset, code = GP0 cmd byte (0 for c1/c2/c3 of
- * a Poly_G4), r/g/b = 8-bit RGB byte values. */
-#define mac_pack_color_word(off, code, r,g,b) \
-		load_upper_i(R_AT, (code) << 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).
- * Migrated from hello_gte_tape.c; takes RGB form per the Phase 3
- * convention. */
-#define mac_format_f3_color(r,g,b) 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).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-#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))
-
-/* 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. */
-#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)
-
-/* 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))
-
-/* 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).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration.
- *
- * History: this macro was named `mac_gte_store_g4_p3` until 2026-07-09
- * when it was discovered to be reading C2_SXY0 (which held v0.screen)
- * instead of C2_SXY2 (which holds v3.screen after RTPS). The rename
- * encodes the pipeline position in the name so the next bug of this
- * class is impossible. */
-#define mac_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.
+// This helps with runtime procedural authoring of mips atoms.
 
 typedef Struct_(FMipsAtom512) { U4 data[512]; U4 used; };
 
-typedef Slice_(MipsCode); typedef Slice_MipsCode MipsAtom; 
 // FArena Related
 typedef Relative_(FArena) Struct_(MipsAtomBuilder) { U4 start; U4 capacity; U4 used; };
 // Whatever the builder is writting to should most likely coresspond
@@ -201,128 +255,14 @@ FI_ void atombuilder_unroll(MipsAtomBuilder_R ab, Slice_MipsCode_R code) {
 
 // When done authoring, utilize this to cap-off the atom
 FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
-	LP_ MipsAtom_(yield) { mac_yield() };
+	mem_copy(ab->start, u4_(ac_yield), S_(ac_yield));
-	mem_copy(ab->start, u4_(code_yield), S_(code_yield));
+	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(code_yield));
 }
 
-#define mipsatom_from_builder(ab) (MipsAtom){ab.start, ab.used}
+#define mipsatom_from_builder(ab) (Slice_MipsCode){ab.start, ab.used}
-
 #pragma endregion Mips Atom Builder
 
 #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)                        /* jalr  $t0, $ra      */
-	, nop                                                 /* 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) {
-	U4 transform;
-};
-internal MipsAtom_(set_gte_world) {
-	/* 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()
-};
-
-// TODO(Ed): I'm not sure yet if the bindings are redundant with the floortri atom yet.
-
-/* DIAGNOSTIC 1: Pure tape loop test */
-internal MipsAtom_(diag_yield) { mac_yield() };
-
-// TODO(Ed): Reduce magic numbers/offsets
-/* 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()
-};
-
-// TODO(Ed): Reduce magic numbers/offsets
-/* 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.c

@@ -0,0 +1,29 @@
+#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);
+
+#pragma region MACs (Mips Atom Component)
+
+FI_ Slice_MipsCode ac_load_v2s2(U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_load_v2s2, {
+	load_half( rs_x, r_base, O_(V3_S2,x)),
+	load_half( rs_y, r_base, O_(V3_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_v2s2(U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_v2s2, {
+	store_half(rt_x, base, offset + O_(V2_S2,x)),
+	store_half(rt_y, base, offset + O_(V2_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_rects2(U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rects2, {
+	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

@@ -11,6 +11,7 @@ enum {
 	v3s2_byteoff = 3, // log2(8), used with shift_left_logical op for index via byte offset.
 };
 
+typedef Array_(U1, 2);
 typedef Array_(U4, 2);
 typedef Array_(S2, 2);
 typedef Array_(S2, 3);
@@ -22,6 +23,7 @@ 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; };
@@ -32,11 +34,12 @@ 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_(Rect_S2) { S2 x; S2 y; S2 width; S2 height; };
+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_(Rect_S4)    { S4 x; S4 y; S4 width; S4 height; };
 
-typedef Struct_(M3_S2) { A3x3_S2 m; A3_S4 t; };
+typedef Struct_(M3_S2)      { A3x3_S2 m; A3_S4 t; };
 
+typedef Array_(V2_S2, 2);
 typedef Array_(V2_S2, 3);
 typedef Array_(V2_S2, 4);
 
@@ -58,10 +61,5 @@ 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) { 
+FI_ void add_v3s4   (V3_S4_R out_a, V3_S4 b) {  add_a3s4   (pcast(A3_S4_R, out_a), pcast(A3_S4, b));  }
-	add_a3s4(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+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 add_v3s4_fp(V3_S4_R out_a, V3_S4 b) { 
-	add_a3s4_fp(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
-}

code/duffle/memory.h

@@ -67,12 +67,13 @@ typedef Slice_(B1);
 #define slice_end(slice)   ((slice).ptr + (slice).len)
 #define S_slice(s)         ((s).len * S_((s).ptr[0]))
 
-#define slice_ut(ptr,len)  slice_ut_(u4_(ptr), u4_(len))
+#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_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) }
 
 FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(s.ptr, s.len); }
 #define  slice_zero(s)        slice_zero_(slice_to_ut(s))
@@ -102,7 +103,7 @@ FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
 	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) {
+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);

code/duffle/mips.atom.c

@@ -0,0 +1,38 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(mips_atom_c);
+
+#pragma region Baked Atoms
+
+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(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/mips.h

@@ -1,38 +1,28 @@
 /* ============================================================================
  *  duffle DSL Suffix Conventions
  *  ============================================================================
- *
  *  Every mnemonic in this header follows the same suffix grammar:
- *
+ *    _i: Immediate value (16-bit constant operand). 
- *    _i        Immediate value (16-bit constant operand). Combine with
+ *        Combine with _u or _s (single-letter modifier + type combined): add_ui, add_si.
- *              _u or _s (single-letter modifier + type combined): add_ui,
+ *        Examples: add_ui, add_si, and_i, or_i, xor_i, load_upper_i. and_i is sign-agnostic (andi zero-extends).
- *              add_si. Examples: add_ui, add_si, and_i, or_i, xor_i,
+ *        load_upper_i is a unique verb; _i is the immediate marker, not a modifier+type combination.
- *              load_upper_i. and_i is sign-agnostic (andi zero-extends).
+ *    _u: Unsigned (no-overflow, no-sign-extension).
- *              load_upper_i is a unique verb; _i is the immediate marker,
+ *        R-type arithmetic examples: add_u, sub_u, mult_u, div_u. I-type (combined with _i): add_ui.
- *              not a modifier+type combination.
+ *    _s: Signed (overflow-traps, sign-extends). 
- *
+ *        R-type: add_s, sub_s, mult_s, div_s, set_lt_s. I-type (combined with _i): add_si.
- *    _u        Unsigned (no-overflow, no-sign-extension). R-type
- *              arithmetic examples: add_u, sub_u, mult_u, div_u. I-type
- *              (combined with _i): add_ui.
- *
- *    _s        Signed (overflow-traps, sign-extends). R-type: add_s,
- *              sub_s, mult_s, div_s, set_lt_s. I-type (combined with _i):
- *              add_si.
  *
  *  --- Shift family (R-type): verb-modifier-direction ---
- *    The shift macros use `shift_<modifier><direction>`. Modifier is
+ *    The shift macros use `shift_<modifier><direction>`.
- *    the single letter `l` (logical) or `a` (arithmetic). Direction
+ *    Modifier is the single letter `l` (logical) or `a` (arithmetic).
- *    is the word `left` or `right`. Combined: `_lleft`, `_lright`,
+ *    Direction is the word `left` or `right`. Combined: `_lleft`, `_lright`, `_aright`.
- *    `_aright`. Examples: shift_lleft( rd, rt, shamt) (= sll)
+ *    Examples: shift_lleft( rd, rt, shamt) (= sll)
- *                         shift_lright(rd, rt, shamt) (= srl)
+ *              shift_lright(rd, rt, shamt) (= srl)
- *                         shift_aright(rd, rt, shamt) (= sra)
+ *              shift_aright(rd, rt, shamt) (= sra)
- *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical
+ *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical to logical-left, so use shift_lleft for that case)
- *    to logical-left, so use shift_lleft for that case)
  *
  *  --- Jump/Call family ---
- *    Simple jumps keep the original short names: jump (j), jump_reg
+ *    Simple jumps keep the original short names: jump (j), jump_reg (jr), jump_link (jalr rs, rd).
- *    (jr), jump_link (jalr rs, rd). The jump-and-link-to variants
+ *    The jump-and-link-to variants (jal, jalr rs with default $ra) get the `call_` verb instead:
- *    (jal, jalr rs with default $ra) get the `call_` verb instead:
  *    call_addr (jal), call_reg (jalr rs, default $ra).
  *    Examples: jump(off)         (= j)
  *              jump_reg(rs)      (= jr)
@@ -40,32 +30,22 @@
  *              call_reg(rs)      (= jalr rs, default $ra)
  *              call_addr(off)    (= jal)
  *
- *    _r        Register marker — used only when the register type needs
+ *    _r: Register marker — used only when the register type needs disambiguation (e.g., GTE data register vs control register).
- *              disambiguation (e.g., GTE data register vs control
+ *        NOT used in plain R-type arithmetic (the R-type is implicit). Examples: gte_mv_to_data_r, gte_mv_to_ctrl_r.
- *              register). NOT used in plain R-type arithmetic (the
+ *    _self:     Destination equals one source operand.
- *              R-type is implicit). Examples: gte_mv_to_data_r,
+ *               Examples: add_ui_self (I-type, to self), add_u_self (R-type, to self).
- *              gte_mv_to_ctrl_r.
+ *    _mv_to_:   Direction: data flows into X.
+ *               Example: gte_mv_to_data_r, gte_mv_to_ctrl_r.
+ *    _mv_from_: Direction: data flows out of X.
+ *               Example: gte_mv_from_data_r, gte_mv_from_ctrl_r.
+ *    _str:      String-form — emits inline-asm string instead of `.word`.
+ *               Example: gte_rtpt_asm_str.
+ *    _2w / _1w: Word count of the emitted sequence.
+ *               Example: load_imm_2w.
  *
- *    _self     Destination equals one source operand.
+ *    _cop2: RESERVED — DO NOT USE in macro names. The `gte_` namespace prefix already implies coprocessor 2. Use `c2` only in:
- *              Examples: add_ui_self (I-type, to self),
+ *      (a) integer opcode enums (op_lwc2 = 0x32, op_swc2 = 0x3A)
- *                        add_u_self (R-type, to self).
+ *      (b) vendor-mnemonic macro aliases (gte_mtc2, gte_mfc2)
- *
- *    _mv_to_   Direction: data flows into X.
- *              Example: gte_mv_to_data_r, gte_mv_to_ctrl_r.
- *
- *    _mv_from_ Direction: data flows out of X.
- *              Example: gte_mv_from_data_r, gte_mv_from_ctrl_r.
- *
- *    _str      String-form — emits inline-asm string instead of `.word`.
- *              Example: gte_rtpt_asm_str.
- *
- *    _2w / _1w Word count of the emitted sequence.
- *              Example: load_imm_2w.
- *
- *    _cop2     RESERVED — DO NOT USE in macro names. The `gte_` namespace
- *              prefix already implies coprocessor 2. Use `c2` only in:
- *                (a) integer opcode enums (op_lwc2 = 0x32, op_swc2 = 0x3A)
- *                (b) vendor-mnemonic macro aliases (gte_mtc2, gte_mfc2)
  *
  *  Primitive commands: gp0_cmd_poly_f3 = 0x20 (byte opcode)
  *  Packed 32-bit cmd:  gp0_word_poly_f3(r, g, b) (32-bit, shifted)
@@ -80,9 +60,8 @@
  *            gte_lw_v0_xy(base) (gte + lw + v0 + xy)
  *            load_upper_i       (load-upper + immediate, unique verb)
  *
- *  Vendor mnemonics (sll, srl, sra, jr, j, jal, jalr) are NOT in this
+ *  Vendor mnemonics (sll, srl, sra, jr, j, jal, jalr) are NOT in this header.
- *  header. They live in the opt-in `mips_vendor_sym.h` for users who
+ *  They live in the opt-in `mips_vendor_sym.h` for users who prefer the textbook MIPS assembly mnemonics.
- *  prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -98,19 +77,17 @@ enum {
 /* ============================================================================
  *  REGISTER INTEGER IDS (preprocessor-visible)
  * ============================================================================
- *  Every R_* enum below has a parallel R_*_Code `#define` so that the
+ *  Every R_* enum below has a parallel R_*_Code `#define` so that the preprocessor can stringify the integer
- *  preprocessor can stringify the integer (e.g. for asm clobber lists and
+ *  (e.g. for asm clobber lists and register-variable declarations via `rgcc(R_X)`).
- *  register-variable declarations via `rgcc(R_X)`). The enum value is
+ *  The enum value is bound to the `#define` so the two forms cannot drift apart.
- *  bound to the `#define` so the two forms cannot drift apart.
  *
- *  Only registers that get stringified need a `_Code` form; the rest are
+ *  Only registers that get stringified need a `_Code` form; the rest are plain enum values.
- *  plain enum values. If you need to add a new one, follow the pattern:
+ *  If you need to add a new one, follow the pattern:
  *      #define R_T7_Code 15
- *      R_T7 = R_T7_Code,        // in the enum
+ *      R_T7 = R_T7_Code,     // in the enum
  *
- *  User code should always reference the enum form (`R_T4`) at arithmetic
+ *  User code should always reference the enum form (`R_T4`) at arithmetic sites and let
- *  sites and let `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify
+ *  `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify cases — never write the bare number `12`.
- *  cases — never write the bare number `12`.
  * ============================================================================ */
 #define R_0_Code    0
 #define R_AT_Code   1
@@ -225,7 +202,6 @@ enum {
 	/* 2F: N/A */
 	// , op_lwc0
 
-
 	// , op_load_addr  = op_la
 	// , op_load_imm   = op_li
 	, op_jump       = op_j
@@ -327,15 +303,15 @@ enum { _BitOffsets = 0
  * Argument order matches the MIPS assembly syntax:
  *   dest-first, then source operands, then immediate last.
  *
- *   load_word(rt, base, off)   → lw   rt, off(base)
+ *   load_word(rt, base, off)    → lw   rt, off(base)
- *   store_word(rt, base, off)  → sw   rt, off(base)
+ *   store_word(rt, base, off)   → sw   rt, off(base)
- *   add_ui(rt, rs, imm)        → addiu rt, rs, imm
+ *   add_ui(rt, rs, imm)         → addiu rt, rs, imm
- *   shift_lleft(rd, rt, shamt)   → sll   rd, rt, shamt
+ *   shift_lleft(rd, rt, shamt)  → sll   rd, rt, shamt
- *   shift_lright(rd, rt, shamt)  → srl   rd, rt, shamt
+ *   shift_lright(rd, rt, shamt) → srl   rd, rt, shamt
- *   shift_aright(rd, rt, shamt)  → sra   rd, rt, shamt
+ *   shift_aright(rd, rt, shamt) → sra   rd, rt, shamt
- *   jump_reg(rs)               → jr    rs
+ *   jump_reg(rs)                → jr    rs
- *   jump_link(rs, rd)          → jalr  rs        (link in rd, default $ra)
+ *   jump_link(rs, rd)           → jalr  rs        (link in rd, default $ra)
- *   nop                        → sll   $0, $0, 0
+ *   nop                         → sll   $0, $0, 0
  */
 #define load_word(rt, base, off)   enc_i(op_lw,    (base), (rt), (off))
 #define load_byte(rt, base, off)   enc_i(op_lb,    (base), (rt), (off))
@@ -360,10 +336,10 @@ enum { _BitOffsets = 0
 
 /* Logic Opcodes */
 
-#define and_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_and)
+#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 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 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 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)
 
@@ -386,10 +362,28 @@ 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.
+ */
 #define jump(off)                  enc_i(op_j,     R_0, R_0, (off))
 
-/* call_addr off — jump-and-link to immediate address. */
+/* 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) --- */
@@ -404,12 +398,9 @@ enum { _BitOffsets = 0
  * 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
+ * NOTE: dsl.h defines `add_s`/`sub_s`/`mut_s`/`gt_s`/etc. as _Generic-based signed integer-arithmetic helpers for U1/U2/U4.
- * _Generic-based signed integer-arithmetic helpers for U1/U2/U4. Those
+ * Those live in a different conceptual layer (generic arithmetic on DSL types) and would collide with the instruction encoders here.
- * live in a different conceptual layer (generic arithmetic on DSL
+ * 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
- * 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
@@ -441,7 +432,7 @@ enum { _BitOffsets = 0
 #define mov_to_low(rs)             enc_r(op_special, (rs), R_0, R_0, 0, fc_mtlo)
 
 /* --- Atomic branches (no pseudos like bgt/bge; compose with slt_* + branch_ne) ---
- * branch_equal  rs, rt, off → beq   rs, rt, off
+ * branch_equal   rs, rt, off → beq   rs, rt, off
  * branch_ne      rs, rt, off → bne   rs, rt, off
  * branch_lt_zero rs, off     → bltz  rs, off
  * branch_gt_zero rs, off     → bgtz  rs, off
@@ -463,7 +454,7 @@ enum { _BitOffsets = 0
 /* --- Shift-amount alias (matches the gas convention `\p3 = shamt`) --- */
 #define shift_amount(rd, rt, n)     shift_lleft(rd, rt, n)
 
-/* nop — canonical sll $0, $0, 0 */
+/* nop — sll $0, $0, 0 */
 #define nop  shift_lleft(rdiscard, rdiscard, 0)
 #define nop2 nop, nop
 
@@ -471,23 +462,16 @@ enum { _BitOffsets = 0
 #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:
- * 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
+ *   - `imm` is not a compile-time constant and you want predictable 2-word emission without the `__builtin_constant_p` branches.
- *     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
+ *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui already cleared bits 15..0)
- *                              already cleared bits 15..0)
+ *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the intended bit pattern)
- *   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
+ * For situations where you need to bypass even this choice (e.g. to force a specific encoding for a known discontiguous high/low pair),
- * 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(...)`.
  */
 #define load_imm_2w(rt, imm) do {                         \
@@ -518,9 +502,8 @@ enum { _BitOffsets = 0
 } while (0)
 
 /* load_imm_2w_addi_forced — force the `lui` + `addi` form regardless of lo16 sign.
- * Use when you know sign-extension is fine (e.g. lo16 is treated as
+ * Use when you know sign-extension is fine (e.g. lo16 is treated as signed downstream) 
- * signed downstream) and you want a smaller effective instruction
+ * and you want a smaller effective instruction (the assembler/MIPS hardware will sign-extend the imm16). */
- * (the assembler/MIPS hardware will sign-extend the imm16). */
 #define load_imm_2w_addi_forced(rt, imm) do {   \
 	/*U4 _li2a_imm_ = (U4)(imm);*/                \
 	asm volatile(asm_words(                       \
@@ -532,23 +515,17 @@ enum { _BitOffsets = 0
 
 /* load_imm rt, imm — true `li` semantics (assembler `li` pseudo)
  *
- * Dispatches at compile time on the immediate's range, picking the
+ * Dispatches at compile time on the immediate's range, picking the smallest single-instruction form when possible:
- * smallest single-instruction form when possible:
+ *   imm in 0       .. 0x7FFF     →  addi  rt, $0, imm   (1 word)
- *
+ *   imm in 0x8000  .. 0xFFFF     →  ori   rt, $0, imm   (1 word; sign-bit must be zeroed)
- *   imm in   0 ..  0x7FFF  →  addi  rt, $0, imm   (1 word)
+ *   imm in 0x10000 .. 0xFFFFFFFF →  lui + (ori | addi)  (2 words)
- *   imm in 0x8000 .. 0xFFFF →  ori   rt, $0, imm   (1 word; sign-bit must be zeroed)
- *   imm in  0x10000 ..  0xFFFFFFFF  →  lui + (ori | addi)  (2 words)
- *
- * Statement-level (not expression-level): the macro emits its own
- * `asm volatile(...)` block with 1 or 2 .word constants. Callers can
- * group multiple `load_imm` calls in a single volatile by using the
- * lower-level encoders directly:
  *
+ * Statement-level (not expression-level): the macro emits its own `asm volatile(...)` block with 1 or 2 .word constants.
+ * Callers can group multiple `load_imm` calls in a single volatile by using the lower-level encoders directly:
  *      load_imm(R_T4, 0x12345678);          // emits 2 .words
  *
- * Falls back to a 2-word form if `imm` is not a compile-time constant,
+ * Falls back to a 2-word form if `imm` is not a compile-time constant, but that path is unusual
- * but that path is unusual (load_imm is most useful with literal
+ * (load_imm is most useful with literal addresses and magic numbers). */
- * addresses and magic numbers). */
 #define load_imm(rt, imm) do {                                     \
 	if (cexpr_(imm) && ((imm) <= 0x7FFFU)) {                         \
 		/* Small positive: addi rt, $0, imm */                         \
@@ -588,9 +565,8 @@ enum { _BitOffsets = 0
 
 
 /* Standard clobber list for pure-MIPS asm volatile blocks: caller-saved
- * GPRs that the kernel treats as volatile (v0/v1/t0/t1/ra) plus the
+ * GPRs that the kernel treats as volatile (v0/v1/t0/t1/ra) plus the "memory" barrier.
- * "memory" barrier. The register ids are passed through `rlit` so
+ * The register ids are passed through `rlit` so the R_*_Code `#define`s are stringified into "$N" at expansion time. */
- * the R_*_Code `#define`s are stringified into "$N" at expansion time. */
 #define clbr_volatile_gprs rlit(R_V0), rlit(R_T0), rlit(R_T1), rlit(R_RA), clb_mem_drain
 
 #define asm_mips_flush_icache() asm volatile( asm_words( \

code/duffle/mips_vendor_sym.h

@@ -2,9 +2,8 @@
  *  duffle DSL — MIPS Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics as thin aliases to the
+ *  Provides the textbook MIPS assembly mnemonics as thin aliases to the duffle macros in mips.h.
- *  canonical duffle macros in mips.h. The duffle names are primary; this
+ *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
- *  header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/mips_vendor_sym.h"   // after mips.h
  *
@@ -21,12 +20,6 @@
  *      jal  -> call_addr         (jump-and-link to immediate address)
  *      jalr -> call_reg          (jump-and-link to register, default $ra)
  *      (for the 2-arg `jalr rs, rd`, use `jump_link(rs, rd)` directly)
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (tape_atom.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries. Verification: V2 (objdump
- *  byte-identical) holds.
- *
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/pad.atom.c

@@ -0,0 +1,184 @@
+#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 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,
+	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(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/pad.h

@@ -0,0 +1,73 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "dsl.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; the active-low-to-active-high inversion is applied bit-by-bit. */
+enum {
+	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,
+};
+
+#define pad0_(btn_id) (btn_id << Pad0)
+#define pad1_(btn_id) (btn_id << Pad1)
+
+/* ============================================================
+ * BIOS pad-buffer subsystem: docs/psx-spx/docs/kernelbios.md (B(12h) + B(13h))
+ * ============================================================ */
+
+enum {
+	PAD_BIOS_RAW_SIZE = 0x22,
+};
+typedef Struct_(PadBiosRaw) {
+	U1 bytes[PAD_BIOS_RAW_SIZE];
+};
+
+typedef Enum_(U4, PadStatus) {
+	PadStatus_Disconnected,
+	PadStatus_Digital,
+	PadStatus_AnalogStick,
+	PadStatus_AnalogPad,
+	PadStatus_Unsupported,
+	PadStatus_Pending,
+	PadStatus_Invalid,
+};
+
+/* PadState — per-port normalized runtime state.
+ * 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.
+ * The struct size stays 12 bytes (unchanged from the prior order,
+ * which left the C compiler to insert 1 byte of trailing pad to reach the 4-byte struct alignment). */
+typedef Struct_(PadState) {
+	PadStatus status;  /* offset 0,  size 4 (U4) */
+	U2        buttons; /* offset 4,  size 2 */
+	U1        id;      /* offset 6,  size 1 */
+	U1        pad;     /* offset 7,  size 1 — explicit pad to align the axes block */
+	U1        left_x;  /* offset 8,  size 1 — store_word target (4-byte aligned) */
+	U1        left_y;  /* offset 9,  size 1 */
+	U1        right_x; /* offset 10, size 1 */
+	U1        right_y; /* offset 11, size 1 */
+};

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 +1,103 @@
+#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
+*/
+
+M3_S2* m3s2_rotation   (V3_S2* vec, M3_S2* mat) asm("RotMatrix");
+M3_S2* m3s2_translation(M3_S2* mat, V3_S4* vec) asm("TransMatrix");
+M3_S2* m3s2_scale      (M3_S2* 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   (M3_S2* mat) asm("SetRotMatrix");
+void gte_matrix_set_translation(M3_S2* mat) asm("SetTransMatrix");

code/duffle/word_count.metadata.h

@@ -1,10 +1,15 @@
-// tape_atom.metadata.h
+// word_count.metadata.h
 // Single source of truth for instruction-word counts.
 // Used by C (to define compile-time constants) AND Python (to count positions).
 //
 // Format: WORD_COUNT(MACRO_NAME, COUNT)
 // 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 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.
+//
 // To regenerate: hand-count the instructions in each macro definition.
 // (You'll only need to do this once per macro — they don't change often.)
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
@@ -17,6 +22,7 @@ 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)
@@ -24,7 +30,9 @@ 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)
@@ -32,6 +40,7 @@ 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)
@@ -47,15 +56,5 @@ WORD_COUNT(gte_cmdw_nclip,                   1)
 WORD_COUNT(gte_avg_sort_z3,                  1)
 WORD_COUNT(sub_u,                            1)
 WORD_COUNT(nop2,                             2)
-WORD_COUNT(mac_load_tri_indices,             3)
-WORD_COUNT(mac_load_tri_verts,              18)
-WORD_COUNT(mac_format_f3_color,              3)
-WORD_COUNT(mac_gte_store_f3_post_rtpt,       3)
-WORD_COUNT(mac_pack_color_word,              3)
-WORD_COUNT(mac_format_g4_color,             12)
-WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
-WORD_COUNT(mac_gte_store_g4_p3_post_rtps,    1)
-WORD_COUNT(mac_insert_ot_tag,               11)
-WORD_COUNT(mac_yield,                        4)
 
