offset corections (dwarf)

Needs heavy curation and problably simplicication.

.gitignore

@@ -1,8 +1,9 @@
 build
 toolchain/armips
+toolchain/luajit-2.1
 toolchain/pcsx-redux
-# toolchain/psyq_iwyu
-# toolchain/PSn00bSDK
+toolchain/psyq_iwyu
+toolchain/PSn00bSDK
 
 *.exe
 *.elf
@@ -14,3 +15,5 @@ toolchain/pcsx-redux
 *.a
 .sentry-native
 .vscode/settings.json
+toolchain/lfs
+toolchain/lpeg

.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

@@ -2,437 +2,139 @@
  * atom_dsl.h
  * ============================================================================
  *
- * ATOM DSL — annotation layer for tape atoms (lottes_tape.h).
+ * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
+ * The metaprogram (scripts/passes/annotation.lua) reads source-as-written and validates:
+ *   - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X), atom_reads(...), atom_writes(...)) in any order and are optional.
+ *   - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a real Binds_* struct declaration.
+ *   - atom word-counts in word_counts.metadata.h match the body's actual .word count.
  *
- * This header turns `__attribute__((annotate(...)))` and `_Pragma(...)` into
- * a small named DSL that the metaprogram can validate against.
- *
- *   The C compiler treats every macro below as a no-op:
- *     - atom_init / atom_terminate / atom_bind / atom_setup / atom_commit /
- *       atom_annot all expand to `__attribute__((annotate("..."))) MipsAtom_(name)`
- *       — accepted by GCC (with -Wno-attributes), absent at runtime.
- *     - atom_resource / atom_region / atom_group / atom_cadence / atom_async
- *       expand to `_Pragma("...")` — accepted by any C11 preprocessor.
- *
- *   The metaprogram (tape_atom_annotation_pass.lua) reads the source-as-written
- *   and validates:
- *     - every MipsAtom_ has one atom_*() annotation (no orphans)
- *     - phase is recognized (init/bind/setup/work/commit/terminate)
- *     - reads/writes reference canonical wave-context registers
- *     - rbind atoms reference a real Binds_* struct declaration
- *     - word-counts in tapre metadata agree with the body's actual .word count
- *     - resource/region/group/cadence/async pragmas are spelled correctly and
- *       reference known enum values
+ * Pure macro anntation.
+ * ---------------
+ * Don't want to constraint the macro usage to some attribute placment constraint, etc, don't want ot dela with the compiler.
+ * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip_over each expand to a C comment or to nothing
+ * (C preprocessor strips them to whitespace).
  *
  * ============================================================================
- *
- *  PUTTING IT ON AN ATOM — the canonical pattern
- *
- *      _tape_resources_
- *      atom_resource(cube_tri, "model_ship_cube")
- *      atom_region  (cube_tri, PRIM_ARENA)
- *      atom_group   (cube_tri, GROUP_RENDER_PRIMS)
- *      atom_cadence (cube_tri, CADENCE_FRAME)
- *
- *      atom_annot(cube_tri, phase_work,
- *                 tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
- *                 tape_regs(R_PrimCursor, R_FaceCursor))
- *      internal MipsAtom_(cube_tri) {
+ *  Usage:
+ *      MipsAtom_(cube_tri) atom_info(
+ *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor)
+ *      ){
  *          atom_label(culling),
  *          // ... atom body ...
+ *          atom_offset(culling, bounds_chk)         // branch target, validated
+ *          // ... atom body ...
  *          atom_label(bounds_chk),
  *      };
  *
- *      atom_offset(culling, bounds_chk)         // ← branch target, validated
  *
- *  RBIND pattern — `Binds_*` is the contract
