Curation pass: reduce nested conditional branching in some defnitions.

Still reviewing. Need to see if gte is handled properly.

.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": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
-				"load hello_psyq.elf",
+				"load hello_gte.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
 				"tbreak main",
 				"continue"
 			]
@@ -36,6 +47,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -44,10 +59,16 @@
 			"osx": {
 				"gdbpath": "gdb"
 			},
-			"executable": "${workspaceRoot}/build/hello_gpu.elf",
+			"executable": "${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_gpu.elf",
+				"load hello_gte.elf",
 				"tbreak main",
 				"continue"
 			]
@@ -60,6 +81,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -69,12 +94,55 @@
 				"gdbpath": "gdb"
 			},
 			"executable": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
 				"load hello_gte.elf",
 				"tbreak main",
 				"continue"
 			]
+		},
+		{
+			"name": "Debug: Hello GTE Psy-Q! (atoms debug — DWARF-injected)",
+			"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",
+				"source build/gen/hello_gte.gdbinit",
+				"tbreak main",
+				"continue"
+			]
 		}
 	]
 }
+

code/duffle/atom_dsl.h

@@ -3,34 +3,18 @@
  * ============================================================================
  *
  * 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.
  *
- *   WHAT THIS HEADER IS
+ * Pure macro anntation.
- *   -------------------
+ * ---------------
- *   The metaprogram (scripts/passes/annotation.lua) reads source-as-written
+ * Don't want to constraint the macro usage to some attribute placment constraint, etc, don't want ot dela with the compiler.
- *   and validates:
+ * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip_over each expand to a C comment or to nothing
- *     - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X),
+ * (C preprocessor strips them to whitespace).
- *       atom_reads(...), atom_writes(...)) in any order. All optional.
- *       (No phase token for now; phases may be reintroduced later.)
- *     - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a
- *       real Binds_* struct declaration.
- *     - wave-context positions only reference the canonical 4-register
- *       set: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
- *     - atom word-counts in word_counts.metadata.h agree with the body's
- *       actual .word count.
- *
- *   WHY A PURE MACRO (atom_info, atom_bind, atom_reads, atom_writes, atom_label)
- *   -----------------------------------------------------------------
- *   Each of these expands to a C comment or to nothing. The C preprocessor
- *   strips them to whitespace. The metaprogram reads the literal token from
- *   source-as-written, NOT from the preprocessed output. This means:
- *     - the C compiler does no work for them (no __attribute__, no
- *       _Pragma, no asm side-effects)
- *     - they can never silently drift from the metaprogram's view
- *       (the metaprogram re-reads the source on every build)
- *     - the annotation is invisible to the linker, debugger, and IDE
  *
  * ============================================================================
- *
  *  Usage:
  *      MipsAtom_(cube_tri) atom_info(
  *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
@@ -60,30 +44,15 @@
  *
  *  Annotation rules
  *  ----------------
- *      1. atom_info(...) is OPTIONAL. Most atoms have no annotation.
+ *      1. atom_info(...) is OPTIONAL. Atoms without atom_info are silently skipped by the metaprogram.
- *         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)
- *      2. If present, atom_info takes up to three sub-calls, all
+ *           - atom_reads(...)
- *         order-independent within the arg list:
+ *           - atom_writes(...)
- *           - atom_bind(Binds_X) (optional; only for rbind atoms)
+ *      3. atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
- *           - atom_reads(...) (optional; wave-context registers)
+ *      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.
- *           - atom_writes(...) (optional; wave-context registers)
+ *      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.
- *      3. atom_bind(Binds_X) pins the ABI-struct shape -- the metaprogram
- *         cross-references Binds_X against the
- *         `typedef struct Binds_X { ... } Binds_X;` declaration.
- *
- *      4. atom_reads(...) and atom_writes(...) args are wave-context
- *         registers: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
- *         Closed set. GTE / SP / DMA / I/O state is declared in source
- *         comments, not in atom_reads/atom_writes.
- *
- *      5. atom_label(name) is an anchor -- the macro is empty in C; the
- *         metaprogram records the marker at the current pos for offset
- *         calculation.
- *
- *      6. atom_offset(F, T) is resolved by gen/atom_offsets.h, generated
- *         from the atom_label markers.
  */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -92,64 +61,69 @@
 #endif
 
 /* ============================================================================
- * WAVE-CONTEXT REGISTERS -- canonical register set for the tape wave model.
+ * atom_reads(...) / atom_writes(...)
- *
- *   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)
- *
- * Closed set. 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 atom_reads/atom_writes.
- *
- * ============================================================================*/
-
-/* ============================================================================
- * atom_reads(...) / atom_writes(...) -- wave-context register list
- *
- *   atom_reads(R_PrimCursor, R_FaceCursor)
- *         ->  (R_PrimCursor, R_FaceCursor)   // comma-evaluated, discarded
- *
- * The macro produces a comma-evaluated expression that the C compiler
- * silently discards (it sits in an unused arg position -- the result is
- * never bound). The Lua tool pattern-matches the "atom_reads(...)" /
- * "atom_writes(...)" token to extract the list.
- *
- * You can have at most one atom_reads(...) and at most one atom_writes(...)
- * in an atom_info(...) call. To declare multiple disjoint sets (rare), just
- * declare the union -- the metaprogram doesn't track which reads need which
- * writes at this granularity.
  *
+ * 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 ANNOTATION MACROS
+ *   atom_info :
- *
- *   atom_info -- single unified annotation. OPTIONAL. Most atoms have none.
- *
  *      MipsAtom_(cube_tri) atom_info(
  *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
  *       , atom_writes(R_PrimCursor, R_FaceCursor)
  *      ){ ... };
- *
+ * 
- *   Shape (sub-args order-independent; all optional):
+ *  - atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
- *     - atom_bind(Binds_X):  at most one; pins the ABI-struct shape
+ *  - atom_reads(...):    comma-list of registers
- *     - atom_reads(...):     at most one; comma-list of wave-context registers
+ *  - atom_writes(...):   comma-list of registers
- *     - atom_writes(...):    at most one; comma-list of wave-context registers
- *
- *   No phase token for now. The metaprogram doesn't check ordering across
- *   atoms -- phases (init / bind / setup / work / commit / terminate) will
- *   be reintroduced when ordering checks are added.
- *
- *   The macro expands to a C comment (or to nothing). The C compiler does
- *   no work. The metaprogram reads the source-as-written directly.
- *
  * ============================================================================*/
 #define atom_info(...)  /* atom_info(__VA_ARGS__) */
 
+/* ----------------------------------------------------------------------------
+ * DEBUG SOURCE-STEP MARKERS
+ *
+ * Place atom_dbg_skip_over() 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.
+ * ----------------------------------------------------------------------------*/
+#define atom_dbg_skip_over() /* atom_dbg_skip_over: 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_over`).
+ * 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
  *
@@ -158,13 +132,7 @@
  *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
  *   ){ ... };
  *
- * The Binds_X MUST be a typedef'd type (declared via
+ * The Binds_X MUST be a typedef'd type (declared via `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
- * `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
- * The Lua tool cross-references this. Missing struct = error.
- *
- * atom_bind is a SUB-CALL of atom_info, not a standalone annotation macro.
- *
- * The macro expands to a C comment. The metaprogram reads source-as-written.
  * ----------------------------------------------------------------------------*/
 #define atom_bind(binds_struct)  /* atom_bind(binds_struct) */
 
@@ -177,25 +145,11 @@
  *
  *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
  *
- * The metaprogram generates gen/atom_offsets.h with one
+ * The metaprogram generates gen/atom_offsets.h with one #define with the offset value per atom_offset(F, T) call.
- *   #define atom_offset__culling__bounds_chk  ((target - branch_pos - 1))
+ * The preprocessor then expands the call to the right immediate value.
- * 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)
  *
+ * 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.
  * ============================================================================*/
-
 #define atom_offset(F, T) atom_offset_ ## F ## _ ## T
-/* atom_label is a pure annotation for the metaprogram's offset calculations.
+// 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/gen/duffle.macs.h

@@ -23,7 +23,6 @@ WORD_COUNT(mac_yield, 4)
 ,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
 WORD_COUNT(mac_load_tri_indices, 3)
 
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
 #define mac_gte_load_tri_verts(...) \
 	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
 ,	add_u_self(R_AT, R_VertBase) \
@@ -77,7 +76,7 @@ WORD_COUNT(mac_insert_ot_tag_g4, 11)
 
 #define mac_pack_color_word(off, cmd, r, g, b) \
 	load_upper_i(R_AT, (cmd) << 8  | (b)) \
-,	or_i_self(   R_AT, ((g)   << 8) | (r)) \
+,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
 ,	store_word(  R_AT, R_PrimCursor, (off))
 WORD_COUNT(mac_pack_color_word, 3)
 

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
@@ -41,15 +39,14 @@
 /* ============================================================================
  *  Hardware MMIO Addresses
  *  ============================================================================
- *
  *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2
  *  0x1F800000+. GP0 (offset 0x10) is the data port (commands + params).
  *  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 $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
+ *   `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
  * Mirrors the `IO_BASE_ADDR equ 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s. */
 enum {
     IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
@@ -75,12 +72,10 @@ 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.
- *  The layer-1 bitfield-layout constants live in the same enum block
+ *  The layer-1 bitfield-layout constants live in the same enum block so the encoder can reference them by name. 
- *  so the encoder can reference them by name. 
+ *  NO macro body past this point uses a raw shift or raw mask.
- *  NO macro body past this point uses a raw shift or raw mask. 
+ *  Every shift/width/mask is named here, named once.
- *  Every shift/width/mask is named here, named once. 
  *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
  * ============================================================================ */
 enum {
@@ -143,9 +138,7 @@ enum {
 /* ============================================================================
  *  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,
- *  and shift it to its own position. 
  *  Mirrors `enc_op` / `enc_rs` / `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.
@@ -186,10 +179,9 @@ enum {
 /* ============================================================================
  *  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,
@@ -202,10 +194,9 @@ enum {
     gp1_cmd_VerticalDisplayRange   = 0x07,
     gp1_cmd_DisplayMode            = 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").
      * Bit positions match the encoder shifts below; values are the
@@ -259,8 +250,7 @@ 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))
 
@@ -282,7 +272,6 @@ enum {
 /* ============================================================================
  *  Pre-baked GPU state words
  *  ============================================================================
- *
  *  Common command words for boot-time GPU init and standard display configurations.
  * ============================================================================ */
 
@@ -356,7 +345,6 @@ 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.
  *
@@ -390,9 +378,9 @@ typedef Struct_(PolyTag) {
  * No raw C-style casts. RHS values are assumed to be `U4` — caller passes a `U4` directly. */
 #define set_len(tag,v)  (C_(PolyTag_R,tag)->len  = u4_(v))
 #define set_addr(tag,v) (C_(PolyTag_R,tag)->addr = u4_(v))
-/* `set_code` is no longer in the new PolyTag design — the code byte lives
+/* `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 `set_poly_*` setters, 
+ *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters,
  *  which set both the tag's length and the code. */
 #define get_len(tag)  C_(U4,C_(PolyTag_R,tag)->len)
 #define get_addr(tag) C_(U4,C_(PolyTag_R,tag)->addr)
@@ -511,7 +499,6 @@ 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)
@@ -575,7 +562,6 @@ typedef Struct_(TexturePage) { U4 raw; };
 /* ============================================================================
  *  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)
@@ -608,7 +594,6 @@ 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)
@@ -626,9 +611,8 @@ 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.
  * ============================================================================ */
 enum {
     tim_file_id_magic = 0x10,
@@ -659,35 +643,24 @@ 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 canonical 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 (word_counts.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 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.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

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 canonical 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
  *

code/duffle/lottes_tape.h

@@ -14,16 +14,14 @@ typedef U4 const MipsCode;
 typedef Slice_(MipsCode);
 typedef Slice_MipsCode MipsAtom;
 
-#define MipsAtom_(sym) MipsCode tmpl(code,sym) [] align_(4) =
+#define MipsAtom_(sym) MipsCode sym [] align_(4) =
 
-// Bare form: file-scope declaration with hardcoded body.
 // 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) =
 
-// Function form: function-body block that returns a MipsAtom slice.
 // Used for components with value-args (e.g., ac_format_f3_color).
 //   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
 // expands to:
@@ -34,16 +32,16 @@ typedef Slice_MipsCode MipsAtom;
 
 /* Register aliases */
 enum {
-	R_AtomJmp    = R_T9,
+	R_AtomJmp    = R_T9 atom_reg,  /* debug-visible; tape yield handshake scratch */
-	R_TapePtr    = R_T8,  /* The Instruction Stream Pointer */
+	R_TapePtr    = R_T8 atom_reg,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,  /* Input data cursor */
+	R_InCursor   = R_T4,
 
-	R_PrimCursor = R_T7,  /* VRAM output cursor (primitive buffer) */
+	R_PrimCursor = R_T7 atom_reg atom_type(U4 *),    /* VRAM output cursor (primitive buffer) */
-	R_FaceCursor = R_T4,  /* Input data cursor (indices/faces) */
+	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,  /* Base address of the vertex array */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2 *), /* Base address of the vertex array */
-	R_OtBase     = R_T6,  /* Base address of the Ordering Table */
+	R_OtBase     = R_T6 atom_reg atom_type(U4 *),    /* Base address of the Ordering Table */
 
-/* Stringification codes for the GCC inline assembler clobber lists */
+/* Stringification codes for the GCC inline assembler clobber lists. */
 #define R_TapePtr_Code     R_T8_Code
 #define R_InCursor_Code    R_T4_Code
 
@@ -62,7 +60,7 @@ enum {
 MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
 /* Generalized Tape Engine Runner */
-FI_ void tape_run(Slice_U4 tape) { register U4* tp rgcc(R_TapePtr) = tape.ptr; asm volatile(
+FI_ void tape_run(Slice_MipsCode tape) { register U4* tp rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
 			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
 		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
@@ -89,13 +87,13 @@ FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start
 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))
+#define tb_emit_(tb, atom) tb_emit(tb, atom)
 FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
 
-FI_ Slice_U4 tb_end  (TapeBuilder* tb) { tb_emit(tb,code_tape_exit); return (Slice_U4){ C_(U4*,tb->ptr), tb->used }; }
+FI_ Slice_MipsCode tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_MipsCode){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Slice_U4){ C_(U4*,tb.ptr),  tb.used }; }
+FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Slice_MipsCode){ C_(U4*,tb.ptr),  tb.used }; }
-#define tb_scope(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,code_tape_exit))
+#define tb_scope(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,tape_exit))
 
 #pragma endregion Tape Drive
 
@@ -106,11 +104,11 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
  * ---------------------------------------------------------------------------*/
 
 // The 'Yield' sequence for Tape Atoms (mac_yield). 
+atom_dbg_skip_over()
 MipsAtomComp_(ac_yield) {
 	load_word(R_AtomJmp, R_TapePtr, 0),
 	add_ui_self(         R_TapePtr, S_(MipsCode)),
-	jump_reg( R_AtomJmp),
+	jump_reg( R_AtomJmp), nop,
-	nop,
 };
 
 /* Words: 3; Loads 3 S2 indices from the face array */
@@ -121,6 +119,7 @@ MipsAtomComp_(ac_load_tri_indices) {
 };
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
+atom_dbg_skip_over()
 MipsAtomComp_(ac_gte_load_tri_verts) {
 	shift_lleft(R_AT, R_T0, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
 	shift_lleft(R_AT, R_T1, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
@@ -162,13 +161,12 @@ MipsAtomComp_(ac_insert_ot_tag_g4) {
 FI_ MipsAtom ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
 MipsAtomComp_Proc_(ac_pack_color_word, {
 	load_upper_i(R_AT, (cmd) << 8  | (b)),
-	or_i_self(   R_AT, ((g)   << 8) | (r)),
+	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).
+ * 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. */
 FI_ MipsAtom ac_format_f3_color(U1 r, U1 g, U1 b)
 MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
 
@@ -237,9 +235,8 @@ 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}
@@ -269,8 +266,7 @@ internal MipsAtom_(mips_flush_icache) {
 	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 
+	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
-	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) 

code/duffle/memory.h

@@ -103,7 +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.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))
@@ -404,12 +380,9 @@ 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 +414,7 @@ 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
@@ -472,22 +445,18 @@ enum { _BitOffsets = 0
 
 /* load_imm_2w — unconditional 2-word `li` form: `lui` + (ori | addi).
  *
- * Granular companion to `load_imm`: skips the compile-time range checks
+ * Granular companion to `load_imm`: skips the compile-time range checks and always emits 2 .words. Use this when:
- * and always emits 2 .words. Use this when:
  *   - you know `imm` is > 0xFFFF (otherwise you're wasting a word), OR
  *   - `imm` is not a compile-time constant and you want predictable
  *     2-word emission without the `__builtin_constant_p` branches.
  *
  * 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 
- * force a specific encoding for a known discontiguous high/low pair),
+ * (e.g. to force a specific encoding for a known discontiguous high/low pair),
  * see `load_imm_2w_ori_forced` and `load_imm_2w_addi_forced` below.
- *
  * Statement-level (not expression-level): emits its own `asm volatile(...)`.
  */
 #define load_imm_2w(rt, imm) do {                         \
@@ -518,9 +487,8 @@ 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 +500,17 @@ 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 +550,8 @@ 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 canonical 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
  *

code/duffle/word_count.metadata.h

@@ -5,11 +5,10 @@
 // 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
+// This file is encoding-macros-only. 
-// macros (mac_X) live in duffle/gen/<dir>.macs.h (included separately
+// The auto-generated component macros (mac_X) live in duffle/gen/<dir>.macs.h (included separately by the unity build).
-// by the unity build). The unity build should include THIS file and
+// The unity build should include THIS file and the .macs.h file in the same TU, with both wrapped 
-// the .macs.h file in the same TU, with both wrapped (or the
+// (or the include guard order handled) to avoid WORD_COUNT redeclaration.
-// 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.)

code/gte_hello/gen/gte_hello.offsets.h

@@ -15,9 +15,9 @@ enum {
     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
 
 enum {

code/gte_hello/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"
@@ -99,8 +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;
@@ -207,6 +212,8 @@ 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)
 	{
@@ -259,9 +266,8 @@ 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));
@@ -269,10 +275,10 @@ 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);
 		}
@@ -344,48 +350,40 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		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);
-			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); 
@@ -394,7 +392,6 @@ 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;
 	}
 }
 

code/gte_hello/hello_gte_tape.c

@@ -22,27 +22,22 @@ typedef Struct_(Binds_CubeTri) {
 	V3_S2* VertBase;
 	U4*    OtBase;
 };
-internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri)
+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_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)), 
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)), 
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)), 
+	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
- *  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),
- *  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.
- */
-internal 
-MipsAtom_(cube_g4_face) atom_info(
 		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
 		atom_writes(R_PrimCursor, R_FaceCursor)
 ){
@@ -94,52 +89,52 @@ typedef Struct_(Binds_FloorTri) {
 	V3_S2* VertBase;
 	U4*    OtBase;
 };
-internal 
+internal
-MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri)
+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_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)), 
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)), 
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)), 
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
 	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
 	mac_yield()
 };
 
-internal 
+// atom_dbg_skip_over()
-MipsAtom_(floor_f3_face) atom_info(
+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)
+	, atom_writes(R_PrimCursor, R_FaceCursr)
 ) {
 	mac_load_tri_indices(  R_T0, R_T1, R_T2),
 	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
-	nop2, gte_cmdw_nclip,
+	gte_cmdw_nclip,
- 
+
 	/* Culling (Branch forward if Backface) */
-	nop2, gte_mv_from_data_r(R_T0, C2_MAC0),
+	gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
-	branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop,
 	/* Format Primitive */
 	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
 	mac_gte_store_f3_post_rtpt(),
-	
+
 	/* Calculate Depth */
-	nop2, gte_avg_sort_z3,            
+	gte_avg_sort_z3,
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),      
+	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),   
+	add_ui(      R_AT, R_0,  OrderingTbl_Len),
-	set_lt_u(    R_AT, R_T1, R_AT), 
+	set_lt_u(    R_AT, R_T1, R_AT),
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop, 
+	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
 	/* Insert into Ordering Table Linked List */
 	mac_insert_ot_tag_f3(),
 	add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
 		// Note(Ed): No bounds checking, should be checked before atom runs.
 