 #undef WORD_COUNT

code/graphics_hello_psyq/hello_gpu.c

@@ -17,19 +17,19 @@ enum {
 };
 
 typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
-typedef def_farray(OrderingTable_Buffer, 2);
+typedef Array_(OrderingTable_Buffer, 2);
 
 typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
-typedef def_farray(PrimitiveBuffer, 2);
+typedef Array_(PrimitiveBuffer, 2);
-typedef def_struct(PrimitiveArena) {
+typedef Struct_(PrimitiveArena) {
 	A2_PrimitiveBuffer buf;
 	U4                 used;
 };
 
 #define Cube_num_verts 8
-typedef def_farray(V3_S2, Cube_num_verts);
+typedef Array_(V3_S2, Cube_num_verts);
 #define Cube_num_faces 6
-typedef def_farray(V4_S2, Cube_num_faces);
+typedef Array_(V4_S2, Cube_num_faces);
 void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
 	memory_copy(verts, & (A8_V3_S2) {
 		{ -128, -128, -128 },
@@ -40,7 +40,7 @@ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
 		{  128,  128, -128 },
 		{  128,  128,  128 },
 		{ -128,  128,  128 }
-	}, size_of(A8_V3_S2) );
+	}, S_(A8_V3_S2) );
 	memory_copy(faces, & (A6_V4_S2) {
 		{ 3, 2, 0, 1 },
 		{ 0, 1, 4, 5 },
@@ -48,10 +48,10 @@ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
 		{ 1, 2, 5, 6 },
 		{ 2, 3, 6, 7 },
 		{ 3, 0, 7, 4 },
-	}, size_of(A6_V4_S2) );
+	}, S_(A6_V4_S2) );
 	return;
 }
-typedef def_struct(Ent_Cube) {
+typedef Struct_(Ent_Cube) {
 	V3_S4 accel;
 	V3_S4 vel;
 	V3_S4 pos;
@@ -62,22 +62,22 @@ typedef def_struct(Ent_Cube) {
 };
 
 #define Floor_num_verts 4
-typedef def_farray(V3_S2, Floor_num_verts);
+typedef Array_(V3_S2, Floor_num_verts);
 #define Floor_num_faces 2
-typedef def_farray(V3_S2, Floor_num_faces);
+typedef Array_(V3_S2, Floor_num_faces);
 void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
 	memory_copy(verts, &(A4_V3_S2) {
 		{ -900, 0, -900 },
 		{ -900, 0,  900 },
 		{  900, 0, -900 },
 		{  900, 0,  900 },
-	}, size_of(A8_V3_S2));
+	}, S_(A8_V3_S2));
 	memory_copy(faces, & (A2_V3_S2) {
 		{ 0, 1, 2 },
 		{ 1, 3, 2 },
-	}, size_of(A2_V3_S2));
+	}, S_(A2_V3_S2));
 };
-typedef def_struct(Ent_Floor) {
+typedef Struct_(Ent_Floor) {
 	V3_S4 accel;
 	V3_S4 pos;
 	V3_S4 scale;
@@ -86,7 +86,7 @@ typedef def_struct(Ent_Floor) {
 	A2_V3_S2 faces;
 };
 
-typedef def_struct(SMemory) {
+typedef Struct_(SMemory) {
 	DoubleBuffer            screen_buf;
 	A2_OrderingTable_Buffer ordering_tbl;
 	PrimitiveArena          primitives;
@@ -108,7 +108,7 @@ B1* prim__alloc(U4 type_width, Str8 type_name) {
 	pa->used += type_width;
 	return next;
 }
-#define prim_alloc(type) (type*)prim__alloc(size_of(type), txt( stringify(type)))
+#define prim_alloc(type) (type*)prim__alloc(S_(type), slit( stringify(type)))
 
 void gp_screen_init_c11(DoubleBuffer* screen_buf, S2* active_buf_id)
 {

code/graphics_hello_psyq/hello_gpu.h

@@ -5,8 +5,8 @@
 #	include "duffle/gp.h"
 #endif
 
-typedef def_struct(DrawEnv_Packed) { U4 tag; U4 code[15]; };
+typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };
-typedef def_struct(DrawEnv) {
+typedef Struct_(DrawEnv) {
 	Rect_S2 clip_area;
 	A2_S2   drawing_offset;
 	Rect_S2 texture_window;
@@ -17,7 +17,7 @@ typedef def_struct(DrawEnv) {
 	RGB8    initial_bg_color;
 	DrawEnv_Packed dr_env; // reserved
 };
-typedef def_struct(DisplayEnv) {
+typedef Struct_(DisplayEnv) {
 	Rect_S2 display_area;
 	Rect_S2 screen;
 	B1      vinterlace;
@@ -25,9 +25,9 @@ typedef def_struct(DisplayEnv) {
 	B1      pad0;
 	B1      pad1;
 };
-typedef def_farray(DrawEnv,    2);
+typedef Array_(DrawEnv,    2);
-typedef def_farray(DisplayEnv, 2);
+typedef Array_(DisplayEnv, 2);
-typedef def_struct(DoubleBuffer) {
+typedef Struct_(DoubleBuffer) {
 	A2_DrawEnv    draw;
 	A2_DisplayEnv display;
 };
@@ -58,7 +58,7 @@ U4 vsync(U4 mode)     __asm__("VSync");
 
 void draw_orderingtbl(U4* buf) __asm__("DrawOTag");
 
-typedef def_struct(PolyTag) {
+typedef Struct_(PolyTag) {
 	U4   addr: 24;
 	U4   len:  8;
 	RGB8 color;
@@ -106,7 +106,7 @@ typedef def_struct(PolyTag) {
 // #define setLineF4(p)	set_len(p, 6),  set_code(p, 0x4c),(p)->pad = 0x55555555
 // #define setLineG4(p)	set_len(p, 9),  set_code(p, 0x5c),(p)->pad = 0x55555555, (p)->p2 = 0, (p)->p3 = 0
 
-typedef def_struct(Poly_F3) {
+typedef Struct_(Poly_F3) {
 	U4   tag;
 	RGB8 color;
 	B1   code;
@@ -120,14 +120,14 @@ typedef def_struct(Poly_F3) {
 	};
 };
 
-typedef def_struct(Poly_G3) {
+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;
 };
 
-typedef def_struct(Poly_F4) {
+typedef Struct_(Poly_F4) {
 	U4   tag;
 	RGB8 color;
 	B1   code;
@@ -142,7 +142,7 @@ typedef def_struct(Poly_F4) {
 	};
 };
 
-typedef def_struct(Poly_G4) {
+typedef Struct_(Poly_G4) {
 	U4    tag; RGB8 c0; B1 code;
 	V2_S2 p0;  RGB8 c1; B1 pad1;
 	V2_S2 p1;  RGB8 c2; B1 pad2;
@@ -150,7 +150,7 @@ typedef def_struct(Poly_G4) {
 	V2_S2 p3;
 };
 
-typedef def_struct(Tile) {
+typedef Struct_(Tile) {
 	U4      tag;
 	RGB8    color;
 	B1      code;
@@ -169,7 +169,7 @@ M3_S2* m3s2_scale      (M3_S2* 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, cast(S4*R_, & xy->x), cast(S4*R_, pp), r_(flag)); }
+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(
@@ -179,8 +179,8 @@ FI_  S4 rtp_avg_nclip_a3_v3s2(
 ){
 	return rtp_avg_nclip_a3_v3s2_raw(
 		v0, v1, v2, 
-		cast(S4*R_, xy0), cast(S4*R_, xy1), cast(S4*R_, xy2),
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2),
-		cast(S4*R_, pp),  cast(S4*R_, otz), cast(S4*R_, flag)
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
 	);
 }
 
@@ -192,8 +192,8 @@ FI_ S4 rtp_avg_nclip_a4_v3s2(
 ){
 	return rtp_avg_nclip_a4_v3s2_raw(
 		v0, v1, v2, v3,
-		cast(S4*R_, xy0), cast(S4*R_, xy1), cast(S4*R_, xy2), cast(S4*R_, xy3),
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2), C_(S4*R_, xy3),
-		cast(S4*R_, pp),  cast(S4*R_, otz), cast(S4*R_, flag)
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
 	);
 }
 

code/gte_hello/hello_gte_tape.c

@@ -1,193 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/lottes_tape.h"
-#	include "duffle/atom_dsl.h"
-#	include "hello_gte.h"
-#	include "tape_atom.metadata.h"
-#	include "gen/hello_gte_tape.offsets.h"
-#endif
-
-#pragma region MACs (Mips Atom components)
-/* The macros mac_format_f3_color and mac_gte_store_f3 moved to
- * lottes_tape.h during the Phase 3 gp.h overhaul. Both are now RGB-form
- * (mac_format_f3_color takes _r, _g, _b byte values rather than raw
- * 16-bit half-words). */
-
-#pragma endregion MACs
-
-#pragma region Baked Atoms
-
-typedef Struct_(Binds_CubeTri) {
-	U4 PrimCursor;
-	U4 FaceCursor;
-	U4 VertBase;
-	U4 OtBase;
-};
-internal MipsAtom_(rbind_cube_g4_face) {
-	/* 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)),
-	// Note(Ed): This entire thing is argument shuffle?
-	// TODO(Ed): Eliminate
-	mac_yield()
-};
-
-/* ============================================================================
- *  cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
- * ============================================================================
- *
- *  Reads 4 indices from R_FaceCur (V4_S2 = 8 bytes), loads 4 vertices into
- *  the GTE, runs the PsyQ RotAverageNclip4 sequence, and renders a Poly_G4.
- */
-	atom_region  (cube_g4_face, REGION_PRIM_ARENA)
-	atom_group   (cube_g4_face, GROUP_RENDER_PRIMS)
-	atom_cadence (cube_g4_face, CADENCE_FRAME)
-	atom_annot(cube_g4_face, phase_work,
-			atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-			atom_writes(R_PrimCursor, R_FaceCursor))
-internal 
-MipsAtom_(cube_g4_face) {
-	/* ── 1. Load 4 face indices from R_FaceCur (V4_S2 = 8 bytes) ───────── */
-	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)),
-
-	/* ── 2. Load V0, V1, V2 into GTE (parallel to mac_load_tri_verts) ── */
-	mac_load_tri_verts(R_T0, R_T1, R_T2),
-
-	/* ── 3. RTPT — transforms V0/V1/V2 → SXY0/SXY1/SXY2 + SZ1/SZ2/SZ3 ─── */
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-
-	/* ── 4. NCLIP — backface culling on SXY0/SXY1/SXY2 (p0,p1,p2) ──────── */
-	/*     MUST be done BEFORE V3-RTPS overwrites SXY0 with p3.             */
-	nop2, gte_cmdw_nclip,
-
-	/* ── 5. Cull check: skip format/insert if MAC0 ≤ 0 (backface) ───────── */
-	nop2,  gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,                                                          /* COP2 stall */
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
-	nop,                                                          /* BD slot    */
-
-	/* ── 6. Format c0..c3 (color+code words) BEFORE V3-RTPS ─────────────── */
-	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),
-
-	/* ── 7. Store p0..p2 BEFORE V3-RTPS overwrites SXY0 ─────────────────── */
-	mac_gte_store_g4_p012_post_rtpt_pre_rtps(),
-
-	/* ── 8. Load V3 = verts[face->w] into V0 ─────────────────────────────── */
-	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),
-
-	/* ── 9. RTPS — transforms V0 (now V3) → SXY0 (p3) + SZ3 ─────────────── */
-	nop2, gte_cmdw_rotate_translate_perspective_single,
-	mac_gte_store_g4_p3_post_rtps(),
-
-	/* ── 10. AVSZ4 — average Z from SZ0/SZ1/SZ2/SZ3 ─────────────────────── */
-	nop2, gte_cmdw_avg_sort_z4,
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
-
-	/* ── 11. Bounds check OTZ < OrderingTbl_Len ─────────────────────────── */
-	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,
-
-	/* ── 12. Insert into Ordering Table (length = 8 words for Poly_G4) ──── */
-	mac_insert_ot_tag(R_T1, Poly_G4),
-
-	/* ── 13. Advance cursors & yield (both branch targets land here) ────── */
-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;
-	U4 FaceCursor;
-	U4 VertBase;
-	U4 OtBase;
-};
-	atom_region(rbind_floor_f3_face, REGION_PRIM_ARENA)
-	atom_group(rbind_floor_f3_face, GROUP_RENDER_FLOOR)
-	atom_cadence(rbind_floor_f3_face, CADENCE_FRAME)
-	atom_annot(rbind_floor_f3_face, phase_bind
-		, atom_reads()
-		, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
-internal 
-MipsAtom_(rbind_floor_f3_face) {
-	/* 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_region( floor_f3_face, REGION_PRIM_ARENA)
-	atom_group(  floor_f3_face, GROUP_RENDER_FLOOR)
-	atom_cadence(floor_f3_face, CADENCE_FRAME)
-	atom_annot(  floor_f3_face, phase_work,
-		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-		atom_writes(R_PrimCursor, R_FaceCursor)) 
-internal 
-MipsAtom_(floor_f3_face) {
-	mac_load_tri_indices(R_T0, R_T1, R_T2),
-	mac_load_tri_verts(  R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-	nop2, gte_cmdw_nclip,
- 
-	/* Culling (Branch forward if Backface) */
-	nop2, gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
-	branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop,
-	/* Format Primitive */
-	// mac_format_f3_color(0x20FF, 0xFFFF),  // works
-	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-	mac_gte_store_f3_post_rtpt(),
-	
-	/* Calculate Depth */
-	nop2, gte_avg_sort_z3,            
-	nop2, 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(R_T1, Poly_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; };
-	atom_region( sync_primitive_arena, REGION_PRIM_ARENA)
-	atom_group(  sync_primitive_arena, GROUP_RENDER_FLOOR)
-	atom_cadence(sync_primitive_arena, CADENCE_FRAME)
-	atom_annot(  sync_primitive_arena, phase_work,
-		atom_reads( R_TapePtr, R_PrimCursor),
-		atom_writes(R_TapePtr)) 
-internal MipsAtom_(sync_primitive_arena) {
-	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/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 +1,50 @@
+// 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_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,
+};
+
+// --- 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,
+};
+
+#pragma endregion hello_camera
+

code/hello_camera/hello_camera.atom.c

@@ -0,0 +1,468 @@
+#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_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(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_disp_env, {
+	// 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(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_draw_env, {
+	/*
+	* 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(),
+};
+
+/* ----- 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(),
+};
+
+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()
+};
+
+#pragma endregion Baked Atoms

code/hello_camera/hello_camera.c

@@ -0,0 +1,336 @@
+#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/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 "hello_camera.h"
+#pragma endregion Hello Camera Headers
+
+#pragma region Hello Joypad TUs
+#include "hello_camera.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 (1) // Pad Input
+	{
+		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 (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 (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_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),
+};
+
+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_gte/gen/gte_hello.offsets.h

@@ -1,24 +1,24 @@
-// Auto-generated by tape_atom_offset_gen.meta.lua — DO NOT EDIT
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
 // Source: C:\projects\Pikuma\ps1\code\gte_hello\hello_gte_tape.c
 #pragma once
 
 #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_cull_cube_g4_face_exit           42
-#define _atom_offset_bounds_chk_cube_g4_face_exit     12
+#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 (66 words) ---
+// --- atom: floor_f3_face (58 words) ---
 
-#define _atom_offset_culling_floor_f3_face_exit       29
+#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

@@ -1,6 +1,6 @@
-#include "stdio.h"
+#include <stdio.h>
 #include <stdlib.h>
-#include "assert.h"
+#include <assert.h>
 // #include "libgpu.h"
 // #include "libetc.h"
 // #include "libgte.h"
@@ -8,19 +8,22 @@
 #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/gen/lottes_tape.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 "tape_atom.metadata.h"
+#	include "gen/offsets.h"
-#	include "gen/hello_gte_tape.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);
@@ -96,8 +99,13 @@ typedef Struct_(Ent_Floor) {
 	A2_V3_S2 faces;
 };
 
-enum { scratchpad_size = 1024, };
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
 typedef Struct_(SMemory) {
+	U4 MemTape[MemTape_Len];
+
 	DoubleBuffer            screen_buf;
 	A2_OrderingTable_Buffer ordering_tbl;
 	PrimitiveArena          primitives;
@@ -114,7 +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;
 }
 
@@ -179,6 +187,7 @@ void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_
 void render(void) {
 }
 
+GCC_OPTIMIZATION_DISABLE
 void update(PrimitiveArena* pa, U4* ordering_buf) 
 {
 	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
@@ -204,6 +213,8 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 	A2_S2 p;    //???
 	S4 flag; //????
 
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
 	// Draw Cube
 	if (0)
 	{
@@ -256,9 +267,8 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
 
-		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
+		tb.used = 0; tb_scope(& tb) {
-		TapeBuilder tb = tb_make_old(&tape_arena); tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
-			tb_emit(& tb, code_rbind_cube_g4_face);
 				tb_data(& tb, prim_cursor);
 				tb_data(& tb, u4_(smem.cube.faces));
 				tb_data(& tb, u4_(smem.cube.verts));
@@ -266,10 +276,10 @@ void update(PrimitiveArena* pa, 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, code_cube_g4_face);
+				tb_emit(& tb, cube_g4_face);
 			}
 
-			tb_emit(& tb, code_sync_primitive_arena);
+			tb_emit(& tb, sync_primitive_arena);
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}
@@ -337,56 +347,44 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
 		m3s2_translation(& smem.tform_world, & smem.floor.pos);
 		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
-		// TODO(Ed): This can either be in the tape or here...
-		// 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;
 
-		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris.
+		// 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)
-		LP_ U4 mem_temp_tape[512];
+		tb.used = 0; tb_scope(& tb) {
-		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
+			tb_emit(& tb, set_gte_world);
-			// TODO(Ed): This is bugged.
-			tb_emit(& tb, code_set_gte_world);
 				tb_data(& tb, u4_(& smem.tform_world));
 
-			tb_emit(& tb, code_rbind_floor_f3_face);
+			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, code_floor_f3_face);
+				tb_emit(& tb, floor_f3_face);
 			}
-			// After code_floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
-			tb_emit(& tb, code_sync_primitive_arena);
+			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;
 	}
 	// --- TAPE DIAGNOSTICS ---
-	if (1)
+	if (0)
 	{
 		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
 		TapeBuilder tb = tb_make_old(& tape_arena); tb_scope(& tb) {
 			// Skip set_gte_world atom for diagnostics to isolate the triangle loop
 			for (U4 i = 0; i < Floor_num_faces; i++) {
-				// =======================================================
-				// SWAP EMIT TO TEST DIFFERENT PARTS OF THE PIPELINE:
-				// =======================================================
-				// 1. code_diag_yield -> Tests Tape Engine jump logic
-				// 2. code_diag_color -> Tests OT and Prim Arena memory
-				// 3. code_diag_gte   -> Tests Vertex arrays and GTE Math
 				// tb_emit(& tb, code_diag_yield);
 				// tb_emit(& tb, code_diag_color);
 				// tb_emit(& tb, code_diag_gte); 
@@ -395,9 +393,9 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		B1* prim_cursor = (B1*)r_(pa->buf)[smem.active_buf_id] + pa->used;
 		tape_run(tb_slice(tb));
 		pa->used = (U4)prim_cursor - (U4)r_(pa->buf)[smem.active_buf_id];
-		smem.floor.rot.y += 5;
 	}
 }
+GCC_OPTIMIZATION_ENABLE
 
 int main(void)
 {

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(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_disp_env, {
+	// 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(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_draw_env, {
+	/*
+	* 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(U4 status_val, U4 state_ptr_reg, U4 scratch_reg)
+MipsAtomComp_Proc_(ac_pad_sio_write_pad_state, {
+	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

@@ -6,18 +6,33 @@ A rest from the usual.
 
 ## Dependencies
 
-I will be programming from a Windows 11 machine:
+I will be programming from a Windows 11 machine (may eventually try this on the Steam Deck...):
-![system_info](./docs/assets/system_info.png)
-
-```ps1
-# not really used yet for scripts (may never)
-scoop install lua
-```
 
 [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)
+
+```
+scoop install luajit
+```
+
+* Used for lua scripts
+* Particularly, ps1_meta.lua which is a staged metaprogram pass for the custom C11 Assembly DSL used in this codebase.
+
+[lpeg](https://github.com/roberto-ieru/LPeg.git)
+
+* Lua is slow (even jitted) so this helps.
+
+[lfs (LuaFileSystem)](https://github.com/lunarmodules/luafilesystem)
+
+* Native directory enumeration + `mkdir` for the build scripts.
+* Used by `passes/word_count_eval.lua :: scan_dir` (native walk vs. `dir /b /s` subprocess,
+  ~2ms vs. ~56ms) and by `duffle.lua :: ensure_dir` + `to_absolute_path` (avoids
+  `cmd.exe mkdir` + `cd` shell spawns, ~50ms each).
 
 [pscx-redux](https://github.com/grumpycoders/pcsx-redux/): A collection of tools, research, hardware design, and libraries aiming at development and reverse engineering on the PlayStation 1.
 
@@ -57,3 +72,10 @@ scoop install lua
 ![polys!](./docs/assets/pcsx-redux.main_2025-08-03_20-45-35.png)  
 ![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...

scratch/gte_debug.lua

@@ -0,0 +1,30 @@
+-- gte_debug.lua — defensive version + prints error context.
+local ok, err = pcall(function()
+    print("[debug] PCSX exists:", PCSX ~= nil)
+    print("[debug] PCSX.WebServer exists:", PCSX and PCSX.WebServer ~= nil)
+    print("[debug] PCSX.WebServer.Handlers exists:", PCSX and PCSX.WebServer and PCSX.WebServer.Handlers ~= nil)
+    if not PCSX.WebServer then
+        print("[debug] creating PCSX.WebServer...")
+        PCSX.WebServer = {}
+    end
+    if not PCSX.WebServer.Handlers then
+        print("[debug] creating PCSX.WebServer.Handlers...")
+        PCSX.WebServer.Handlers = {}
+    end
+    print("[debug] type of Handlers:", type(PCSX.WebServer.Handlers))
+
+    PCSX.WebServer.Handlers.gte = function(req)
+        local r = PCSX.getRegisters()
+        local out = { "pc=0x" .. string.format("%x", r.pc) }
+        for i = 0, 31 do
+            out[#out + 1] = string.format("D[%d]=0x%08x C[%d]=0x%08x",
+                i, r.CP2D.r[i], i, r.CP2C.r[i])
+        end
+        return table.concat(out, "\n")
+    end
+    print("[debug] handler registered")
+end)
+
+if not ok then
+    print("[debug] ERROR: " .. tostring(err))
+end