+ *  Data Binding pattern -- atom_bind as a sub-call of atom_info
  *
  *      // Wave-context register layout (declarative):
- *      typedef struct Binds_TrackFaceBatch {
- *          U4 R_PrimCursor,  R_FaceCursor,
- *            R_VertBase,    R_OtBase;
- *      } Binds_TrackFaceBatch;
- *
- *      atom_resource(rbind_track_face_batch, "track_face_batch_42")
- *      atom_region  (rbind_track_face_batch, HEAP_3D)
- *      atom_group   (rbind_track_face_batch, GROUP_LOAD_FACES)
- *      atom_cadence (rbind_track_face_batch, CADENCE_ONDEMAND)
- *      atom_async   (rbind_track_face_batch, true)
- *
- *      atom_bind(rbind_track_face_batch, Binds_TrackFaceBatch,
- *                tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
- *      internal MipsAtom_(rbind_track_face_batch) { ... };
+ *      typedef Struct_(Binds_TrackFaceBatch) { 
+ * 					U4 PrimCursor; 
+ * 					U4 FaceCursor;
+ *          U4 VertBase;
+ * 					U4 OtBase;
+ *      };
+ *      MipsAtom_(rbind_track_face_batch) atom_info(
+ *         atom_bind(Binds_TrackFaceBatch)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *      ){ ... };
  *
  *  Annotation rules
  *  ----------------
- *      1. Each MipsAtom_(name) needs EXACTLY ONE atom_*() macro on the line
- *         immediately above. No annotation = orphan (warning). Two annotations
- *         on the same name = duplicate (error).
- *
- *      2. atom_init and atom_terminate take only the name.
- *
- *      3. atom_setup and atom_commit take name + reads.
- *
- *      4. atom_bind takes name + Binds_* type + writes.
- *
- *      5. atom_annot takes name + phase token + reads + writes.
- *         Phase tokens: phase_init / phase_bind / phase_setup / phase_work /
- *         phase_commit / phase_terminate.
- *
- *      6. Optional pragmas (atom_resource / atom_region / atom_group /
- *         atom_cadence / atom_async) attach metadata to the atom. They can
- *         appear in any order, with one per atom. They're independent of the
- *         atom_*() macro — multiple pragmatics are fine.
- *
- * ============================================================================
- *
- *  WHY A SEPARATE LAYER (not just put everything in source comments)?
- *
- *      Source comments are invisible to the compiler. Annotations live in the
- *      source as actual C tokens, so:
- *        - they can never silently get out of sync with the code (the build
- *          fails at preprocessing if the metaprogram disagrees)
- *        - they can be cross-validated against metadata (build fails if a
- *          WORD_COUNT entry drifts away from the .word count in source)
- *        - they make the C compiler a witness ("there's a marker here, and
- *          it's labelled, and it has arguments") without making the C compile
- *          itself do any work
- *
- * ============================================================================
+ *      1. atom_info(...) is OPTIONAL. Atoms without atom_info are silently skipped by the metaprogram.
+ *      2. If present, atom_info takes up to three sub-calls, all order-independent within the arg list:
+ *           - atom_bind(Binds_X)
+ *           - atom_reads(...)
+ *           - atom_writes(...)
+ *      3. atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *      4. atom_reads(...) and atom_writes(...): Used to to check if registers are used correctly in macros: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
+ *      5. atom_label(name: Utilize with atom_offset as a target location.
+ *      6. atom_offset(F, T): Resolved by gen/atom_offsets.h, generated from the atom_label markers. Calculated during the offset pass of the lua  metaprogram.
  */
 