 /* Advance Input Cursor & Yield (Both branch targets land here) */
-atom_label(floor_f3_face_exit) 
+atom_label(floor_f3_face_exit)
 	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
 	mac_yield()
 };
@@ -150,7 +145,7 @@ internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitive
 	, atom_writes(R_TapePtr)
 ){
 	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
-	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),      
+	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

readme.md

@@ -27,6 +27,13 @@ scoop install luajit
 
 * 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.
 
 * Used as the runtime sandbox emulated the ps1

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

@@ -168,28 +168,18 @@ if arg and arg[1] then
 	end
 
 	-- Accept either a directory or a file path. Directory args are
-	-- expanded via `dir /b *.lua` (Windows) or `ls *.lua` (Unix).
+	-- expanded via lfs.dir (native, no subprocess).
+	local lfs = require("lfs")
 	local function is_dir(p)
-		local f = io.open(p, "r")
+		return lfs.attributes(p, "mode") == "directory"
-		if    f then f:close() return false end
-		return true
 	end
 	local function list_lua(dir)
 		local out = {}
-		local cmd
+		if not is_dir(dir) then return out end
-		if package.config:sub(1, 1) == "\\" then
+		for entry in lfs.dir(dir) do
-			cmd = 'dir /b "' .. dir .. '\\*.lua" 2>nul'
+			if entry:match("%.lua$") then
-		else
+				out[#out + 1] = dir .. "/" .. entry
-			cmd = 'ls -1 "' .. dir .. '"/*.lua 2>/dev/null'
-		end
-		local p = io.popen(cmd)
-		if p then
-			for line in p:lines() do
-				if line:match("%.lua$") then
-					out[#out + 1] = dir .. "/" .. line
-				end
 			end
-			p:close()
 		end
 		return out
 	end

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]  $link_module,
@@ -243,6 +230,52 @@ function make-binary { param([string]$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 build-hello_psyqo {
 	$includes += @()
 
@@ -317,33 +350,16 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function ps1-meta { param(
-		[Parameter(Mandatory=$true)][string[]]$sources,
-		[Parameter(Mandatory=$true)][string]$metadata,
-		[string]$out_root = (join-path $path_build 'gen'),
-		[string[]]$passes = @('--all')
-	)
-	$script = join-path $path_scripts 'ps1_meta.lua'
-	write-host "ps1-meta  $($sources.Count) source(s), passes=$($passes -join ',')" ` -ForegroundColor Magenta
-	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root)
-	foreach ($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 build-gte_hello {
 	$includes += @()
 
 	$path_module        = join-path $path_code   'gte_hello'
 	$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'
 
-	$source_dirs  = @($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
-	ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -359,7 +375,6 @@ function build-gte_hello {
 
 	# 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 = @()
@@ -378,11 +393,70 @@ function build-gte_hello {
 	$link_args += $f_debug
 	# $link_args += $f_optimize_size
 	$link_modules = @(
-		$module_asm_crt, 
+		$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)
+
+	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
+	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
+	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
+	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
+	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
+	{
+		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
+		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
+		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
+		$f_args = @(
+			"--update-section=.debug_line=$dwarfLineBin",
+			"--update-section=.debug_aranges=$dwarfArangesBin",
+			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
+		)
+		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
+		if ($LASTEXITCODE -ne 0) {
+			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
+			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
+			return;
+		}
+
+		$dwarfInfoBin     = join-path $path_build_gen 'hello_gte.dwarf_info.bin'
+		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_gte.dwarf_abbrev.bin'
+		$dwarfStrBin      = join-path $path_build_gen 'hello_gte.dwarf_str.bin'
+		$dwarfLocBin      = join-path $path_build_gen 'hello_gte.dwarf_loc.bin'
+		$dwarfLoclistsBin = join-path $path_build_gen 'hello_gte.dwarf_loclists.bin'
+		$g_args = @(
+			"--update-section=.debug_info=$dwarfInfoBin",
+			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
+			"--update-section=.debug_str=$dwarfStrBin",
+			"--add-section=.debug_loc=$dwarfLocBin",
+			"--add-section=.debug_loclists=$dwarfLoclistsBin"
+		)
+		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
+		if ($LASTEXITCODE -ne 0) {
+			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
+			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
+			return;
+		}
+
+		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
+		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
+		# The original ELF and PS-EXE remain byte/flag unchanged.
+		& $Objcopy `
+			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
+			--set-section-flags ".data=alloc,load,data,code,contents" `
+			$injectElf 2>&1 | Out-Null
+		if ($LASTEXITCODE -ne 0) {
+			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
+			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
+		} 
+		else {
+			Write-Host "[build] DWARF-injected ELF: $injectElf"
+		}
+	}
 }
 build-gte_hello
 

scripts/duffle_paths.lua

@@ -11,11 +11,10 @@
 ---   ```
 ---
 --- That small bootstrap: (a) locates this helper via `arg[0]` / `debug.getinfo`,
---- (b) loads it (which sets `package.path` + `package.cpath` via cached `git rev-parse`),
+--- (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.
 ---
---- Replaces the prior 2-line (entry) or 4-line (pass) pattern that had the call site do its own path resolution + duplicated setup.
 
 local M = {}
 
@@ -27,50 +26,25 @@ local CACHE_KEY = "__duffle_repo_root__"
 --- parent of the directory containing this script. We derive it directly from `debug.getinfo(1, "S").source`
 --- (returns `@<path>` for the currently-running chunk).
 ---
---- Replaces the prior `io.popen("git rev-parse --show-toplevel")` approach, which cost ~100-180ms per
+--- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source),
---- LuaJIT process on Windows due to git's CLI startup. The path-derive approach costs <1ms.
+--- return nil and let `M.setup()` fail loud.
----
---- If this script's path can't be parsed (shouldn't happen — dofile/debug.getinfo always populates source),
---- fall back to a defensive walk: starting from this script's directory, walk UP until we find a parent that
---- contains a `scripts/` directory. The first match is the repo root.
 --- @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/` (with trailing slash or not).
+	-- What remains is the directory containing this script, i.e. `<repo>/scripts/`.
 	local scripts_dir = source and source:match("^@?(.*)[/\\]duffle_paths%.lua$")
-	if scripts_dir then
+	if not scripts_dir then return nil end
-		-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
-		local root = scripts_dir:gsub("scripts[\\/]?$", "")
-		root = root:gsub("\\", "/")
-		if root == "" then root = "./" end
-		if not root:match("/$") then root = root .. "/" end
-		package.loaded[CACHE_KEY] = root
-		return root
-	end
 
-	-- Defensive fallback: walk UP from this script's directory until we find a parent that contains `scripts/`.
+	-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
-	-- In practice this branch never fires — debug.getinfo always returns a source for dofile()'d chunks.
+	local root = scripts_dir:gsub("scripts[\\/]?$", "")
-	local lfs = pcall(require, "lfs") and require("lfs") or nil
+	root = root:gsub("\\", "/")
-	if lfs then
+	if root == "" then root = "./" end
-		local dir = source and source:match("^@?(.*[/\\])") or "./"
+	if not root:match("/$") then root = root .. "/" end
-		dir = dir:gsub("\\", "/")
+	package.loaded[CACHE_KEY] = root
-		while dir and dir ~= "" do
+	return root
-			local candidate_scripts = dir .. "scripts"
-			if lfs.attributes(candidate_scripts, "mode") == "directory" then
-				dir = dir:gsub("/$", "")
-				package.loaded[CACHE_KEY] = dir .. "/"
-				return dir .. "/"
-			end
-			local parent = dir:match("^(.*)/[^/]+/$")
-			if not parent then break end
-			dir = parent .. "/"
-		end
-	end
-
-	return nil
 end
 
 --- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
@@ -83,7 +57,12 @@ end
 function M.setup()
 	local repo_root = find_repo_root()
 	if not repo_root then
-		io.stderr:write("[duffle_paths] git rev-parse failed -- not in a git repo?\n")
+		-- 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
 