scripts/audit_lua_nesting.lua

@@ -0,0 +1,217 @@
+--- audit_lua_nesting.lua — Walk Lua source files and flag any block nesting deeper than 5 levels.
+---
+--- Usage:
+---   luajit scripts/audit_lua_nesting.lua scripts/duffle.lua scripts/ps1_meta.lua
+---   luajit scripts/audit_lua_nesting.lua scripts/passes/
+---
+--- Output: for each file, a list of {line, depth} entries where depth > 5.
+--- Returns exit code 1 if any violations found, 0 if clean.
+---
+--- **Implementation**: a hand-rolled depth tracker that counts:
+---   - `do`, `function`, `if`, `for`, `while`, `repeat`  -> depth +1
+---   - `end`, `until`                                    -> depth -1
+---   - `else`, `elseif`                                  -> depth unchanged
+---
+--- **Caveats**: doesn't fully handle string/comment state (will miscount braces inside multi-line strings or block comments). 
+--- For our metaprogram files (no embedded code generation), this is acceptable.
+
+local M = {}
+
+local BLOCK_OPEN = {
+	["do"]       = true,
+	["function"] = true,
+	["if"]       = true,
+	["for"]      = true,
+	["while"]    = true,
+	["repeat"]   = true,
+}
+
+local function is_block_close(token) return token == "end" or token == "until" end
+
+-- (internal) Walk one source file and return a list of
+-- {line, depth, token} entries where depth > max_nesting.
+local function audit_file(path, max_nesting)
+	local  f = io.open(path, "r")
+	if not f then error("Cannot open " .. path) end
+	local content = f:read("*a")
+	f:close()
+
+	local violations = {}
+	local depth      = 0
+	local line       = 1
+	local pos        = 1
+	local src_len    = #content
+	local token_idx  = 0
+
+	local function read_ident_at(start_pos)
+		local ident_start = start_pos
+		if ident_start > src_len then return nil end
+		local first_ch = content:sub(ident_start, ident_start)
+		if not (first_ch:match("[%a_]")) then return nil end
+		local scan = start_pos + 1
+		while scan <= src_len do
+			local ch = content:sub(scan, scan)
+			if not (ch:match("[%w_]")) then break end
+			scan = scan + 1
+		end
+		return content:sub(ident_start, scan - 1), scan
+	end
+
+	-- Skip past a string literal or comment starting at `start_pos`.
+	-- Returns the position just past the construct, or nil if `start_pos`
+	-- is not the start of a string/comment.
+	local function skip_string_or_comment(start_pos)
+		local ch = content:sub(start_pos, start_pos)
+		if    ch == '"' or ch == "'" then
+			local scan = start_pos + 1
+			while scan <= src_len do
+				local  c = content:sub(scan, scan)
+				if     c == "\\" then        scan = scan + 2
+				elseif c == ch   then return scan + 1
+				else                         scan = scan + 1
+				end
+			end
+			return src_len + 1
+		elseif ch == "-" and content:sub(start_pos + 1, start_pos + 1) == "-" then
+			local scan = start_pos + 2
+			if content:sub(scan, scan + 1) == "[[" and content:sub(scan + 2, scan + 3) == "[" then
+				-- Long bracket comment [==[ ... ]==]
+				scan = scan + 2
+				local eq = ""
+				while content:sub(scan, scan) == "=" do
+					eq = eq .. "="
+					scan = scan + 1
+				end
+				local close_marker = "]" .. eq .. "]"
+				local close_pos = content:find(close_marker, scan, true)
+				if close_pos then
+					return close_pos + #close_marker
+				else
+					return src_len + 1
+				end
+			else
+				while scan <= src_len and content:sub(scan, scan) ~= "\n" do scan = scan + 1 end
+				return scan + 1
+			end
+		elseif ch == "[" and content:sub(start_pos + 1, start_pos + 1) == "[" then
+			local scan = start_pos + 2
+			local eq = ""
+			while content:sub(scan, scan) == "=" do
+				eq   = eq .. "="
+				scan = scan + 1
+			end
+			local close_marker = "]" .. eq .. "]"
+			local close_pos = content:find(close_marker, scan, true)
+			if close_pos then
+				return close_pos + #close_marker
+			else
+				return src_len + 1
+			end
+		end
+		return nil
+	end
+
+	while pos <= src_len do
+		local ch = content:sub(pos, pos)
+		if ch == "\n" then line = line + 1 end
+
+		local skip_to = skip_string_or_comment(pos)
+		if skip_to then
+			for scan = pos, skip_to - 1 do
+				if content:sub(scan, scan) == "\n" then line = line + 1 end
+			end
+			pos = skip_to
+		elseif ch:match("[%a_]") then
+			local tok, next_pos = read_ident_at(pos)
+			token_idx = token_idx + 1
+			if BLOCK_OPEN[tok] then
+				depth = depth + 1
+				if depth > max_nesting then
+					violations[#violations + 1] = {
+						line  = line,
+						depth = depth,
+						token = tok,
+					}
+				end
+			elseif is_block_close(tok) then
+				depth = depth - 1
+			end
+			pos = next_pos
+		else
+			pos = pos + 1
+		end
+	end
+
+	return violations
+end
+
+--- Audit one file. Returns nil if clean, else a list of violations.
+--- @param path string
+--- @param max_nesting integer  -- default 5
+--- @return table|nil
+function M.audit(path, max_nesting)
+	local  violations = audit_file(path, max_nesting or 5)
+	if     #violations == 0 then return nil end
+	return violations
+end
+
+-- Module CLI.
+if arg and arg[1] then
+	local max_nesting = 5
+	local files = {}
+	for arg_idx = 1, #arg do
+		if arg[arg_idx] == "--max" and arg[arg_idx + 1] then
+			max_nesting = tonumber(arg[arg_idx + 1]) or 5
+		else
+			files[#files + 1] = arg[arg_idx]
+		end
+	end
+
+	-- Accept either a directory or a file path. Directory args are
+	-- expanded via lfs.dir (native, no subprocess).
+	local lfs = require("lfs")
+	local function is_dir(p)
+		return lfs.attributes(p, "mode") == "directory"
+	end
+	local function list_lua(dir)
+		local out = {}
+		if not is_dir(dir) then return out end
+		for entry in lfs.dir(dir) do
+			if entry:match("%.lua$") then
+				out[#out + 1] = dir .. "/" .. entry
+			end
+		end
+		return out
+	end
+
+	local to_check = {}
+	for _, f in ipairs(files) do
+		if is_dir(f) then
+			for _, sub in ipairs(list_lua(f)) do to_check[#to_check + 1] = sub end
+		else
+			to_check[#to_check + 1] = f
+		end
+	end
+
+	local total_violations = 0
+	for _, f in ipairs(to_check) do
+		local v = M.audit(f, max_nesting)
+		if v then
+			io.write(string.format("\n%s\n", f))
+			for _, x in ipairs(v) do
+				io.write(string.format("  line %d: depth %d (after '%s')\n", x.line, x.depth, x.token))
+			end
+			total_violations = total_violations + #v
+		end
+	end
+
+	if total_violations == 0 then
+		io.write("OK: no files exceed max nesting of " .. max_nesting .. "\n")
+		os.exit(0)
+	else
+		io.write(string.format("\n%d nesting violation(s) found.\n", total_violations))
+		os.exit(1)
+	end
+end
+
+return M

scripts/build_psyq.ps1

@@ -81,19 +81,6 @@ $path_psyq          = join-path $path_toolchain  'psyq-4_7'
 $path_psyq_iwyu     = join-path $path_toolchain  'psyq_iwyu'
 $path_psyq_imyu_inc = join-path $path_psyq_iwyu  'include'
 
-function Get-SourceFiles { param([Parameter(Mandatory=$true)] [string[]]$paths, [Parameter(Mandatory=$true)] [string[]]$extensions)
-	$files = @()
-	foreach ($p in $paths) {
-		if (-not (test-path $p)) { continue }
-		foreach ($ext in $extensions) {
-			Get-ChildItem -Path $p -File -Recurse -Filter "*$ext" -ErrorAction SilentlyContinue | ForEach-Object {
-				$files += $_.FullName
-			}
-		}
-	}
-	return ($files | Sort-Object -Unique)
-}
-
 function assemble-unit { param(
 	[string]  $unit,
 	[string]  $link_module,
@@ -153,7 +140,7 @@ function compile-unit { param(
 		$f_arch_no_shared,
 		$f_arch_no_stack_prot
 	)
-	# $compile_args    += $f_std_c23
+	$compile_args    += $f_std_c11
 	$compile_args    += ($f_include + $path_psyq_imyu_inc)
 	$compile_args    += ($f_include + $path_nugget)
 
@@ -193,29 +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.).
 	$libraries = @(
 		"api",
 		"c",
-		"c2",
-		"card",
-		"cd",
-		"comb",
-		"ds",
 		"etc",
 		"gpu",
-		"gs",
+		"gte"
-		"gte",
-		"gun",
-		"hmd",
-		"math",
-		"mcrd",
-		"mcx",
-		"pad",
-		"press",
-		"sio",
-		"snd",
-		"spu",
-		"tap"
 	)
 	foreach ($lib in $libraries) {
 		$link_args += ($f_link_lib + $lib)
@@ -243,6 +216,117 @@ function make-binary { param([string]$elf, [string]$exe)
 	if ($LASTEXITCODE -ne 0) { Write-Error "Objcopy failed. Aborting."; exit 1 }
 }
 
+function ps1-meta { param(
+		[string]$unity_root,
+		[string[]]$sources,
+		[Parameter(Mandatory=$true)][string]$metadata,
+		[string]$out_root = (join-path $path_build 'gen'),
+		[string[]]$passes = @('--pre-link'),
+		[string[]]$extra_args = @()
+	)
+	# `--unity-root` and `--source` are mutually exclusive. Exactly one of `$unity_root` / `$sources` must be supplied; the other must be absent.
+	if ($null -ne $unity_root -and $unity_root -ne '') 
+	{
+		if ($null -ne $sources -and $sources.Count -gt 0) {
+			write-error 'ps1-meta: -unity_root and -sources are mutually exclusive'
+			exit 2
+		}
+	} 
+	elseif ($null -eq $sources -or $sources.Count -eq 0) {
+		write-error 'ps1-meta: either -unity_root <file> or -sources <file...> is required'
+		exit 2
+	}
+
+	$script = join-path $path_scripts 'ps1_meta.lua'
+	$input_summary = if ($null -ne $unity_root -and $unity_root -ne '') {
+		"unity=$unity_root"
+	}
+	else {
+		"$($sources.Count) source(s)"
+	}
+	write-host "ps1-meta  $input_summary, passes=$($passes -join ',')" ` -ForegroundColor Magenta
+
+	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root) + @($extra_args)
+	if ($null -ne $unity_root -and $unity_root -ne '') {
+		$arg_list += @('--unity-root', $unity_root)
+	}
+	else {
+		foreach ($s in $sources) { $arg_list += @('--source', $s) }
+	}
+	& luajit $script @arg_list
+	if ($LASTEXITCODE -ne 0) {
+		write-error "ps1-meta failed (exit $LASTEXITCODE). Aborting."
+		exit $LASTEXITCODE
+	}
+}
+
+function inject-dwarf { param(
+	[string]$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 += @()
 
@@ -287,7 +371,7 @@ function build-graphis_hello {
 
 	$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
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
 
 	$src_asm    = join-path $path_module 'hello_gpu.s'
 	$module_asm = join-path $path_build  'hello_gpu.o'
@@ -317,114 +401,16 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function generate-TapeAtomOffsets {param([Parameter(Mandatory=$true)] [string[]]$sources, [Parameter(Mandatory=$true)] [string]$metadata)
+function build-hello_gte {
-	$gen_atom_offsets_script = join-path $path_scripts 'tape_atom.offset_gen.meta.lua'
-
-    $any_stale = $false
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $out      = join-path $gen_dir "$basename.offsets.h"
-
-        if (-not (test-path $out)) { $any_stale = $true; break }
-        $src_mtime  = (get-item $src).LastWriteTimeUtc
-        $out_mtime  = (get-item $out).LastWriteTimeUtc
-        $meta_mtime = (get-item $metadata).LastWriteTimeUtc
-        if (($src_mtime -gt $out_mtime) -or ($meta_mtime -gt $out_mtime)) {
-            $any_stale = $true
-            break
-        }
-    }
-
-    if (-not $any_stale) {
-        write-host "AtomOffsets  all $($sources.Count) source(s) up-to-date" -ForegroundColor DarkGray
-        return
-    }
-
-    write-host "AtomOffsets  $($sources.Count) source(s)" -ForegroundColor Magenta
-    & luajit $gen_atom_offsets_script $metadata @sources
-    if ($LASTEXITCODE -ne 0) {
-        write-error "Atom offset generation failed. Aborting."
-        exit 1
-    }
-}
-
-function generate-TapeAtomAnnotations {param([Parameter(Mandatory=$true)] [string[]]$sources, [Parameter(Mandatory=$true)] [string]$metadata)
-	# Sibling to generate-TapeAtomOffsets. Validates TAPE_ATOM_* / TAPE_WORDS
-	# annotations against the metadata manifest. Emits gen/<basename>.errors.h
-	# containing #error directives for the C build to fail on annotation drift.
-	$gen_atom_annot_script = join-path $path_scripts 'tape_atom_annotation_pass.lua'
-
-    $any_stale = $false
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $out_txt  = join-path $gen_dir "$basename.annotations.txt"
-        $out_err  = join-path $gen_dir "$basename.errors.h"
-
-        if (-not (test-path $out_txt) -or -not (test-path $out_err)) { $any_stale = $true; break }
-        $src_mtime     = (get-item $src).LastWriteTimeUtc
-        $out_txt_mtime = (get-item $out_txt).LastWriteTimeUtc
-        $out_err_mtime = (get-item $out_err).LastWriteTimeUtc
-        $out_mtime     = if ($out_txt_mtime -gt $out_err_mtime) { $out_txt_mtime } else { $out_err_mtime }
-        $meta_mtime    = (get-item $metadata).LastWriteTimeUtc
-        if (($src_mtime -gt $out_mtime) -or ($meta_mtime -gt $out_mtime)) {
-            $any_stale = $true
-            break
-        }
-    }
-
-    if (-not $any_stale) {
-        write-host "AtomAnnotations all $($sources.Count) source(s) up-to-date" -ForegroundColor DarkGray
-        return
-    }
-
-    write-host "AtomAnnotations $($sources.Count) source(s)" -ForegroundColor Magenta
-    & luajit $gen_atom_annot_script $metadata @sources
-    if ($LASTEXITCODE -ne 0) {
-        write-error "Atom annotation generation failed. Aborting."
-        exit 1
-    }
-
-    # If any source produced annotation errors, surface them now and halt the
-    # build. The errors.h files are also #include'd via -include below, so
-    # the C build would fail at preprocessing time anyway — failing here gives
-    # a more readable error in the build log.
-    $err_count = 0
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $ann_txt  = join-path $gen_dir "$basename.annotations.txt"
-        $err_h    = join-path $gen_dir "$basename.errors.h"
-        if ((test-path $ann_txt) -and (test-path $err_h)) {
-            $txt = get-content $ann_txt -raw
-            if ($txt -match 'Errors:\s+([1-9]\d*)') {
-                $err_count += [int]$Matches[1]
-                write-warning "Annotation errors in $src — see $ann_txt"
-            }
-        }
-    }
-    if ($err_count -gt 0) {
-        write-error "Annotation pass failed: $err_count error(s) across $($sources.Count) source(s). Aborting."
-        exit 1
-    }
-}
-
-function build-gte_hello {
 	$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_module 'tape_atom.metadata.h'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build   'gen'
 
-	$source_dirs  = @($path_duffle, $path_module)
+	$src_c = join-path $path_module 'hello_gte.c'
-	$atom_sources = Get-SourceFiles -paths $source_dirs -extensions @('.h', '.c')
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
-
-	generate-TapeAtomAnnotations -sources $atom_sources -metadata $path_atom_metadata
-	generate-TapeAtomOffsets     -sources $atom_sources -metadata $path_atom_metadata
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -433,14 +419,13 @@ function build-gte_hello {
 
 	$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
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
 
 	# $src_asm    = join-path $path_module 'hello_gte.s'
 	# $module_asm = join-path $path_build  'hello_gte.o'
 
 	# assemble-unit $src_asm $module_asm $includes $assemble_args
 
-	$src_c    = join-path $path_module 'hello_gte.c'
 	$module_c = join-path $path_build  'hello_gte_c.o'
 
 	$compile_args = @()
@@ -458,25 +443,130 @@ function build-gte_hello {
 	$link_args = @()
 	$link_args += $f_debug
 	# $link_args += $f_optimize_size
-	link-modules @($module_asm_crt, $module_c) $elf $link_args
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
 	make-binary $elf $exe
-}
-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)
+
+	inject-dwarf $elf $path_build_gen
+}
+# 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_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(
+function Send-ToEmulator { param( [string]$exePath )
-    [string]$exePath
-)
     $uri = "http://localhost:8080/api/v1/load-exec"
 
     # Absolute path is safest for the emulator web server
     $absolutePath = [System.IO.Path]::GetFullPath($exePath)
 
     # Create JSON payload pointing to your compiled .ps-exe
-    $body = @{
+    $body = @{ filename = $absolutePath } | ConvertTo-Json
-        filename = $absolutePath
-    } | ConvertTo-Json
 
     Write-Host "Pushing hot-reload to PCSX-Redux..." -ForegroundColor Magenta
     try {

scripts/duffle_paths.lua

@@ -0,0 +1,92 @@
+--- duffle_paths.lua — Single-line bootstrap helper for the tape-atom Lua scripts.
+---
+--- Each entry script (ps1_meta.lua + the 7 passes/*.lua files) starts with one of:
+---   ```lua
+---   -- Entry script (ps1_meta.lua — `arg[0]` is set):
+---   local duffle = dofile((arg[0]:match("(.*[/\\])") or "./") .. "duffle_paths.lua")
+---
+---   -- Pass module (debug.getinfo path resolution; works both standalone and when require'd):
+---   local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+---   local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+---   ```
+---
+--- That small bootstrap: (a) locates this helper via `arg[0]` / `debug.getinfo`,
+--- (b) loads it (which sets `package.path` + `package.cpath`),
+--- (c) at the bottom calls `require("duffle")` (now resolvable since `package.path` was just set) and returns the duffle M.
+--- Net effect: the caller gets the duffle module in one statement; no separate `dofile(...)` + `require("duffle")` dance.
+---
+
+local M = {}
+
+-- Cache key for the repo root. Stored in `package.loaded` (process-global) so all 8 entry scripts + passes scripts share one resolution.
+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).
+---
+--- 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
+
+	local source = debug.getinfo(1, "S").source
+	-- Strip the leading `@` (Lua's dofile marker) and the trailing `/duffle_paths.lua` filename.
+	-- What remains is the directory containing this script, i.e. `<repo>/scripts/`.
+	local scripts_dir = source and source:match("^@?(.*)[/\\]duffle_paths%.lua$")
+	if not scripts_dir then return nil end
+
+	-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
+	local root = scripts_dir:gsub("scripts[\\/]?$", "")
+	root = root:gsub("\\", "/")
+	if root == "" then root = "./" end
+	if not root:match("/$") then root = root .. "/" end
+	package.loaded[CACHE_KEY] = root
+	return root
+end
+
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
+--- `package.cpath` (for `lpeg.dll`).
+---
+--- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$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).
+function M.setup()
+	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.
+		os.exit(2)
+	end
+
+	local scripts_dir = repo_root .. "scripts/"
+	local passes_dir  = repo_root .. "scripts/passes/"
+	package.path = scripts_dir .. "?.lua;"
+		.. scripts_dir .. "?/init.lua;"
+		.. passes_dir  .. "?.lua;"
+		.. passes_dir  .. "?/init.lua;"
+		.. package.path
+
+	-- lpeg: built by `update_deps.ps1` to `toolchain/lpeg/lpeg.dll`.
+	-- lfs: compiled from pcsx-redux's vendored luafilesystem source to `toolchain/lfs/lfs.dll`.
+	-- Wire both directories into cpath so `require("lpeg")` and `require("lfs")` resolve.
+	local lpeg_dir = repo_root .. "toolchain/lpeg/"
+	local lfs_dir  = repo_root .. "toolchain/lfs/"
+	package.cpath = lpeg_dir .. "?.dll;"
+		.. lfs_dir  .. "?.dll;"
+		.. package.cpath
+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.
+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/gdb/gdb_tape_atoms.gdb

@@ -0,0 +1,105 @@
+# scripts/gdb/gdb_tape_atoms.gdb
+#
+# 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/gdb_tape_atoms_runtime.gdb`.
+# Sourcing that file RE-DEFINES the commands with real implementations.
+#
+# 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. ??
+
+define tape_atoms
+  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.
+  STUB state: runtime file not sourced. Run build_psyq.ps1 to regenerate.
+end
+
+define break_atom
+  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/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/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/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/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.
+end
+
+define show_c2
+  printf "C2[ 0]   0x%08x\n", $c2_data[0]
+  printf "C2[ 7]   0x%08x  [otz]\n", $c2_data[7]
+  printf "C2[12]   0x%08x  [sxy0]\n", $c2_data[12]
+  printf "C2[13]   0x%08x  [sxy1]\n", $c2_data[13]
+  printf "C2[14]   0x%08x  [sxy2]\n", $c2_data[14]
+  printf "C2[24]   0x%08x  [mac0]\n", $c2_data[24]
+  printf "...\n"
+   echo "(STUB state: only 7 representative regs shown. Run build_psyq.ps1 for full dump.)"
+end
+document show_c2
+  Pretty-print all 32 C2 data registers as hex + named alias. STUB state (7 reg subset).
+end
+
+define show_c2ctl
+  printf "C2CTL[ 0]  0x%08x\n", $c2_control[0]
+  printf "...\n"
+  echo "(STUB state: only 1 reg shown. Run build_psyq.ps1 for full dump.)"
+end
+document show_c2ctl
+  Pretty-print all 32 C2 control registers. STUB state (1 reg subset).
+end
+
+define wave_ctx
+  printf "$t4 = R_FaceCursor     0x%08x\n", $t4
+  printf "$t5 = R_VertBase       0x%08x\n", $t5
+  printf "$t6 = R_OtBase         0x%08x\n", $t6
+  printf "$t7 = R_PrimCursor     0x%08x\n", $t7
+end
+document wave_ctx
+  Pretty-print the 4 wave-context GPRs ($t4..$t7). (wave_ctx works in stub state too.)
+end
+
+
+# ?? 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/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.
+set confirm off
+
+# Source the runtime file (re-defines commands with real impls + data).
+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

@@ -0,0 +1,94 @@
+# 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).
+#
+# usage:
+#   .\scripts\launch_pcsx_debug.ps1
+#   .\scripts\launch_pcsx_debug.ps1 -ExePath build\hello_gte.ps-exe
+#   .\scripts\launch_pcsx_debug.ps1 -HelperZip scripts\pcsx_debug_helper.zip
+#
+# After launch:
+#   - gdb: target remote localhost:3333
+#   - web: curl http://localhost:8080/api/v1/lua/gte
+#
+# Companion: scripts/debug_psyq.ps1 (bare launch — no .ps-exe, no helper).
+
+[CmdletBinding()]
+param(
+    [string]$PcsxPath = (Join-Path $PSScriptRoot '..\toolchain\pcsx-redux\vsprojects\x64\Release\pcsx-redux.exe'),
+    [string]$ExePath  = (Join-Path $PSScriptRoot '..\build\hello_gte.ps-exe'),
+    [string]$HelperZip = (Join-Path $PSScriptRoot 'pcsx_debug_helper.zip'),
+    [int]   $GdbPort  = 3333,
+    [int]   $WebPort  = 8080
+)
+
+$ErrorActionPreference = 'Stop'
+
+# ── Pre-checks ──
+foreach ($p in @($PcsxPath, $ExePath, $HelperZip)) {
+    if (-not (Test-Path $p)) {
+        Write-Error "Missing: $p"
+        exit 1
+    }
+}
+
+# Kill any existing pcsx-redux so the archive file isn't locked.
+Get-Process pcsx-redux -ErrorAction SilentlyContinue | Stop-Process -Force
+Start-Sleep -Seconds 2
+
+# ── Launch ──
+$absExe  = [System.IO.Path]::GetFullPath($ExePath)
+$absZip  = [System.IO.Path]::GetFullPath($HelperZip)
+
+$args = @(
+    '-gdb', '-run'
+    '-loadexe', "`"$absExe`""
+    '-archive', "`"$absZip`""
+)
+
+Write-Host "Launching pcsx-redux..." -ForegroundColor Cyan
+Write-Host "  ps-exe    : $absExe"
+Write-Host "  helper zip: $absZip"
+Write-Host "  gdb       : localhost:$GdbPort"
+Write-Host "  web       : localhost:$WebPort/api/v1/lua/gte"
+Write-Host ""
+
+Start-Process -FilePath $PcsxPath -ArgumentList $args | Out-Null
+
+# ── Wait for both endpoints to come up ──
+$deadline = (Get-Date).AddSeconds(15)
+while ((Get-Date) -lt $deadline) {
+    $gdbUp = $false
+    $webUp = $false
+    try {
+        $tcp = New-Object System.Net.Sockets.TcpClient
+        $tcp.BeginConnect('localhost', $GdbPort, $null, $null) | Out-Null
+        Start-Sleep -Milliseconds 100
+        $gdbUp = $tcp.Connected
+        $tcp.Close()
+    } catch { $gdbUp = $false }
+    try {
+        $r = Invoke-WebRequest -Uri "http://localhost:$WebPort/" -UseBasicParsing -TimeoutSec 1 -ErrorAction SilentlyContinue
+        $webUp = $r.StatusCode -ne 0
+    } catch { $webUp = $false }
+    if ($gdbUp -and $webUp) { break }
+    Start-Sleep -Milliseconds 500
+}
+
+# ── Smoke-test the gte handler ──
+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 {
+    Write-Warning "GTE handler NOT responding: $_"
+    Write-Host "Check the pcsx-redux Lua Console for debug cli messages." -ForegroundColor Yellow
+}
+
+Write-Host ""
+Write-Host "pcsx-redux running. PIDs:" -ForegroundColor Cyan
+Get-Process pcsx-redux | Select-Object Id, ProcessName | Format-Table

scripts/passes/annotation.lua

@@ -0,0 +1,639 @@
+--- passes/annotation.lua — Atom-annotation DSL validator.
+---
+--- Validates `MipsAtom_(name) atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...)) { ... }` declarations in source files.
+--- Also reads `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`).
+---
+--- `duffle.scan_source()` scans each source once upstream; `ps1_meta.lua` stores that result in `src.scan`.
+---
+--- 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.
+
+-- 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")
+
+-- 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
+--   * pipe_ctx.type_name_registry      — for atom_dbg_reg_default(<T>, ...) and atom_reg_types(<T>, ...) type-identity checks
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @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[]
+--- @field metadata_path      string
+--- @field shared             table
+--- @field shared.word_counts table<string, integer>
+--- @field out_root           string
+--- @field project_root       string
+--- @field upstream           table<string, table>
+--- @field flags              table
+--- @field verbose            boolean
+
+--- @class PassResult
+--- @field outputs  table[]
+--- @field errors   table[]
+--- @field warnings table[]
+
+--- @class AtomAnnotation
+--- @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)
+
+--- @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 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 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
+
+--- @class Findings
+--- @field errors   Finding[]
+--- @field warnings Finding[]
+--- @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 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) 
+
+--- @class AnnotatedResult
+--- @field atoms    AtomEntry[]
+--- @field annots   AtomAnnotation[]
+--- @field macros   MacroEntry[]
+--- @field binds    BindsEntry[]
+--- @field errors   Finding[]
+--- @field warnings Finding[]
+--- @field info     Finding[]
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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.
+--- @param a        AtomAnnotation
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_decl_exists(a, pipe_ctx, findings)
+	if not pipe_ctx.atom_index[a.name] then
+		findings.errors[#findings.errors + 1] = {
+			line = a.line,
+			msg  = string.format("annotation for '%s' has no matching MipsAtom_(%s) { ... }", a.name, a.name),
+		}
+	end
+end
+
+--- 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)
+	for name, n in pairs(pipe_ctx.annot_counts) do
+		if n > 1 then
+			findings.errors[#findings.errors + 1] = {
+				line = pipe_ctx.atom_index[name] and pipe_ctx.atom_index[name].line or 0,
+				msg  = string.format("MipsAtom_(%s) has %d annotations (expected at most 1)", name, n),
+			}
+		end
+	end
+end
+
+--- Check: BIND atoms must reference a real Binds_* struct.
+--- I keep this as a warning so the annotation pass can report the common test-fixture case; `check_abi_handoff` in static analysis supplies the build-stopping error.
+--- @param a        AtomAnnotation
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_binds_struct_exists(a, pipe_ctx, findings)
+	if not a.binds                   then return end
+	if pipe_ctx.binds_index[a.binds] then return end
+	findings.warnings[#findings.warnings + 1] = {
+		line = a.line,
+		msg  = string.format("'%s' binds '%s' but no Struct_(%s) { ... } " 
+			.. "declaration found (also flagged as an error by check_abi_handoff in the static-analysis pass)"
+			, a.name, a.binds, a.binds),
+	}
+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> -- Shared word-count table (from ctx.shared.word_counts)
+--- @param findings Findings
+local function check_macro_word_drift(m, wc, findings)
+	local  declared = wc[m.name]
+	if not declared then
+		findings.errors[#findings.errors + 1] = {
+			line = m.line,
+			msg  = string.format("TAPE_WORDS(%s, %d) but '%s' is not in metadata.h", m.name, m.words, m.name),
+		}
+		return
+	end
+	if declared ~= m.words then
+		findings.errors[#findings.errors + 1] = {
+			line = m.line,
+			msg  = string.format("DRIFT: TAPE_WORDS(%s, %d) but metadata.h declares WORD_COUNT(%s, %d)", m.name, m.words, m.name, declared),
+		}
+		return
+	end
+	findings.info[#findings.info + 1] = {
+		line = m.line,
+		msg  = string.format("OK: %s = %d words", m.name, m.words),
+	}
+end
+
+--- Check: atom_dbg_reg_default(R_X, <type>) targets an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
+--- Pointer depth remains bounded to 0 or 1, and duplicate defaults remain errors.
+--- @param _src     SourceFile -- unused (kept for the per_source shape)
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_semantic_reg_defaults(_src, pipe_ctx, findings)
+	-- Detect duplicate defaults using the ordered occurrence list (the out.types hash only retains the last declaration).
+	local seen_first_line = {}
+	for _, occ in ipairs(pipe_ctx.type_occurrences or {}) do
+		if seen_first_line[occ.reg] == nil then
+			seen_first_line[occ.reg] = occ.source_line
+		else
+			findings.errors[#findings.errors + 1] = {
+				line = occ.source_line,
+				msg  = string.format(
+					"duplicate atom_dbg_reg_default for %q at line %d (first declared at line %d); one default per register",
+					occ.reg, occ.source_line, seen_first_line[occ.reg]),
+			}
+		end
+	end
+	local reg_registry  = pipe_ctx.register_alias_registry or {}
+	local type_registry = pipe_ctx.type_name_registry      or {}
+	for reg, def in pairs(pipe_ctx.types or {}) do
+		if not reg_registry[reg] then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d references unknown register %q (not in register_alias_registry)",
+					def.source_line, reg),
+			}
+		end
+		if def.pointer_depth == nil or def.pointer_depth < 0 or def.pointer_depth > 1 then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d for %q has unsupported pointer depth %d (expected 0 or 1)",
+					def.source_line, reg, def.pointer_depth or -1),
+			}
+		end
+		if not def.type_name or not type_registry[def.type_name] then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d for %q uses unknown type %q (not in type_name_registry)",
+					def.source_line, reg, tostring(def.type_name)),
+			}
+		end
+	end
+end
+
+--- Check: atom_reg_types(R_X, <type>) entries target an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
+--- A bare `atom_reg` marker opts the `R_<n>` alias into GPR identity; references to R_T0..R_T3 require the same explicit marker.
+--- @param _src     SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_reg_types(_src, pipe_ctx, findings)
+	local reg_registry  = pipe_ctx.register_alias_registry or {}
+	local type_registry = pipe_ctx.type_name_registry      or {}
+	for _, ai in ipairs(pipe_ctx.atom_infos_list or {}) do
+		if ai.reg_type_overrides then
+			for reg, ov in pairs(ai.reg_type_overrides) do
+				if not reg_registry[reg] then
+					findings.errors[#findings.errors + 1] = {
+						line = ai.info_line,
+						msg  = string.format(
+							"atom '%s' has atom_reg_types for %q; compute-register types are restricted to opt-in aliases (%q not in register_alias_registry)",
+							ai.atom_name, reg, reg),
+					}
+				end
+				if not ov.type_name or not type_registry[ov.type_name] then
+					findings.errors[#findings.errors + 1] = {
+						line = ai.info_line,
+						msg  = string.format(
+							"atom '%s' atom_reg_types for %q uses unknown compute type %q (not in type_name_registry)",
+							ai.atom_name, reg, tostring(ov.type_name)),
+					}
+				end
+			end
+		end
+	end
+end
+
+--- Check: atom_view(Binds_X) entries reference a Binds_* struct with at least one field.
+--- @param _src     SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_view_layout(_src, pipe_ctx, findings)
+	for atom_name, view in pairs(pipe_ctx.atom_views or {}) do
+		if not view.binds_name then
+			-- The atom had atom_reg_types but no atom_view; no layout check needed.
+		else
+			local bs = pipe_ctx.binds_index[view.binds_name]
+			if not bs then
+				findings.errors[#findings.errors + 1] = {
+					line = view.info_line,
+					msg  = string.format(
+						"atom '%s' has atom_view(%s) but no Struct_(%s) { ... } declaration was found",
+						atom_name, view.binds_name, view.binds_name),
+				}
+			elseif not bs.fields or #bs.fields == 0 then
+				findings.errors[#findings.errors + 1] = {
+					line = bs.line,
+					msg  = string.format(
+						"atom '%s' has atom_view(%s) but that struct declares zero typed fields",
+						atom_name, view.binds_name),
+				}
+			end
+		end
+	end
+end
+
+--- Check: Binds_* structs require unique field names because atom_view uses those names for typed-field lookup in gdb.
+--- @param _src SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
+	for _, bs in ipairs(pipe_ctx.binds_list or {}) do
+		local seen = {}
+		for _, f in ipairs(bs.fields or {}) do
+			seen[f.name] = (seen[f.name] or 0) + 1
+		end
+		for name, count in pairs(seen) do
+			if count > 1 then
+				findings.errors[#findings.errors + 1] = {
+					line = bs.line,
+					msg  = string.format(
+						"%s has duplicate field name %q (count %d); the typed-view contract requires unique field names",
+						bs.name, name, count),
+				}
+			end
+		end
+	end
+end
+
+-- 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`)
+---   2. marker_kind == "atom_dbg_skip" AND has_parens -> parenthesized form (the marker is bare-only)
+---   3. args  ~= ""                 -> takes no arguments
+---   4. superseded_by_marker_line   -> duplicate marker (cite superseding line)
+---   5. pending + no target_kind    -> dangling (no following declaration)
+---   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; 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
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d must be bare; the parenthesized form is no longer accepted (use `atom_dbg_skip MipsAtom_(name) { ... }`)",
+				kind, line),
+		}
+		return
+	end
+
+	if marker.args ~= nil and marker.args ~= "" then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d takes no arguments; found %q", kind, line, marker.args),
+		}
+		return
+	end
+
+	if marker.superseded_by_marker_line then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("duplicate %s marker at line %d; superseded by another %s marker at line %d"
+				, kind, line, kind, marker.superseded_by_marker_line),
+		}
+		return
+	end
+
+	if marker.pending and not marker.target_kind then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("dangling %s marker at line %d: no following MipsAtom_/MipsAtomComp_/MipsAtomComp_Proc_ declaration"
+				, kind, line),
+		}
+		return
+	end
+
+	if marker.target_kind
+		and marker.target_kind ~= "atom"
+		and marker.target_kind ~= "comp_bare"
+		and marker.target_kind ~= "comp_proc" then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d must precede MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_; found an unrelated declaration"
+				, kind, line),
+		}
+	end
+end
+
+--- Warn when a source references an unregistered alias.
+--- 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
+--- @param findings Findings
+local function check_wave_context_migration(_src, pipe_ctx, findings)
+	if not (pipe_ctx.types and next(pipe_ctx.types)) then return end
+	if not (pipe_ctx.atom_infos_list) then return end
+	local reg_registry = pipe_ctx.register_alias_registry or {}
+	for _, ai in ipairs(pipe_ctx.atom_infos_list) do
+		if ai.reg_type_overrides then
+			for reg, _ in pairs(ai.reg_type_overrides) do
+				if not reg_registry[reg] then
+					findings.warnings[#findings.warnings + 1] = {
+						line = 0,
+						msg  = "wave-context removed; opt in via #define atom_reg in mips.h "
+							.. "(every R_<alias> that should be visible to the annotation pass "
+							.. "must be enum-declared with the bare atom_reg marker)",
+					}
+					return
+				end
+			end
+		end
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- CHECK_RULES — data-driven check dispatch (the plex pattern)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Each rule entry picks one of four "shapes" of dispatch:
+--   per_annot(annot, pipe_ctx, findings)        -- runs once per AtomAnnotation
+--   post(pipe_ctx, findings)                    -- runs once after all per_annot calls complete (full-corpus aggregation)
+--   per_macro(macro, wc, findings)              -- runs once per TAPE_WORDS / _Pragma macro declaration
+--   per_skip_marker(marker, pipe_ctx, findings) -- runs once per src.scan.debug_skip_markers entry
+--
+-- Adding a new check = 1 row here + 1 function above. The `validate()` dispatch loop never needs editing.
+
+local CHECK_RULES = {
+	{ name = "atom_decl_exists",          per_annot       = check_atom_decl_exists          },
+	{ name = "binds_struct_exists",       per_annot       = check_binds_struct_exists       },
+	{ name = "unique_annotation",         post            = check_unique_annotation         },
+	{ name = "macro_word_drift",          per_macro       = check_macro_word_drift          },
+	{ name = "skip_marker_validation",    per_skip_marker = check_skip_marker               },
+	{ name = "semantic_reg_defaults",     per_source      = check_semantic_reg_defaults     },
+	{ name = "atom_reg_types",            per_source      = check_atom_reg_types            },
+	{ name = "atom_view_layout",          per_source      = check_atom_view_layout          },
+	{ name = "binds_no_duplicate_fields", per_source      = check_binds_no_duplicate_fields },
+	{ name = "wave_context_migration",    per_source      = check_wave_context_migration    },
+}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Validation
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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 annot_counts = {}
+	for _, info in ipairs(corpus.atom_infos or {}) 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 {},
+	}
+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.
+--- @return AnnotatedResult
+local function validate(ctx, src, corpus_pipe_ctx)
+	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
+	local scan = src.scan
+
+	-- 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 }
+		end
+	end
+
+	-- Project the pre-scanned atom_infos to AtomAnnotation shape.
+	local annots = {}
+	for _, info in ipairs(scan.atom_infos) do
+		annots[#annots + 1] = {
+			line   = info.info_line,
+			macro  = "atom_info",
+			name   = info.atom_name,
+			kind   = "info",
+			binds  = info.binds,
+			reads  = info.reads or {},
+			writes = info.writes or {},
+			errors = info.errors,
+		}
+	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 pipe_ctx = {
+		atom_index               = {},
+		binds_index              = {},
+		annot_counts             = corpus_pipe_ctx.annot_counts,
+		types                    = scan.types or {},
+		type_occurrences         = scan.type_occurrences or {},
+		atom_views               = scan.atom_views or {},
+		seen_defaults            = seen_defaults,
+		atom_infos_list          = atom_infos_list,
+		binds_list               = scan.binds or {},
+		-- See the module ownership contract; these shared lookup tables come from corpus_pipe_ctx.
+		register_alias_registry  = corpus_pipe_ctx.register_alias_registry,
+		type_name_registry       = corpus_pipe_ctx.type_name_registry,
+	}
+	for _, a in ipairs(atoms)      do pipe_ctx.atom_index [a.name] = a end
+	for _, b in ipairs(scan.binds) do pipe_ctx.binds_index[b.name] = b end
+
+	-- Findings live in a single struct with three lists (errors / warnings / info).
+	-- Each check writes to the list appropriate for its severity.
+	local findings = { errors = {}, warnings = {}, info = {} }
+
+	-- Lift parse-time errors already recorded in scan_source's atom_info payload into this pass's findings list.
+	for _, a in ipairs(annots) do
+		if a.errors then
+			for _, msg in ipairs(a.errors) do
+				findings.errors[#findings.errors + 1] = {
+					line = a.line,
+					msg  = string.format("'%s': %s", a.name, msg),
+				}
+			end
+		end
+	end
+
+	-- 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
+	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
+
+	-- 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
+	end
+
+	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
+	local wc = 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
+	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
+
+	-- Information summary (always emitted).
+	findings.info[#findings.info + 1] = {
+		line = 0,
+		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)"
+			, #atoms, #annots, #scan.macros, #scan.binds),
+	}
+
+	return {
+		atoms    = atoms,
+		annots   = annots,
+		macros   = scan.macros,
+		binds    = scan.binds,
+		errors   = findings.errors,
+		warnings = findings.warnings,
+		info     = findings.info,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M.run — orchestrator entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class M
+
+local M = {}
+
+-- Expose `validate` for downstream passes (e.g. report.lua) that need to re-render the per-source results into a per-MODULE report.
+M.validate = validate
+
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	-- Build the shared pipe_ctx once for this run; every validate() call sees the same cross-source registries.
+	-- The corpus owns the canonical cross-source registries; per-source scans retain body / declaration ownership.
+	local corpus_pipe_ctx = build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared.corpus
+
+	-- Group `corpus.sources_by_dir` by module, validate every source in each bucket, and emit one errors.h per directory.
+	local by_dir = (corpus and corpus.sources_by_dir) or {}
+
+	for dir, dir_sources in pairs(by_dir) do
+		local dir_basename = dir:match("([^/\\]+)$") or dir
+		local dir_atoms    = 0
+		local dir_errors   = {}
+		local dir_warnings = {}
+		for _, src in ipairs(dir_sources) do
+			local result  = validate(ctx, src, corpus_pipe_ctx)
+			result.source = src.path                           -- tag for downstream rendering
+			dir_atoms = dir_atoms + #result.atoms
+			for _, e in ipairs(result.errors) do
+				dir_errors[#dir_errors + 1] = { line = e.line, msg = e.msg, source = src.path }
+				errors    [#errors     + 1] = { line = e.line, msg = e.msg }
+			end
+			for _, w in ipairs(result.warnings) do
+				dir_warnings[#dir_warnings + 1] = { line = w.line, msg = w.msg }
+				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
+			end
+		end
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/atoms_source_map.lua

@@ -0,0 +1,568 @@
+--- 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>`).
+---
+--- 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:
+---   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_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 forma (sourcemap.txt form):
+---   ```
+---   # FORMAT_VERSION 1
+---   # auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT
+---   ATOM <name> "<abs-source-path>" <total_words>
+---   WORD 0  LINE 49  TEXT load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+---   WORD 1  LINE 49  TEXT load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+---   ... (one WORD line per .word emitted by the atom body) ...
+---   ENDATOM
+---   ATOM <next-name> "<abs-source-path>" <total_words>
+---   ...
+---   ENDATOM
+---   ```
+--- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- 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 elf_dwarf      = require("elf_dwarf")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Format version emitted as the first line. Bump + add a migration test if the format changes; 
+-- the gdb runtime loader rejects mismatches (E2).
+local FORMAT_VERSION = 1
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class AtomSourceMapCtx
+--- @field shared             table   -- `ctx.shared`
+--- @field shared.corpus      table   -- source-order registry; single writer is build_ctx
+--- @field shared.word_counts table
+--- @field out_root           string  -- output root (e.g. "build/gen")
+--- @field flags              table   -- `ctx.flags`; reads `flags.gdb_runtime` + `flags.elf_path`
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Atom-path renderers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Join word boundaries (from `items`) to per-word call text + source lines (from `word_events`).
+--- @param atom table
+--- @return table[], integer
+local function canonical_word_entries(atom)
+	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
+	end
+
+	local entries = {}
+	for index, event in ipairs(events) do
+		local item = word_items[index] or {}
+		entries[#entries + 1] = {
+			pos        = event.i or (index - 1),
+			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,
+			invocation = (event.outermost_invocation_id
+				and paths.invocations
+				and paths.invocations[event.outermost_invocation_id]) or nil,
+		}
+	end
+	return entries, #events
+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
+--- @param atom table
+--- @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 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 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)
+		else
+			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[#lines + 1] = "ENDATOM"
+	return lines, total
+end
+
+--- Render the full provenance file content for one source.
+--- @param src table
+--- @param wc  table
+--- @return string
+local function render_provenance(src, wc)
+	local lines = {}
+	lines[#lines + 1] = "# FORMAT_VERSION 1"
+	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
+	lines[#lines + 1] = "# Per-.word provenance: maps each emitted .word to its call site (atom body"
+	lines[#lines + 1] = "# file:line) and, when the word was emitted by a `mac_X(...)` component invocation,"
+	lines[#lines + 1] = "# the component's definition file:line + the per-word BODY line. Used by"
+	lines[#lines + 1] = "# dwarf_injection to synthesize DW_TAG_inlined_subroutine instances + per-word"
+	lines[#lines + 1] = "# line program rows for native source-level step into component bodies."
+
+	local function append(atom)
+		local stanza = emit_provenance_stanza(src, atom, wc)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render one atom's stanza for the sourcemap.txt form (ATOM header line, N WORD lines, ENDATOM marker).
+--- Returns (lines, total_words).
+--- @param src    table
+--- @param atom   table
+--- @param wc     table
+--- @return string[], integer
+local function emit_atom_stanza(src, atom)
+	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
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+	end
+	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
+	lines[#lines + 1] = "ENDATOM"
+	return lines, total
+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
+--- @return string
+local function render_source_map(src)
+	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"
+
+	local function append(atom)
+		local stanza = emit_atom_stanza(src, atom)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- gdb-runtime emission (post-link, addresses via nm)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Escape a string for embedding in a gdb `set $var = "..."` literal.
+--- gdb uses C-style escaping; we escape `\` and `"` (newlines were flattened earlier).
+--- @param s string
+--- @return string
+local function gdb_escape(s)
+	return (s:gsub("\\", "\\\\"):gsub('"', '\\"'))
+end
+
+--- Build the list of atoms with addresses + word entries. Shared helper for the gdb-runtime file emission.
+--- @param ctx PassCtx
+--- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
+local function build_atom_table(ctx)
+	local addrs   = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local corpus  = ctx.shared and ctx.shared.corpus
+	local matched = {}
+
+	for _, src in ipairs(corpus.source_order or {}) do
+		local file_base = src.path:match("([^/\\\\]+)$") 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]
+			if not info then return end
+			local entries, total = canonical_word_entries(atom)
+			matched[#matched + 1] = {
+				name       = name,
+				src_path   = src.path,
+				file_base  = file_base,
+				addr       = info[1],
+				size_bytes = info[2],
+				words      = total,
+				entries    = entries,
+			}
+		end
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do append(atom) end
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do append(atom) end
+	end
+
+	-- Deterministic order: sort by address (matches `nm` output ordering).
+	table.sort(matched, function(a, b) return a.addr < b.addr end)
+	for i, a in ipairs(matched) do a.idx = i - 1 end
+	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.
+---
+--- 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.
+--- @param lines   table  -- output line buffer (mutated in place)
+--- @param matched table  -- list of atom records from `build_atom_table`
+local function append_gdb_commands(lines, matched)
+	-- ── tape_atoms ──
+	-- Hardcoded one printf per atom. No loop.
+	lines[#lines + 1] = "define tape_atoms"
+	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',
+			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] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── break_atom (generic) + per-atom break_atom_X ──
+	lines[#lines + 1] = "define break_atom"
+	lines[#lines + 1] = '  echo "Usage: break_atom_<exact_name>  (pick from the list below)"'
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format('  printf "  break_atom_%%-32s\\n", $__atom_name_%d', a.idx)
+	end
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document break_atom"
+	lines[#lines + 1] = "  Generic help: lists the per-atom break_atom_<name> commands."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format("define break_atom_%s", a.name)
+		lines[#lines + 1] = string.format("  break *$__atom_addr_%d", a.idx)
+		lines[#lines + 1] = string.format('  printf "  Breakpoint set at code_%s (0x%%08x)\\n", $__atom_addr_%d', a.name, a.idx)
+		lines[#lines + 1] = "end"
+		lines[#lines + 1] = string.format("document break_atom_%s", a.name)
+		lines[#lines + 1] = string.format("  Set a breakpoint at code_%s.", a.name)
+		lines[#lines + 1] = "end"
+		lines[#lines + 1] = ""
+	end
+
+	-- ── step_atom / next_atom ──
+	-- Hardcoded one tbreak per atom. No loop.
+	lines[#lines + 1] = "define step_atom"
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format("  tbreak *$__atom_addr_%d", a.idx)
+	end
+	lines[#lines + 1] = "  continue"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document step_atom"
+	lines[#lines + 1] = "  Set one-shot BPs at every atom + continue. Stops at the next atom boundary."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	lines[#lines + 1] = "define next_atom"
+	lines[#lines + 1] = "  step_atom"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document next_atom"
+	lines[#lines + 1] = "  Alias for step_atom."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── where_in_atom ──
+	-- Hardcoded one outer-if per atom; inside, one inner-if per WORD entry.
+	lines[#lines + 1] = "define where_in_atom"
+	lines[#lines + 1] = "  set $__pc = (unsigned int)$pc"
+	lines[#lines + 1] = "  set $__matched = 0"
+	for _, a in ipairs(matched) do
+		-- 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] = '    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)
+		-- One inner-if per WORD entry. Each word's line + text hardcoded.
+		for _, we in ipairs(a.entries) do
+			lines[#lines + 1] = string.format("    if $__word == %d", we.pos)
+			-- Escape TEXT for printf format string.
+			local escaped_text = we.text:gsub("%%", "%%%%"):gsub('"', '\\"')
+			lines[#lines + 1] = string.format('      printf "source:  %%s:%%d  %%s\\n", $__atom_file_%d, %d, "%s"', a.idx, we.line, escaped_text)
+			lines[#lines + 1] = "    end"
+		end
+		-- Fallback for words beyond the source map (shouldn't happen if nm matches).
+		local max_word = 0
+		if #a.entries > 0 then max_word = a.entries[#a.entries].pos end
+		lines[#lines + 1] = string.format('    if $__word > %d', max_word)
+		lines[#lines + 1] = '      printf "source:  (no source-map entry for word %%d; map may be stale)\\n", $__word'
+		lines[#lines + 1] = "    end"
+		lines[#lines + 1] = "    set $__matched = 1"
+		lines[#lines + 1] = "  end"
+	end
+	lines[#lines + 1] = "  if !$__matched"
+	lines[#lines + 1] = '    echo PC is not inside any known atom (in .text or unmapped region).'
+	lines[#lines + 1] = "  end"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document where_in_atom"
+	lines[#lines + 1] = "  Report current atom name, .rodata addr, word offset, and source line."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── stepi_inside_atom ──
+	-- Hardcoded one if-containment-check per atom (no loop).
+	-- Precompute end_addr in Lua so we don't ask gdb to evaluate `addr + words*4` inside the if condition
+	-- (gdb 12.1's expression evaluator chokes on the `*` and emits a misleading 'function malloc' error in some gdb builds).
+	lines[#lines + 1] = "define stepi_inside_atom"
+	lines[#lines + 1] = "  set $__in_atom = 0"
+	lines[#lines + 1] = "  set $__did_step = 0"
+	lines[#lines + 1] = "  set $__pc = (unsigned int)$pc"
+	for _, a in ipairs(matched) do
+		-- Precompute end_addr in the convenience var (single expression gdb handles).
+		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] = "    set $__in_atom = 1"
+		lines[#lines + 1] = "    stepi"
+		lines[#lines + 1] = "    set $__did_step = 1"
+		lines[#lines + 1] = "  end"
+	end
+	lines[#lines + 1] = "  if !$__did_step"
+	lines[#lines + 1] = '    echo [gdb_tape_atoms] stepi_inside_atom: PC is not inside any atom; refusing to step.'
+	lines[#lines + 1] = "  end"
+	lines[#lines + 1] = "  where_in_atom"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document stepi_inside_atom"
+	lines[#lines + 1] = "  One MIPS-instruction step, then where_in_atom. The step-and-see-source-line workflow."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── wave_ctx ──
+	lines[#lines + 1] = "define wave_ctx"
+	lines[#lines + 1] = '  printf "$t4 = R_FaceCursor     0x%08x\\n", $t4'
+	lines[#lines + 1] = '  printf "$t5 = R_VertBase       0x%08x\\n", $t5'
+	lines[#lines + 1] = '  printf "$t6 = R_OtBase         0x%08x\\n", $t6'
+	lines[#lines + 1] = '  printf "$t7 = R_PrimCursor     0x%08x\\n", $t7'
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document wave_ctx"
+	lines[#lines + 1] = "  Pretty-print the 4 wave-context GPRs ($t4=R_FaceCursor, $t5=R_VertBase, $t6=R_OtBase, $t7=R_PrimCursor). Requires target attached."
+	lines[#lines + 1] = "end"
+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 $var = ...` + `define ... end` blocks at gdb source-time.
+--- @param ctx PassCtx
+local function emit_gdb_runtime(ctx)
+	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
+	local  elf_path = ctx.flags.elf_path
+	if not elf_path or elf_path == "" then
+		io.stderr:write("[atoms_source_map] --gdb-runtime requires --elf <elf>\n")
+		return
+	end
+	if lfs.attributes(elf_path, "mode") ~= "file" then
+		io.stderr:write(string.format(
+			"[atoms_source_map] --gdb-runtime: ELF not found at %s\n", elf_path))
+		return
+	end
+
+	local matched = build_atom_table(ctx)
+	if #matched == 0 then
+		io.stderr:write("[atoms_source_map] --gdb-runtime: no atoms matched against nm symbols (stale scan?).\n")
+		return
+	end
+
+	local lines = {}
+	lines[#lines + 1] = "# Auto-generated by ps1_meta.lua (passes/atoms_source_map.lua)"
+	lines[#lines + 1] = "# DO NOT EDIT — re-run ps1_meta.lua --atoms-source-map --gdb-runtime to regenerate"
+	lines[#lines + 1] = "# Sourced by scripts/gdb/gdb_tape_atoms.gdb (the wrapper)."
+	lines[#lines + 1] = "# Pure gdb scripting — no Python, no Tcl, no Guile required."
+	lines[#lines + 1] = "# Commands are FULLY HARDCODED per-atom because gdb doesn't do nested"
+	lines[#lines + 1] = "# `$` substitution in var names (`$foo_$i` is one literal identifier)."
+	lines[#lines + 1] = "# Per-atom convenience vars ($__atom_name_<i> etc.) are set so gdb's"
+	lines[#lines + 1] = "# `printf` has valid expression args (gdb 12.1 quirks: literals in"
+	lines[#lines + 1] = "# printf args require an attached target; convenience-var args do not)."
+	lines[#lines + 1] = string.format("# %d atoms from ELF: %s", #matched, elf_path)
+	lines[#lines + 1] = ""
+
+	-- Format version + count + ELF path (the latter is referenced by the load-line).
+	lines[#lines + 1] = "set $__atom_format_version = " .. FORMAT_VERSION
+	lines[#lines + 1] = string.format("set $__atom_count = %d", #matched)
+	lines[#lines + 1] = string.format('set $__elf_path = "%s"', gdb_escape(elf_path))
+	lines[#lines + 1] = ""
+
+	-- Per-atom convenience vars (used as printf args; literals aren't accepted
+	-- without an attached target on gdb 12.1).
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format('set $__atom_name_%d = "%s"', a.idx, gdb_escape(a.name))
+		lines[#lines + 1] = string.format("set $__atom_addr_%d = 0x%x", a.idx, a.addr)
+		lines[#lines + 1] = string.format("set $__atom_words_%d = %d", a.idx, a.words)
+		lines[#lines + 1] = string.format('set $__atom_file_%d = "%s"', a.idx, gdb_escape(a.file_base))
+	end
+	lines[#lines + 1] = ""
+
+	-- The 9 commands (each `define ... end` overrides the wrapper's stub).
+	lines[#lines + 1] = "# ── 9 user commands (overrides wrapper stubs) ──"
+	append_gdb_commands(lines, matched)
+	lines[#lines + 1] = ""
+
+	-- Confirmation line for the source operator.
+	lines[#lines + 1] = 'printf "[gdb_tape_atoms] runtime loaded %d atoms from %s\\n", $__atom_count, $__elf_path'
+
+	local out_path
+	-- Move out of `<out_root>/gdb_tape_atoms_runtime.gdb` to `<out_root>/../gdb_tape_atoms_runtime.gdb` when the conventional `<out_root>` is `<build>/gen`
+	-- (any equivalent spelling — relative, absolute backslash, absolute forward-slash, trailing-separator variants).
+	-- This puts the gdb runtime alongside the ELF at `build/` rather than under the report subdir.
+	local function ends_with_gen_dir(p)
+		if type(p) ~= "string" then return false end
+		return p:match("[/\\]gen[/\\]?$") ~= 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))
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- 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
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+	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 `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`.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+		error("atoms_source_map.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	end
+
+	-- Word counts come from `corpus.word_counts` (populated by word_count_eval + components passes).
+	local wc = corpus.word_counts or {}
+	if not next(wc) then
+		warnings[#warnings + 1] = {
+			line = 0,
+			msg  = "atoms_source_map: corpus.word_counts is empty; the word-counts + components passes may not have populated it. Check the PASSES dep edges.",
+		}
+	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
+		emit_gdb_runtime(ctx)
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/components.lua