scripts/elf_dwarf.lua

@@ -0,0 +1,794 @@
+--- elf_dwarf.lua — ELF32 + DWARF + atoms source-map utilities.
+--- All ELF32 + DWARF-specific code lives here.
+---
+--- **What this module contains:**
+---   - **Format-constant tables** (the byte-offset / opcode / size encyclopedias for ELF32, DWARF4 aranges, DWARF5 rnglists, DWARF line-program, MIPS).
+---     Every constant carries a spec:` comment naming the spec section that defines it.
+---   - **I/O helpers**: little-endian byte read/write, ELF32 section walker, nm symbol reader, source-map parser, native directory glob.
+---
+--- **Conventions:** tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Native dependencies
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- lfs is wired into package.cpath by `duffle_paths.lua` (vendored under `toolchain/lfs/lfs.dll`). 
+local lfs = require("lfs")
+
+local M = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- DWARF tag + form constants
+-- ════════════════════════════════════════════
+-- (DWARF5 §7.5.5 "Tag Encodings" + Table 7.1; gcc emits these exact values for the DWARF3-extension and DWARF5 line units.)
+
+M.DW_TAG = {
+	compile_unit       = 0x11,
+	subprogram         = 0x2E,
+	variable           = 0x34,
+	structure_type     = 0x13,
+	member             = 0x0D,
+	base_type          = 0x24,
+	typedef            = 0x2A,
+	pointer_type       = 0x0F,
+	const_type         = 0x26,
+	volatile_type      = 0x27,
+	inlined_subroutine = 0x1D,
+	-- We index the canonical gcc-emitted tags. Anything else falls through.
+}
+
+M.DW_AT = {
+	name                  = 0x03,
+	low_pc                = 0x11,
+	high_pc               = 0x12,
+	language              = 0x13,
+	location              = 0x02,
+	comp_dir              = 0x1B,
+	byte_size             = 0x0B,
+	encoding              = 0x3E,
+	data_member_location  = 0x38,
+	type                  = 0x49,
+	linkage_name          = 0x6E,
+	external              = 0x3F,
+	abstract_origin       = 0x31,
+	call_file             = 0x58,
+	call_line             = 0x59,
+	inline                = 0x20,
+	decl_file             = 0x3A,
+	decl_line             = 0x3B,
+}
+
+M.DW_FORM = {
+	addr        = 0x01,
+	data1       = 0x0B,
+	data2       = 0x05,
+	data4       = 0x06,
+	string      = 0x08,
+	strp        = 0x0E,
+	exprloc     = 0x18,
+	ref4        = 0x13,
+	udata       = 0x0F,
+	ref_sig8    = 0x20,
+	implicit_const = 0x21,
+	flag_present = 0x19,
+	sec_offset  = 0x17,
+}
+
+M.DW_ATE = {
+	address       = 0x01,
+	boolean       = 0x02,
+	complex_float = 0x03,
+	float         = 0x04,
+	signed        = 0x05,
+	signed_char   = 0x06,
+	unsigned      = 0x07,
+	unsigned_char = 0x08,
+}
+
+-- DWARF5 §7.5.6 DW_FORM_implicit_const
+local DW_FORM_implicit_const = 0x21
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Format-constant tables
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- ----------------------------------------------------------------------------
+-- MIPS sizes
+-- ----------------------------------------------------------------------------
+
+--- spec: MIPS o32 ABI §"Register Usage" — 32-bit general-purpose registers
+M.MIPS_BYTES_PER_WORD = 0x04
+
+
+-- ----------------------------------------------------------------------------
+-- ELF32 (System V ABI gABI v1.2)
+-- ----------------------------------------------------------------------------
+--- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors,
+--- and section-relative values are zero-based wire offsets. Only Lua string APIs receive
+--- a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+---
+--- ELF/DWARF field offsets are expressed in hex so they map directly to the
+--- zero-based byte positions in the binary file.
+
+
+--- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
+M.ELF32 = {
+	magic_offset          = 0x00,         -- 4-byte magic "\127ELF" at file offset 0x00
+	magic                 = "\127ELF",
+	class_offset          = 0x04,         -- 1-byte; 1 = ELF32, 2 = ELF64
+	class_elf32           = 1,
+	endian_offset         = 0x05,         -- 1-byte; 1 = little-endian, 2 = big-endian
+	endian_little         = 1,
+	header_bytes          = 0x34,         -- spec: gABI v1.2 §"ELF Header" — ELF32 header is 52 bytes total
+	e_shoff_offset        = 0x20,         -- 4-byte LE; section-header table file offset
+	e_shentsize_offset    = 0x2E,         -- 2-byte LE; section-header entry size in bytes
+	e_shnum_offset        = 0x30,         -- 2-byte LE; number of section headers
+	e_shstrndx_offset     = 0x32,         -- 2-byte LE; index of section-name string table
+	sh_size_bytes         = 0x28,         -- spec: gABI v1.2 §"Section Header Table" — each entry is 40 bytes
+	sh_name_offset        = 0x00,         -- 4-byte LE; offset into .shstrtab
+	sh_type_offset        = 0x04,         -- 4-byte LE; section type (SHT_*)
+	sh_offset_offset      = 0x10,         -- 4-byte LE; section's file offset
+	sh_size_offset        = 0x14,         -- 4-byte LE; section's size in bytes
+	dw_dwarf32_terminator = 0xFFFFFFFF,   -- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
+}
+
+-- ----------------------------------------------------------------------------
+-- DWARF4 .debug_aranges (per DWARF5 spec §7.4 — Address Range Table)
+-- ----------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets.
+
+--- spec: DWARF5 spec §7.4 (Address Range Table) — 32-bit DWARF form
+M.DWARF4_ARANGES = {
+	unit_length_offset = 0x00,    -- 4-byte LE; length of unit body (excludes these 4 bytes)
+	version_offset     = 0x04,    -- 2-byte LE; expected = 2
+	cu_offset_offset   = 0x06,    -- 4-byte LE; CU DIE offset in .debug_info
+	addr_size_offset   = 0x0A,    -- 1-byte;  expected = 4 (32-bit MIPS)
+	seg_size_offset    = 0x0B,    -- 1-byte;  expected = 0
+	entry_size         = 0x08,    -- 4-byte addr + 4-byte length (per §7.4)
+	terminator_size    = 0x08,    -- 8 zero bytes (per §7.4 end-of-list marker)
+	version_expected   = 2,
+	addr_size_expected = 4,
+	seg_size_expected  = 0,
+}
+
+-- ----------------------------------------------------------------------------
+-- DWARF5 .debug_rnglists (per DWARF5 spec §2.17 + §7.21)
+-- ----------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets.
+
+--- spec: DWARF5 spec §2.17 + §7.21 (Range List Table) — 32-bit DWARF form
+M.DWARF5_RNGLISTS = {
+	unit_length_offset    = 0x00,    -- 4-byte LE
+	version_offset        = 0x04,    -- 2-byte LE; expected = 5
+	addr_size_offset      = 0x06,    -- 1-byte;  expected = 4
+	seg_size_offset       = 0x07,    -- 1-byte;  expected = 0
+	offset_count_offset   = 0x08,    -- 4-byte LE; expected = 0
+	first_entry_offset    = 0x0C,
+	end_of_list           = 0x00,    -- spec: DWARF5 §7.7 — DW_RLE_end_of_list byte value
+	start_length          = 0x07,    -- spec: DWARF5 §7.7 — DW_RLE_start_length byte value
+	version_expected      = 5,
+	addr_size_expected    = 4,
+	seg_size_expected     = 0,
+	offset_count_expected = 0,
+}
+
+-- ----------------------------------------------------------------------------
+-- DWARF line-program opcodes (per DWARF5 spec §6.2.5)
+-- ----------------------------------------------------------------------------
+-- Opcode VALUES stay in decimal — they're identifiers (DW_LNS_copy = 1), not binary positions.
+-- Compare to the *_offset fields above which are hex.
+
+--- spec: DWARF5 spec §6.2.5 (Line Number Program Opcodes)
+M.DWARF_LINE_OPS = {
+	-- Standard opcodes (§6.2.5.2)
+	DW_LNS_extended     = 0,    -- spec: §6.2.5.2 — extended opcode marker byte
+	DW_LNS_copy         = 1,
+	DW_LNS_advance_pc   = 2,
+	DW_LNS_advance_line = 3,
+	DW_LNS_set_file     = 4,
+	DW_LNS_negate_stmt  = 6,    -- spec: §6.2.5.2 — toggle the line-state is_stmt register
+	-- Extended sub-opcodes (§6.2.5.3)
+	DW_LNE_end_sequence = 1,    -- spec: §6.2.5.3
+	DW_LNE_set_address  = 2,    -- spec: §6.2.5.3
+	-- Standard opcode header (§6.2.5.1)
+	-- opcode_base + line_range are 1-byte header fields; hex so they map
+	-- directly to their position in the line-program header byte sequence.
+	-- line_base stays signed decimal (=-5) since 0xFB obscures the spec semantics.
+	opcode_base              = 0x0D,
+	line_base                = -5,
+	line_range               = 0x0E,
+	-- Extended opcode payload sizes (include the sub-opcode byte; §6.2.5.3)
+	-- Hex so they match the byte positions in the line-program wire format.
+	end_sequence_payload_size = 0x01,   -- size = sub_opcode only
+	set_address_payload_size  = 0x05,   -- size = sub_opcode(1) + addr(4)
+}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- I/O helpers: little-endian byte read/write
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read a 4-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
+--- Equivalent to `string.unpack("<I4", buf, off + 1)` but avoids the table-return shape + works under LuaJIT 2.1
+--- (which has partial `string.unpack` coverage).
+---
+--- **Convention:** `off` is a zero-based wire offset; `+ 1` is applied only at the `string.byte` boundary.
+---
+--- **Byte weights** are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
+--- byte 0 contributes its value directly; byte 1 is shifted left by 8
+--- (= 0x100); byte 2 by 16 (= 0x10000); byte 3 by 24 (= 0x1000000).
+--- @param buf string
+--- @param off integer  -- zero-based wire offset
+--- @return integer
+function M.read_u32_le(buf, off)
+	local byte_off = off + 1
+	return buf:byte(byte_off)
+		+ buf:byte(byte_off + 0x01) * 0x00000100
+		+ buf:byte(byte_off + 0x02) * 0x00010000
+		+ buf:byte(byte_off + 0x03) * 0x01000000
+end
+
+--- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
+--- (`off` is zero-based; `+ 1` is applied only at the `string.byte` boundary.)
+--- @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
+
+-- Pure-Lua 5.3 LEB128 readers (no `bit` library). `2^shift` arithmetic matches the existing parser.
+-- Offsets are 0-based; returns (value, next_pos).
+-- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua
+-- can import them as file-scope locals per the 2nd-caller lift precedent
+-- (the uleb128 + sleb128 encoders were promoted the same way).
+function M.read_uleb128_at(buf, pos)
+	local value, shift = 0, 0
+	local len = #buf
+	while pos < len do
+		local b = buf:byte(pos + 1)
+		value = value + (b % 0x80) * (2 ^ shift)
+		shift = shift + 7
+		pos   = pos + 1
+		if b < 0x80 then return value, pos end
+	end
+	return nil, pos
+end
+
+function M.read_sleb128_at(buf, pos)
+	local value, shift = 0, 0
+	local len = #buf
+	while pos < len do
+		local b = buf:byte(pos + 1)
+		value = value + (b % 0x80) * (2 ^ shift)
+		shift = shift + 7
+		pos   = pos + 1
+		if b < 0x80 then
+			if b >= 0x40 then value = value - (2 ^ shift) end
+			return value, pos
+		end
+	end
+	return nil, pos
+end
+
+-- Find the 0-based offset of the table-terminator byte (a single 0) for the abbrev table starting at `table_start`.
+-- Returns nil on truncated input. Walks declaration headers
+-- (code, tag, has_children, attr/form pairs, DW_FORM_implicit_const constant) until it finds a 0 byte that follows a complete declaration.
+function M.find_abbrev_table_end(table_bytes, table_start)
+	local pos, len = table_start, #table_bytes
+	if    pos >= len or table_bytes:byte(pos + 1) == 0 then return pos end
+	while pos < len do
+		local  _code, code_end = M.read_uleb128_at(table_bytes, pos)
+		if not _code then return nil end
+		pos = code_end
+		local  _tag, tag_end = M.read_uleb128_at(table_bytes, pos)
+		if not _tag then return nil end
+		pos = tag_end
+		if pos >= len then return nil end
+		pos = pos + 1   -- has_children byte
+		while pos < len do
+			local  attr, attr_end = M.read_uleb128_at(table_bytes, pos)
+			if not attr then return nil end
+			pos = attr_end
+			local  form, form_end = M.read_uleb128_at(table_bytes, pos)
+			if not form then return nil end
+			pos = form_end
+			if attr == 0 and form == 0 then break end
+			if form == DW_FORM_implicit_const then
+				local _c, ce = M.read_sleb128_at(table_bytes, pos)
+				if not _c then return nil end
+				pos = ce
+			end
+		end
+		if pos >= len then return nil end
+		if table_bytes:byte(pos + 1) == 0 then return pos end
+	end
+	return nil
+end
+
+-- Read the null-terminated C string at 0-based offset `off` in `buf`.
+-- Stops at the first 0 byte or end of buffer.
+local function read_c_string_at(buf, off)
+	local len = #buf
+	local start = off
+	while off < len and buf:byte(off + 1) ~= 0 do off = off + 1 end
+	return buf:sub(start + 1, off)
+end
+
+-- Walk the .debug_abbrev table starting at 0-based offset `table_start` and return a list of declarations:
+-- {code, tag, has_children, attrs={ {name, form}, ... }}.
+-- Stops at the table terminator.
+local function parse_abbrev_table(table_bytes, table_start)
+	local table_end = M.find_abbrev_table_end(table_bytes, table_start)
+	if not table_end then return nil, "no terminator" end
+	local decls = {}
+	local pos = table_start
+	while pos < table_end do
+		local  code, code_end = M.read_uleb128_at(table_bytes, pos)
+		if not code then return nil, "truncated code" end
+		pos = code_end
+		local  tag, tag_end = M.read_uleb128_at(table_bytes, pos)
+		if not tag then return nil, "truncated tag" end
+		pos = tag_end
+		local has_children = table_bytes:byte(pos + 1)
+		pos = pos + 1
+		local attrs = {}
+		while true do
+			local  attr, attr_end = M.read_uleb128_at(table_bytes, pos)
+			if not attr then return nil, "truncated attr" end
+			pos = attr_end
+			local form, form_end = M.read_uleb128_at(table_bytes, pos)
+			if not form then return nil, "truncated form" end
+			pos = form_end
+			if attr == 0 and form == 0 then break end
+			attrs[#attrs + 1] = { name = attr, form = form }
+			if form == DW_FORM_implicit_const then
+				local  _c, ce = M.read_sleb128_at(table_bytes, pos)
+				if not _c then return nil, "truncated const" end
+				pos = ce
+			end
+		end
+		decls[#decls + 1] = { code = code, tag = tag, has_children = has_children, attrs = attrs }
+	end
+	return decls
+end
+
+-- Read a ULEB attribute value at 0-based offset `pos` for the given `form`.
+-- Returns (value, next_pos). For DW_FORM_string we return the inline string.
+-- For DW_FORM_strp we return the inline string resolved from `str_buf`.
+-- For DW_FORM_ref4 we return the absolute CU-relative offset.
+-- The caller decides whether to interpret that as a section offset.
+local function read_form_value(buf, str_buf, pos, form)
+	if form == M.DW_FORM.addr then
+		return M.read_u32_le(buf, pos), pos + 4
+	elseif form == M.DW_FORM.string then
+		local s = read_c_string_at(buf, pos)
+		return s, pos + #s + 1
+	elseif form == M.DW_FORM.strp then
+		-- DW_FORM_strp: 4-byte offset into .debug_str.
+		local strp_off = M.read_u32_le(buf, pos)
+		return read_c_string_at(str_buf, strp_off), pos + 4
+	elseif form == M.DW_FORM.udata then return M.read_uleb128_at(buf, pos)
+	elseif form == M.DW_FORM.data1 then return buf:byte(pos + 1), pos + 1
+	elseif form == M.DW_FORM.data2 then return M.read_u16_le(buf, pos), pos + 2
+	elseif form == M.DW_FORM.data4 then return M.read_u32_le(buf, pos), pos + 4
+	elseif form == M.DW_FORM.ref4  then return M.read_u32_le(buf, pos), pos + 4
+	elseif form == M.DW_FORM.sec_offset then
+		-- DW_FORM_sec_offset: 4-byte offset (size depends on DWARF version;
+		-- on DWARF5 32-bit it's always 4 bytes).
+		return M.read_u32_le(buf, pos), pos + 4
+	elseif form == M.DW_FORM.flag_present then
+		return 1, pos
+	elseif form == M.DW_FORM.exprloc then
+		-- DW_FORM_exprloc: ULEB byte count + that many bytes of DW_OP_*.
+		local  len, ne = M.read_uleb128_at(buf, pos)
+		if not len then return nil, pos end
+		return nil, ne + len
+	elseif form == DW_FORM_implicit_const then
+		-- The constant is declared in the abbrev; no value bytes in the DIE.
+		return nil, pos
+	elseif form == M.DW_FORM.ref_sig8 then
+		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): an 8-byte value identifying a type
+		-- by signature. The low 4 bytes (LE) are the type signature (content hash);
+		-- the high 4 bytes (LE) are a CU-relative offset into the matching type unit.
+		-- Consumers use the low 4 to look up the type unit (see M.find_type_unit_by_signature)
+		-- then the high 4 to resolve the specific type within it.
+		-- Return the low 4 as the primary value to preserve the (value, next_pos) shape;
+		-- the high 4 is exposed via M.read_ref_sig8 (which returns both halves).
+		local _, _, next_pos = M.read_ref_sig8(buf, pos)
+		return M.read_u32_le(buf, pos), next_pos
+	else
+		return nil, pos
+	end
+end
+
+--- Read a `DW_FORM_ref_sig8` value at 0-based offset `pos` from `buf`.
+--- Returns the low 4 bytes (LE) as `low`, the high 4 bytes (LE) as `high`, and the cursor position after the 8-byte value as `next_pos`.
+--- Callers that need the full type-unit + type-offset pair (e.g. to resolve a type identifier embedded as a signature)
+--- should use this directly rather than going through `read_form_value`,
+--- which only exposes the low 4 bytes to preserve its existing (value, next_pos) return shape.
+--- @param buf string
+--- @param pos integer  -- zero-based wire offset
+--- @return integer  -- low 4 bytes (LE), the type signature
+--- @return integer  -- high 4 bytes (LE), the offset within the matching type unit
+--- @return integer  -- cursor after the 8-byte value
+function M.read_ref_sig8(buf, pos)
+	return M.read_u32_le(buf, pos), M.read_u32_le(buf, pos + 4), pos + 8
+end
+
+-- DWARF5 §7.5.6 (Type Entries).
+-- Walk all units in `info` and return the 0-based offset of the first unit
+-- whose `DW_AT_type_signature` (8-byte value at the end of the unit header) equals `target_sig`.
+-- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
+-- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
+--
+-- Unit header layout (from pos 0):
+--   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
+--   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
+-- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
+-- @param info string  -- the .debug_info section bytes
+-- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
+-- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
+-- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
+function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
+	local pos = 0
+	local section_len = #info
+	while pos + 4 < section_len do
+		local unit_length = M.read_u32_le(info, pos)
+		if unit_length == 0xFFFFFFFF then
+			return nil, nil  -- DWARF64 not supported
+		end
+		-- unit_length is the body size, NOT including the 4-byte unit_length field itself.
+		local body_start = pos + 4
+		local body_end   = body_start + unit_length
+		if body_end > section_len then
+			return nil, nil  -- malformed
+		end
+		-- Per DWARF5 §7.5.6, the type_unit (DW_UT_type = 0x02) body layout is:
+		--   0:  version (2)
+		--   2:  unit_type (1) -- DW_UT_type = 0x02
+		--   3:  address_size (1)
+		--   4:  debug_abbrev_offset (4)
+		--   8:  type_signature (8)
+		--   16: type_offset (4)
+		--   20: <children>
+		if body_end - body_start >= 20 then
+			-- read_ref_sig8 / write_u32_le / etc. are 1-indexed (string:byte);
+			-- pos / body_start / body_end are 0-based wire offsets, so the
+			-- 1-indexed byte at 0-based wire offset X is string:byte(X + 1).
+			-- Per DWARF5 §7.5.6, the type_unit body is laid out as:
+			--   byte 0-1: version (2)
+			--   byte 2:   unit_type (1) -- DW_UT_type = 0x02
+			--   byte 3:   address_size (1)
+			--   byte 4-7: debug_abbrev_offset (4)
+			--   byte 8-15: type_signature (8)
+			--   byte 16-19: type_offset (4)
+			local unit_type = info:byte(body_start + 2 + 1)  -- 0-based +2 = unit_type in 1-indexed
+			if unit_type == 0x02 then  -- DW_UT_type
+				local sig_lo, sig_hi, _ = M.read_ref_sig8(info, body_start + 8)  -- 0-based +8 = type_signature in 1-indexed
+				if sig_lo == target_sig_lo and sig_hi == target_sig_hi then
+					local type_offset = M.read_u32_le(info, body_start + 16)  -- 0-based +16 = type_offset in 1-indexed
+					return pos, type_offset
+				end
+			end
+		end
+		-- Advance to the next unit (the 4-byte unit_length + the body).
+		pos = body_end
+	end
+	return nil, nil
+end
+
+--- Return a 4-byte little-endian byte string for `value`.
+--- Caller concatenates with `..` if composing multi-word blobs.
+--- **Byte weights** written as `0x100` etc. (see `M.read_u32_le` for rationale).
+--- @param value integer  -- 0 ≤ value ≤ 0xFFFFFFFF
+--- @return string
+function M.write_u32_le(value)
+	return string.char(
+		value                          % 0x00000100,
+		math.floor(value / 0x00000100) % 0x00000100,
+		math.floor(value / 0x00010000) % 0x00000100,
+		math.floor(value / 0x01000000) % 0x00000100)
+end
+
+--- Return a 2-byte little-endian byte string for `value`.
+--- @param value integer  -- 0 ≤ value ≤ 0xFFFF
+--- @return string
+function M.write_u16_le(value)
+	return string.char(value % 0x00000100, math.floor(value / 0x00000100) % 0x00000100)
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- I/O helpers: ELF32 / DWARF / symbols
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read the named sections from a post-link ELF32 by walking the ELF32 section-header table directly
+--- (no subprocess; lfs only for the existence check). Returns `{[name] = bytes_or_empty_string, ...}`.
+---
+--- **Convention:** ELF/DWARF offsets are zero-based wire offsets. Direct Lua string APIs add `+ 1` at the boundary.
+--- Every requested name has an entry in the returned dict;
+--- missing sections have an empty string (NOT nil) so callers can do `sections[".debug_x"] or ""` for the missing case.
+---
+--- **Cost:** one file open + one `f:seek` + one `f:read` per section header 
+--- (we walk all `e_shnum` headers regardless of how many names are requested, to find the .shstrtab first).
+--- For frequent callers, pass the union of all needed sections in one call.
+--  Can add `.debug_info` + `.debug_loc` + `.debug_str_offsets` to the list without writing a 2nd ELF walker.
+--- @param elf_path Path
+--- @param section_names string[]  -- list of section names to read
+--- @return table<string, string>
+function M.read_elf_sections(elf_path, section_names)
+	-- Initialize result with all requested names set to "" so callers can do `sections[X] 
+	-- or ""` for missing sections without nil-checks.
+	local result = {}
+	for _, name in ipairs(section_names) do result[name] = "" end
+
+	-- O(1) lookup set.
+	local wanted = {}
+	for _, name in ipairs(section_names) do wanted[name] = true end
+
+	-- Existence check (lfs.attributes avoids an io.open-vs-fail race).
+	if lfs.attributes(elf_path, "mode") ~= "file" then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] ELF not found: %s\n", elf_path))
+		return result
+	end
+
+	local f = io.open(elf_path, "rb")
+	if not f then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] io.open failed: %s\n", elf_path))
+		return result
+	end
+
+	-- Read the ELF32 header.
+	local header = f:read(M.ELF32.header_bytes)
+	if not header or #header < M.ELF32.header_bytes then
+		io.stderr:write("[elf_dwarf.read_elf_sections] ELF too small for ELF32 header\n")
+		f:close()
+		return result
+	end
+
+	-- Sanity-check magic + class + endianness.
+	if header:sub(M.ELF32.magic_offset + 1, M.ELF32.magic_offset + 0x04) ~= M.ELF32.magic then
+		io.stderr:write("[elf_dwarf.read_elf_sections] not an ELF file\n")
+		f:close()
+		return result
+	end
+	if header:byte(M.ELF32.class_offset + 1) ~= M.ELF32.class_elf32 then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] not ELF32 (class=%d)\n", header:byte(M.ELF32.class_offset + 1)))
+		f:close()
+		return result
+	end
+	if header:byte(M.ELF32.endian_offset + 1) ~= M.ELF32.endian_little then
+		io.stderr:write("[elf_dwarf.read_elf_sections] not little-endian; unsupported\n")
+		f:close()
+		return result
+	end
+
+	-- Parse section-header table location + dimensions from the header.
+	local e_shoff     = M.read_u32_le(header, M.ELF32.e_shoff_offset)
+	local e_shentsize = M.read_u16_le(header, M.ELF32.e_shentsize_offset)
+	local e_shnum     = M.read_u16_le(header, M.ELF32.e_shnum_offset)
+	local e_shstrndx  = M.read_u16_le(header, M.ELF32.e_shstrndx_offset)
+
+	-- Read the section-header string table (.shstrtab) so we can resolve section names from their `sh_name` offsets.
+	f:seek("set", e_shoff + e_shstrndx * e_shentsize)
+	local strtab_hdr = f:read(e_shentsize)
+	if not strtab_hdr or #strtab_hdr < e_shentsize then
+		io.stderr:write("[elf_dwarf.read_elf_sections] could not read .shstrtab header\n")
+		f:close()
+		return result
+	end
+	local strtab_offset = M.read_u32_le(strtab_hdr, M.ELF32.sh_offset_offset)
+	local strtab_size   = M.read_u32_le(strtab_hdr, M.ELF32.sh_size_offset)
+	f:seek("set", strtab_offset)
+	local strtab = f:read(strtab_size) or ""
+
+	-- Walk all section headers; collect (offset, size) for the wanted names.
+	local function read_section_bytes(sh_offset, sh_size)
+		f:seek("set", sh_offset)
+		return f:read(sh_size) or ""
+	end
+
+	for sh_idx = 0, e_shnum - 1 do
+		f:seek("set", e_shoff + sh_idx * e_shentsize)
+		local sh = f:read(e_shentsize)
+		if not sh or #sh < e_shentsize then break end
+		local sh_name   = M.read_u32_le(sh, M.ELF32.sh_name_offset)
+		local sh_offset = M.read_u32_le(sh, M.ELF32.sh_offset_offset)
+		local sh_size   = M.read_u32_le(sh, M.ELF32.sh_size_offset)
+
+		-- Extract the name (null-terminated C string in strtab).
+		local name_end = strtab:find("\0", sh_name + 1, true) or (sh_name + 1)
+		local name     = strtab:sub(sh_name + 1, name_end - 1)
+		if wanted[name] then
+			result[name] = read_section_bytes(sh_offset, sh_size)
+		end
+	end
+
+	f:close()
+	return result
+end
+
+--- Read ELF symbol addresses by walking the `.symtab` + `.strtab` sections directly (no `nm` subprocess). 
+--- Returns a map `{name -> {addr, size_bytes}}` for every `code_<name>` symbol.
+---
+--- **Conventions:**
+--- - ELF32 symtab entry = 16 bytes (`st_name:4 + st_value:4 + st_size:4 + st_info:1 + st_other:1 + st_shndx:2`); offsets within each entry are zero-based wire offsets.
+--- - Direct Lua `string.byte`/`string.sub`/`string.find` boundaries receive `+ 1`.
+--- - We filter on STB_GLOBAL (high nibble of st_info = 1) to match `nm`'s default (external symbols only). STB_WEAK excluded.
+--- - The `code_` prefix is stripped (MipsAtom_ macros emit bare atom names, no `code_` prefix).
+--- - `st_size > 0` filter excludes undefined/imported symbols.
+--- @param elf_path Path
+--- @return table<string, {integer, integer}>
+function M.read_nm(elf_path)
+	local addrs = {}
+
+	-- Read .symtab + .strtab via the existing ELF walker (no subprocess).
+	local sections = M.read_elf_sections(elf_path, {".symtab", ".strtab"})
+	local symtab   = sections[".symtab"]
+	local strtab   = sections[".strtab"]
+	if not symtab or not strtab or #symtab == 0 or #strtab == 0 then
+		-- No symbol table (e.g. stripped ELF). Return empty.
+		return addrs
+	end
+
+	-- Iterate the 16-byte ELF32 symtab entries.
+	-- Each entry (zero-based): st_name at 0, st_value at 4, st_size at 8, st_info at 12, st_other at 13, st_shndx at 14.
+	local SYM_ENTRY_BYTES = 0x10
+	local SYM_ST_NAME     = 0x00
+	local SYM_ST_VALUE    = 0x04
+	local SYM_ST_SIZE     = 0x08
+	local SYM_ST_INFO     = 0x0C
+	local n_syms = #symtab / SYM_ENTRY_BYTES
+	for i = 0, n_syms - 1 do
+		local entry_off  = i * SYM_ENTRY_BYTES
+		local st_info    = symtab:byte(entry_off + SYM_ST_INFO + 1)
+		-- High nibble = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
+		-- Use math.floor(/16) instead of bit.rshift for LuaJIT 2.1 compat
+		-- (LuaJIT's `>>` is 5.3+, but math.floor(x/16) works on all versions).
+		local binding    = math.floor(st_info / 16)
+		if binding == 0 or binding == 1 then            -- STB_LOCAL or STB_GLOBAL
+			local st_size   = M.read_u32_le(symtab, entry_off + SYM_ST_SIZE)
+			if st_size > 0 then
+				local st_name_off = M.read_u32_le(symtab, entry_off + SYM_ST_NAME)
+				-- Extract the name from .strtab (null-terminated C string).
+				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
+				local name     = strtab:sub(st_name_off + 1, name_end - 1)
+				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — MipsAtom_ macros strip the `code_` prefix).
+				-- The atoms_source_map pass already filters out non-atom symbols via the source-map.txt cross-ref.
+				if name and #name > 0 then
+					local st_value = M.read_u32_le(symtab, entry_off + SYM_ST_VALUE)
+					addrs[name] = { st_value, st_size }
+				end
+			end
+		end
+	end
+
+	return addrs
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- LEB128 encoders (Unsigned + Signed Little-Endian Base 128)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- DWARF uses LEB128 to encode variable-length integers in its wire format (line-program opcodes, DW_AT values, etc.).
+-- Both encoders pack 7 bits of data per byte + 1 bit of "more bytes follow" signaling.
+--
+-- Per-byte layout:
+--     bit:   7 6 5 4 3 2 1 0
+--           │ └───── 7-bit data ─────┘
+--           └─ continuation flag (LEB_CONT_BIT = 0x80)
+--
+-- For SLEB128 (signed), bit 6 of the 7-bit data is the sign bit that the
+-- decoder uses for sign extension:
+--     bit 6 = 0 → value is positive (or zero); zero-extend on decode
+--     bit 6 = 1 → value is negative;            one-extend on decode
+--
+-- The signed encoder must emit the MINIMUM number of bytes whose final 7-bit payload already has the correct sign bit set 
+-- (otherwise the decoder would round-trip to a different value).
+--
+-- Spec: DWARF5 §7.6 "Variable-Length Data" / Appendix C.
+
+-- Top bit of each LEB128 byte. Set if more bytes follow in the encoding.
+local LEB_CONT_BIT = 0x80
+
+-- Low 7 bits of each LEB128 byte. The actual data payload.
+local LEB_DATA_MASK = 0x7F
+
+-- Bit 6 of the 7-bit data (i.e. 0x40). For SLEB128: the sign-bit position used by the decoder for sign extension.
+-- Encoders MUST stop when the next byte would be redundant AND the sign bit in the last byte matches the value's sign.
+local SLEB_SIGN_BIT = 0x40
+
+--- ULEB128 (Unsigned Little-Endian Base 128) encoder. Returns the byte string for the non-negative integer `n`.
+--- Algorithm:
+---   - Extract the low 7 bits of `n` (LEB_DATA_MASK = 0x7F).
+---   - Shift `n` right by 7 bits.
+---   - If more bytes remain, OR in the continuation flag (LEB_CONT_BIT).
+---   - Repeat until `n` is fully consumed.
+--- @param n integer  -- non-negative
+--- @return string
+function M.uleb128(n)
+	if n == nil or type(n) ~= "number" then
+		io.stderr:write("[elf_dwarf.uleb128] got " .. type(n) .. ": " .. tostring(n) .. "\n")
+		io.stderr:write(debug.traceback() .. "\n")
+		error("uleb128 requires non-negative number")
+	end
+	assert(n >= 0, "uleb128 requires non-negative input")
+	local bytes = {}
+	repeat
+		local b = n % (LEB_DATA_MASK + 1)      -- extract low 7 bits
+		n = (n - b) / (LEB_DATA_MASK + 1)      -- shift right by 7 bits
+		if n > 0 then b = b + LEB_CONT_BIT end -- set continuation bit if more bytes follow
+		bytes[#bytes + 1] = string.char(b)
+	until n == 0
+	return table.concat(bytes)
+end
+
+--- SLEB128 (Signed Little-Endian Base 128) encoder. Returns the byte string for the integer `n` (may be negative).
+--- Algorithm differs from ULEB128 by the termination condition: 
+--- stop when the remaining bits can be inferred from the sign bit in the last byte's 7-bit data payload.
+---   - If `n == 0`  (no more value bits)   AND bit 6 of the data = 0 → positive terminator (sign bit says "zero-extend").
+---   - If `n == -1` (sign-extended all-1s) AND bit 6 of the data = 1 → negative terminator (sign bit says "one-extend").
+---
+--- Without these checks, the decoder would round-trip to a different value
+--- (e.g. encoding `0` as `0x80 0x00` decodes to `0` correctly but is 2 bytes long; the termination check picks the 1-byte `0x00` form).
+--- @param n integer  -- any integer (negative allowed)
+--- @return string
+function M.sleb128(n)
+	local bytes = {}
+	local more = true
+	while more do
+		local b = n % (LEB_DATA_MASK + 1) -- extract low 7 bits
+		n = (n - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
+		-- Termination: remaining value bits fit in the sign bit of the last byte.
+		if     n == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if     n == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if more then b = b + LEB_CONT_BIT end
+		bytes[#bytes + 1] = string.char(b)
+	end
+	return table.concat(bytes)
+end
+
+--- ULEB128 byte-length: number of bytes the encoder M.uleb128 would produce for `n`.
+--- Used by callers that need to size a buffer before encoding (e.g. compute_loclists_offsets
+--- needs the encoded length of an `uleb128(4)` for a `DW_OP_piece + uleb128(U4_BYTE_SIZE)` tail).
+--- @param n integer  -- non-negative
+--- @return integer  -- 1..5 for n in [0, 2^32)
+function M.uleb128_size(n)
+	assert(n >= 0, "uleb128_size requires non-negative input")
+	if n == 0 then return 1 end
+	local bytes = 1
+	while n >= 0x80 do
+		n = (n - (n % (LEB_DATA_MASK + 1))) / (LEB_DATA_MASK + 1)  -- arithmetic shift right by 7
+		bytes = bytes + 1
+	end
+	return bytes
+end
+
+--- SLEB128 byte-length: number of bytes the encoder M.sleb128 would produce for `n`.
+--- Used by callers that need to size a buffer before encoding.
+--- (e.g. compute_loclists_offsets needs the encoded length of an `sleb128(field.offset)` in a tape piece).
+--- Handles the signed DWARF5 termination: positive terminator if (n == 0) and bit 6 of last byte is unset;
+--- negative terminator if (n == -1) and bit 6 of last byte is set.
+--- @param n integer  -- any integer (negative allowed)
+--- @return integer
+function M.sleb128_size(n)
+	local more = true
+	local bytes = 0
+	local v = n
+	while more do
+		local b = v % (LEB_DATA_MASK + 1) -- extract low 7 bits
+		v = (v - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
+		if     v == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if     v == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if more then b = b + LEB_CONT_BIT end
+		bytes = bytes + 1
+	end
+	return bytes
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- I/O helpers: atoms source-map + native directory glob
+-- ════════════════════════════════════════════════════════════════════════════
+
+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/gen/gdb_tape_atoms_runtime.gdb`.
+# Sourcing that file RE-DEFINES the commands with real implementations.
+#
+# If `build/gen/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
+# 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/gen/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/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document break_atom
+  Set a breakpoint at the start of tape atom <name>. STUB state.
+end
+
+define step_atom
+  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document step_atom
+  Resume execution until the next atom boundary. STUB state.
+end
+
+define next_atom
+  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document next_atom
+  Alias for step_atom. STUB state.
+end
+
+define where_in_atom
+  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document where_in_atom
+  Report current atom name, .rodata addr, word offset, and source line (if known). STUB state.
+end
+
+define stepi_inside_atom
+  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+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/gen/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/gen/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 +1,99 @@
+# 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'
+
+$gdbInitPath = [System.IO.Path]::GetFullPath((Join-Path $PSScriptRoot '..\build\gen\hello_gte.gdbinit'))
+if (-not (Test-Path -LiteralPath $gdbInitPath -PathType Leaf)) {
+    Write-Warning "Generated GDB skip sidecar missing (non-fatal): $gdbInitPath. Run the GTE build to regenerate it; debugger launch will continue without generated skip-over commands."
+}
+
+# ── Pre-checks ──
+foreach ($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

@@ -3,33 +3,27 @@
 --- 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) { ... };`)
 ---
---- Source scanning: done ONCE upstream by `duffle.scan_source()` (ps1_meta.lua pre-scans each
+--- Source scanning: done ONCE upstream by `duffle.scan_source()` (ps1_meta.lua pre-scans each source and stashes the result in `src.scan`).
---- source and stashes the result in `src.scan`). This pass is pure: read from the scan, run
---- checks, emit findings. No source re-walking.
 ---
 --- Writes:
 ---   - `<ctx.out_root>/<dir_basename>.errors.h` — one per module, with `#error` directives on findings (the C compile will surface the error)
----   - The annotations.txt report is rendered by `passes/report.lua` from the per-module results stashed in `ctx.flags._annot_results`
+---   - The annotations.txt report is rendered by `passes/report.lua` from the canonical `corpus.sources_by_dir` projection (re-validating each source via `M.validate()`).
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible
---- Lua 5.3 compatible
 
 -- 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
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local write_file          = duffle.write_file
+local write_file     = duffle.write_file
-local ensure_dir          = duffle.ensure_dir
+local ensure_dir     = duffle.ensure_dir
 
--- Domain tables (single source of truth in duffle.lua).
+-- The annotation pass reads the source-derived registries from scan_source:
-local WAVE_CONTEXT_REGS = duffle.WAVE_CONTEXT_REGS
+--   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
-local TAPE_ATOM_MACROS  = duffle.TAPE_ATOM_MACROS
+--   * pipe_ctx.type_name_registry      — for atom_dbg_reg_default(<T>, ...) and atom_reg_types(<T>, ...) type-identity checks
-
-local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
@@ -51,8 +45,6 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @field project_root       string
 --- @field upstream           table<string, table>
 --- @field flags              table
---- @field flags._annot_results table[]   -- stashed by annotation pass; consumed by report.lua
---- @field dry_run            boolean
 --- @field verbose            boolean
 
 --- @class PassResult
@@ -68,10 +60,36 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @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 SkipOverMarker  -- sub-shape of scan_source.lua's @class SkipOverMarker
+--- @field marker_kind    string                 -- exact marker ident (always "atom_dbg_skip_over")
+--- @field marker_line    integer
+--- @field args           string|nil             -- trimmed text inside the parens (nil when has_parens is false)
+--- @field has_parens     boolean
+--- @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
+--- @field declaration_line integer|nil
 
 --- @class Finding
---- @field line integer  -- source line (or 0 for pass-level)
+--- @field line integer -- source line (or 0 for pass-level)
---- @field msg  string   -- finding message
+--- @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[]
@@ -82,18 +100,408 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @field warnings Finding[]
 --- @field info     Finding[]
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-check functions (the CHECK_RULES table's payload)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Each check has a uniform `append_to_findings` shape (errors[] / warnings[] / info[]).
+-- The dispatcher in `validate()` decides which findings list each check writes to — by convention,
+-- "existence" checks (declaration must exist, struct must exist) write errors[]; "shape" checks (writes/reads must be wave-context) write warnings[].
+-- The `macro_word_drift` check writes both errors[] (missing/mismatch) and info[] (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.
+--- Emitting a warning here keeps the annotation pass from being stop-on-error for the common test-fixture case,
+--- while still surfacing the issue in the report.
+--- The static-analysis report remains the source of truth for build-stopping errors.
+--- @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> -- the 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>) must target a register declared as a debug-visible alias in `pipe_ctx.register_alias_registry`,
+--- with a type name found in `pipe_ctx.type_name_registry`.
+--- Pointer depth is still bounded to 0 or 1. Duplicate defaults are still detected.
+--- @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 must point to a register declared in `pipe_ctx.register_alias_registry`, with a type name found in `pipe_ctx.type_name_registry`. 
+--- The alias ident `R_<n>` now encodes the GPR identity only for entries that are explicitly opted in via the bare `atom_reg` marker.
+--- R_T0..R_T3 are intentionally NOT auto-included (per the prototype principle: no auto-include of wave-context; explicit opt-in only). 
+--- The check fires for any R_T0..R_T3 reference that hasn't been opted in via `#define atom_reg`.
+--- @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 must reference a real Binds_* struct and that struct must declare 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 may not have duplicate field names
+--- (they would defeat the typed-field name lookup that atom_view exposes 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: skip-over markers must satisfy shape + placement constraints.
+--- Walks the priority list once; at most one error is appended per marker so that a single source-level defect does not cascade into multiple findings.
+--- Priority order (first defect wins):
+---   1. has_parens == false         -> requires parentheses: marker()
+---   2. args  ~= ""                 -> takes no arguments
+---   3. superseded_by_marker_line   -> duplicate marker (cite superseding line)
+---   4. pending + no target_kind    -> dangling (no following declaration)
+---   5. unsupported target_kind     -> marker precedes an unrelated declaration
+--- Valid markers before whole-atom / bare-component / proc-component declarations emit no error and remain in src.scan.skip_over.atoms / .components.
+--- @param marker SkipOverMarker
+--- @param _pipe_ctx PipeCtx     -- unused today; kept for plex-shape consistency with per_annot
+--- @param findings Findings
+local function check_skip_marker(marker, _pipe_ctx, findings)
+	local kind = marker.marker_kind
+	local line = marker.marker_line
+
+	if not marker.has_parens then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d requires parentheses: marker()", 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.
+---
+--- R_TapePtr / R_AtomJmp / R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase are the context aliases opted in via `#define atom_reg` in lottes_tape.h.
+--- A source referencing an unregistered R_X emits one pass-level info entry
+--- (emitted only when at least one such rejection lands in this source) tells users where to look.
+---
+--- This check directs raw 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.skip_over.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.
+-- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
--- No source walking; no parsing. The scan was done once upstream.
 