@@ -0,0 +1,786 @@
+--- passes/components.lua — Component-macro header generator.
+---
+--- 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_ Slice_MipsCode ac_X(...)` declaration via a backward walk.
+---
+--- 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
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- 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.
+-- 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")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Atom component declaration identifiers.
+local ATOM_COMP_PROC  = "MipsAtomComp_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
+local AC_PREFIX_LEN   = 3
+local MAC_PREFIX      = "mac_" -- prefix on generated macros; the rest is the atom name
+local MAC_PREFIX_LEN  = 4
+
+-- ASCII byte values used in tokenization.
+local BYTE_NEWLINE    = 10
+local BYTE_SLASH      = 47
+
+-- 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  -- 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 flags              table                  -- CLI flags
+--- @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
+
+--- @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
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Local helpers (file I/O + path normalization)
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Back-walk helpers (composed into the entry point below: find_function_args_for)
+--
+-- Only the function-args lookup for proc components occurs here.
+-- The preceding-comment walk occur in `scan_source.lua` — `a.declaration_comment` carries the resolved comment,
+-- 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.
+--- 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_ Slice_MipsCode 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 `Slice_MipsCode`
+--- (the function-decl keyword with possible qualifiers between).
+---
+--- @param source     string
+--- @param name       string
+--- @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
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Argument-name extraction
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Extract just the parameter NAMES from a function-args string (stripping type annotations). E.g.,
+---   `"U4 off, U4 code, U1 r, U1 g, U1 b"`  ->  `{"off", "code", "r", "g", "b"}`
+---   `"U4 *ptr"`                            ->  `{"ptr"}`
+---   `""`                                   ->  nil
+--- @param args_str string|nil
+--- @return string[]|nil
+local function extract_arg_names(args_str)
+	if not args_str or args_str == "" then return nil end
+	local names  = {}
+	local tokens = duffle.split_top_level_commas(args_str)
+	for _, tok in ipairs(tokens) do
+		local trimmed = duffle.trim(tok)
+		if trimmed ~= "" then
+			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
+			-- then walk back over the identifier chars (alnum + `_`).
+			local ident_end = #trimmed
+			while ident_end > 0 do
+				local ch = trimmed:sub(ident_end, ident_end)
+				if ch == " " or ch == "\t" or ch == "*" or ch == "]" or ch == "[" then
+					ident_end = ident_end - 1
+				else
+					break
+				end
+			end
+			local ident_start = ident_end
+			while ident_start > 0 do
+				local ch = trimmed:sub(ident_start, ident_start)
+				if duffle.is_alnum_byte(string.byte(ch)) or ch == "_" then
+					ident_start = ident_start - 1
+				else
+					break
+				end
+			end
+			ident_start = ident_start + 1
+			local name = trimmed:sub(ident_start, ident_end)
+			if name ~= "" then names[#names + 1] = name end
+		end
+	end
+	if #names == 0 then return nil end
+	return names
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Component projection (read from pre-scanned SourceScan)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
+--- Reads the scanner-owned `declaration_comment` (resolved by scan_source.lua, skipping backward across an associated bare `atom_dbg_skip` marker when present).
+--- Per-source backward lookups remain in place only for the function `args` of proc components.
+--- That lookup is unique to components.lua and stays separate from the declaration-comment walk.
+--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table instead of calling duffle.tokenize_body again.
+--- Carries the scanner-owned `debug_skip` flag forward so the generated projection can emit `/* atom_dbg_skip */`
+--- before the authored comment and so `update_canonical_components` can mirror the same field onto `corpus.components[name]`.
+--- @param source string  -- the full source text (needed for backward lookups)
+--- @param scan   table   -- SourceScan from duffle.scan_source
+--- @return Component[]
+local function project_components(source, scan)
+	local out = {}
+	for _, a in ipairs(scan.atoms) do
+		if a.kind == "comp_bare" or a.kind == "comp_proc" then
+			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.
+			local comment = a.declaration_comment or ""
+			out[#out + 1] = {
+				line        = a.line,
+				name        = a.name,
+				body        = a.body,
+				body_off    = a.body_off,
+				body_tokens = a.body_tokens,
+				args        = args,
+				comment     = comment,
+				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
+				debug_skip  = a.debug_skip == true,
+			}
+		end
+	end
+	return out
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Line-comment → block-comment conversion
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Convert `//` line comments to `/* */` block comments in a token.
+-- C macros use `\` line-continuations; a `//` comment before `\` would consume the continuation,
+-- breaking the macro. We convert `//` to `/* */` so the multi-line macro structure is preserved.
+--
+-- 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)
+	local result = s
+	local pos    = 1
+	local len    = #result
+	while pos <= len do
+		local is_double_slash = result:byte(pos) == BYTE_SLASH
+			and pos + 1 <= len and result:byte(pos + 1) == BYTE_SLASH
+		if not is_double_slash then
+			pos = pos + 1
+		else
+			-- Find end of line.
+			local eol = pos
+			while eol <= len and result:byte(eol) ~= BYTE_NEWLINE do
+				eol = eol + 1
+			end
+			local before  = result:sub(1, pos - 1)
+			local comment = result:sub(pos + 2, eol - 1)  -- skip the `//`
+			local after
+			if eol <= len and result:byte(eol) == BYTE_NEWLINE then
+				after = " */" .. result:sub(eol)  -- keep the newline
+			else
+				after = " */"
+			end
+			result = before .. "/*" .. comment .. after
+			pos = #before + 2 + #comment + 3  -- skip past converted comment
+		end
+	end
+	return result
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Word-count computation (memoized recursive lookup)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
+--- Returns the ident unchanged if it doesn't start with the prefix 
+--- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
+--- @param ident string|nil
+--- @return string|nil
+local function strip_mac_prefix(ident)
+	if not ident then return nil end
+	if ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+		return ident:sub(MAC_PREFIX_LEN + 1)
+	end
+	return ident
+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_`)
+--- @param comp_by_name table<string, Component>
+--- @param wc           table<string, integer>
+--- @param cache        table<string, integer>
+--- @return integer
+local function word_count_rec(name, comp_by_name, wc, cache)
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1  -- mark in-progress (cycle detection)
+	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 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
+				end
+			end
+		end
+	else
+		-- Not a known component: assume 1 word (regular instruction).
+		n = 1
+	end
+	cache[name] = n
+	return n
+end
+
+--- Compute word counts for every component in `components` in a single pass.
+--- The name-lookup table + memoization cache are built ONCE (per source) instead of per-component,
+--- so the cache survives across siblings and a component's recursive `mac_Y(...)` 
+--- references hit memoized values instead of re-walking the body.
+--- Cycle detection (A -> B -> A) is preserved via the in-progress `-1` sentinel in `cache`.
+--- @param components Component[]
+--- @param wc         table<string, integer>
+--- @return table<string, integer>  -- map of component name (without `mac_`) -> word count
+local function count_all_components(components, wc)
+	local comp_by_name = {}
+	for _, cc in ipairs(components) do comp_by_name[cc.name] = cc end
+	local cache        = {}
+	local counts       = {}
+	for _, c in ipairs(components) do
+		counts[c.name] = word_count_rec(c.name, comp_by_name, wc, cache)
+	end
+	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. Look up in INSTRUCTION_LATENCY; default 1.
+						n = n + (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 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 == "store_word" or ident == "store_half" or ident == "store_byte" then
+					if trimmed:find("R_PrimCursor", 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
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Split a (possibly multi-line) comment into per-line entries.
+--- Hand-rolled (no regex patterns used).
+--- @param s string
+--- @return string[]
+local function split_comment_lines(s)
+	local out   = {}
+	local pos   = 1
+	local s_len = #s
+	while pos <= s_len do
+		local nl = s:find("\n", pos, true)
+		if not nl then
+			out[#out + 1] = s:sub(pos)
+			break
+		end
+		out[#out + 1] = s:sub(pos, nl - 1)
+		pos = nl + 1
+	end
+	return out
+end
+
+--- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
+--- @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, ", ")
+	end
+	return "..."
+end
+
+--- Strip the trailing `" \"` (space + backslash) line continuation from the last body line.
+--- The last 2 chars are always that pair.
+local function strip_trailing_continuation(lines)
+	local last = lines[#lines]
+	if    last:sub(-2) == " \\" then
+		lines[#lines] = last:sub(1, -3)
+	end
+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.
+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
+	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] .. " \\"
+	end
+	strip_trailing_continuation(lines)
+end
+
+--- Build the list of lines for one component
+--- (signature comment, `#define mac_X(...)` line with backslash-continued tokens, then `WORD_COUNT(mac_X, N)` entry).
+--- For skipped components, a `/* atom_dbg_skip */` marker comment is emitted immediately before the authored comment block.
+--- The marker is a single line, the comment comes next, and the `#define` line follows. The `debug_skip` stamp is scanner-owned
+--- (`a.debug_skip == true` on the declaration record); the components pass projects it directly.
+--- @param c          Component
+--- @param components Component[]
+--- @param wc         table<string, integer>
+--- @return string[]  -- list of lines for this component
+local function build_component_lines(c, counts)
+	local lines = {}
+
+	-- 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.
+	if c.debug_skip then
+		lines[#lines + 1] = "/* atom_dbg_skip */"
+	end
+
+	if c.comment and c.comment ~= "" then
+		for _, line in ipairs(split_comment_lines(c.comment)) do
+			lines[#lines + 1] = line
+		end
+	end
+
+	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)
+	-- Direct lookup against the per-source precomputed `counts` table (built once by count_all_components).
+	local n   = counts[c.name]
+
+	if n > 0 then
+		emit_macro_body(lines, c, sig, tokens)
+	end
+
+	-- Emit the WORD_COUNT(mac_<X>, N) entry.
+	lines[#lines + 1] = "WORD_COUNT(mac_" .. c.name .. ", " .. n .. ")"
+	lines[#lines + 1] = ""
+
+	return lines
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-source emit logic
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- 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 dir     string       -- the absolute source directory
+--- @param sources SourceFile[] -- sources contributing to this directory (for the header comment)
+--- @return string[]
+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.
+		"#ifdef INTELLISENSE_DIRECTIVES",
+		"#pragma once",
+		"#endif",
+		"// Auto-generated by ps1_meta.lua — DO NOT EDIT",
+		source_blob,
+		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
+		"",
+		-- Self-contained: define WORD_COUNT if not already defined.
+		-- We use the same definition here so the auto-generated entries below expand
+		-- to compile-time constants whether the metadata file is included first or not.
+		"#ifndef WORD_COUNT",
+		"#define WORD_COUNT(name, count)  enum { words_##name = (count) };",
+		"#endif",
+		"",
+	}
+end
+
+--- 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  -- the absolute source directory
+--- @return string  -- the output directory
+--- @return string  -- the 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-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 dir        string                  -- the 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(dir)
+	local lines             = header_boilerplate(dir, sources)
+
+	for _, c in ipairs(components) do
+		for _, l in ipairs(build_component_lines(c, counts)) do
+			lines[#lines + 1] = l
+		end
+	end
+
+	local content = table.concat(lines, "\n") .. "\n"
+	duffle.ensure_dir(out_dir)
+	duffle.write_file_lf(out_path, content)
+	print(string.format("  -> %s", out_path))
+	return out_path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Pass entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- (internal) Extend `corpus.word_counts` with this source's component macros so offsets sees them without re-reading the file.
+--- First declaration wins: a later caller's count is dropped (the existing entry from the first source is preserved).
+--- @param corpus     table  -- the corpus
+--- @param components Component[]
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
+local function update_canonical_word_counts(corpus, components, counts)
+	local wc = corpus.word_counts
+	for _, c in ipairs(components) do
+		local key = "mac_" .. c.name
+		if wc[key] == nil then
+			wc[key] = counts[c.name]
+		end
+	end
+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
+
+--- (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`.
+--- No parallel skip map is built here; consumers that need the per-component skip state read `corpus.components[name].debug_skip` directly.
+--- 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[]
+--- @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,
+				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 first_kind = existing.kind or "comp_bare"
+				corpus.collisions[#corpus.collisions + 1] = {
+					kind              = "component",
+					name              = c.name,
+					first_site        = { path = existing.path, line = existing.line },
+					conflicting_site  = { path = rel_path,       line = c.line },
+					first_shape       = "kind=" .. first_kind,
+					conflicting_shape = "kind=" .. kind,
+				}
+			end
+		end
+	end
+end
+
+--- (internal) Populate `corpus.component_body_index` with this source's body index entries.
+--- First declaration wins; later declarations are dropped (no separate collision record: the components collision is already surfaced by `update_canonical_components`).
+--- The pass writes to `corpus.component_body_index` only (the corpus owns this projection).
+--- @param corpus     table  -- the corpus
+--- @param src        SourceFile
+--- @param components Component[]
+--- @param scan       table  -- the SourceScan payload (for line_of)
+local function update_canonical_component_body_index(corpus, src, components, scan)
+	local line_of = scan and scan.line_of
+	for _, c in ipairs(components) do
+		if corpus.component_body_index[c.name] == nil then
+			corpus.component_body_index[c.name] = {
+				body_tokens = c.body_tokens,
+				body_off    = c.body_off,
+				line_of     = line_of,
+				source      = src.path,
+				declaration = c.line,
+				kind        = c.kind,
+			}
+		end
+	end
+end
+
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	-- Corpus ownership gate.
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("components.run requires ctx.shared.corpus.", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("components.run requires ctx.shared.corpus.source_order.", 0)
+	end
+	if type(corpus.word_counts) ~= "table" then
+		error("components.run requires ctx.shared.corpus.word_counts; "
+			.. "word_count_eval.run must run before components.run "
+			.. "(see PASSES deps).", 0)
+	end
+
+	-- Projection ownership:
+	--   * `corpus.word_counts["mac_"..name]` — current component count
+	--   * `corpus.components[name]`         — bare-name component definition
+	--   * `corpus.component_body_index[name]` — body / line_of / source index
+	-- The pass writes to the corpus only; consumers read from the corpus directly.
+
+	-- 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, duffle.INSTRUCTION_LATENCY)
+			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 (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
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/emission_model.lua

@@ -0,0 +1,237 @@
+--- passes/emission_model.lua: Per-atom emission projection.
+---
+--- The `emission-model` pass owns `atom.paths`, the canonical per-atom mutable surface for atoms and raw atoms with bodies in `ctx.shared.corpus.source_order`.
+--- For each atom, the pass invokes `duffle.project_emission(body_text, component_index, word_counts, components)`.
+--- It stores the ordered `items` stream plus the dense `word_events` / `markers` / `invocations` views on `atom.paths`.
+---
+--- 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.
+---
+--- Source-order discipline:
+---   * `corpus.source_order` sets the source-record order.
+---   * Within each source, the pass visits `src.scan.atoms` and `src.scan.raw_atoms` in declaration order.
+---
+--- Per-atom projection fields on `atom.paths`:
+---   `tokens`, `line_in_body`, `items`, `word_events`, `markers`, `invocations`, `errors`, `warnings`.
+---   The construction walk appends `items` and derives each dense view from that ordered stream.
+---
+--- Component expansion and construction validation:
+---   * known `mac_X(...)` calls recursively expand component bodies;
+---   * invocation records retain monotonic IDs, parent IDs, immediate call text, and the immutable outermost root call text;
+---   * invocation construction stamps `debug_skip` from `corpus.components[name].debug_skip` at the construction site (no second pass, no source parse, no parallel lookup);
+---   * component cycles close balanced invocation boundaries and emit a `cycle` construction error at the recursive edge;
+---   * declared-vs-measured component word counts emit `count_mismatch` construction errors; opaque uncounted macros emit warnings.
+---
+--- `passes.scan_source` strips its private `_code_macros` / `_code_macro_bodies` tables before this pass runs.
+
+local M = {}
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Bootstrap: load `duffle_paths.lua` via debug.getinfo so the module works standalone (run as `luajit passes/emission_model.lua`) and when require'd from the orchestrator.
+-- ─────────────────────────────────────────────────────────────────────────
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Helpers
+-- ─────────────────────────────────────────────────────────────────────────
+
+-- 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.
+--   * INNER invocations (`inv.parent_id ~= 0`) receive physical `call_line` values directly from the COMPONENT's `line_of` in the walker.
+--     Recursive descent forwards that closure through `corpus.component_body_index[name].line_of`; those values arrive physical and remain unchanged.
+--
+-- After this function, every `inv.call_line` is physical. DWARF and provenance output read it directly.
+-- The word-event loop forwards the already-physical `outer_inv.call_line` into `we.call_line` for words inside an invocation.
+local function stamp_root_provenance(projection, atom_record, src, corpus)
+	local root_line_of   = src.scan and src.scan.line_of
+	assert(type(root_line_of) == "function"
+		, "emission_model: src.scan.line_of is required (canonical LineIndex closure over the source text) to stamp physical provenance")
+	assert(type(atom_record.body_off) == "number"
+		, "emission_model: atom_record.body_off (byte offset of the body's first byte in source) is required to derive `root_body_line`. The scanner must populate body_off for every atom record.")
+	-- `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 component_index = corpus.component_body_index or {}
+	local word_items      = {}
+
+	for _, item in ipairs(projection.items) do
+		if item.kind == "word" then word_items[#word_items + 1] = item end
+	end
+
+	-- Resolve one word's physical body line, where the byte containing that word appears in source.
+	--   * Component expansions carry `invocation_ids`; the component's full-file `line_of` leaves `item.line` physical.
+	--   * Raw tokens in the root atom body carry an empty `invocation_ids` list and a body-relative `item.line`; convert them here.
+	local function body_line_for(event, item)
+		local ids = event.invocation_ids or {}
+		-- The innermost open invocation identifies which line index the walker used.
+		-- A component `line_of` makes `item.line` physical; the atom's `body_text` line index makes it body-relative.
+		if ids and #ids > 0 then
+			local inner_id  = ids[#ids]
+			local inner_inv = inner_id and projection.invocations[inner_id]
+			if inner_inv then
+				local component = component_index[inner_inv.component_name]
+				if component and component.line_of then
+					-- Walker used `comp.line_of`, which is the source's physical LineIndex. item.line is already physical.
+					return item.line or 0
+				end
+			end
+		end
+		-- RAW root-body word: item.line is body-text's 1-based line number (the first content line is line 2 because line 1 is the trailing `\n` after `{`).
+		-- Convert body-text-relative → physical using `root_body_line + (item.line - 1)`.
+		return (root_body_line or 0) + (item.line or 1) - 1
+	end
+
+	-- Stamp the root source path onto invocation records whose `call_path` the walker left empty.
+	-- The walker passes `body_entry.source` to `emit_invoke_begin`; `M.project_emission` creates the root `body_entry` with source `""`, leaving its `call_path` empty.
+	-- This stamp gives every invocation a physical `call_path` matching `passes/atoms_source_map.lua`'s in-memory provenance projection.
+	local root_path = src.path or ""
+	for _, inv in ipairs(projection.invocations) do
+		if inv.call_path == nil or inv.call_path == "" then
+			inv.call_path = root_path
+		end
+	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`.
+	--   * 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
+			inv.call_line = (root_body_line or 0) + (inv.call_line or 1) - 1
+		end
+	end
+
+	-- Build `body_lines` for each invocation.
+	-- `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 bls = {}
+		for i = sw, ew do
+			local it = projection.items and projection.items[i]
+			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
+			end
+		end
+		inv.body_lines = bls
+	end
+
+	-- Resolve each `word_event`'s physical `body_line` and `call_line`.
+	-- For words inside an invocation, `we.call_line` identifies the OUTER atom source line containing the `mac_X(...)` token that triggered expansion.
+	-- The root-invocation conversion above makes every `inv.call_line` physical; forward it directly and use each raw word's `body_line` as the fallback.
+	for index, we in ipairs(projection.word_events) do
+		local item      = word_items[index] or {}
+		local body_line = body_line_for(we, item)
+		item.line    = body_line
+		we.body_line = body_line
+
+		local call_line = body_line
+		local outer_id  = we.outermost_invocation_id or 0
+		local outer_inv = projection.invocations[outer_id]
+		if outer_inv then
+			-- `outer_inv.call_line` is physical after the conversion loop above, so use it directly.
+			call_line = outer_inv.call_line
+		end
+		we.call_line = call_line
+
+		if we.def_path  == nil or we.def_path  == "" then we.def_path  = src.path         or "" end
+		if we.def_line  == nil or we.def_line  == 0  then we.def_line  = atom_record.line or 0  end
+		if we.call_path == nil or we.call_path == "" then we.call_path = src.path         or "" end
+	end
+end
+
+-- Project one atom record into `atom.paths`.
+-- Mutates the atom record in-place and returns the projection (for pass-level error/warning accumulation).
+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 {}
+	-- 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 paths = {
+		tokens       = atom_record.body_tokens or {},
+		line_in_body = duffle.build_body_line_index(body),
+		items        = proj.items,
+		word_events  = proj.word_events,
+		markers      = proj.markers,
+		invocations  = proj.invocations,
+		errors       = proj.errors,
+		warnings     = proj.warnings,
+	}
+	stamp_root_provenance(proj, atom_record, src, corpus)
+	atom_record.paths = paths
+	return proj
+end
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Run the emission-model pass.
+-- ─────────────────────────────────────────────────────────────────────────
+
+--- @param ctx PassCtx  -- { shared = { corpus = ... }, out_root, ... }
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx and ctx.shared and ctx.shared.corpus
+	if type(corpus)              ~= "table" then error("emission_model: ctx.shared.corpus is required (canonical projection)", 0) end
+	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.
+	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
+			return
+		end
+		local proj = project_atom(atom, src, corpus)
+		for _, e in ipairs(proj.errors) do
+			-- Preserve `kind` (cycle / count_mismatch / unbalanced) so readers dispatch on the diagnostic class and leave the message string as display text.
+			errors[#errors + 1] = {
+				kind   = e.kind,
+				line   = e.line,
+				msg    = e.msg,
+				source = e.source or src.path,
+			}
+		end
+		for _, w in ipairs(proj.warnings) do
+			warnings[#warnings + 1] = {
+				kind = w.kind,
+				line = w.line,
+				msg  = w.msg,
+			}
+		end
+	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.
+	-- 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 {}
+		for _, atom in ipairs(scan.atoms or {}) do
+			process_atom(atom, src)
+		end
+		for _, atom in ipairs(scan.raw_atoms or {}) do
+			process_atom(atom, src)
+		end
+	end
+
+	return {
+		outputs  = outputs,
+		errors   = errors,
+		warnings = warnings,
+	}
+end
+
+return M