---- Validate one source against its pre-scanned SourceScan payload.
+--- Build the corpus-wide pipe_ctx ONCE per pass run.
+--- Reads the merged `corpus.*` registries (canonical cross-source lookups),
+--- and the corpus-wide `atom_infos` list (preserving source order + duplicates).
+--- The corpus is the source of truth; per-source scans retain body / declaration
+--- ownership via `src.scan` and the per-source `atoms` / `atom_infos` projections.
+---
+--- Canonical ownership: a context without `ctx.shared.corpus` is rejected with an explicit canonical-corpus message.
+--- No per-source fallback synthesis is performed; callers MUST construct a canonical ctx through `build_ctx`.
 --- @param ctx PassCtx
---- @param src SourceFile
+--- @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 + duplicates;
+	-- the per-check `check_unique_annotation` post-rule still flags duplicate annotation
+	-- names within this list. We pre-compute the annot_counts map here so the per_source checks can iterate it without re-walking.
+	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
+
+	-- The pipe_ctx views REFERENCE the corpus tables directly (no copies).
+	-- Every consumer of these fields observes mutations via the canonical corpus without independently mutable registry construction.
+	return {
+		-- Cross-source lookup tables (canonical corpus projections).
+		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 {},
+		-- wc still consumed by check_macro_word_drift; reads from the canonical
+		-- `corpus.word_counts` table (built 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 = self-build (canonical projection).
 --- @return AnnotatedResult
-local function validate(ctx, src)
+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.
@@ -115,141 +523,102 @@ local function validate(ctx, src)
 			binds  = info.binds,
 			reads  = info.reads or {},
 			writes = info.writes or {},
-			errors = {},
+			errors = info.errors,
 		}
 	end
 
-	-- Index atoms by name for lookup.
+	-- Build the per-source pipe_ctx (Fleury: expose structure).
-	local atom_index = {}
+	-- Cross-source visibility comes from `corpus_pipe_ctx`;
-	for _, a in ipairs(atoms) do atom_index[a.name] = a end
+	-- per-source declaration / body ownership comes from `src.scan`.
+	-- pipe_ctx.types / pipe_ctx.atom_views / pipe_ctx.seen_defaults / pipe_ctx.type_occurrences 
+	-- are projected from the per-source scan so the per_source check rules can iterate the source-local occurrences.
+	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
 
-	-- Index binds by name for lookup.
+	local pipe_ctx = {
-	local binds_index = {}
+		atom_index               = {},
-	for _, b in ipairs(scan.binds) do binds_index[b.name] = b end
+		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 {},
+		-- Source-derived registries: still populated from the scan payload as a convenience for callers that want source-local visibility.
+		-- The canonical cross-source lookup tables live in 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
 
-	local errors   = {}
+	-- Findings live in a single struct with three lists (errors / warnings / info).
-	local warnings = {}
+	-- Each check writes to the list appropriate for its severity.
-	local info     = {}
+	local findings = { errors = {}, warnings = {}, info = {} }
 
-	-- 1. Every annotated atom must exist as a real MipsAtom_ declaration.
+	-- Propagate parse-time errors from scan_source's atom_info parsing.
+	-- These are errors found in the atom_info(...) body itself (e.g., malformed args).
+	-- They are pre-existing in the scan payload — we just lift them into our findings list.
 	for _, a in ipairs(annots) do
-		if not atom_index[a.name] then
-			errors[#errors + 1] = {
-				line = a.line,
-				msg  = string.format("annotation for '%s' has no matching MipsAtom_(%s) { ... }", a.name, a.name),
-			}
-		end
 		if a.errors then
 			for _, msg in ipairs(a.errors) do
-				errors[#errors + 1] = {line = a.line, msg = string.format("'%s': %s", a.name, msg)}
+				findings.errors[#findings.errors + 1] = {
-			end
-		end
-	end
-
-	-- 2. Every atom may have AT MOST ONE annotation (no duplicates).
-	-- (Atoms with ZERO annotations are valid in the new minimal shape.)
-	local count_per_atom = {}
-	for _, a in ipairs(annots) do
-		if a.name then
-			count_per_atom[a.name] = (count_per_atom[a.name] or 0) + 1
-		end
-	end
-	for name, n in pairs(count_per_atom) do
-		if n > 1 then
-			errors[#errors + 1] = {
-				line = atom_index[name] and atom_index[name].line or 0,
-				msg  = string.format("MipsAtom_(%s) has %d annotations (expected at most 1)", name, n),
-			}
-		end
-	end
-
-	-- 3. (Phase validity check DROPPED. Phases were removed from the annotation DSL.)
-
-	-- 4. BIND atoms must reference a real Binds_* struct.
-	for _, a in ipairs(annots) do
-		if a.binds then
-			if not binds_index[a.binds] then
-				-- Demoted from error to warning (2026-07-10): the same condition is now caught by passes/static_analysis.lua's
-				-- check_abi_handoff() as an error. Emitting a warning here keeps the annotation pass from being stop-on-error
-				-- for the common test-fixture case, while still surfacing the issue in the report.
-				-- The static-analysis report remains the source of truth for build-stopping errors.
-				warnings[#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),
+					msg  = string.format("'%s': %s", a.name, msg),
 				}
 			end
 		end
 	end
 
-	-- 5. BIND writes must be wave-context registers that match Binds_ fields.
+	-- THE per-annotation pipeline. ONE loop. CHECK_RULES dispatches per_annot rules.
 	for _, a in ipairs(annots) do
-		if a.binds and binds_index[a.binds] then
+		for _, rule in ipairs(CHECK_RULES) do
-			local bs = binds_index[a.binds]
+			if rule.per_annot then rule.per_annot(a, pipe_ctx, findings) end
-
-			for _, f in ipairs(bs.fields) do
-				local candidate = "R_" .. f.name
-				if not is_wave_context_reg(candidate) then
-					warnings[#warnings + 1] = {
-						line = bs.line,
-						msg  = string.format("%s field '%s' doesn't match a known wave-context register (candidate '%s')", a.binds, f.name, candidate),
-					}
-				end
-			end
-
-			for _, w in ipairs(a.writes) do
-				if not is_wave_context_reg(w) then
-					warnings[#warnings + 1] = {
-						line = a.line,
-						msg  = string.format("%s writes '%s' which is not a known wave-context register", a.name, w),
-					}
-				end
-			end
 		end
 	end
 
-	-- 6. INFO reads should be wave-context registers (or R_TapePtr for rbind).
+	-- Post-loop rules (one-shot checks that need full-corpus aggregation in pipe_ctx).
-	for _, a in ipairs(annots) do
+	for _, rule in ipairs(CHECK_RULES) do
-		for _, r in ipairs(a.reads) do
+		if rule.post then rule.post(pipe_ctx, findings) end
-			if not is_wave_context_reg(r) and r ~= "R_TapePtr" then
+	end
-				warnings[#warnings + 1] = {
+
-					line = a.line,
+	-- Per-skip-marker rules.
-					msg  = string.format("atom '%s' reads '%s' which is not a known wave-context register", a.name, r),
+	-- Each raw marker recorded by scan_source (in scan.skip_over.markers) is validated independently;
-				}
+	-- the check emits at most one error per marker.
-			end
+	-- Valid markers stay attached to scan.skip_over.atoms /.components for dwarf_injection.lua consumer.
+	local skip_markers = scan.skip_over and scan.skip_over.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
 
-	-- 7. TAPE_WORDS(mac_X, N) ↔ WORD_COUNT(mac_X, N) drift.
+	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
-	--    Three outcomes: missing (error), mismatch (error), match (info).
+	local wc = corpus_pipe_ctx.word_counts
-	local function check_macro_drift(m, declared)
-		if not declared then
-			errors[#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
-			errors[#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
-		info[#info + 1] = {
-			line = m.line,
-			msg  = string.format("OK: %s = %d words", m.name, m.words),
-		}
-	end
 	for _, m in ipairs(scan.macros) do
-		check_macro_drift(m, ctx.shared.word_counts[m.name])
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_macro then rule.per_macro(m, wc, findings) end
+		end
 	end
 
-	-- 8. Information summary.
+	-- Per-source rules (reg defaults, atom_view layout, compute-register type overrides, Binds_* field uniqueness).
-	info[#info + 1] = {
+	-- 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)",
+		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),
+			, #atoms, #annots, #scan.macros, #scan.binds),
 	}
 
 	return {
@@ -257,9 +626,9 @@ local function validate(ctx, src)
 		annots   = annots,
 		macros   = scan.macros,
 		binds    = scan.binds,
-		errors   = errors,
+		errors   = findings.errors,
-		warnings = warnings,
+		warnings = findings.warnings,
-		info     = info,
+		info     = findings.info,
 	}
 end
 
@@ -270,7 +639,6 @@ end
 --- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
 --- Empty directories (no errors, no atoms) produce no file.
 local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
-	if ctx.dry_run then return nil end
 	if atoms_count == 0 and #errors == 0 then
 		return nil
 	end
@@ -298,17 +666,6 @@ local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sour
 	return out_path
 end
 
---- Stash aggregated per-module results for the report pass to consume.
-local function emit_module_annotations_stub(ctx, dir, dir_basename, atoms_count)
-	ctx.flags         = ctx.flags or {}
-	ctx.flags._annot_results = ctx.flags._annot_results or {}
-	ctx.flags._annot_results[#ctx.flags._annot_results + 1] = {
-		dir          = dir,
-		dir_basename = dir_basename,
-		atoms_count  = atoms_count,
-	}
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- M.run — orchestrator entry
 -- ════════════════════════════════════════════════════════════════════════════
@@ -327,32 +684,33 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- Per-DIRECTORY (per-module) aggregation. Group sources by `src.dir`,
+	-- Build the corpus-wide pipe_ctx ONCE per pass run.
-	-- validate every source in the dir, then emit ONE errors.h per dir.
+	-- The corpus owns the canonical cross-source registries; per-source scans retain body / declaration ownership.
-	-- `ctx.by_dir` is pre-computed in build_ctx (shared across all passes).
+	-- The pipe_ctx is shared across every validate() invocation in this M.run so cross-source visibility is constant.
-	local by_dir = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	local corpus_pipe_ctx = build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared.corpus
+
+	-- Per-DIRECTORY (per-module) aggregation.
+	-- Group sources by `src.dir`, validate every source in the dir, then emit ONE errors.h per dir.
+	-- The corpus owns `sources_by_dir`; this pass reads the corpus bucket directly.
+	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 = {}
-		-- Per-source validate() results, cached for the report pass (it reads from this instead of re-validating each source).
-		ctx.flags                       = ctx.flags or {}
-		ctx.flags._annot_source_results = ctx.flags._annot_source_results or {}
 		for _, src in ipairs(dir_sources) do
-			local result = validate(ctx, src)
+			local result  = validate(ctx, src, corpus_pipe_ctx)
-			result.source = src.path                                  -- tag for downstream rendering
+			result.source = src.path                           -- tag for downstream rendering
-			ctx.flags._annot_source_results[src.path] = result          -- stash so report.lua reads from cache instead of re-running validate()
 			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 }
+				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 }
+				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
 			end
 		end
 
@@ -360,8 +718,6 @@ function M.run(ctx)
 		if err_path then
 			table.insert(outputs, { errors_h = err_path })
 		end
-
-		emit_module_annotations_stub(ctx, dir, dir_basename, dir_atoms)
 	end
 
 	return { outputs = outputs, errors = errors, warnings = warnings }

scripts/passes/atoms_source_map.lua

@@ -0,0 +1,541 @@
+--- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
+---
+--- Reads the canonical `atom.paths` projection produced by the upstream `emission_model` pass.
+--- The ordered `items` stream, dense `word_events`, and `invocations` views are the only semantic inputs to this pass;
+--- it emits one `WORD N LINE L TEXT T` line per emitted `.word`.
+---
+--- **Two output forms** (per the workspace's per-emission-form pattern from
+--- `guide_metaprogram_ssdl.md`):
+---   1. **Canonical text form** — `<out_root>/<basename>.atoms.sourcemap.txt`.
+---      Format-version-tagged for forward-compat. 
+---      Lives in `<out_root>/` (build/gen).
+---      Matches the convention used by `annotation.lua` (`<out_root>/<basename>.errors.h`) + `static_analysis.lua` (`<out_root>/<basename>.static_analysis.txt`).
+---      Compile artifacts (`*.macs.h`, `*.offsets.h`) stay in `<source_dir>/gen/`.
+---   2. **gdb-runtime form** — `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`
+---      (pure gdb command script; addresses pre-computed via `nm`; the 9 user commands defined as `define ... end` blocks).
+---      Emitted ONLY when `ctx.flags.gdb_runtime` is true AND `ctx.flags.elf_path` points to an existing ELF.
+---      The gdb runtime form lets `gdb-multiarch --without-python` users (the common case on Windows MinGW builds)
+---      load the source-map data via `source <path>` — no Python/Tcl/Guile required.
+---
+--- **Output format** (canonical text 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`; they do not appear in
+--- the dense word view and therefore emit no WORD rows.
+---
+--- **Conventions:** tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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   -- canonical source-order corpus
+--- @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`
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Canonical atom-path renderers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Join canonical words to canonical word items. `items` supplies the ordered
+--- word boundaries, while `word_events` supplies call text and source lines.
+--- @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 remains:
+---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>" BODY <line>`
+---   `WORD N  CALL <src-path>:<src-line>  RAW`
+--- Component identity comes from the canonical outermost invocation record;
+--- the count-table lookup is the canonical component declaration witness.
+--- @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 canonical text 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 no Python, no Tcl, no Guile required. 
+--- **Fully hardcoded per-atom** because gdb doesn't do nested `$` substitution in var names
+--- `$__atom_name_$__i` inside a `while` loop is treated as one literal identifier, not a concat.
+---
+--- Each command is a static sequence of `printf` / `tbreak` / `if ... end` blocks.
+--- The Lua pass emits N atoms' worth of lines — no runtime iteration. 
+--- @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 — no Python.
+--- Reads ELF addresses via `mipsel-none-elf-nm -S`, embeds them in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`
+--- so gdb loads the data via `set $var = ...` + `define ... end` blocks at source-time.
+--- @param ctx PassCtx
+local function emit_gdb_runtime(ctx)
+	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
+	local  elf_path = ctx.flags.elf_path
+	if not elf_path or elf_path == "" then
+		io.stderr:write("[atoms_source_map] --gdb-runtime requires --elf <elf>\n")
+		return
+	end
+	if lfs.attributes(elf_path, "mode") ~= "file" then
+		io.stderr:write(string.format(
+			"[atoms_source_map] --gdb-runtime: ELF not found at %s\n", elf_path))
+		return
+	end
+
+	local matched = build_atom_table(ctx)
+	if #matched == 0 then
+		io.stderr:write("[atoms_source_map] --gdb-runtime: no atoms matched against nm symbols (stale scan?).\n")
+		return
+	end
+
+	local lines = {}
+	lines[#lines + 1] = "# Auto-generated by ps1_meta.lua (passes/atoms_source_map.lua)"
+	lines[#lines + 1] = "# DO NOT EDIT — re-run ps1_meta.lua --atoms-source-map --gdb-runtime to regenerate"
+	lines[#lines + 1] = "# Sourced by scripts/gdb/gdb_tape_atoms.gdb (the wrapper)."
+	lines[#lines + 1] = "# Pure gdb scripting — no Python, no Tcl, no Guile required."
+	lines[#lines + 1] = "# Commands are FULLY HARDCODED per-atom because gdb doesn't do nested"
+	lines[#lines + 1] = "# `$` 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 = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	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 = {}
+
+--- Pass entry: emit one `<out_root>/<basename>.atoms.sourcemap.txt` per source file that contains at least one `MipsAtom_(name)` / `MipsCode code_<name>` declaration.
+--- Also emits `<out_root>/<basename>.atoms.provenance.txt`:
+--- per-.word provenance with `mac_X(...)` component resolution back to the component's definition file:line + the per-word body line.
+--- Optionally also emit `<ctx.out_root>/gdb_tape_atoms_runtime.gdb` when `ctx.flags.gdb_runtime` is true.
+--- @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 are owned by `corpus.word_counts`.
+	-- The canonical owner is `corpus.word_counts` (populated by `passes/word_count_eval.lua` + `passes/components.lua`).
+	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
+
+	-- Always emit the canonical text form (per-source).
+	for _, src in ipairs(corpus.source_order) do
+		local has_projection = false
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do
+			if (atom.kind == "atom" or atom.kind == "raw_atom") and atom.paths then
+				has_projection = true; break
+			end
+		end
+		if not has_projection then
+			for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
+				if atom.paths then has_projection = true; break end
+			end
+		end
+		if has_projection then
+			local basename = duffle.basename_no_ext(src.path)
+			-- (1) atoms.sourcemap.txt — format-1 per-word call-site map.
+			local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
+			local sourcemap_body = render_source_map(src)
+			-- (2) atoms.provenance.txt — format-1 per-word definition/body map.
+			local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
+			local prov_body = render_provenance(src, wc)
+			duffle.ensure_dir(duffle.dirname(sourcemap_path))
+			duffle.write_file_lf(sourcemap_path, sourcemap_body)
+			duffle.write_file_lf(prov_path,    prov_body)
+			outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
+			outputs[#outputs + 1] = { kind = "report", path = prov_path }
+		end
+	end
+
+	-- 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

@@ -1,38 +1,27 @@
 --- passes/components.lua — Component-macro header generator.
 ---
 --- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
---- for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations, then does
+--- for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations, then does per-source backward lookups
---- per-source backward lookups for the function-args string (from the preceding `FI_ MipsAtom ac_X(...)`
+--- for the function-args string (from the preceding `FI_ MipsAtom ac_X(...)` function declaration)
---- function declaration) and the preceding comment block (for LSP/IntelliSense signature docs).
+--- and the preceding comment block (for LSP/IntelliSense signature docs).
 ---
---- Emits a per-directory `<dir_basename>.macs.h` containing one `#define mac_X(sig) \` macro per component
+--- Emits a per-directory `<dir_basename>.macs.h` containing one `#define mac_X(sig) \` macro per component + `WORD_COUNT(mac_X, N)`
---- + `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
+--- entries for downstream offset computation.
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
 
---- @class Component
---- @field name    string
---- @field body    string
---- @field args    string|nil
---- @field line    integer
---- @field comment string|nil
-
---- @class M
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- 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
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- 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 _bootstrap_dir  = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle          = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -43,9 +32,9 @@ local ATOM_COMP_PROC  = "MipsAtomComp_Proc_"
 local MIPS_ATOM       = "MipsAtom"  -- 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       = "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      = "mac_" -- prefix on generated macros; the rest is the atom name
 local MAC_PREFIX_LEN  = 4
 
 -- ASCII byte values used in tokenization.
@@ -75,7 +64,6 @@ local GEN_SUBDIR      = "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 dry_run            boolean                -- if true, compute but don't write
 --- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
@@ -84,11 +72,11 @@ local GEN_SUBDIR      = "gen"
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class Component
---- @field name    string        -- atom name (without `ac_` prefix)
+--- @field name    string     -- atom name (without `ac_` prefix)
---- @field body    string        -- brace-delimited body (without the braces)
+--- @field body    string     -- brace-delimited body (without the braces)
---- @field args    string|nil    -- function-args string (function form only)
+--- @field args    string|nil -- function-args string (function form only)
---- @field line    integer       -- source line of the declaration
+--- @field line    integer    -- source line of the declaration
---- @field comment string|nil    -- preceding `/* */` or `//` comment block (signature doc)
+--- @field comment string|nil -- preceding `/* */` or `//` comment block (signature doc)
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Local helpers (file I/O + path normalization)
@@ -97,44 +85,35 @@ local GEN_SUBDIR      = "gen"
 local M = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Function-args extraction (precedes MipsAtomComp_Proc_ invocations)
+-- Back-walk helpers (composed into the 2 entry points below: find_function_args_for + preceding_comment_block)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
--- Returns the position of the open paren, or nil if not found.
--- @param source string
--- @param name string
--- @param before_pos integer
--- @return integer|nil
-local function find_last_name_open_paren(source, name, before_pos)
-	local search     = source:sub(1, before_pos)
-	local name_open  = name .. "("
-	local last_idx   = nil
-	local scan_pos   = 1
-	while true do
-		local found = search:find(name_open, scan_pos, true)  -- plain (no regex)
-		if not found then break end
-		last_idx = found
-		scan_pos = found + #name_open
-	end
-	return last_idx
-end
-
 --- 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_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
 --- We find the LAST occurrence of `"ac_X("` before `before_pos` and extract the args from inside the parens.
---- We then verify the preceding context ends with `MipsAtom` (the function-decl keyword
+--- We then verify the preceding context ends with `MipsAtom`
---- with possible qualifiers between).
+--- (the function-decl keyword with possible qualifiers between).
 ---
---- @param source string
+--- @param source     string
---- @param name string
+--- @param name       string
 --- @param before_pos integer
 --- @return string|nil
 local function find_function_args_for(source, name, before_pos)
-	local last_idx = find_last_name_open_paren(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).
@@ -153,125 +132,63 @@ local function find_function_args_for(source, name, before_pos)
 	return inner
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Preceding-comment-block extraction
--- ════════════════════════════════════════════════════════════════════════════
-
--- Skip whitespace (space/tab/newline/CR) backward from `pos`, returning the position of the first non-whitespace char.
--- @param source string
--- @param pos integer
--- @return integer
-local function skip_ws_backward(source, pos)
-	local back = pos - 1
-	while back > 0 do
-		local ch = source:sub(back, back)
-		if ch == " " or ch == "\t" or ch == "\n" or ch == "\r" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
--- Find the opening `/*` for a block comment whose `*/` ends at `close_pos`.
--- Returns the position of `/`, or nil if not found.
--- @param source string
--- @param close_pos integer  -- position of the closing `*` of `*/`
--- @return integer|nil
-local function find_block_comment_open(source, close_pos)
-	local prefix  = source:sub(1, close_pos - 1)
-	local open_at = nil
-	for scan = #prefix - 1, 1, -1 do
-		if prefix:sub(scan, scan + 1) == "/*" then
-			open_at = scan
-			break
-		end
-	end
-	return open_at
-end
-
--- Walk back from `open_at` over leading spaces + tabs to include the indentation before the `/*` in the captured comment.
--- @param source string
--- @param open_at integer
--- @return integer
-local function extend_left_over_indent(source, open_at)
-	local start = open_at
-	while start > 1 do
-		local ch = source:sub(start - 1, start - 1)
-		if ch == " " or ch == "\t" then
-			start = start - 1
-		else
-			break
-		end
-	end
-	return start
-end
-
--- Walk back from `line_end` to the start of the source line (the most recent `\n` or position 1).
--- @param source string
--- @param line_end integer
--- @return integer
-local function find_line_start(source, line_end)
-	local start = line_end
-	while start > 1 and source:sub(start - 1, start - 1) ~= "\n" do
-		start = start - 1
-	end
-	return start
-end
-
--- (internal) Capture one `/* ... */` block comment whose closing `*/`
--- ends at `close_end_pos`. Returns (block_text, new_scan_pos) where `new_scan_pos`
--- is where to continue scanning for more comments, or nil if no block comment was found.
-local function capture_block_comment(source, close_end_pos)
-	local open_at = find_block_comment_open(source, close_end_pos)
-	if not open_at then return nil end
-	local block_start = extend_left_over_indent(source, open_at)
-	return source:sub(block_start, close_end_pos), block_start
-end
-
--- (internal) Capture one `// ...` line comment ending at `line_end_pos`.
--- Returns (comment_text, new_scan_pos) or nil if the line is not a `//` comment.
-local function capture_line_comment(source, line_end_pos)
-	local line_start = find_line_start(source, line_end_pos)
-	local line       = source:sub(line_start, line_end_pos)
-	if line:sub(1, 2) == "//" then
-		return line, line_start - 1
-	end
-	return nil
-end
-
 --- Find the contiguous comment block immediately preceding `pos` in `source`.
 --- Returns the comment text (with the `/* */` or `//` markers preserved) or an empty string if no comment is adjacent.
 ---
 --- Used to copy signature comments from the source declaration (`MipsAtomComp_` / `MipsAtomComp_Proc_` / function decl)
 --- over to the generated `mac_X` macro, so LSP/IntelliSense displays the args doc.
----
 --- @param source string
---- @param pos integer
+--- @param pos    integer
 --- @return string
 local function preceding_comment_block(source, pos)
 	local scan_pos = pos
 	local pieces   = {}
 	while true do
-		local non_ws = skip_ws_backward(source, scan_pos)
+		-- skip whitespace backward; land on the next non-ws character.
+		local non_ws = scan_pos - 1
+		while non_ws > 0 do
+			local ch = source:sub(non_ws, non_ws)
+			if ch == " " or ch == "\t" or ch == "\n" or ch == "\r" then
+				non_ws = non_ws - 1
+			else
+				break
+			end
+		end
 		if non_ws == 0 then break end
-
+		if non_ws >= 2 and source:sub(non_ws - 1, non_ws) == "*/" then
-		local is_block_close = non_ws >= 2 and source:sub(non_ws - 1, non_ws) == "*/"
+			-- block comment close: find the opening /* by walking back over /* candidates
-		local is_line_end    = source:sub(non_ws, non_ws) == "\n" or source:sub(non_ws, non_ws) == "\r"
+			-- in source[1..non_ws-1].
-
+			local prefix  = source:sub(1, non_ws - 1)
-		if is_block_close then
+			local open_at = nil
-			local block_text, new_scan_pos = capture_block_comment(source, non_ws)
+			for scan = #prefix - 1, 1, -1 do
-			if not block_text then break end
+				if prefix:sub(scan, scan + 1) == "/*" then
-			table.insert(pieces, 1, block_text)
+					open_at = scan
-			scan_pos = new_scan_pos
+					break
-		elseif is_line_end then
+				end
-			local line_text, new_scan_pos = capture_line_comment(source, non_ws)
+			end
-			if not line_text then break end
+			if not open_at then break end
-			table.insert(pieces, 1, line_text)
+			-- include the indentation before the /* by walking back over leading spaces + tabs.
-			scan_pos = new_scan_pos
+			local block_start = open_at
+			while block_start > 1 do
+				local ch = source:sub(block_start - 1, block_start - 1)
+				if ch ~= " " and ch ~= "\t" then break end
+				block_start = block_start - 1
+			end
+			table.insert(pieces, 1, source:sub(block_start, non_ws))
+			scan_pos = block_start
 		else
-			break
+			-- line comment path: must end in newline, must start with //.
+			local ch = source:sub(non_ws, non_ws)
+			if ch ~= "\n" and ch ~= "\r" then break end
+			-- walk back from non_ws to the start of the source line (most recent \n or position 1).
+			local line_start = non_ws
+			while line_start > 1 and source:sub(line_start - 1, line_start - 1) ~= "\n" do
+				line_start = line_start - 1
+			end
+			local line = source:sub(line_start, non_ws)
+			if line:sub(1, 2) ~= "//" then break end
+			table.insert(pieces, 1, line)
+			scan_pos = line_start - 1
 		end
 	end
 	if #pieces == 0 then return "" end
@@ -282,49 +199,10 @@ end
 -- Argument-name extraction
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Walk `trimmed` backward from `pos` over trailing whitespace / asterisks / brackets, returning the position of the first
--- non-trailer character (i.e. the end of the identifier).
--- @param trimmed string
--- @param pos integer
--- @return integer
-local function trim_trailer_back(trimmed, pos)
-	local back = pos
-	while back > 0 do
-		local ch = trimmed:sub(back, back)
-		if ch == " " or ch == "\t" or ch == "*" or ch == "]" or ch == "[" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
--- Walk `trimmed` backward from `pos` over identifier chars (alnum + `_`),
--- returning the position just before the identifier starts.
--- @param trimmed string
--- @param pos integer
--- @return integer
-local function trim_ident_back(trimmed, pos)
-	local back = pos
-	while back > 0 do
-		local ch = trimmed:sub(back, back)
-		if duffle.is_alnum(ch) or ch == "_" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
 --- 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
----
---- No regex — uses `duffle.is_alnum` + plain string ops.
----
 --- @param args_str string|nil
 --- @return string[]|nil
 local function extract_arg_names(args_str)
@@ -334,8 +212,27 @@ local function extract_arg_names(args_str)
 	for _, tok in ipairs(tokens) do
 		local trimmed = duffle.trim(tok)
 		if trimmed ~= "" then
-			local ident_end = trim_trailer_back(trimmed, #trimmed)
+			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
-			local ident_start = trim_ident_back(trimmed, ident_end) + 1
+			-- 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
@@ -348,13 +245,12 @@ end
 -- Component projection (read from pre-scanned SourceScan)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
+--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
--- Does per-source backward lookups for args (preceding function decl) and comment (preceding comment block).
+--- Does per-source backward lookups for args (preceding function decl) and comment (preceding comment block).
--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table
+--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table instead of calling duffle.tokenize_body again.
--- instead of calling duffle.tokenize_body again.
+--- @param source string  -- the full source text (needed for backward lookups)
--- @param source string  -- the full source text (needed for backward lookups)
+--- @param scan   table   -- SourceScan from duffle.scan_source
--- @param scan  table   -- SourceScan from duffle.scan_source
+--- @return Component[]
--- @return Component[]
 local function project_components(source, scan)
 	local out = {}
 	for _, a in ipairs(scan.atoms) do
@@ -365,9 +261,11 @@ local function project_components(source, scan)
 				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.
 			}
 		end
 	end
@@ -379,7 +277,6 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- 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.
 --
@@ -403,8 +300,8 @@ local function convert_line_comments_to_block(s)
 			while eol <= len and result:byte(eol) ~= BYTE_NEWLINE do
 				eol = eol + 1
 			end
-			local before   = result:sub(1, pos - 1)
+			local before  = result:sub(1, pos - 1)
-			local comment  = result:sub(pos + 2, eol - 1)  -- skip the `//`
+			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
@@ -422,10 +319,11 @@ 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
+--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
--- if it doesn't start with the prefix (so a non-component ident like `mask_upper` falls through to the wc-table branch).
+--- Returns the ident unchanged if it doesn't start with the prefix 
--- @param ident string|nil
+--- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
--- @return string|nil
+--- @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
@@ -434,13 +332,13 @@ local function strip_mac_prefix(ident)
 	return ident
 end
 
--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
+--- (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).
+--- 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 name         string -- the component name (without `mac_`)
--- @param comp_by_name table<string, Component>
+--- @param comp_by_name table<string, Component>
--- @param wc table<string, integer>
+--- @param wc           table<string, integer>
--- @param cache table<string, integer>
+--- @param cache        table<string, integer>
--- @return 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)
@@ -475,13 +373,11 @@ 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
+--- so the cache survives across siblings and a component's recursive `mac_Y(...)` 
---- instead of re-walking the body. Previously each call rebuilt both tables (O(N) tables per call → O(N^2)).
+--- 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>
+--- @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 = {}
@@ -518,19 +414,12 @@ local function split_comment_lines(s)
 	return out
 end
 
---- Split an atom body by top-level commas; drop empty tokens.
---- @param body string
---- @return string[]
-local function tokens_from_body(body)
-	return duffle.tokenize_body_simple(body)
-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
+	if    arg_names and #arg_names > 0 then
 		return table.concat(arg_names, ", ")
 	end
 	return "..."
@@ -540,7 +429,7 @@ end
 --- The last 2 chars are always that pair.
 local function strip_trailing_continuation(lines)
 	local last = lines[#lines]
-	if last:sub(-2) == " \\" then
+	if    last:sub(-2) == " \\" then
 		lines[#lines] = last:sub(1, -3)
 	end
 end
@@ -561,9 +450,9 @@ 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).
---- @param c Component
+--- @param c          Component
 --- @param components Component[]
---- @param wc table<string, integer>
+--- @param wc         table<string, integer>
 --- @return string[]  -- list of lines for this component
 local function build_component_lines(c, counts)
 	local lines = {}
@@ -574,7 +463,8 @@ local function build_component_lines(c, counts)
 		end
 	end
 
-	local tokens = tokens_from_body(c.body)
+	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]
@@ -594,10 +484,10 @@ end
 -- Per-source emit logic
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
+--- 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).
+--- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
--- @param src SourceFile
+--- @param src SourceFile
--- @return string[]
+--- @return string[]
 local function header_boilerplate(src)
 	return {
 		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
@@ -610,8 +500,8 @@ local function header_boilerplate(src)
 		"// 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
+		-- We use the same definition here so the auto-generated entries below expand
-		-- the metadata file is included first or not.
+		-- 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",
@@ -619,32 +509,28 @@ local function header_boilerplate(src)
 	}
 end
 
--- Compute the output path for one source's `.macs.h` file.
+--- Compute the output path for one source's `.macs.h` file.
--- The pre-rework convention uses the *directory* basename
+--- The pre-rework convention uses the *directory* basename (not the source file basename)
--- (not the source file basename) — e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+---  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
--- This matches what the C codebase #includes.
+--- This matches what the C codebase #includes.
--- @param src SourceFile
+--- @param src SourceFile
--- @return string  -- the output directory
+--- @return string  -- the output directory
--- @return string  -- the full output path
+--- @return string  -- the full output path
 local function compute_macs_h_path(src)
-	local out_dir  = src.dir .. "/" .. GEN_SUBDIR
+	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
-	local out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
 	return out_dir, out_path
 end
 
---- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. Writes in BINARY mode so LF line endings are
+--- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
---- preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
+--- 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
---- Honors `ctx.dry_run`: prints the intended path but does not write the file.
+--- @param src        SourceFile
----
---- @param ctx PassCtx
---- @param src SourceFile
 --- @param components Component[]
---- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
 --- @return string|nil  -- path to the written file (nil if no components)
 local function emit_component_macros_h(ctx, src, components, counts)
 	if #components == 0 then return nil end
-
 	local out_dir, out_path = compute_macs_h_path(src)
 	local lines             = header_boilerplate(src)
 
@@ -655,12 +541,6 @@ local function emit_component_macros_h(ctx, src, components, counts)
 	end
 
 	local content = table.concat(lines, "\n") .. "\n"
-
-	if ctx.dry_run then
-		print(string.format("  -> %s (dry-run)", out_path))
-		return out_path
-	end
-
 	duffle.ensure_dir(out_dir)
 	duffle.write_file_lf(out_path, content)
 	print(string.format("  -> %s", out_path))
@@ -671,15 +551,86 @@ end
 -- Pass entry
 -- ════════════════════════════════════════════════════════════════════════════
 
--- (internal) Extend `ctx.shared.word_counts` with this source's component macros
+--- (internal) Extend the canonical `corpus.word_counts` with this source's component macros so offsets sees them without re-reading the file.
--- 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 ctx PassCtx
+--- @param corpus     table  -- the canonical corpus
--- @param components Component[]
+--- @param components Component[]
--- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
-local function update_shared_word_counts(ctx, components, counts)
+local function update_canonical_word_counts(corpus, components, counts)
-	local wc = ctx.shared.word_counts
+	local wc = corpus.word_counts
 	for _, c in ipairs(components) do
-		wc["mac_" .. c.name] = counts[c.name]
+		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"
+
+--- (internal) Populate the canonical `corpus.components` projection with this source's components-by-name map.
+--- First declaration wins; later declarations of the same bare name are dropped and recorded as a collision via `corpus.collisions` (kind = "component").
+--- The pass does NOT write to `ctx.shared.components` (ownership follows the canonical contract).
+--- @param corpus     table  -- the canonical corpus
+--- @param src        SourceFile
+--- @param components Component[]
+local function update_canonical_components(corpus, src, components)
+	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 canonical corpus;
+		-- `mac_` prefix lives at the call-site identifier and is stripped before lookup.
+		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",
+			}
+		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) do NOT record a collision (the first-wins entry already covers the case).
+			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 the canonical `corpus.component_body_index` projection 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 does NOT write to `ctx.shared.component_body_index` (the corpus owns this projection).
+--- @param corpus     table  -- the canonical 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
 
@@ -690,16 +641,42 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	for _, src in ipairs(ctx.sources) do
+	-- Canonical-corpus ownership gate.
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("components.run requires ctx.shared.corpus (canonical corpus).", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("components.run requires ctx.shared.corpus.source_order (canonical corpus).", 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
+
+	-- Canonical 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 does NOT mutate `ctx.shared.components` or `ctx.shared.component_body_index`
+	-- (ownership follows the canonical corpus; consumers read from the corpus directly).
+
+	for _, src in ipairs(corpus.source_order) do
 		-- project_components reads from src.scan + does backward lookups on src.text
 		local components = project_components(src.text, src.scan)
 		if #components > 0 then
-			-- Compute word counts for ALL components once (was: rebuilt per call inside the helpers).
+			-- Compute all component word counts once per source.
-			local counts = count_all_components(components, ctx.shared.word_counts)
+			-- Use `corpus.word_counts` (the canonical count table) so the recursive lookup sees both authored-metadata entries
+			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
+			local counts = count_all_components(components, corpus.word_counts)
 			local macs_path = emit_component_macros_h(ctx, src, components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				update_shared_word_counts(ctx, components, counts)
+				-- Populate the canonical projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
+				update_canonical_word_counts(corpus, components, counts)
+				update_canonical_components(corpus, src, components)
+				update_canonical_component_body_index(corpus, src, components, src.scan)
 			end
 		end
 	end

scripts/passes/emission_model.lua

@@ -0,0 +1,184 @@
+--- passes/emission_model.lua: Per-atom emission projection.
+---
+--- The `emission-model` pass owns `atom.paths` (the canonical per-atom mutable surface)
+--- for every atom-with-body and every raw atom-with-body declared in `ctx.shared.corpus.source_order`.
+--- For each such atom, the pass invokes `duffle.project_emission(body_text, component_index, word_counts)`
+--- and 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` keeps build-stopping semantics (no policy change in this task).
+---
+--- Source-order discipline:
+---   * `corpus.source_order` is the canonical ordering of source records.
+---   * For each source, the pass iterates `src.scan.atoms` and `src.scan.raw_atoms` IN SOURCE ORDER, preserving declaration order.
+---
+--- Per-atom projection fields on `atom.paths`:
+---   `tokens`, `line_in_body`, `items`, `word_events`, `markers`, `invocations`, `errors`, `warnings`.
+---   The dense views are built from `items` only; the pass never re-walks source text or tokens.
+---
+--- 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;
+---   * component cycles retain balanced invocation boundaries and emit a `cycle` construction error without recursing indefinitely;
+---   * declared-vs-measured component word counts emit `count_mismatch` construction errors; opaque uncounted macros emit warnings.
+---
+--- The pass does NOT consult `_code_macros` / `_code_macro_bodies`. Those private tables are owned by `passes.scan_source` and stripped 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.
+-- Consumers read these canonical fields rather than rebuilding line state or tokenizing source again.
+local function stamp_root_provenance(projection, atom_record, src, corpus)
+	local root_line_of   = src.scan and src.scan.line_of
+	local root_body_line = root_line_of and root_line_of((atom_record.body_off or 1) - 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
+
+	local function body_line_for(event, item)
+		local body_line_of = root_line_of
+		local body_off     = atom_record.body_off or 0
+		local ids          = event.invocation_ids or {}
+		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
+				-- Component-body walkers already receive the declaration source's full line index, so their item.line is physical.
+				return item.line or 0
+			end
+		end
+		local first_line = body_line_of and body_line_of(math.max(1, body_off - 1)) or root_body_line
+		return (first_line or 0) + (item.line or 1) - 1
+	end
+
+	-- Stamp root-source path onto invocation records whose `call_path` was left empty by the walker.
+	-- The walker passes `body_entry.source` to `emit_invoke_begin` as the call_path argument; for the root body_entry created by `M.project_emission` that source is ""
+	-- (the caller passes only the body text).
+	-- After this stamp every invocation record has a physical call_path that matches what `passes/atoms_source_map.lua` matches the 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
+
+	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 call_line = (root_body_line or 0) + (outer_inv.call_line or 1) - 1 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 {}
+	local proj  = duffle.project_emission(body, cbi, wc)
+	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 can dispatch on the diagnostic class without re-parsing the message string.
+			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 every source in canonical order; for each source, iterate atoms + 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