scripts/passes/offsets.lua

@@ -0,0 +1,289 @@
+--- 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
+--- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
+--- `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.
+---
+--- The offset is `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding: branch_offset = relative_pc_in_words - 1).
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- 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.
+-- 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")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Offset macro/enum naming prefixes (the emitted header uses these).
+local OFFSET_MACRO_PREFIX = "_atom_offset_"
+local OFFSET_ENUM_PREFIX  = "atom_offset_"
+
+-- Column width for the `#define _atom_offset_F_T = N` alignment.
+local OFFSET_MACRO_COL = 44
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @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 -- Corpus projection
+--- @field shared.word_counts table
+--- @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
+
+--- @class BranchOffset
+--- @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
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Canonical marker projection
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- MARKER_PROJECTORS is the marker-kind data table.
+-- 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)
+		state.labels[marker.name] = marker.word_index
+	end,
+	offset = function(state, marker)
+		state.branches[#state.branches + 1] = {
+			tag                = marker.name,
+			target             = marker.target,
+			branch_word        = marker.word_index,
+			consuming_encoder  = marker.consuming_encoder,
+			consuming_arg_pos  = marker.consuming_arg_pos,
+		}
+	end,
+}
+
+--- Project canonical marker records into the two lookup tables used by the offset renderer.
+--- No source text, body text, or body token is inspected.
+--- @param markers table[] -- atom.paths.markers
+--- @return table<string, integer>, table[]
+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
+	end
+	return state.labels, state.branches
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Offset computation + header generation
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- 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[]
+--- @return BranchOffset[]
+local function compute_offsets(labels, branches)
+	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 .. ")")
+		end
+		local consuming = br.consuming_encoder
+		local offset
+		if consuming == "jump_reg" or consuming == "call_reg" or consuming == "jump_link" then
+			-- Register-form jumps have no offset field. `atom_offset` cannot be used here.
+			error("atom_offset cannot be used with " .. consuming
+				.. " (register-form jumps have no offset field); at word " .. br.branch_word)
+		end
+		-- 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.
+		offset = target - br.branch_word - 1
+		results[#results + 1] = {
+			target             = br.target,
+			tag                = br.tag,
+			branch_word        = br.branch_word,
+			offset             = offset,
+			consuming_encoder  = br.consuming_encoder,
+			consuming_arg_pos  = br.consuming_arg_pos,
+		}
+	end
+	return results
+end
+
+--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
+--- @param s string
+--- @param w integer
+--- @return string
+local function pad_right(s, w)
+	return s .. string.rep(" ", math.max(0, w - #s))
+end
+
+--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
+--- @param bo BranchOffset
+--- @return table
+local function make_offset_const(bo)
+	return {
+		macro_name = OFFSET_MACRO_PREFIX .. bo.tag .. "_" .. bo.target,
+		enum_name  = OFFSET_ENUM_PREFIX  .. bo.tag .. "_" .. bo.target,
+		value      = bo.offset,
+	}
+end
+
+--- (internal) Emit one atom's offset constants + enum into the lines buffer.
+--- @param add  fun(s: string)
+--- @param atom AtomData
+local function emit_atom_offsets(add, atom)
+	if #atom.offsets == 0 then return end
+	add("// --- atom: " .. atom.name .. " (" .. atom.total_words .. " words) ---")
+	add("")
+	local consts = {}
+	for _, r in ipairs(atom.offsets) do
+		consts[#consts + 1] = make_offset_const(r)
+	end
+	for _, c in ipairs(consts) do
+		add("#define " .. pad_right(c.macro_name, OFFSET_MACRO_COL) .. "  " .. c.value)
+	end
+	add("")
+	add("enum {")
+	for _, c in ipairs(consts) do
+		add("    " .. c.enum_name .. " = " .. c.macro_name .. ",")
+	end
+	add("};")
+	add("")
+end
+
+--- 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(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("// Directory: " .. dir:gsub("/", "\\") .. "\\")
+	for _, src in ipairs(sources) do
+		add("//   source: " .. src.path:gsub("/", "\\"))
+	end
+	add("#pragma once")
+	add("")
+	add("#pragma region " .. dir_basename)
+	add("")
+	add("")
+	for _, atom in ipairs(atoms_data) do
+		emit_atom_offsets(add, atom)
+	end
+	add("#pragma endregion " .. dir_basename)
+	add("")
+	return table.concat(lines, "\n") .. "\n"
+end
+
+local M = {}
+
+--- (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 dir     string       -- the absolute source directory
+--- @param sources SourceFile[] -- sources in this directory
+--- @return string|nil  -- the offsets_h path
+local function process_directory(ctx, dir, sources)
+	local atoms_data = {}
+
+	local function append_atom(atom)
+		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),
+		}
+	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 = dir .. "/gen/offsets.h"
+	duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file(out_path, generate_header(dir, sources, atoms_data))
+	return out_path
+end
+
+--- Run the offsets pass.
+--- 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)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx.shared and ctx.shared.corpus
+	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
+
+	-- 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)
+		if out_path then
+			outputs[#outputs + 1] = { offsets_h = out_path }
+		end
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/report.lua

@@ -0,0 +1,599 @@
+--- passes/report.lua — Per-MODULE annotation report renderer + project-wide summary writer.
+---
+--- Two output files per build:
+---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
+---   - `build/gen/annotation_validation.txt` — the project summary.
+---
+--- The annotation pass emits `errors.h` files per module and the canonical `corpus.sources_by_dir` projection groups sources by directory.
+--- This pass iterates the dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
+-- Bootstrap: 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.
+-- 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")
+
+-- Load the annotation pass so we can re-validate each source against the canonical corpus projection.
+-- The annotation pass exposes `M.validate`, which returns the per-source AnnotationResult (atoms / annots / macros / binds / errors / warnings)
+-- that the report pass renders into the per-module `<dir_basename>.annotations.txt` output.
+local annotation = dofile(_bootstrap_dir .. "annotation.lua")
+
+-- Load atoms_source_map for the `render_source_map` / `render_provenance` module functions (used by `render_module_atoms_md` to produce `<module>.atoms.md` without re-walking source tokens).
+-- The pass itself emits no per-source files anymore; we only consume the two pure renderers here.
+-- Defined BEFORE the renderer functions below so their upvalues resolve to this local (not the global `atoms_source_map`, which is nil).
+local atoms_source_map = dofile(_bootstrap_dir .. "atoms_source_map.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Section separators used in the rendered text reports.
+-- The thin rules are hand-tuned to align with the per-section content width; do not change without also checking the section renderers below.
+local RULE_THICK               = "========================================================"
+local SECTION_HEADER_ATOMS     = "── Atoms ────────────────────────────────────────────────"
+local SECTION_HEADER_ANNOTS    = "── Annotations ──────────────────────────────────────────"
+local SECTION_HEADER_BINDS     = "── Binds_* structs ──────────────────────────────────────"
+local SECTION_HEADER_MACROS    = "── Macro word-count declarations ─────────────────────────"
+local SECTION_HEADER_ERRORS    = "── Errors ──────────────────────────────────────────────"
+local SECTION_HEADER_WARNINGS  = "── Warnings ────────────────────────────────────────────"
+
+-- Lua pattern that captures the basename (last path segment) of a forward- or back-slash separated path.
+local BASENAME_PATTERN = "([^/\\]+)$"
+
+-- Debug flag name — set to truthy in `_G` to enable verbose logging.
+local DEBUG_FLAG = "_DEBUG_REPORT"
+
+-- Pass identifier for log messages.
+local PASS_NAME = "report"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @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 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 + per-pass stash
+--- @field verbose            boolean              -- If true, 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
+
+-- Shapes produced by `passes/annotation.lua`'s `M.validate()`.
+
+--- @class AtomEntry
+--- @field name string   -- Atom name (e.g. "cube_g4_face")
+--- @field line integer  -- Source line of the atom declaration
+
+--- @class AnnotEntry
+--- @field line    integer     -- Source line
+--- @field macro   string      -- Macro name (e.g. "atom_reads")
+--- @field name    string      -- Atom name (if a `name(...)` was given)
+--- @field kind    string      -- "atom_info" | "atom_bind" | ...
+--- @field binds   string|nil  -- Binds_X name if any
+--- @field reads   string[]    -- R_* names (read targets)
+--- @field writes  string[]    -- R_* names (write targets)
+--- @field error   string|nil  -- Error message if annotation was malformed
+
+--- @class BindsField
+--- @field name   string   -- Field name
+--- @field offset integer  -- Byte offset within the Binds_X struct
+
+--- @class BindsStruct
+--- @field name   string         -- Struct name (e.g. "Binds_Floor")
+--- @field line   integer        -- Source line of the typedef
+--- @field bytes  integer        -- Total byte size
+--- @field fields BindsField[]   -- The field list
+
+--- @class MacroEntry
+--- @field name  string   -- Macro name (e.g. "WORD_COUNT(my_macro, 4)")
+--- @field line  integer  -- Source line
+--- @field words integer  -- Declared word count
+
+--- @class Finding
+--- @field line integer  -- Source line
+--- @field msg  string   -- Finding message
+
+--- @class AnnotationResult
+--- @field source   string         -- Set by this pass; original source path
+--- @field atoms    AtomEntry[]    -- Atom declarations in this source
+--- @field annots   AnnotEntry[]   -- Annotation entries
+--- @field macros   MacroEntry[]   -- Macro word-count declarations
+--- @field binds    BindsStruct[]  -- Binds_* struct declarations
+--- @field errors   Finding[]      -- Errors from validation
+--- @field warnings Finding[]      -- Warnings from validation
+--- @field info     table          -- Info summary (not rendered here)
+
+--- @class ModuleEntry
+--- @field dir          string   -- Absolute directory path
+--- @field dir_basename string   -- Basename (e.g. "duffle", "gte_hello")
+--- @field atoms_count  integer  -- Pre-counted atoms for filtering
+
+--- @class ModuleReport
+--- @field dir       string             -- Module directory
+--- @field sources   SourceFile[]       -- Sources in this module
+--- @field results   AnnotationResult[] -- Per-source validate() results
+
+--- @class ProjectReport
+--- @field results AnnotationResult[]  -- All per-source results
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-MODULE annotation report (aggregated across all sources in a dir)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
+--- @param path string
+--- @return string
+local function source_basename(path)
+	return path:match(BASENAME_PATTERN) or path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Markdown renderers (consolidated-report-files refactor, 2026-07-26)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render the thin project-wide summary (`build/atom_meta_report.summary.md`).
+--- @param all_results { 
+--- 	module:string,
+--- 	atoms:integer, 
+--- 	annots:integer, 
+--- 	binds:integer,
+---		macros:integer,   
+---		findings:integer,
+---		errors:integer,
+--- 	warnings:integer,
+--- 	info:integer }[]
+--- @return string
+local function render_project_summary(all_results)
+	local lines = {
+		"# Project summary",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua).",
+		"",
+		"| module | atoms | annots | binds | macros | findings | errors | warnings | info |",
+		"|--------|-------|--------|-------|--------|----------|--------|----------|------|",
+	}
+	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0, findings = 0, errors = 0, warnings = 0, info = 0 }
+	for _, e in ipairs(all_results) do
+		lines[#lines + 1] = string.format("| %s | %d | %d | %d | %d | %d | %d | %d | %d |"
+			, e.module, e.atoms, e.annots, e.binds, e.macros, e.findings, e.errors, e.warnings, e.info)
+		totals.atoms    = totals.atoms    + e.atoms
+		totals.annots   = totals.annots   + e.annots
+		totals.binds    = totals.binds    + e.binds
+		totals.macros   = totals.macros   + e.macros
+		totals.findings = totals.findings + e.findings
+		totals.errors   = totals.errors   + e.errors
+		totals.warnings = totals.warnings + e.warnings
+		totals.info     = totals.info     + e.info
+	end
+	lines[#lines + 1] = string.format("| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |"
+		, totals.atoms, totals.annots, totals.binds, totals.macros, totals.findings, totals.errors, totals.warnings, totals.info)
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render the per-module verbose source-map markdown (`build/<module>.atoms.md`).
+--- Per-source sub-section, per-atom stanza with sourcemap + provenance rows.
+--- Pulls sourcemap + provenance from `atoms_source_map` (no second source walk).
+--- @param dir         string
+--- @param dir_sources SourceFile[]
+--- @param wc          table<string, integer>
+--- @return string
+local function render_module_atoms_md(dir, dir_sources, wc)
+	local dir_basename = source_basename(dir)
+	local lines        = {
+		"# " .. dir_basename .. " — atoms (verbose source map)",
+		"> Per-word call-site + provenance. Auto-generated.",
+		"",
+	}
+	for _, src in ipairs(dir_sources) do
+		local src_name = source_basename(src.path)
+		lines[#lines + 1] = "## " .. src_name
+		lines[#lines + 1] = ""
+		-- For each atom with a projection, render its sourcemap + provenance.
+		local atoms_list = {}
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
+		end
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
+		end
+		if #atoms_list == 0 then
+			lines[#lines + 1] = "_(no atom projections)_"
+			lines[#lines + 1] = ""
+		else
+			-- Per-source forward-slash path (same one `emit_atom_stanza` / `emit_provenance_stanza` would derive;
+			-- computed once per `## <source>` heading and reused by each atom's `WORD N CALL ...` field).
+			local rel_path = src.path:gsub("\\\\", "/")
+			for _, atom in ipairs(atoms_list) do
+				lines[#lines + 1] = string.format(
+					"### atom: %s (line %d, %d words)",
+					atom.name, atom.line or 0, #(atom.paths.items or {}))
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Sourcemap** — per-word call site:"
+				lines[#lines + 1] = "```"
+				-- Per-atom invariant: call the per-atom renderers, NOT the per-source ones.
+				-- The per-source renderers enumerate every atom in `src`;
+				-- calling them in a per-atom loop would repeat the whole source under every `### atom:` heading.
+				lines[#lines + 1] = atoms_source_map.render_atom_source_map(atom):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Provenance** — per-word definition + body:"
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = atoms_source_map.render_atom_provenance(atom, wc, rel_path):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+			end
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render the consolidated per-module markdown (`build/<module>.atom_meta_report.md`).
+--- Aggregates annotation + static-analysis content across all sources in `dir`.
+--- Annotations come from re-running `annotation.validate()` per source (the existing pattern);
+--- static-analysis comes from `corpus.static_analysis_results[dir_basename]` (populated by `static_analysis.lua` — no second corpus_pipe_ctx build).
+--- @param dir           string
+--- @param dir_sources   SourceFile[]
+--- @param annot_results AnnotationResult[]
+--- @param sa_results    table -- corpus.static_analysis_results[dir_basename]
+--- @return string
+local function render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+	local dir_basename = source_basename(dir)
+	local lines        = {
+		"# " .. dir_basename .. " — atom meta report",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua). Do not edit.",
+		"",
+	}
+	local function add(s) lines[#lines + 1] = s end
+
+	-- Module summary table.
+	local n_atoms        = 0
+	local n_annot        = 0
+	local n_binds        = 0
+	local n_macros       = 0
+	local n_bare, n_proc = 0, 0
+	for _, r in ipairs(annot_results) do
+		n_atoms  = n_atoms  + #r.atoms
+		n_annot  = n_annot  + #r.annots
+		n_binds  = n_binds  + #r.binds
+		n_macros = n_macros + #r.macros
+	end
+	for _, a in ipairs(sa_results.atoms or {}) do
+		if     a.kind == "comp_bare" then n_bare = n_bare + 1
+		elseif a.kind == "comp_proc" then n_proc = n_proc + 1
+		end
+	end
+
+	add("## Module summary"); add("")
+	add("| metric | value |"); add("|--------|-------|")
+	add(string.format("| sources | %d |", #dir_sources))
+	add(string.format("| atoms | %d (atoms: %d, comp_bare: %d, comp_proc: %d) |",
+		#(sa_results.atoms or {}),
+		#(sa_results.atoms or {}) - n_bare - n_proc, n_bare, n_proc))
+	add(string.format("| annotations | %d |", n_annot))
+	add(string.format("| binds structs | %d |", n_binds))
+	add(string.format("| macro decls | %d |", n_macros))
+	add(string.format("| findings | %d (errors: %d, warnings: %d, info: %d) |",
+		#(sa_results.findings or {}),
+		#(sa_results.errors   or {}),
+		#(sa_results.warnings or {}),
+		#(sa_results.info     or {})))
+	add("")
+
+	-- Sources
+	add("## Sources"); add("")
+	for _, s in ipairs(dir_sources) do add("- `" .. s.path .. "`") end
+	add("")
+
+	-- Atoms (annotation)
+	add("## Atoms"); add("")
+	add("| kind | name | source | line |"); add("|------|------|--------|------|")
+	for _, r in ipairs(annot_results) do
+		local src_name = source_basename(r.source)
+		for _, a in ipairs(r.atoms) do
+			add(string.format("| atom | %s | %s | %d |", a.name, src_name, a.line))
+		end
+	end
+	add("")
+
+	-- Annotations
+	add("## Annotations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | name | binds | reads | writes |")
+		add("|--------|------|------|-------|-------|--------|")
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, a in ipairs(r.annots) do
+				local binds  = a.binds  or "—"
+				local reads  = (#a.reads  > 0 and table.concat(a.reads,  ",")) or "—"
+				local writes = (#a.writes > 0 and table.concat(a.writes, ",")) or "—"
+				add(string.format("| %s | %d | %s | %s | %s | %s |"
+					, src_name, a.line, a.name, binds, reads, writes))
+			end
+		end
+	end
+	add("")
+
+	-- Binds_* structs
+	add("## Binds_* structs"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, b in ipairs(r.binds) do
+				add(string.format("### %s (%s:%d, %d bytes)",
+					b.name, src_name, b.line, b.bytes))
+				for _, f in ipairs(b.fields) do
+					add(string.format("- `+%d %s`", f.offset, f.name))
+				end
+				add("")
+			end
+		end
+	end
+
+	-- Macro decls
+	add("## Macro word-count declarations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | macro declaration |")
+		add("|--------|------|-------------------|")
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, m in ipairs(r.macros) do
+				add(string.format("| %s | %d | %s |",
+					src_name, m.line, m.name))
+			end
+		end
+	end
+	add("")
+
+	-- Findings by atom (static-analysis)
+	add("## Static analysis — findings by atom"); add("")
+	local by_atom = {}
+	for _, f in ipairs(sa_results.findings or {}) do
+		by_atom[f.atom] = by_atom[f.atom] or {}
+		by_atom[f.atom][#by_atom[f.atom] + 1] = f
+	end
+	if next(by_atom) == nil then
+		add("_(no findings)_")
+	else
+		for _, a in ipairs(sa_results.atoms or {}) do
+			local fs = by_atom[a.name]
+			if fs then
+				add(string.format("### %s", a.name))
+				for _, f in ipairs(fs) do
+					add(string.format("- `[%s] %s`", f.check, f.msg))
+				end
+				add("")
+			end
+		end
+	end
+
+	-- Errors / Warnings / Info
+	local function add_findings(label, entries)
+		add(string.format("## %s", label))
+		if #entries == 0 then
+			add("_(none)_")
+		else
+			for _, e in ipairs(entries) do
+				add(string.format("- line %d  %s", e.line, e.msg))
+			end
+		end
+		add("")
+	end
+	add_findings("Errors",   sa_results.errors   or {})
+	add_findings("Warnings", sa_results.warnings or {})
+	add_findings("Info",     sa_results.info     or {})
+
+	-- Per-atom cycle counts (path-aware)
+	add("## Per-atom cycle counts (path-aware, best case, no stalls)"); add("")
+	add("| atom | source | min | max | branches | paths | notes |")
+	add("|------|--------|-----|-----|----------|-------|-------|")
+	local sorted = {}
+	for _, a in ipairs(sa_results.atoms or {}) do sorted[#sorted + 1] = a end
+	table.sort(sorted, function(x, y)
+		return ((x.paths or {}).cycles_max or 0) > ((y.paths or {}).cycles_max or 0)
+	end)
+	for _, a in ipairs(sorted) do
+		local p        = a.paths or {}
+		local src_name = a.source_path and source_basename(a.source_path) or ""
+		local notes    = ""
+		if p.has_loops then notes = notes .. " [loop!]" end
+		if p.unknown_macros and #p.unknown_macros > 0 then
+			notes = notes .. " [unknown: " .. table.concat(p.unknown_macros, ", ") .. "]"
+		end
+		add(string.format("| %s | %s | %d | %d | %d | %d | %s |",
+			a.name, src_name,
+			p.cycles_min or 0, p.cycles_max or 0,
+			p.branches   or 0, p.paths      or 0, notes))
+	end
+	add("")
+
+	-- Per-source scan summary
+	add("## Per-source scan summary"); add("")
+	for _, src in ipairs(dir_sources) do
+		local src_atoms = {}
+		for _, a in ipairs(sa_results.atoms or {}) do
+			if a.source_path == src.path then src_atoms[#src_atoms + 1] = a end
+		end
+		if #src_atoms > 0 then
+			local mn, mx = math.huge, -1
+			for _, a in ipairs(src_atoms) do
+				local p = a.paths or {}
+				if (p.cycles_min or 0) < mn then mn = p.cycles_min or 0 end
+				if (p.cycles_max or 0) > mx then mx = p.cycles_max or 0 end
+			end
+			local path_str
+			if mx > 0 then
+				path_str = string.format("  cycles=%d..%d", mn, mx)
+			else
+				path_str = string.format("  %d cycles", mn)
+			end
+			add(string.format("- `%s` — %d atom%s%s",
+				src.basename, #src_atoms,
+				#src_atoms == 1 and "" or "s", path_str))
+		end
+	end
+	add("")
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- REPORT_RENDERERS — data-driven report dispatch (one row per file kind)
+-- ════════════════════════════════════════════════════════════════════════════
+-- `once = true` means render once at the project level (not per-module).
+-- `basename(dir_basename)` yields the file's basename for that kind.
+-- `gather(ctx, dir, dir_sources [, all_modules])` returns the rendered string.
+local REPORT_RENDERERS = {
+	{
+		name     = "atom_meta_report",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atom_meta_report" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			-- Annotations: re-run `annotation.validate()` per source (the existing pattern).
+			local annot_results = {}
+			for _, src in ipairs(dir_sources) do
+				if src.scan then
+					local r = annotation.validate(ctx, src, nil)
+					r.source = src.path
+					annot_results[#annot_results + 1] = r
+				end
+			end
+			-- Static-analysis: read stashed projection (no re-validate).
+			local dir_basename = dir:match("([^/\\]+)$") or dir
+			local sa_results   = (ctx.shared.corpus.static_analysis_results or {})[dir_basename] or {}
+			return render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+		end,
+	},
+	{
+		name     = "atoms",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atoms" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			return render_module_atoms_md(dir, dir_sources,
+				ctx.shared.corpus.word_counts or {})
+		end,
+	},
+	{
+		name     = "summary",
+		ext      = "md",
+		basename = function(_dir_basename) return "atom_meta_report.summary" end,
+		once     = true,
+		gather   = function(_ctx, _dir, _dir_sources, all_modules)
+			return render_project_summary(all_modules)
+		end,
+	},
+}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — public pass surface
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+--- Run the report pass. Emits 1 `atom_meta_report.summary.md` per build + 2 `atom_meta_report.md` + 2 `atoms.md` files per module (duffle + gte_hello).
+--- Reads `corpus.static_analysis_results` (added in Phase 1) to populate per-module findings without re-running validate().
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs = {}
+	local corpus  = ctx.shared and ctx.shared.corpus
+	local by_dir  = (corpus and corpus.sources_by_dir) or {}
+
+	-- `out_path_root`: when the conventional `out_root` is `build/gen` (any spelling — relative, absolute, separator variants).
+	-- Write the md files to `build/` (parent of `gen/`) instead of nested under `gen/`.
+	-- Mirrors the `gdb_tape_atoms_runtime.gdb` relocation.
+	local function ends_with_gen(p)
+		return type(p) == "string" and (p:match("[/\\]gen[/\\]?$") ~= nil
+			or p == "build/gen" or p == "build\\gen")
+	end
+	local out_root_effective = ends_with_gen(ctx.out_root)
+		and ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		or ctx.out_root
+
+	duffle.ensure_dir(out_root_effective)
+
+	-- Aggregator for the project-wide `once = true` summary renderer.
+	local all_modules = {}
+
+	for dir, dir_sources in pairs(by_dir) do
+		local dir_basename = dir:match("([^/\\]+)$") or dir
+
+		-- Per-renderer dispatch for the per-module renderers (once = false).
+		for _, renderer in ipairs(REPORT_RENDERERS) do
+			if not renderer.once then
+				local body     = renderer.gather(ctx, dir, dir_sources)
+				local out_path = out_root_effective .. "/" .. renderer.basename(dir_basename) .. "." .. renderer.ext
+				duffle.write_file(out_path, body)
+				outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
+			end
+		end
+
+		-- For the summary, compute per-module totals once (re-validating annotations per source — same pattern as the meta_report renderer).
+		local annot_results = {}
+		for _, src in ipairs(dir_sources) do
+			if src.scan then
+				local r = annotation.validate(ctx, src, nil)
+				r.source = src.path
+				annot_results[#annot_results + 1] = r
+			end
+		end
+		local n_annot, n_binds, n_macros = 0, 0, 0
+		for _, r in ipairs(annot_results) do
+			n_annot  = n_annot  + #r.annots
+			n_binds  = n_binds  + #r.binds
+			n_macros = n_macros + #r.macros
+		end
+		local sa_results = (corpus.static_analysis_results or {})[dir_basename] or {}
+		all_modules[#all_modules + 1] = {
+			module   = dir_basename,
+			atoms    = #(sa_results.atoms or {}),
+			annots   = n_annot,
+			binds    = n_binds,
+			macros   = n_macros,
+			findings = #(sa_results.findings or {}),
+			errors   = #(sa_results.errors   or {}),
+			warnings = #(sa_results.warnings or {}),
+			info     = #(sa_results.info     or {}),
+		}
+	end
+
+	-- Project-wide renderer (once = true): write the summary file.
+	for _, renderer in ipairs(REPORT_RENDERERS) do
+		if renderer.once then
+			local body     = renderer.gather(ctx, nil, nil, all_modules)
+			local out_path = out_root_effective .. "/" .. renderer.basename("") .. "." .. renderer.ext
+			duffle.write_file(out_path, body)
+			outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
+		end
+	end
+
+	return { outputs = outputs, errors = {}, warnings = {} }
+end
+
+return M