@@ -21,18 +21,12 @@
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
-local count_token_words = word_count_eval.count_token_words
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Marker-call identifiers inside atom bodies.
-local LABEL_MARKER    = "atom_label"
-local OFFSET_MARKER   = "atom_offset"
-
 -- Offset macro/enum naming prefixes (the emitted header uses these).
 local OFFSET_MACRO_PREFIX = "_atom_offset_"
 local OFFSET_ENUM_PREFIX  = "atom_offset_"
@@ -52,16 +46,10 @@ local OFFSET_MACRO_COL = 44
 --- @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 shared.word_counts table                  -- macro name -> word count
+--- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared.word_counts table
 --- @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 dry_run            boolean                -- if true, compute but don't write
---- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -69,10 +57,10 @@ local OFFSET_MACRO_COL = 44
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class BranchOffset
---- @field tag    string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field tag         string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
---- @field target string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field target      string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field pos    integer  -- the branch's word position within the atom body
+--- @field branch_word integer  -- branch word position within the atom body
---- @field offset integer  -- computed `target_word - branch_word - 1`
+--- @field offset      integer  -- computed `target_word - branch_word - 1`
 
 --- @class AtomData
 --- @field name        string         -- atom name
@@ -80,188 +68,85 @@ local OFFSET_MACRO_COL = 44
 --- @field offsets     BranchOffset[] -- per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Per-token marker-call helpers (atom_label / atom_offset inside bodies)
+-- Canonical marker projection
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Extract comma-separated identifier args from a parenthesized group after a function-like macro call.
+-- MARKER_PROJECTORS is the marker-kind data table.
--- Returns (args, after_paren) where `after_paren` is the position just past the closing `)`, or nil if `token` did not start with `(`.
+-- The emission-model pass already records marker word positions;
--- @param token string
+-- this pass only projects those records into the label/branch lookup shape needed by offset computation.
--- @param after_ident integer
+local MARKER_PROJECTORS = {
--- @return string[], integer|nil
+	label = function(state, marker)
-local function extract_ident_args(token, after_ident)
+		state.labels[marker.name] = marker.word_index
-	local arg_start = duffle.skip_ws_and_cmt(token, after_ident)
+	end,
-	if token:sub(arg_start, arg_start) ~= "(" then return {}, nil end
+	offset = function(state, marker)
-	local inner, after_paren = duffle.read_parens(token, arg_start)
+		state.branches[#state.branches + 1] = {
-	-- scan: <marker>(<args>)
+			tag         = marker.name,
+			target      = marker.target,
+			branch_word = marker.word_index,
+		}
+	end,
+}
 
-	local args      = {}
+--- Project canonical marker records into the two lookup tables used by the offset renderer.
-	local pos       = 1
+--- No source text, body text, or body token is inspected.
-	local inner_len = #inner
+--- @param markers table[] -- atom.paths.markers
-	while pos <= inner_len do
+--- @return table<string, integer>, table[]
-		pos = duffle.skip_ws_and_cmt(inner, pos)
+local function project_markers(markers)
-		if pos > inner_len then break end
+	local state = { labels = {}, branches = {} }
-		local ident, after = duffle.read_ident(inner, pos)
+	for _, marker in ipairs(markers or {}) do
-		if ident and ident ~= "" then
+		local project = MARKER_PROJECTORS[marker.kind]
-			table.insert(args, ident)
+		if project then project(state, marker) end
-			pos = after
-		else
-			pos = pos + 1
-		end
-		pos = duffle.skip_ws_and_cmt(inner, pos)
-		if pos <= inner_len and inner:sub(pos, pos) == "," then pos = pos + 1 end
 	end
-
+	return state.labels, state.branches
-	return args, after_paren
-end
-
--- (internal) Record a `atom_label(name)` marker — `at_pos` is the branch-free word position within the atom body.
--- @param labels table<string, integer>
--- @param args string[]
--- @param at_pos integer
-local function record_label_marker(labels, args, at_pos)
-	if #args >= 1 then labels[args[1]] = at_pos end
-end
-
--- (internal) Record a `atom_offset(tag, target)` marker.
--- @param branches table[]  -- list of {pos=, target=, tag=}
--- @param args string[]
--- @param at_pos integer
-local function record_offset_marker(branches, args, at_pos)
-	if #args >= 2 then
-		table.insert(branches, { pos = at_pos, target = args[2], tag = args[1] })
-	end
-end
-
---- Scan a single token for atom_label/atom_offset markers, walking through balanced groups transparently (so nested calls are found).
---- @param token string
---- @param at_pos integer    -- the branch-free word position of this token in the body
---- @param labels table<string, integer>
---- @param branches table[]
-local function scan_for_atom_markers(token, at_pos, labels, branches)
-	local pos     = 1
-	local tok_len = #token
-	while pos <= tok_len do
-		pos = duffle.skip_ws_and_cmt(token, pos)
-		if pos > tok_len then break end
-		local ch = token:sub(pos, pos)
-		if duffle.is_alpha(ch) then
-			local ident, after = duffle.read_ident(token, pos)
-			if ident == LABEL_MARKER then
-				local args, after_paren = extract_ident_args(token, after)
-				record_label_marker(labels, args, at_pos)
-				pos = after_paren or after
-			elseif ident == OFFSET_MARKER then
-				local args, after_paren = extract_ident_args(token, after)
-				record_offset_marker(branches, args, at_pos)
-				pos = after_paren or after
-			else
-				pos = after
-			end
-		else
-			local nx = duffle.skip_str_or_cmt(token, pos)
-			pos = (nx > pos) and nx or (pos + 1)
-		end
-	end
-end
-
--- (internal) Count words emitted by the rest of `tok` after a marker call
--- (the marker call itself emits 0 words, but the source pattern may bundle the marker with the next instruction on the same line,
--- separated by no top-level comma).
--- Returns the word count contributed by that rest.
--- @param tok string
--- @param word_counts table
--- @return integer
-local function count_marker_rest(tok, word_counts)
-	-- duffle.find_marker_call_end returns the position PAST the closing `)` of the marker call
-	-- (or nil if `tok` isn't a marker call). Canonical impl in duffle.lua is faster than the
-	-- file-local copy that used to live here (byte-indexed, no `tok:sub` per char).
-	local marker_end = duffle.find_marker_call_end(tok)
-	if not marker_end or marker_end >= #tok then return 0 end
-	local rest = duffle.trim(tok:sub(marker_end))
-	if rest == "" then return 0 end
-	return count_token_words(rest, word_counts)
-end
-
--- (internal) Is this token a marker call (`atom_label` or `atom_offset`)?
--- @param tok string
--- @return boolean
-local function is_marker_token(tok)
-	local leading_ident = duffle.read_ident(tok, 1)
-	return leading_ident == LABEL_MARKER or leading_ident == OFFSET_MARKER
-end
-
---- Scan an atom body for labels + branches, count total words.
---- Returns (labels, branches, total_words).
---- @param body string
---- @param word_counts table
---- @return table<string, integer>, table[], integer
--- scan_atom_body: walk pre-tokenized body for atom_label/atom_offset markers + word counts.
--- Uses `atom.body_tokens` from the SourceScan payload (pre-tokenized by scan-source pass).
--- @param body_tokens table[]  -- {{tok=string, rel=integer}, ...} from duffle.tokenize_body
--- @param word_counts table
--- @return table, table, integer  -- labels, branches, total_words
-local function scan_atom_body(body_tokens, word_counts)
-	local pos      = 0
-	local labels   = {}
-	local branches = {}
-	for _, t in ipairs(body_tokens) do
-		local tok = t.tok
-		if is_marker_token(tok) then
-			-- Marker call: record at the current pos, do NOT advance pos.
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + count_marker_rest(tok, word_counts)
-		else
-			local words = count_token_words(tok, word_counts)
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + words
-		end
-	end
-	return labels, branches, pos
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Compute branch offsets as `target_word - branch_word - 1`
+--- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
--- (the standard MIPS branch-immediate encoding).
+--- @param labels table<string, integer>
--- @param labels table<string, integer>
+--- @param branches table[]
--- @param branches table[]
+--- @return BranchOffset[]
--- @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.pos .. ")")
+			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")")
 		end
-		results[#results + 1] = { target = br.target, tag = br.tag, offset = target - br.pos - 1 }
+		results[#results + 1] = {
+			target      = br.target,
+			tag         = br.tag,
+			branch_word = br.branch_word,
+			offset      = target - br.branch_word - 1,
+		}
 	end
 	return results
 end
 
--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
+--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
--- @param s string
+--- @param s string
--- @param w integer
+--- @param w integer
--- @return string
+--- @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.
+--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
--- @param r BranchOffset
+--- @param bo BranchOffset
--- @return table
+--- @return table
-local function make_offset_const(r)
+local function make_offset_const(bo)
 	return {
-		macro_name = OFFSET_MACRO_PREFIX .. r.tag .. "_" .. r.target,
+		macro_name = OFFSET_MACRO_PREFIX .. bo.tag .. "_" .. bo.target,
-		enum_name  = OFFSET_ENUM_PREFIX  .. r.tag .. "_" .. r.target,
+		enum_name  = OFFSET_ENUM_PREFIX  .. bo.tag .. "_" .. bo.target,
-		value      = r.offset,
+		value      = bo.offset,
 	}
 end
 
--- (internal) Emit one atom's offset constants + enum into the lines buffer.
+--- (internal) Emit one atom's offset constants + enum into the lines buffer.
--- @param add fun(s: string)
+--- @param add  fun(s: string)
--- @param atom AtomData
+--- @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) ---")
@@ -282,10 +167,10 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
--- Generate the per-source .offsets.h header.
+--- Generate the per-source .offsets.h header.
--- @param source_path string
+--- @param source_path string
--- @param atoms_data AtomData[]
+--- @param atoms_data  AtomData[]
--- @return string
+--- @return string
 local function generate_header(source_path, atoms_data)
 	local basename = duffle.basename_no_ext(source_path)
 
@@ -307,59 +192,41 @@ local function generate_header(source_path, atoms_data)
 	return table.concat(lines, "\n") .. "\n"
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- M — module exports
--- ════════════════════════════════════════════════════════════════════════════
-
 local M = {}
 
--- Project the pre-scanned SourceScan entries into the {name, body, body_tokens} shape this pass needs.
+--- (internal) Process one source: render offsets from canonical atom paths.
--- MipsAtom_ entries have kind="atom"; MipsCode code_<name> entries have kind="raw_atom".
+--- Returns the offsets_h path if a header was written, or nil.
--- `body_tokens` is set by scan-source on every `scan.atoms[i]` / `scan.raw_atoms[i]`; we carry it forward
+--- @param ctx PassCtx
--- so `scan_atom_body` reads from the precomputed table directly (no per-atom tokenize_body fallback).
+--- @param src SourceFile
--- @param scan table  -- SourceScan from duffle.scan_source
+--- @return string|nil  -- the offsets_h path
--- @return table[]  -- list of {name=, body=, body_tokens=}
-local function project_atoms(scan)
-	local out = {}
-	for _, a in ipairs(scan.atoms) do
-		out[#out + 1] = { name = a.raw_name, body = a.body, body_tokens = a.body_tokens }
-	end
-	for _, a in ipairs(scan.raw_atoms) do
-		out[#out + 1] = { name = a.name, body = a.body, body_tokens = a.body_tokens }
-	end
-	return out
-end
-
--- (internal) Process one source: project atoms from scan, scan bodies, write header.
--- Returns the offsets_h path if a header was written, or nil.
--- @param ctx PassCtx
--- @param src SourceFile
--- @return string|nil  -- the offsets_h path
 local function process_source(ctx, src)
-	local atoms = project_atoms(src.scan)
-	if #atoms == 0 then return nil end
-
 	local atoms_data = {}
-	for _, atom in ipairs(atoms) do
+	local scan = src.scan or {}
-		local labels, branches, total = scan_atom_body(atom.body_tokens, ctx.shared.word_counts)
+
+	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.name,
+			name        = atom.raw_name or atom.name,
-			total_words = total,
+			total_words = #(paths.word_events or {}),
 			offsets     = compute_offsets(labels, branches),
 		}
 	end
 
+	for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+	for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	if #atoms_data == 0 then return nil end
+
 	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
-	if not ctx.dry_run then
+	duffle.ensure_dir(duffle.dirname(out_path))
-		duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file(out_path, generate_header(src.path:gsub("/", "\\"), atoms_data))
-		duffle.write_file(out_path, generate_header(src.path, atoms_data))
-	end
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each source, emits a per-module `<dir_basename>.offsets.h` containing `#define _atom_offset_F_T = N` constants 
+--- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
---- for every `atom_offset(F, T)` reference in the source's atoms.
+--- containing constants for every marker recorded in atom.paths.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -367,7 +234,12 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	for _, src in ipairs(ctx.sources) do
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	end
+
+	for _, src in ipairs(corpus.source_order) do
 		local out_path = process_source(ctx, src)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }

scripts/passes/report.lua

@@ -5,8 +5,10 @@
 ---   - `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 stashes per-MODULE summary entries in `ctx.flags._annot_results` (set by `passes/annotation.lua`). 
+--- The annotation pass emits `errors.h` files per module and the canonical
---- This pass re-validates each source via `annotation.validate()` to get the detailed per-source results needed for the report. 
+--- `corpus.sources_by_dir` projection groups sources by directory. This pass
+--- iterates the canonical dir projection directly and re-validates each source
+--- via `annotation.validate()` to get the detailed per-source results.
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
@@ -16,7 +18,6 @@
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
--- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
 -- 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
@@ -26,6 +27,13 @@
 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")
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
@@ -61,16 +69,14 @@ local PASS_NAME = "report"
 --- @field basename string  -- filename without extension
 
 --- @class PassCtx
---- @field sources            SourceFile[]            -- all source files in the build
+--- @field sources            SourceFile[]         -- all source files in the build
---- @field metadata_path      string                  -- path to word_count.metadata.h
+--- @field metadata_path      string               -- path to word_count.metadata.h
---- @field shared             table                   -- cross-pass shared state
+--- @field shared             table                -- cross-pass shared state
---- @field out_root           string                  -- output root (e.g. "build/gen")
+--- @field out_root           string               -- output root (e.g. "build/gen")
---- @field project_root       string                  -- project root (e.g. "code/")
+--- @field project_root       string               -- project root (e.g. "code/")
---- @field upstream           table<string, table>    -- per-pass upstream outputs
+--- @field upstream           table<string, table> -- per-pass upstream outputs
---- @field flags              table                   -- CLI flags + per-pass stash
+--- @field flags              table                -- CLI flags + per-pass stash
---- @field flags._annot_results ModuleEntry[]         -- stashed by annotation pass
+--- @field verbose            boolean              -- if true, log diagnostic info
---- @field dry_run            boolean                 -- if true, compute but don't write
---- @field verbose            boolean                 -- if true, log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -139,31 +145,31 @@ local PASS_NAME = "report"
 -- 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.
+--- 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
+--- @param path string
--- @return string
+--- @return string
 local function source_basename(path)
 	return path:match(BASENAME_PATTERN) or path
 end
 
--- (internal) Format a single annotation entry as one rendered line.
+--- (internal) Format a single annotation entry as one rendered line.
--- @param a AnnotEntry
+--- @param a AnnotEntry
--- @param src_name string
+--- @param src_name string
--- @return string
+--- @return string
 local function format_annot_line(a, src_name)
 	if a.error then
 		return string.format("  ✗ line %d  %s  [ERROR: %s]  [%s]", a.line, a.macro or "?", a.error, src_name)
 	end
 	local line = string.format("  ●  line %d  %s  [%s]", a.line, a.name, src_name)
-	if a.binds               then line = line .. "  binds=" .. a.binds end
+	if a.binds       then line = line .. "  binds="   .. a.binds end
-	if #a.reads > 0          then line = line .. "  reads={" .. table.concat(a.reads, ",") .. "}" end
+	if #a.reads  > 0 then line = line .. "  reads={"  .. table.concat(a.reads,  ",") .. "}" end
-	if #a.writes > 0         then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
+	if #a.writes > 0 then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
 	return line
 end
 
--- (internal) Tally totals across all results in a module.
+--- (internal) Tally totals across all results in a module.
--- @param results AnnotationResult[]
+--- @param results AnnotationResult[]
--- @return integer, integer, integer, integer, integer, integer
+--- @return integer, integer, integer, integer, integer, integer
 local function tally_module_totals(results)
 	local total_atoms, total_annots, total_binds, total_macros = 0, 0, 0, 0
 	local total_errors, total_warnings = 0, 0
@@ -261,6 +267,28 @@ local function render_module_warnings_section(add, results, total_warnings)
 	add("")
 end
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- SECTION_RENDERERS — data-driven section dispatch (the plex pattern)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Each entry maps a section to its (header, render_fn). The render_fn signature:
+--   render_fn(add, results, totals)
+--     add    -- the `add(line)` closure from the surrounding report renderer
+--     results -- AnnotationResult[] (per-source results)
+--     totals -- {atoms, annots, binds, macros, errors, warnings} counts
+--
+-- Sections that need to render "(none)" vs iterate use totals.errors / totals.warnings;
+-- other sections ignore the totals arg.
+-- Adding a new section = 1 row here + 1 render_<thing>_section function.
+local SECTION_RENDERERS = {
+	{ header = SECTION_HEADER_ATOMS,    render = render_module_atoms_section     },
+	{ header = SECTION_HEADER_ANNOTS,   render = render_module_annots_section    },
+	{ header = SECTION_HEADER_BINDS,    render = render_module_binds_section     },
+	{ header = SECTION_HEADER_MACROS,   render = render_module_macros_section    },
+	{ header = SECTION_HEADER_ERRORS,   render = function(add, results, totals) return render_module_errors_section(add, results, totals.errors)   end },
+	{ header = SECTION_HEADER_WARNINGS, render = function(add, results, totals) return render_module_warnings_section(add, results, totals.warnings) end },
+}
+
 --- Render the per-MODULE annotation report (one `<dir_basename>.annotations.txt`).
 --- @param dir string                  -- module directory path
 --- @param sources SourceFile[]        -- sources in this module
@@ -282,12 +310,20 @@ local function render_module_report(dir, sources, results)
 		total_atoms, total_annots, total_binds, total_macros))
 	add("")
 
-	render_module_atoms_section(add, results)
+	-- Bundle the totals so the section renderers don't need separate parameter lists.
-	render_module_annots_section(add, results)
+	-- Errors/warnings sections need their total count to decide "(none)" vs iterate.
-	render_module_binds_section(add, results)
+	-- Sections without totals (atoms/annots/binds/macros) ignore this arg.
-	render_module_macros_section(add, results)
+	local totals = {
-	render_module_errors_section(add, results, total_errors)
+		atoms    = total_atoms,  annots = total_annots,  binds  = total_binds,
-	render_module_warnings_section(add, results, total_warnings)
+		macros   = total_macros, errors = total_errors,  warnings = total_warnings,
+	}
+
+	-- THE per-section dispatch. ONE loop over SECTION_RENDERERS.
+	-- Each renderer writes its header + content via the `add` closure (pre-bound above).
+	-- Adding a new section = 1 row here + 1 render_<thing>_section function.
+	for _, section in ipairs(SECTION_RENDERERS) do
+		section.render(add, results, totals)
+	end
 
 	return table.concat(lines, "\n") .. "\n"
 end
@@ -346,20 +382,22 @@ end
 -- Orchestration helpers
 -- ════════════════════════════════════════════════════════════════════════════
 
--- (internal) Pull per-source validate() results from the annotation pass's stash.
+--- (internal) Re-validate every source in a directory against the canonical
--- The annotation pass runs first in the dep chain and caches results in `ctx.flags._annot_source_results`; we read from there instead of re-validating each source.
+--- corpus projection. Calls `annotation.validate()` per source to produce the
--- Returns the list of module results + the flat list of all results (for the project-wide summary).
+--- per-source AnnotationResult (atoms / annots / macros / binds / errors /
--- @param ctx PassCtx
+--- warnings) that the report renderer consumes. This is the canonical path —
--- @param dir_sources SourceFile[]
+--- no private stash; each report pass run is reproducible from the corpus.
--- @return AnnotationResult[], AnnotationResult[]
+--- Returns the list of module results + the flat list of all results (for the project-wide summary).
+--- @param ctx PassCtx
+--- @param dir_sources SourceFile[]
+--- @return AnnotationResult[], AnnotationResult[]
 local function lookup_module_results(ctx, dir_sources)
-	local src_cache      = (ctx.flags and ctx.flags._annot_source_results) or {}
 	local module_results = {}
 	local all_results    = {}
 	for _, src in ipairs(dir_sources) do
-		local result = src_cache[src.path]
+		if src.scan then
-		if result then
+			local result = annotation.validate(ctx, src, nil)
-			result.source = src.path                                    -- defensive (annotation tags it too; this guards against cache misses from earlier iterations)
+			result.source = src.path                                    -- tag for downstream rendering
 			module_results[#module_results + 1] = result
 			all_results[#all_results + 1] = result
 		end
@@ -367,9 +405,9 @@ local function lookup_module_results(ctx, dir_sources)
 	return module_results, all_results
 end
 
--- (internal) Does this module's results contain anything worth emitting?
+--- (internal) Does this module's results contain anything worth emitting?
--- @param module_results AnnotationResult[]
+--- @param module_results AnnotationResult[]
--- @return boolean
+--- @return boolean
 local function module_has_content(module_results)
 	for _, r in ipairs(module_results) do
 		if #r.atoms  > 0 or #r.annots > 0 or #r.binds    > 0
@@ -380,8 +418,8 @@ local function module_has_content(module_results)
 	return false
 end
 
--- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
+--- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
--- @param fmt string
+--- @param fmt string
 local function debug_log(fmt, ...)
 	if _G[DEBUG_FLAG] then
 		io.stderr:write(string.format("[%s] " .. fmt, PASS_NAME, ...))
@@ -403,27 +441,27 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	local module_entries = (ctx.flags and ctx.flags._annot_results) or {}
+	-- Module grouping comes from `corpus.sources_by_dir` (the canonical projection).
-	local by_dir         = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	-- Iterate it directly; no private cache, no per-pass stash.
+	local corpus = ctx.shared and ctx.shared.corpus
+	local by_dir = (corpus and corpus.sources_by_dir) or {}
 
-	if not ctx.dry_run then duffle.ensure_dir(ctx.out_root) end
+	duffle.ensure_dir(ctx.out_root)
 
 	local all_results_for_summary = {}
-	for _, entry in ipairs(module_entries) do
+	for dir, dir_sources in pairs(by_dir) do
-		debug_log("entry: dir=%s basename=%s atoms_count=%d dir_sources=%d\n", entry.dir, entry.dir_basename, entry.atoms_count, #(by_dir[entry.dir] or {}))
+		local dir_basename = dir:match("([^/\\]+)$") or dir
+		debug_log("dir=%s basename=%s sources=%d\n", dir, dir_basename, #dir_sources)
 
-		if entry.atoms_count > 0 or #(by_dir[entry.dir] or {}) > 0 then
+		if #dir_sources > 0 then
-			local dir_sources                 = by_dir[entry.dir] or {}
 			local module_results, all_results = lookup_module_results(ctx, dir_sources)
 			for _, r in ipairs(all_results) do
 				all_results_for_summary[#all_results_for_summary + 1] = r
 			end
 
 			if module_has_content(module_results) then
-				local out_path = ctx.out_root .. "/" .. entry.dir_basename .. ".annotations.txt"
+				local out_path = ctx.out_root .. "/" .. dir_basename .. ".annotations.txt"
-				if not ctx.dry_run then
+				duffle.write_file(out_path, render_module_report(dir, dir_sources, module_results))
-					duffle.write_file(out_path, render_module_report(entry.dir, dir_sources, module_results))
-				end
 				outputs[#outputs + 1] = { annotations_txt = out_path }
 			else
 				debug_log("  -> no content; skipping\n")
@@ -431,7 +469,7 @@ function M.run(ctx)
 		end
 	end
 
-	if not ctx.dry_run and #all_results_for_summary > 0 then
+	if #all_results_for_summary > 0 then
 		local summary_path = ctx.out_root .. "/annotation_validation.txt"
 		duffle.write_file(summary_path, render_project_report(all_results_for_summary))
 		outputs[#outputs + 1] = { summary_txt = summary_path }

scripts/passes/word_count_eval.lua

@@ -1,11 +1,17 @@
 --- word_count_eval.lua — Word-counting logic for the tape-atom metaprogram pipeline.
 ---
---- Three responsibilities:
+--- Two responsibilities:
----   1. **Public utilities** (used by `passes/components.lua`, `passes/offsets.lua`, `passes/annotation.lua`):
+---   1. **Public utility** `M.count_token_words(token, wc)`: Used by `passes/offsets.lua`, `passes/annotation.lua`, and other passes.
----        - `M.count_token_words(token, wc)` — words emitted by one token
+---   2. **Pass entry** `M.run(ctx)`: Loads the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts` for downstream passes.
----        - `M.scan_dir(dir, suffix)` — glob walk for *.macs.h
+---      The generated `.macs.h` files are OUTPUT artifacts and are NOT inputs to this pass;
----   2. **Pass entry** `M.run(ctx)` — loads metadata.h + *.macs.h into `ctx.shared.word_counts` for downstream passes.
+---      Current component counts are owned by `passes/components.lua` (which populates `corpus.word_counts` and `corpus.component_body_index`
----   3. **Internal helpers** for the body scanner.
+---      AFTER computing each current count from the just-built body + `corpus.word_counts`).
+---
+--- **Canonical contract**:
+---   * `ctx.shared.corpus.word_counts` is the canonical 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 (canonical 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.
@@ -14,31 +20,11 @@
 -- 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.
--- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
--- 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")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-
--- ════════════════════════════════════════════════════════════════════════════
--- Constants
--- ════════════════════════════════════════════════════════════════════════════
-
--- Windows separator chars — used to convert `dir` output (which uses `\`) into POSIX paths (which our scripts expect).
-local PATH_SEP_BACKSLASH = "\\"
-local PATH_SEP_FORWARD   = "/"
-
--- Fallback glob command (subprocess). Used when `lfs` (LuaFileSystem) is not available.
--- Scoped to `code\` to avoid walking `.git/`, `toolchain/`, `build/`, etc.
-local DIR_GLOB_CMD = 'dir /b /s "%s\\code\\%s" 2>nul'
-
--- Try to load lfs (LuaFileSystem). If available, scan_dir uses native directory enumeration (~2ms)
--- instead of spawning `dir /b /s` as a subprocess (~56ms). Built by update_deps.ps1 into toolchain/lfs/lfs.dll.
-local lfs = pcall(require, "lfs") and require("lfs") or nil
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
@@ -57,12 +43,12 @@ local lfs = pcall(require, "lfs") and require("lfs") or nil
 --- @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.word_counts WordCounts    -- populated by this pass
+--- @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 dry_run            boolean       -- if true, compute but don't write
 --- @field verbose            boolean       -- if true, log diagnostic info
 
 --- @class PassResult
@@ -84,9 +70,8 @@ local M = {}
 --- 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 token string        -- a single token from split_top_level_commas
+--- @param wc    WordCounts -- the shared word-count table
---- @param wc WordCounts       -- the shared word-count table
 --- @return integer
 function M.count_token_words(token, wc)
 	local s = duffle.trim(token)
@@ -101,112 +86,42 @@ function M.count_token_words(token, wc)
 	return 1
 end
 
--- ┌────────────────────────────────────────────────────────────────────┐
--- │ Shared utility: scan_dir                                           │
--- └────────────────────────────────────────────────────────────────────┘
-
---- Recursively scan a directory for files matching a glob suffix.
---- No regex per the no_regex constraint — uses plain byte matching via `dir /b /s` on Windows.
----
---- The `.macs.h` files produced by the components pass always live at `<project_root>/<module>/gen/`.
---- We can shortcut the `dir /b /s` walk by listing modules first (one `dir /b /ad`), then walking each `<module>/gen/`
---- (one `dir /b` per module, no recursion). 
---- For projects with 2 modules and 0 .macs.h files, this drops the cost from ~52ms 
---- (full recursive walk of the entire project tree) to ~5ms.
----
---- @param dir string        -- directory to scan (absolute or relative)
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
--- Cache the scan_dir result per (dir, suffix) in package.loaded. 
--- Each `io.popen` call on Windows is ~50-100ms of subprocess overhead, so caching the result saves a fixed cost on every build.
--- The cache persists for the lifetime of the Lua process (cleared when ps1_meta.lua exits).
--- If a build removes/creates .macs.h files mid-process, the caller can invalidate by calling `M._invalidate_scan_cache()`.
-local SCAN_CACHE_KEY = "__word_count_eval_scan_cache__"
-
---- Scan `code/` for files matching `suffix` (e.g. `*.macs.h`).
---- Uses `lfs` (LuaFileSystem) when available — native directory enumeration at ~2ms.
---- Falls back to `dir /b /s` subprocess (~56ms) when `lfs` is not compiled.
----
---- @param dir string        -- project root directory
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
-function M.scan_dir(dir, suffix)
-	local key = dir .. "\0" .. suffix
-
-	local cache = package.loaded[SCAN_CACHE_KEY]
-	if    cache and cache[key] then return cache[key] end
-
-	local results = {}
-
-	if lfs then
-		-- Native walk: list code/<module>/gen/ for matching files. Zero subprocess spawns.
-		local code_dir = dir .. "/code"
-		if lfs.attributes(code_dir, "mode") == "directory" then
-			for mod_name in lfs.dir(code_dir) do
-				if mod_name ~= "." and mod_name ~= ".." then
-					local gen_path = code_dir .. "/" .. mod_name .. "/gen"
-					if lfs.attributes(gen_path, "mode") == "directory" then
-						for fname in lfs.dir(gen_path) do
-							if fname:match("%.macs%.h$") then
-								results[#results + 1] = gen_path .. "/" .. fname
-							end
-						end
-					end
-				end
-			end
-		end
-	else
-		-- Fallback: single `dir /b /s` subprocess scoped to code\.
-		local pipe = io.popen(DIR_GLOB_CMD:format(dir, suffix))
-		if pipe then
-			for raw_line in pipe:lines() do
-				results[#results + 1] = raw_line:gsub(PATH_SEP_BACKSLASH, PATH_SEP_FORWARD)
-			end
-			pipe:close()
-		end
-	end
-
-	-- Cache the result (including empty results).
-	cache      = cache or {}
-	cache[key] = results
-	package.loaded[SCAN_CACHE_KEY] = cache
-
-	return results
-end
-
---- Invalidate the scan cache (call after creating new .macs.h files in the same Lua process — usually not needed).
-function M._invalidate_scan_cache() package.loaded[SCAN_CACHE_KEY] = nil end
-
 -- ┌────────────────────────────────────────────────────────────────────┐
 -- │ Pass entry: M.run(ctx) — "word-counts" pass                        │
 -- └────────────────────────────────────────────────────────────────────┘
 
---- Load metadata.h + scan for existing *.macs.h files into ctx.shared.word_counts.
+--- Load the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts`.
---- Loading the .macs.h files is idempotent: entries from later (current-build) .macs.h files override metadata.h entries of the same name.
+--- Generated `.macs.h` files are OUTPUT artifacts and are NOT scanned as inputs.
+--- Current component counts are computed and inserted by `passes/components.lua`
+--- after the components pass iterates `corpus.source_order` and writes each source's `<dir_basename>.macs.h` file.
 ---
+--- 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)
-	local wc = {}
+	-- 1. Canonical-corpus ownership gate.
-
+	local corpus = ctx.shared and ctx.shared.corpus
-	-- 1. Load metadata.h (the encoding-macro source of truth).
+	if type(corpus) ~= "table" then
-	local meta_counts = duffle.load_word_counts(ctx.metadata_path)
+		error("word_count_eval.run requires ctx.shared.corpus (canonical corpus). The fixture must install the corpus before running this pass.", 0)
-	for name, count in pairs(meta_counts) do wc[name] = count end
-
-	-- 2. Scan project_root recursively for *.macs.h files (component-macro source).
-	local macs_files = M.scan_dir(ctx.project_root, "*.macs.h")
-	for _, macs_path in ipairs(macs_files) do
-		local ok, mc = pcall(duffle.load_word_counts, macs_path)
-		if not ok then
-			io.stderr:write(string.format("[word_count_eval] parse error in '%s': %s\n", macs_path, tostring(mc)))
-		elseif type(mc) ~= "table" then
-			io.stderr:write(string.format("[word_count_eval] '%s' did not return a table (got %s)\n", macs_path, type(mc)))
-		else
-			for name, count in pairs(mc) do wc[name] = count end
-		end
 	end
 
-	ctx.shared.word_counts = wc
+	-- 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 canonical 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 canonical count table. ONE assignment, no copy. The assignment creates no secondary alias.
+	corpus.word_counts = wc
+
 	return { outputs = {}, errors = {}, warnings = {} }
 end
 

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/update_deps.ps1

@@ -39,13 +39,7 @@ pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
 # PCSX-Redux — built via MSBuild (VS2022)
-#
 # Requires: Visual Studio 2022 with the C++ desktop workload.
-# The .vcxproj files target platform toolset v145, but VS2022 ships v143;
-# we pass /p:PlatformToolset=v143 to retarget at build time (no file edits).
-# NuGet packages (glfw, luajit.native, libFFmpeg-lite, x64sentry) are
-# restored automatically by MSBuild on first build.
-#
 # Output: toolchain\pcsx-redux\vsprojects\x64\Debug\pcsx-redux.exe
 # ════════════════════════════════════════════════════════════════════════════
 
@@ -65,8 +59,7 @@ $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
+# we use `scoop prefix` to find the install root for the include dir (needed to compile lpeg against luajit's headers).
-# (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'))) {
@@ -87,7 +80,6 @@ if (-not $lua_inc_dir) {
 # 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
-# (per the lpeg makefile — no `make.lua` template generator in this version).
 # 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')
@@ -103,8 +95,6 @@ pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
 # lfs (LuaFileSystem) — compiled from pcsx-redux's vendored luafilesystem source.
-# Used by word_count_eval.lua :: scan_dir for native directory enumeration (~2ms)
-# instead of spawning `dir /b /s` as a subprocess (~56ms).
 # Source: toolchain/pcsx-redux/third_party/luafilesystem/src/lfs.c
 # Output: toolchain/lfs/lfs.dll
 # ════════════════════════════════════════════════════════════════════════════
@@ -118,11 +108,6 @@ $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 
 # ════════════════════════════════════════════════════════════════════════════
 # OpenBIOS — built from the PCSX-Redux source tree via make + mipsel-none-elf
-#
-# OpenBIOS is an open-source PS1 BIOS implementation (no retail BIOS dump needed). 
-# It builds with the MIPS cross-toolchain (`mipsel-none-elf-gcc`, on PATH via the `mips` toolchain installer) 
-# + `make` (on PATH via scoop).
-#
 # Output: toolchain\pcsx-redux\src\mips\openbios\openbios.bin
 # ════════════════════════════════════════════════════════════════════════════