scripts/passes/word_count_eval.lua

@@ -0,0 +1,128 @@
+--- word_count_eval.lua — Word-counting logic for the tape-atom metaprogram pipeline.
+---
+--- Two responsibilities:
+---   1. **Public utility** `M.count_token_words(token, wc)`: Used by `passes/offsets.lua`, `passes/annotation.lua`, and other passes.
+---   2. **Pass entry** `M.run(ctx)`: Loads the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts` for downstream passes.
+---      The generated `.macs.h` files are OUTPUT artifacts and are NOT inputs to this pass;
+---      Current component counts are owned by `passes/components.lua` (which populates `corpus.word_counts` and `corpus.component_body_index`
+---      AFTER computing each current count from the just-built body + `corpus.word_counts`).
+---
+--- **Canonical contract**:
+---   * `ctx.shared.corpus.word_counts` is the count table.
+---   * `corpus.word_counts` is the sole count table. Consumers read `corpus.word_counts` directly.
+---   * `ctx.shared.components` and `ctx.shared.component_body_index` are NOT created by this pass (projections only).
+---   * No `.macs.h` recursive discovery (no `scan_dir`, no scan cache, no `_invalidate_scan_cache`).
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- 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.
+-- 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")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class WordCounts
+--- @field [string] integer  -- macro name -> word count
+
+--- @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
+
+--- @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 shared.corpus      table         -- canonical corpus (required)
+--- @field shared.corpus.word_counts WordCounts -- canonical count table (populated by this pass)
+--- @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       -- if true, 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
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- ┌────────────────────────────────────────────────────────────────────┐
+-- │ Shared utility: count_token_words                                  │
+-- └────────────────────────────────────────────────────────────────────┘
+
+--- Count words emitted by a single comma-separated token inside an atom body.
+--- For most tokens (regular MIPS instructions) this returns 1.
+--- For `mac_X(...)` calls, this returns the resolved word count from `wc` (recursively if needed). For `nop2` etc., returns wc[name].
+--- For unknown macros, returns 1 and (optionally) warns.
+--- @param token string     -- a single token from split_top_level_commas
+--- @param wc    WordCounts -- the shared word-count table
+--- @return integer
+function M.count_token_words(token, wc)
+	local s = duffle.trim(token)
+	if    s == "" then return 0 end
+	local name, after = duffle.read_ident(s, 1)
+	if not name then return 1 end
+	if wc[name] then return wc[name] end
+	local paren_pos = duffle.skip_ws_and_cmt(s, after)
+	if s:sub(paren_pos, paren_pos) == "(" then
+		io.stderr:write("  warning: unknown macro '" .. name .. "', assuming 1 word\n")
+	end
+	return 1
+end
+
+-- ┌────────────────────────────────────────────────────────────────────┐
+-- │ Pass entry: M.run(ctx) — "word-counts" pass                        │
+-- └────────────────────────────────────────────────────────────────────┘
+
+--- 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-directory's `gen/macs.h` file.
+---
+--- Contract:
+---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).
+---   * `ctx.metadata_path` MUST be a readable file path to the authored `word_count.metadata.h`.
+---   * The pass assigns exactly one table to `corpus.word_counts`.
+---     Consumers read the corpus-owned table directly.
+---     Consumers must read `corpus.word_counts` directly.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	-- 1. Canonical-corpus ownership gate.
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("word_count_eval.run requires ctx.shared.corpus (canonical corpus). The fixture must install the corpus before running this pass.", 0)
+	end
+
+	-- 2. metadata_path gate.
+	if type(ctx.metadata_path) ~= "string" or ctx.metadata_path == "" then
+		error("word_count_eval.run requires ctx.metadata_path (path to the authored word_count.metadata.h).", 0)
+	end
+
+	-- 3. Load authored metadata. Generated .macs.h files are NOT scanned
+	--    (the pass computes their counts from the just-built bodies after disk emission; see passes/components.lua).
+	local wc = duffle.load_word_counts(ctx.metadata_path)
+
+	-- 4. Assign the count table. ONE assignment, no copy. The assignment creates no secondary alias.
+	corpus.word_counts = wc
+
+	return { outputs = {}, errors = {}, warnings = {} }
+end
+
+return M

scripts/pcsx_debug_helper/autoexec.lua

@@ -0,0 +1,51 @@
+-- autoexec.lua - pcsx_debug_helper plugin entry point.
+-- Packaged in scripts/pcsx_debug_helper.zip. Loaded by pcsx-redux via the -archive CLI flag (see scripts/launch_pcsx_debug.ps1).
+--
+-- Registers two web handlers for external CLI tools:
+--   /api/v1/lua/gte - full GTE state (32 data + 32 control regs + PC)
+--   /api/v1/lua/gp  - GP state summary (screenshot endpoint + VRAM endpoint refs)
+--
+-- The GTE handler reads COP2 regs via PCSX.getRegisters().CP2D/CP2C.
+-- The pcsx-redux gdb stub doesn't expose COP2, so this is the only way for external tools to see GTE state.
+--
+-- The GP handler is a thin pointer: 
+-- pcsx-redux's Lua API exposes only PCSX.GPU.takeScreenShot() (no GPUSTAT, no GP0/GP1 command log, no display state). For richer GP state, the existing web endpoints are the practical path:
+--   /api/v1/state/still      - PNG screenshot
+--   /api/v1/gpu/vram/raw     - VRAM raw bytes (1MB)
+--
+-- Companion: scripts/gdb/gdb_tape_atoms.gdb (covers GPRs + atom-aware stepping).
+
+local function register_handlers()
+    if not PCSX.WebServer then PCSX.WebServer = {} end
+    if not PCSX.WebServer.Handlers then PCSX.WebServer.Handlers = {} end
+
+    -- ── GTE state ──
+    PCSX.WebServer.Handlers.gte = function(req)
+        local r = PCSX.getRegisters()
+        local out = { "pc=0x" .. string.format("%x", r.pc) }
+        for i = 0, 31 do
+            out[#out + 1] = string.format("D[%d]=0x%08x C[%d]=0x%08x",
+                i, r.CP2D.r[i], i, r.CP2C.r[i])
+        end
+        return table.concat(out, "\n")
+    end
+
+		-- ── GP state (pointer to existing endpoints) ──
+		-- pcsx-redux's Lua GPU API exposes only takeScreenShot(); no GPUSTAT / GP0 / GP1 command log / display state.
+		-- We point to the existing web endpoints that DO expose those (when the emulator is actually rendering. Paused-at-BP frames won't have a fresh frame).
+    PCSX.WebServer.Handlers.gp = function(req)
+        local out = {
+            "gpu_screenshot_png=http://localhost:8080/api/v1/state/still",
+            "vram_raw=http://localhost:8080/api/v1/gpu/vram/raw (1MB VRAM)",
+            "gpustat=NOT_AVAILABLE_VIA_LUA",
+            "gp_command_log=NOT_AVAILABLE_VIA_LUA (use pcsx-redux Debug > GPU Logger)",
+            "hint_run_emulator_unpaused_for_screenshot",
+        }
+        return table.concat(out, "\n")
+    end
+end
+
+local ok, err = pcall(register_handlers)
+if ok then print("[pcsx_debug_helper] handlers registered: gte, gp")
+else       print("[pcsx_debug_helper] registration failed: " .. tostring(err))
+end

scripts/ps1_meta.lua

@@ -0,0 +1,758 @@
+--- ps1_meta.lua — Orchestrator entry point for the tape-atom metaprogram.
+---
+--- 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) → 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. 
+---
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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).
+-- 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
+local _bootstrap_src
+if _is_entry_script then
+	_bootstrap_src = arg[0]
+else
+	-- debug.getinfo(1, "S").source returns "@<path>" for the current chunk;
+	-- strip the leading "@" so the directory match works in both cases.
+	_bootstrap_src = debug.getinfo(1, "S").source:sub(2)
+end
+local duffle = dofile((_bootstrap_src:match("(.*[/\\])") or "./") .. "duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Exit codes (per the --help text and the post-build summary convention).
+local EXIT_OK                  = 0
+local EXIT_VALIDATION_ERRORS   = 1
+local EXIT_INTERNAL_ERROR      = 2
+
+-- Default --out-root value if not provided.
+local DEFAULT_OUT_ROOT         = "build/gen"
+
+-- Sentinel for "all passes" in `PASS_FLAG_TO_NAME`. Distinguishes `--all` from the per-pass flags (which map to individual pass names).
+local ALL_PASSES_SENTINEL      = "__all__"
+
+-- Sentinel key for the pass-flag dispatcher in `FLAG_HANDLERS`. 
+-- The actual pass names are looked up via `PASS_FLAG_TO_NAME`, not direct dispatch, so this key never matches a real flag.
+local PASS_FLAG_DISPATCH_KEY   = "__pass__"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class PassDescriptor
+--- @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
+
+--- @class SourceFile
+--- @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                  -- 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
+
+--- @class Finding
+--- @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
+
+--- @class ParsedArgs
+--- @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
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- PASSES Table
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Build-phase groups: Each PASSES row may declare membership in one or more named groups via `groups = { ... }`.
+-- The CLI flags --pre-link and --post-link request the *roots* of their group; topo_sort then closes transitive dependencies from those roots,
+-- and dispatch_passes runs every pass in the resulting closure without phase-filtering.
+--
+-- A row without a `groups` entry is dependency-only: it runs only when a transitive dep requests it, 
+-- but it remains directly requestable through its explicit CLI flag (e.g. --atoms-source-map, --scan-source).
+
+local PASSES = {
+	["scan-source"] = {
+		module = "passes.scan_source",
+		kind   = "shared", deps = {},
+	},
+	["word-counts"] = {
+		module = "passes.word_count_eval",
+		kind   = "shared", deps = {},
+	},
+	components = {
+		module = "passes.components",
+		kind   = "header-output",
+		deps   = {"scan-source", "word-counts"},
+	},
+	["emission-model"] = {
+		module = "passes.emission_model",
+		kind   = "validation",
+		deps   = {"components"},
+	},
+	annotation = {
+		module = "passes.annotation",
+		kind   = "validation",
+		deps   = {"scan-source", "word-counts"},
+	},
+	offsets = {
+		module = "passes.offsets",
+		kind   = "header-output",
+		deps   = {"scan-source", "word-counts", "components", "emission-model"},
+		groups = { "pre-link" },
+	},
+	["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).
+		-- Report severity is independent from process exit policy.
+		kind   = "diagnostic",
+		deps   = {"scan-source", "word-counts", "components", "emission-model"},
+	},
+	["atoms-source-map"] = {
+		module = "passes.atoms_source_map",
+		kind   = "header-output",
+		deps   = {"word-counts", "components", "emission-model"},
+	},
+	["dwarf-injection"] = {
+		module = "passes.dwarf_injection",
+		kind   = "shared",
+		deps   = {"scan-source", "atoms-source-map"},
+		groups = { "post-link" },
+	},
+	report = {
+		module = "passes.report",
+		kind   = "report",
+		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`.
+-- topo_sort closes the transitive deps from there; dispatch_passes runs every resolved pass without phase-filtering.
+-- ────────────────────────────────────────────────────────────────────────────
+
+--- @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
+		if pass.groups then
+			for _, g in ipairs(pass.groups) do
+				if g == group_name then
+					names[#names + 1] = name
+					break
+				end
+			end
+		end
+	end
+	table.sort(names)
+	return names
+end
+
+--- Append every root belonging to `group_name` to `args.requested_set`.
+--- Errors loudly if no PASSES row declares the group, so a typo'd or future-removed group name
+--- cannot silently fall through to pre-link (or any other default) and dispatch nothing.
+--- @param args ParsedArgs
+--- @param group_name string
+local function request_roots_for_group(args, group_name)
+	local roots = roots_for_group(group_name)
+	if #roots == 0 then
+		error(string.format("ps1_meta: build-phase group %q has zero roots in PASSES; check PASSES rows for a `groups = { %q }` field"
+			, group_name, group_name))
+	end
+	for _, name in ipairs(roots) do
+		args.requested_set[#args.requested_set + 1] = 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.
+-- 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,
+	["diagnostic"]    = false,
+	["report"]        = false,
+}
+
+-- Closed set of CLI flags -> pass names.
+-- 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 = {
+	["--word-counts"]       = "word-counts",
+	["--components"]        = "components",
+	["--validate"]          = "annotation",
+	["--offsets"]           = "offsets",
+	["--static-analysis"]   = "static-analysis",
+	["--atoms-source-map"]  = "atoms-source-map",
+	["--report"]            = "report",
+	["--scan-source"]       = "scan-source",
+	["--all"]               = ALL_PASSES_SENTINEL,
+}
+
+--- Append every pass name to args.requested_set.
+--- Names are derived from PASSES (no parallel name list); used by --all and by any caller that wants the full closure.
+--- @param args ParsedArgs
+local function request_all_passes(args)
+	local names = {}
+	for name in pairs(PASSES) do names[#names + 1] = name end
+	table.sort(names)
+	for _, n in ipairs(names) do
+		args.requested_set[#args.requested_set + 1] = n
+	end
+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 = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- CLI parsing
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Print the CLI usage to stdout and exit 0.
+local function print_help()
+	io.write([[
+ps1_meta.lua - Tape-atom metaprogram orchestrator
+
+USAGE:
+  ps1_meta.lua [PASS_FLAGS] [COMMON_FLAGS]
+
+PASS_FLAGS:
+  Pick a phase or one-or-more individual passes:
+    --pre-link           [phase; default] Run the pre-link group + transitive deps.
+                         The root set is data-driven from each PASSES row's groups` field; no parallel name list is maintained.
+    --post-link          [phase] Run the post-link group + transitive deps.
+                         Requires --elf. Sets --gdb-runtime and --dwarf-injection opt-in flags as well.
+    --all                Select every row of the PASSES table. Pass-local opt-in guards remain active, so --dwarf-injection still requires
+                         --elf and --gdb-runtime still requires a runtime emission.
+  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 <srcdir>/gen/macs.h (per-directory aggregation)
+    --validate            Run atom annotation DSL validation
+    --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
+    --report              Render per-project summary
+
+COMMON_FLAGS:
+  --unity-root FILE     Unity source root: load root + direct quoted authored includes only. Mutually exclusive with --source.
+  --source FILE         Exact source file to process (repeatable, never expands includes). Mutually exclusive with --unity-root.
+  --metadata PATH       Path to metadata.h (required)
+  --out-root DIR        Output root for reports (default: build/gen)
+  --project-root DIR    PS1 repository root (default: derived from <repo>/code/duffle/word_count.metadata.h)
+  --gdb-runtime         Also emit <out_root>/gdb_tape_atoms_runtime.gdb (post-link, requires --elf)
+  --elf PATH            Path to linked .elf (for --gdb-runtime / --dwarf-injection)
+  --verbose             Print per-pass debug output
+  --help                Show this help and exit
+
+EXIT CODES:
+  0  All requested passes succeeded
+  1  Validation errors found
+  2  Metaprogram internal error
+
+EXAMPLES:
+  ps1_meta.lua --pre-link  --metadata code/duffle/word_count.metadata.h --unity-root code/gte_hello/hello_gte.c
+  ps1_meta.lua --post-link --metadata code/duffle/word_count.metadata.h --unity-root code/gte_hello/hello_gte.c --elf build/hello_gte.elf
+  ps1_meta.lua --all       --metadata metadata.h --source code/foo.c --source code/bar.c
+]])
+end
+
+local FLAG_VALUE_NAMES = {
+	["--source"]       = "FILE",
+	["--unity-root"]   = "FILE",
+	["--metadata"]     = "PATH",
+	["--out-root"]     = "DIR",
+	["--project-root"] = "DIR",
+	["--elf"]          = "PATH",
+}
+
+local function require_flag_value(argv, arg_idx, flag)
+	local value      = argv[arg_idx + 1]
+	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")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+	return value, arg_idx + 1
+end
+
+-- Per-flag handlers. Each takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it).
+-- 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["--source"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--source")
+	args.sources[#args.sources + 1] = value
+	return value_idx
+end
+FLAG_HANDLERS["--unity-root"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--unity-root")
+	args.unity_root = value
+	return value_idx
+end
+FLAG_HANDLERS["--metadata"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--metadata")
+	args.metadata = value
+	return value_idx
+end
+FLAG_HANDLERS["--out-root"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--out-root")
+	args.out_root = value
+	return value_idx
+end
+FLAG_HANDLERS["--project-root"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--project-root")
+	args.project_root = value
+	return value_idx
+end
+
+-- Per-pass stash flags. Read by `passes/atoms_source_map.lua` to opt into the post-link gdb-runtime emission.
+-- Same shape as the existing per-flag handlers. mutates `args.flags` (which propagates into `ctx.flags`).
+FLAG_HANDLERS["--gdb-runtime"] = function(args)
+	args.flags = args.flags or {}
+	args.flags.gdb_runtime = true
+end
+FLAG_HANDLERS["--elf"] = function(args, argv, arg_idx)
+	local value, value_idx = require_flag_value(argv, arg_idx, "--elf")
+	args.flags = args.flags or {}
+	args.flags.elf_path = value
+	return value_idx
+end
+-- Enable DWARF injection (default OFF). Opts in to the post-link pass and sets the flag in one shot.
+-- The explicit handler below owns both selection and opt-in state, so --dwarf-injection is intentionally absent from PASS_FLAG_TO_NAME.
+FLAG_HANDLERS["--dwarf-injection"] = function(args)
+	args.flags                 = args.flags or {}
+	args.flags.dwarf_injection = true
+	args.requested_set[#args.requested_set + 1] = "dwarf-injection"
+end
+-- Build-phase flags: --pre-link and --post-link request the roots of their declared groups (see roots_for_group).
+-- topo_sort closes transitive deps from those roots; dispatch_passes runs every pass in the resolved closure without phase-filtering.
+FLAG_HANDLERS["--pre-link"] = function(args)
+	request_roots_for_group(args, "pre-link")
+end
+-- Batch post-link phase: gdb-runtime + dwarf-injection in one luajit cold start.
+-- Sets the same opt-in flags as --gdb-runtime + --dwarf-injection and selects the post-link build-phase group.
+-- elf is required; parse_args enforces it after all flags are parsed.
+FLAG_HANDLERS["--post-link"]      = function(args)
+	args.flags = args.flags or {}
+	args.flags.gdb_runtime     = true
+	args.flags.dwarf_injection = true
+	request_roots_for_group(args, "post-link")
+end
+
+-- `--dwarf-injection` also emits atom-local debug data.
+
+-- Pass-flag handler. Reads the closed-set table, expands --all, appends to requested_set.
+FLAG_HANDLERS[PASS_FLAG_DISPATCH_KEY] = function(args, a)
+	local name = PASS_FLAG_TO_NAME[a]
+	if name == ALL_PASSES_SENTINEL then
+		request_all_passes(args)
+		return
+	end
+	args.requested_set[#args.requested_set + 1] = name
+end
+
+--- Parse argv into a structured table. Validates against a closed enum.
+--- @param argv string[]
+--- @return ParsedArgs
+local function parse_args(argv)
+	local args = {
+		requested_set = {},
+		sources       = {},
+		unity_root    = nil,
+		metadata      = nil,
+		out_root      = DEFAULT_OUT_ROOT,
+		project_root  = nil,
+		verbose       = false,
+	}
+
+	local pos = 1
+	while pos <= #argv do
+		local a       = argv[pos]
+		local handler = FLAG_HANDLERS[a]
+		if handler then
+			pos = handler(args, argv, pos) or pos
+		elseif PASS_FLAG_TO_NAME[a] then
+			FLAG_HANDLERS[PASS_FLAG_DISPATCH_KEY](args, a)
+		else
+			io.stderr:write("ps1_meta: unknown flag '" .. a .. "'\n")
+			io.stderr:write("Run with --help for usage.\n")
+			os.exit(EXIT_INTERNAL_ERROR)
+		end
+		pos = pos + 1
+	end
+
+	-- Default: --pre-link if no explicit pass flags were given.
+	-- The first invocation of a build is always pre-link, so this avoids silently also invoking post-link work in builds without an ELF artifact.
+	if #args.requested_set == 0 then request_roots_for_group(args, "pre-link") end
+
+	if not args.metadata then
+		io.stderr:write("ps1_meta: --metadata PATH is required\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	-- `<repo>/code/duffle/word_count.metadata.h` is the canonical metadata location. 
+	-- `project_root` names `<repo>`; the resolver derives `<project_root>/code` separately.
+	if not args.project_root then
+		local metadata_dir = duffle.dirname(duffle.normalize_path(args.metadata))
+		local code_root    = duffle.dirname(metadata_dir)
+		args.project_root  = duffle.dirname(code_root)
+	else
+		args.project_root = duffle.normalize_path(args.project_root)
+	end
+
+	local has_unity = type(args.unity_root) == "string" and args.unity_root ~= ""
+	if has_unity and #args.sources > 0 then
+		io.stderr:write("ps1_meta: --unity-root FILE and --source FILE are mutually exclusive\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+	if not has_unity and #args.sources == 0 then
+		io.stderr:write("ps1_meta: either --unity-root FILE or at least one --source FILE is required\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	-- Post-link opt-ins (--gdb-runtime, --dwarf-injection) write output that depends on the linked ELF.
+	-- Without --elf the metaprogram can't satisfy those requests, so refuse loud and early.
+	-- This covers the explicit --post-link batch, --dwarf-injection by itself, and --gdb-runtime by itself.
+	local flags      = args.flags or {}
+	local elf_path   = flags.elf_path
+	local has_elf    = type(elf_path) == "string" and #elf_path > 0
+	local post_links = flags.gdb_runtime or flags.dwarf_injection
+	if post_links and not has_elf then
+		io.stderr:write("ps1_meta: --elf PATH is required for post-link output\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	return args
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Build ctx from parsed args
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Build the PassCtx from parsed args. Exact mode opens only the repeated `--source` inputs;
+--- unity mode delegates direct-include resolution to duffle.resolve_source_corpus`. 
+--- Scanning remains pass-owned (`src.scan`).
+--- @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 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.
+		duffle.canonical_path_key(normalized_project_root)
+		project_root = duffle.normalize_path(duffle.to_absolute_path(normalized_project_root))
+	else
+		-- Do not route POSIX/UNC/drive-absolute paths through to_absolute_path.
+		duffle.canonical_path_key(project_root)
+	end
+	local resolution
+	if args.unity_root then
+		local ok_resolve, resolved = pcall(duffle.resolve_source_corpus, {
+			unity_root   = args.unity_root,
+			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")
+			os.exit(EXIT_INTERNAL_ERROR)
+		end
+		resolution = resolved
+	else
+		local source_order    = {}
+		local sources_by_path = {}
+		local resolver = {
+			resolved = {},
+			skipped  = {},
+			shadowed = {},
+		}
+		for _, input_path in ipairs(args.sources) do
+			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)
+			end
+			local file = io.open(path, "r")
+			if not file then
+				io.stderr:write("ps1_meta: cannot open --source " .. input_path .. "\n")
+				os.exit(EXIT_INTERNAL_ERROR)
+			end
+			local text = file:read("*a")
+			file:close()
+
+			local source = {
+				path     = path,
+				text     = text,
+				dir      = duffle.dirname(path),
+				basename = duffle.basename_no_ext(path),
+			}
+			source_order[#source_order + 1] = source
+			local key = key_or_error
+			if not sources_by_path[key] then sources_by_path[key] = source end
+			resolver.resolved[#resolver.resolved + 1] = {
+				include_path  = path,
+				include_text  = nil,
+				root_source   = nil,
+				root_line     = nil,
+				candidate_a   = path,
+				candidate_b   = nil,
+				selected_path = path,
+				disposition   = "exact",
+			}
+		end
+		resolution = {
+			unity_root      = nil,
+			project_root    = project_root,
+			code_root       = duffle.normalize_path(project_root .. "/code"),
+			source_order    = source_order,
+			sources_by_path = sources_by_path,
+			sources_by_dir  = duffle.group_sources_by_dir(source_order),
+			resolver        = resolver,
+		}
+	end
+
+	local corpus = {
+		unity_root              = resolution.unity_root,
+		project_root            = resolution.project_root,
+		code_root               = resolution.code_root,
+		source_order            = resolution.source_order,
+		sources_by_path         = resolution.sources_by_path,
+		sources_by_dir          = resolution.sources_by_dir,
+		atoms_by_name           = {},
+		binds_by_name           = {},
+		atom_infos              = {},
+		register_alias_registry = {},
+		type_name_registry      = {},
+		atom_views              = {},
+		atom_ctxs               = {},
+		atom_phases             = {},
+		word_counts             = {},
+		components              = {},
+		component_body_index    = {},
+		collisions              = {},
+		resolver                = resolution.resolver,
+	}
+	local ctx = {
+		metadata_path = args.metadata,
+		shared        = { corpus = corpus },
+		out_root      = args.out_root,
+		project_root  = corpus.project_root,
+		flags         = args.flags or {},
+		verbose       = args.verbose,
+	}
+
+	-- Source records and directory buckets are owned by the corpus.
+	-- Consumers read `corpus.source_order` and `corpus.sources_by_dir` directly.
+	-- The corpus is the sole source of truth for source records and module grouping; `ctx` only holds per-pass execution state.
+	return ctx
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Topological sort (Kahn's algorithm + cycle detection)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Topologically sort the requested pass set, augmented with all transitive deps.
+--- Detects cycles and errors out with details.
+--- @param passes        table<string, PassDescriptor>
+--- @param requested_set string[]
+--- @return string[]  -- execution order
+---
+--- Dependency closure, in-degree calculation, queue seeding, and sorting are local blocks.
+--- Keeping these blocks local makes the topological sort self-contained.
+local function topo_sort(passes, requested_set)
+	-- Dependency closure: include every pass transitively required by `requested_set`.
+	local needed = {}
+	for _, name in ipairs(requested_set) do needed[name] = true end
+	local changed = true
+	while changed do
+		changed = false
+		for name, _ in pairs(needed) do
+			local pass = passes[name]
+			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
+					changed = true
+				end
+			end
+		end
+	end
+
+	-- In-degree calculation: count each needed pass's needed dependencies.
+	local in_degree = {}
+	for name, _ in pairs(needed) do in_degree[name] = 0 end
+	for name, _ in pairs(needed) do
+		for _, dep in ipairs(passes[name].deps) do
+			if needed[dep] then
+				in_degree[name] = in_degree[name] + 1
+			end
+		end
+	end
+
+	-- Ready-queue seeding: add zero-in-degree passes in deterministic order.
+	local ready = {}
+	for name, deg in pairs(in_degree) do
+		if deg == 0 then ready[#ready + 1] = name end
+	end
+	table.sort(ready)
+
+	-- Ready-queue drain: decrement dependents when each pass is emitted.
+	-- Newly-zero-degree passes are inserted back into the ready queue (kept sorted).
+	local order = {}
+	while #ready > 0 do
+		local just_finished = table.remove(ready, 1)
+		order[#order + 1] = just_finished
+		for name, _ in pairs(needed) do
+			if name ~= just_finished then
+				for _, dep in ipairs(passes[name].deps) do
+					if dep == just_finished then
+						in_degree[name] = in_degree[name] - 1
+						if in_degree[name] == 0 then
+							ready[#ready + 1] = name
+							table.sort(ready)
+						end
+					end
+				end
+			end
+		end
+	end
+
+	-- Cycle detection: if `order` doesn't include all needed passes, some are stuck with in_degree > 0
+	-- (the cycle closed on itself before Kahn could process them).
+	-- Without this check, a fully-closed cycle (e.g. A -> B -> A) would silently return an empty order list, leaving the orchestrator to dispatch nothing.
+	local needed_count = 0
+	for _ in pairs(needed) do needed_count = needed_count + 1 end  -- count hash entries; Lua's #t doesn't work
+	if #order ~= needed_count then
+		for name, deg in pairs(in_degree) do
+			if deg > 0 then
+				error("dependency cycle detected involving pass '" .. name .. "'")
+			end
+		end
+	end
+
+	return order
+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.
+--- @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
+	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
+end
+
+--- (internal) Run each pass in `order` in topological sequence.
+--- @param ctx   PassCtx
+--- @param order string[]
+--- @return boolean  -- true if any validation errors were reported
+local function dispatch_passes(ctx, order)
+	local had_errors = false
+	for _, pass_name in ipairs(order) do
+		local pass   = PASSES[pass_name]
+		local mod    = require(pass.module)
+		local result = mod.run(ctx)
+		if report_validation_errors(pass_name, pass, result) then
+			had_errors = true
+		end
+	end
+	return had_errors
+end
+
+--- Main entry point. Runs the requested passes in dep-topological order.
+--- @param argv string[]
+local function main(argv)
+	local ok, err = pcall(function()
+		local args = parse_args(argv)
+		local ctx  = build_ctx(args)
+
+		local requested = args.requested_set
+		local closed    = topo_sort(PASSES, requested)
+
+		local had_errors = dispatch_passes(ctx, closed)
+		if had_errors then os.exit(EXIT_VALIDATION_ERRORS) end
+	end)
+
+	if not ok then
+		io.stderr:write("[ps1_meta] internal error: " .. tostring(err) .. "\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	os.exit(EXIT_OK)
+end
+
+-- Module export for in-process consumers (tests that dofile this script).
+-- The conditional `main(...)` call below only fires when this file is invoked as the entry script (arg[0] ends in "ps1_meta.lua");
+-- in dofile() mode (test's arg[0] does not match), main() is skipped and the chunk returns `_M` to the caller.
+local _M = {
+	PASSES                   = PASSES,
+	PASS_KIND_STOP_ON_ERROR  = PASS_KIND_STOP_ON_ERROR,
+	parse_args               = parse_args,
+	build_ctx                = build_ctx,
+}
+
+if arg and arg[0] and arg[0]:match("ps1_meta%.lua$") then
+	main({...})
+end
+
+return _M

scripts/tape_atom.offset_gen.meta.lua

@@ -1,437 +0,0 @@
-#!/usr/bin/env lua
--- tape_atom_offset_gen.lua
---
--- Finds every `MipsAtom_(name) { ... }` declaration in the given sources,
--- counts the words in each body using the WORD_COUNT manifest, computes
--- branch offsets for atom_label(name) / atom_offset(tag, name) markers,
--- and writes one header per source into <source_dir>/gen/<basename>.offsets.h
---
--- Generated header layout (per source):
---   #pragma region <basename>
---   #undef atom_offset
---   #define atom_offset(tag, name)  atom_offset_##tag##_##name
---   // --- atom: <name> (<n> words) ---
---   #define atom_offset_<tag>_<target>  (N)        // preprocessor form
---   #undef atom_offset_<tag>_<target>              // (so enum can reuse)
---   enum {
---       atom_offset_<tag>_<target> = N,            // C enum form
---   };
---   #define atom_offset_<tag>_<target>  (N)        // re-define for preprocessor
---   #pragma endregion <basename>
---
--- Usage:
---   luajit gen_atom_offsets.lua <metadata.h> <source1> [source2 ...]
-
--- Make require("duffle") resolve to the sibling duffle.lua in this dir,
--- AND make require("lpeg") find the vendored LPeg DLL in the toolchain.
--- Both prepends are explicit (no :match / no Lua pattern — plain byte scan).
-local script_path = arg[0]
-local last_sep = 0
-for i = 1, #script_path do
-    local c = script_path:sub(i, i)
-    if c == "/" or c == "\\" then last_sep = i end
-end
-local script_dir = last_sep == 0 and "./" or script_path:sub(1, last_sep)
-package.path   = script_dir .. "?.lua;" .. script_dir .. "?/init.lua;" .. package.path
-package.cpath = "C:\\projects\\Pikuma\\ps1\\toolchain\\luajit-2.1\\lib\\lua\\5.1\\?.dll;" .. package.cpath
-
--- Shared primitives + domain tables live in scripts/duffle.lua.
-local duffle = require("duffle")
-
--- Local aliases so the rest of this file reads cleanly. These resolve
--- to the same functions in duffle.lua (5.3-compatible, no regex).
-local is_space        = duffle.is_space
-local is_alpha        = duffle.is_alpha
-local is_alnum        = duffle.is_alnum
-local trim            = duffle.trim
-local find_byte       = duffle.find_byte
-local read_file       = duffle.read_file
-local write_file      = duffle.write_file
-local ensure_dir      = duffle.ensure_dir
-local dirname         = duffle.dirname
-local basename_no_ext = duffle.basename_no_ext
-local skip_str_or_cmt = duffle.skip_str_or_cmt
-local skip_ws_and_cmt = duffle.skip_ws_and_cmt
-local read_ident      = duffle.read_ident
-local read_parens     = duffle.read_parens
-local read_braces     = duffle.read_braces
-local read_brackets   = duffle.read_brackets
-local scan_to_char    = duffle.scan_to_char
-local split_top_level_commas = duffle.split_top_level_commas
-local load_word_counts = duffle.load_word_counts
-
--- ============================================================
--- Offset-gen-specific helpers (not in duffle.lua)
--- ============================================================
-
-local function starts_with(s, prefix)
-    if #s < #prefix then return false end
-    for i = 1, #prefix do
-        if s:sub(i, i) ~= prefix:sub(i, i) then return false end
-    end
-    return true
-end
-
-local function ends_with(s, suffix)
-    if #s < #suffix then return false end
-    local off = #s - #suffix
-    for i = 1, #suffix do
-        if s:sub(off + i, off + i) ~= suffix:sub(i, i) then return false end
-    end
-    return true
-end
-
-local function to_upper(s) return s:upper() end
-local function to_alnum_underscore(s)
-    local out = ""
-    for i = 1, #s do
-        local c = s:sub(i, i)
-        if is_alnum(c) then out = out .. c
-        else out = out .. "_" end
-    end
-    return out
-end
-local function pad_right(s, w) return s .. string.rep(" ", w - #s) end
-
--- ============================================================
--- Extract comma-separated identifier args from a parenthesized group
--- after a function-like macro call.
--- ============================================================
-
-local function extract_ident_args(token, after_ident)
-    local arg_start = skip_ws_and_cmt(token, after_ident)
-    if token:sub(arg_start, arg_start) ~= "(" then return {}, nil end
-    local inner, after_paren = read_parens(token, arg_start)
-
-    local args = {}
-    local n    = 1
-    local len  = #inner
-    while n <= len do
-           n = skip_ws_and_cmt(inner, n)
-        if n > len then break end
-        local ident, after = read_ident(inner, n)
-        if ident and ident ~= "" then
-            table.insert(args, ident)
-            n = after
-        else
-            n = n + 1
-        end
-        n = skip_ws_and_cmt(inner, n)
-        if n <= len and inner:sub(n, n) == "," then n = n + 1 end
-    end
-
-    return args, after_paren
-end
-
--- ============================================================
--- Count words for a single comma-separated token
--- ============================================================
-
-local function word_count_of_token(token, wc)
-    local s = trim(token)
-    if    s == "" then return 0 end
-    local name, after = read_ident(s, 1)
-    if not name then return 1        end
-    if wc[name] then return wc[name] end
-    local j = skip_ws_and_cmt(s, after)
-    if s:sub(j, j) == "(" then
-        io.stderr:write("  warning: unknown macro '" .. name .. "', assuming 1 word\n")
-    end
-    return 1
-end
-
--- ============================================================
--- Scan token for atom_label/atom_offset markers, walking through
--- balanced groups transparently (so nested calls are found)
--- ============================================================
-
-local function scan_for_atom_markers(token, at_pos, labels, branches)
-    local i   = 1
-    local len = #token
-    while i <= len do
-           i = skip_ws_and_cmt(token, i)
-        if i > len then break end
-        local c = token:sub(i, i)
-        if is_alpha(c) then
-            local ident, after = read_ident(token, i)
-            if ident == "atom_label" then
-                local args, after_paren = extract_ident_args(token, after)
-                if #args >= 1 then labels[args[1]] = at_pos end
-                if after_paren then i = after_paren else i = after end
-            elseif ident == "atom_offset" then
-                local args, after_paren = extract_ident_args(token, after)
-                if #args >= 2 then table.insert(branches, {pos = at_pos, target = args[2], tag = args[1]}) end
-                if after_paren then i = after_paren else i = after end
-            else
-                i = after
-            end
-        else
-            local nx = skip_str_or_cmt(token, i)
-            if nx > i then i = nx else i = i + 1 end
-        end
-    end
-end
-
--- ============================================================
--- Scan atom body, count words, find markers
--- ============================================================
-
--- Find the end position (just past the closing ')') of the first
--- atom_label/atom_offset call in `tok`. Returns 0 if no such call.
-local function find_marker_call_end(tok)
-    local i   = 1
-    local len = #tok
-    while i <= len do
-        i = skip_ws_and_cmt(tok, i)
-        if i > len then break end
-        local c = tok:sub(i, i)
-        if is_alpha(c) then
-            local ident, after = read_ident(tok, i)
-            if ident == "atom_label" or ident == "atom_offset" then
-                local j = skip_ws_and_cmt(tok, after)
-                if tok:sub(j, j) == "(" then
-                    local _, end_paren = read_parens(tok, j)
-                    return end_paren - 1
-                end
-                return 0
-            end
-            i = after
-        else
-            local nx = skip_str_or_cmt(tok, i)
-            if nx > i then i = nx else i = i + 1 end
-        end
-    end
-    return 0
-end
-
-local function scan_atom_body(body, word_counts)
-    local pos      = 0
-    local labels   = {}
-    local branches = {}
-    for _, tok in ipairs(split_top_level_commas(body)) do
-        local k    = 1
-        local tlen = #tok
-        while k <= tlen and is_space(tok:sub(k, k)) do k = k + 1 end
-        local leading_ident = read_ident(tok, k)
-        if leading_ident == "atom_label" or leading_ident == "atom_offset" then
-            -- Marker call: record at the current pos, do NOT advance pos.
-            -- But the source pattern may bundle the marker with the next
-            -- instruction on a new line (no top-level comma between them).
-            -- In that case, the rest of `tok` after the marker call is a
-            -- real instruction that must still be counted.
-            scan_for_atom_markers(tok, pos, labels, branches)
-            local marker_end = find_marker_call_end(tok)
-            if marker_end > 0 and marker_end < #tok then
-                local rest = trim(tok:sub(marker_end + 1))
-                if rest ~= "" then
-                    local rest_words = word_count_of_token(rest, word_counts)
-                    pos = pos + rest_words
-                end
-            end
-        else
-            local words = word_count_of_token(tok, word_counts)
-            scan_for_atom_markers(tok, pos, labels, branches)
-            pos = pos + words
-        end
-    end
-    return labels, branches, pos
-end
-
--- ============================================================
--- Find every MipsAtom_(name) { ... } in a source
--- ============================================================
-
-local function skip_qualifiers(source, i)
-    local keywords = {
-        ["static"]  = true, ["const"]    = true, ["volatile"] = true,
-        ["extern"]  = true, ["register"] = true, ["auto"]     = true,
-        ["inline"]  = true, ["typedef"]  = true,
-        ["internal"]= true, ["LP_"]      = true, ["global"] = true, ["gkknown"] = true
-    }
-    while true do
-        i = skip_ws_and_cmt(source, i)
-        local ident, after = read_ident(source, i)
-        if not ident       then                return i end
-        if keywords[ident] then i = after else return i end
-    end
-end
-
-local function find_atoms(source_text)
-    local atoms = {}
-    local len   = #source_text
-    local i     = 1
-
-    local function try_wrapped(after_pos)
-        local paren_pos = skip_ws_and_cmt(source_text, after_pos)
-        if source_text:sub(paren_pos, paren_pos) ~= "(" then return nil end
-        local inner, after_paren = read_parens(source_text, paren_pos)
-        local n = 1
-        while n <= #inner and is_space(inner:sub(n, n)) do n = n + 1 end
-        local ns = n
-        while n <= #inner and is_alnum(inner:sub(n, n)) do n = n + 1 end
-        local name = inner:sub(ns, n - 1)
-        if name == "" then return nil end
-        local brace_pos = scan_to_char(source_text, "{", after_paren)
-        if not brace_pos then return nil end
-        local body, after_brace = read_braces(source_text, brace_pos)
-        return {name = name, body = body, after_brace = after_brace}
-    end
-
-    local function try_raw(after_pos)
-        local next_pos = skip_ws_and_cmt(source_text, after_pos)
-        local next_ident, next_after = read_ident(source_text, next_pos)
-        if not next_ident then return nil end
-        if not starts_with(next_ident, "code_") then return nil end
-        if #next_ident <= 5 then return nil end
-        local atom_name = next_ident:sub(6)
-        local brace_pos = scan_to_char(source_text, "{", next_after)
-        if not brace_pos then return nil end
-        local body, after_brace = read_braces(source_text, brace_pos)
-        return {name = atom_name, body = body, after_brace = after_brace}
-    end
-
-    while i <= len do
-        i = skip_ws_and_cmt(source_text, i); if i > len then break end
-        i = skip_qualifiers(source_text, i); if i > len then break end
-        local ident, after = read_ident(source_text, i)
-        if not ident then
-            i = i + 1
-        elseif ident == "MipsAtom_" then
-            local atom = try_wrapped(after)
-            if atom then
-                table.insert(atoms, {name = atom.name, body = atom.body})
-                i = atom.after_brace
-            else
-                i = i + 1
-            end
-        elseif ident == "MipsCode" then
-            local atom = try_raw(after)
-            if atom then
-                table.insert(atoms, {name = atom.name, body = atom.body})
-                i = atom.after_brace
-            else
-                i = after
-            end
-        else
-            i = after
-        end
-    end
-    return atoms
-end
-
--- ============================================================
--- Compute branch offsets (target - branch - 1)
--- ============================================================
-
-local function compute_offsets(labels, branches)
-    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.pos .. ")")
-        end
-        table.insert(results, {target = br.target, tag = br.tag, offset = target - br.pos - 1 })
-    end
-    return results
-end
-
--- ============================================================
--- Generate header for one source
--- ============================================================
-
-local function generate_header(source_path, atoms_data)
-    local basename = basename_no_ext(source_path)
-    local guard    = to_alnum_underscore(to_upper(basename)) .. "_OFFSETS_H"
-
-    local lines = {}
-    local function add(s) table.insert(lines, s) end
-
-    add("// Auto-generated by tape_atom_offset_gen.meta.lua — DO NOT EDIT")
-    add("// Source: " .. source_path)
-    add("#pragma once")
-    add("")
-    add("#pragma region " .. basename)
-    add("")
-    -- add("// Dispatch macro: token-pastes <tag>_<target> to the enum name")
-    -- add("#undef atom_offset")
-    -- add("#define atom_offset(tag, name)  atom_offset_##tag##_##name")
-    add("")
-    for _, atom in ipairs(atoms_data) do
-        if #atom.offsets > 0 then
-            add("// --- atom: " .. atom.name .. " (" .. atom.total_words .. " words) ---")
-            add("")
-            local consts = {}
-            for _, r in ipairs(atom.offsets) do
-                table.insert(consts, {
-                    macro_name = "_atom_offset_" .. r.tag .. "_" .. r.target,
-                    enum_name  = "atom_offset_"  .. r.tag .. "_" .. r.target,
-                    value      = r.offset
-                })
-            end
-            for _, c in ipairs(consts) do add("#define " .. pad_right(c.macro_name, 44) .. "  " .. c.value .. "") end
-            add("")
-            add("enum {")
-            for _, c in ipairs(consts) do add("    " .. c.enum_name .. " = " .. c.macro_name .. ",") end
-            add("};")
-            add("")
-        end
-    end
-    add("#pragma endregion " .. basename)
-    add("")
-    return table.concat(lines, "\n") .. "\n"
-end
-
--- ============================================================
--- Process one source
--- ============================================================
-
-local function process_source(source_path, word_counts)
-    local source    = read_file(source_path)
-    local atoms_raw = find_atoms(source)
-
-    if #atoms_raw == 0 then
-        -- io.stderr:write("  note: no MipsAtom_ declarations in " .. source_path .. "\n")
-        return
-    end
-
-    local atoms_data = {}
-    for _, atom in ipairs(atoms_raw) do
-        local labels, branches, total = scan_atom_body(atom.body, word_counts)
-        local offsets = compute_offsets(labels, branches)
-        table.insert(atoms_data, {
-            name = atom.name,
-            total_words = total,
-            offsets = offsets
-        })
-    end
-
-    local basename = basename_no_ext(source_path)
-    local out_dir  = dirname(source_path) .. "/gen"
-    ensure_dir(out_dir)
-    local out_path = out_dir .. "/" .. basename .. ".offsets.h"
-    write_file(out_path, generate_header(source_path, atoms_data))
-
-    local total_branches = 0
-    for _, a in ipairs(atoms_data) do total_branches = total_branches + #a.offsets end
-    print("  " .. basename .. ": " .. #atoms_data .. " atom(s), " .. total_branches .. " branch(es)")
-    for _, a in ipairs(atoms_data) do
-        for _, r in ipairs(a.offsets) do
-            print("    " .. a.name .. " -> " .. r.tag .. ":" .. r.target .. " : " .. r.offset)
-        end
-    end
-end
-
--- ============================================================
--- Main
--- ============================================================
-
-local function main(args)
-    if #args < 2 then
-        print("Usage: luajit gen_atom_offsets.lua <metadata.h> <source1> [source2 ...]")
-        os.exit(1)
-    end
-    local word_counts = load_word_counts(args[1])
-    for i = 2, #args do process_source(args[i], word_counts) end
-end
-
-main({...})

scripts/update_deps.ps1

@@ -4,25 +4,24 @@ $path_code      = join-path $path_root 'code'
 $path_scripts   = join-path $path_root 'scripts'
 $path_toolchain = join-path $path_root 'toolchain'
 
+# Halt on any error (instead of PowerShell's default `Continue`).
+$ErrorActionPreference = 'Stop'
+
 $misc = join-path $PSScriptRoot 'helpers/misc.ps1'
 . $misc
 
-# TODO(Ed): Review usage of these deps
-# I orgiinally cloned them when starting to get to the C runtime usage of the course
-# However, based on the heavy reliance of the PSX.Dev extension I might fallback; also
-# The gdb server doesn't need the full repo and were only using the src/mips 
-# which has a standalone repo (nuggets)
-# armips may not be used at all but I'm not sure...
-
 $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'
 
 $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'
 
 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
 
@@ -37,3 +36,83 @@ pop-location
 # $path_pcsx_redux_binaries   = join-path $path_pcsx_redux_vsprojects 'x64/Release'
 
 # $psyq_obj_parser = join-path $path_pcsx_redux_binaries 'psyq-obj-parser.exe'
+
+# ════════════════════════════════════════════════════════════════════════════
+# PCSX-Redux — built via MSBuild (VS2022)
+# Requires: Visual Studio 2022 with the C++ desktop workload.
+# Output: toolchain\pcsx-redux\vsprojects\x64\Debug\pcsx-redux.exe
+# ════════════════════════════════════════════════════════════════════════════
+
+# Locate MSBuild from the VS2022 install (no hardcoded path — uses vswhere).
+$vswhere = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
+if (-not (Test-Path $vswhere)) {
+	write-error "vswhere not found at '$vswhere'. Install Visual Studio 2022 with the C++ desktop workload."
+	exit 1
+}
+$msbuild_exe = & $vswhere -latest -products * -requires Microsoft.Component.MSBuild -find "MSBuild\**\Bin\MSBuild.exe" 2>$null | Select-Object -First 1
+if (-not $msbuild_exe) {
+	write-error "MSBuild not found via vswhere. Install Visual Studio 2022 with the C++ desktop workload."
+	exit 1
+}
+
+$path_pcsx_sln = join-path $path_pcsx_redux 'vsprojects\pcsx-redux.sln'
+& $msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
+
+# Locate luajit via scoop. `luajit.exe` is on PATH via scoop's shim;
+# we use `scoop prefix` to find the install root for the include dir (needed to compile lpeg against luajit's headers).
+# If scoop or luajit is missing, fail fast with an actionable message.
+$luajit_prefix = & scoop prefix luajit 2>$null
+if (-not $luajit_prefix -or -not (Test-Path (Join-Path $luajit_prefix 'bin/luajit.exe'))) {
+	write-error "luajit not found via 'scoop prefix luajit'. Install via: scoop install luajit"
+	exit 1
+}
+
+# Discover the luajit include dir by globbing `include/luajit-*`.
+# This avoids hardcoding a specific version (e.g. `luajit-2.1`).
+$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
+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
+}
+
+# Generate lpeg.dll by compiling the 6 source files directly.
+# `gcc` is on PATH (scoop's shim puts it there).
+# The source files: lpcap.c lpcode.c lpcset.c lpprint.c lptree.c lpvm.c
+# Link against luajit's import library (`libluajit-5.1.a`) for the Lua C API symbols (lua_*, luaL_*).
+$luajit_lib_dir = Join-Path $luajit_prefix 'lib'
+$lpeg_sources = @('lpcap.c', 'lpcode.c', 'lpcset.c', 'lpprint.c', 'lptree.c', 'lpvm.c')
+$lpeg_compile_args = @(
+	'-O2', '-shared',
+	"-I$lua_inc_dir",
+	"-L$luajit_lib_dir",
+	'-o', 'lpeg.dll'
+) + $lpeg_sources + @('-lluajit-5.1')
+push-location $path_lpeg
+& gcc @lpeg_compile_args
+pop-location
+
+# ════════════════════════════════════════════════════════════════════════════
+# lfs (LuaFileSystem) — compiled from pcsx-redux's vendored luafilesystem source.
+# Source: toolchain/pcsx-redux/third_party/luafilesystem/src/lfs.c
+# Output: toolchain/lfs/lfs.dll
+# ════════════════════════════════════════════════════════════════════════════
+
+$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_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
+& gcc -O2 -shared "-I$lua_inc_dir" -o $lfs_dll $lfs_src $lfs_dll_import
+
+# ════════════════════════════════════════════════════════════════════════════
+# OpenBIOS — built from the PCSX-Redux source tree via make + mipsel-none-elf
+# Output: toolchain\pcsx-redux\src\mips\openbios\openbios.bin
+# ════════════════════════════════════════════════════════════════════════════
+
+$path_openbios = join-path $path_pcsx_redux 'src\mips\openbios'
+push-location $path_openbios
+&	make clean
+&	make
+pop-location