- #ifdef INTELLISENSE_DIRECTIVES
+#ifdef INTELLISENSE_DIRECTIVES
 #pragma once
 // #include <stdint.h>
 #endif
 
 /* ============================================================================
- * PHASE TOKENS — strings, used as the second arg to atom_annot(...)
- *
- * Why strings? They preserve the metaprogram's ability to read phase directly
- * from the source-as-written, even when the macro isn't expanded. The Lua
- * tool also has a MACRO_EXPANSION table for resolving phase_* source-level
- * references.
- *
- *   atom_annot(cube_tri, phase_work, ...)    ← legal
- *   atom_annot(cube_tri, "work", ...)        ← legal (and equivalent)
- *   atom_annot(cube_tri, phase_setup, ...)   ← legal
- *
- * ============================================================================*/
-
-#define phase_init       "init"
-#define phase_bind       "bind"
-#define phase_setup      "setup"
-#define phase_work       "work"
-#define phase_commit     "commit"
-#define phase_terminate  "terminate"
-
-/* ============================================================================
- * WAVE-CONTEXT REGISTERS — canonical register set for the tape wave model.
- *
- * The tape-atom runtime carries four registers across a wave:
- *
- *   R_PrimCursor  output pointer into the prim arena (next OT entry to write)
- *   R_FaceCursor  input pointer into the face array (next face to consume)
- *   R_VertBase    base pointer into the vertex arena (this wave's vertices)
- *   R_OtBase      base pointer into the ordering table (this wave's OT slot)
- *
- * Each atom declares its reads/writes against this canonical set. The Lua
- * tool rejects wave-context positions that reference any other register
- * (warning today — the C compiler's R_T4..R_T7 / R_RA / etc. aliases are
- * implementation details and not part of the typed surface).
- *
- * If your atom needs to touch GTE / SP / DMA / other side state, declare it
- * at the source level as you normally would — but DO NOT put those registers
- * in tape_regs(...). Wave-context is a closed set.
- *
- * ============================================================================*/
-
-/* ============================================================================
- * REGION TOKENS — memory regions atoms may allocate from or write into.
- *
- * Use atom_region(name, REGION) to declare. The Lua tool validates that the
- * region is in this set, AND that:
- *   - rbind atoms declare the source region (usually HEAP_3D or CDROM_STREAM)
- *   - work atoms declare the destination region (the arena they push to)
- *   - commit atoms must declare a region equal to what setup wrote, so the
- *     C-side mirror is consistent
- *
- * Add new regions by extending this list and the metaprogram's KNOWN_REGIONS.
- * Don't add regions ad-hoc — every new region becomes part of the contract.
- *
- * ============================================================================*/
-#define REGION_PRIM_ARENA     prim_arena       /* OT/prim packet arena */
-#define REGION_FACE_ARENA     face_arena       /* face index array */
-#define REGION_VERTEX_ARENA   vertex_arena     /* vertex pool */
-#define REGION_OT_ARENA       ot_arena         /* ordering-table array */
-#define REGION_HEAP_3D        heap_3d_models   /* loaded model heap */
-#define REGION_CDROM_STREAM   cdrom_stream     /* CDROM read buffer */
-#define REGION_VRAM           vram_heap        /* VRAM texture/GPU buffer */
-
-/* ============================================================================
- * CADENCE TOKENS — how often the atom runs.
- *
- *   frame    runs every vsync (rendering, input poll)
- *   once     runs exactly once per process lifetime (init, terminate)
- *   ondemand runs when triggered by event (CDROM load, async DMA complete)
- *
- * Used as a hint for the metaprogram to flag:
- *   - frame-cadence atoms that have side effects (they'll be hit many times,
- *     so avoid global state mutation unless it's idempotent)
- *   - once-cadence atoms inside "if (frame_count == 0)" guards (the guard
- *     is then provably one-shot, the metaprogram can lift initialization)
- *   - ondemand atoms that are missed by the wave scheduler (forces async
- *     and discards yield results without further processing)
- *
- * ============================================================================*/
-#define CADENCE_FRAME     frame
-#define CADENCE_ONCE      once
-#define CADENCE_ONDEMAND  ondemand
-
-/* ============================================================================
- * tape_regs(...) — wave-context register list
- *
- *   tape_regs(R_PrimCursor, R_FaceCursor)  →  (R_PrimCursor, R_FaceCursor)
- *
- * The macro produces a comma-evaluated expression that the C compiler
- * silently discards (it's wrapped in parentheses in the call argument
- * position — the result is never bound). The Lua tool pattern-matches the
- * "tape_regs(...)" token to extract the list.
- *
- * You can have at most one tape_regs(...) in the reads slot and one in the
- * writes slot of atom_annot. To declare multiple disjoint sets (rare), just
- * declare the union — the metaprogram doesn't track which reads need which
- * writes at this granularity.
+ * atom_reads(...) / atom_writes(...)
  *
+ * Used during the static analysis pass of the metaprogram to do
  * ============================================================================*/
 #define atom_reads(...)  (__VA_ARGS__)
 #define atom_writes(...) (__VA_ARGS__)
 
-/* ============================================================================
- * ATOM ANNOTATION MACROS
- *
- *   Each expands to `__attribute__((annotate("kind"))) MipsAtom_(name)` —
- *   the GCC attribute is accepted under -Wno-attributes (already in your
- *   build flags) and stripped at runtime. The annotation string is just the
- *   macro kind ("atom_annot", "atom_bind", etc.) — the metaprogram reads
- *   the macro call's full args list from the source-as-written.
- *
- * ============================================================================*/
-
 /* ----------------------------------------------------------------------------
- * atom_init — entry into tape_runtime_main
+ * atom_reg (per-enum opt-in marker for the DWARF register-alias registry)
  *
- *   atom_init(tape_main)
- *   internal MipsAtom_(tape_main) { ... };
- *
- * Implies: no reads, no writes (wave-context not established yet).
+ * 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_init(name) __attribute__((annotate("atom_init")))
-
-/* ----------------------------------------------------------------------------
- * atom_terminate — exit from tape_runtime_main
- *
- *   atom_terminate(tape_exit)
- *   internal MipsAtom_(tape_exit) { ... };
- *
- * Implies: no reads, no writes (wave-context destroyed at this point).
- * ----------------------------------------------------------------------------*/
-#define atom_terminate(name) __attribute__((annotate("atom_terminate")))
-
-/* ----------------------------------------------------------------------------
- * atom_setup — pre-work atom: prepares engine state (e.g., set_gte_world)
- *
- *   atom_setup(set_gte_world, tape_regs(R_TapePtr))
- *   internal MipsAtom_(set_gte_world) { ... };
- *
- * Reads:   anything (the engine state you're reading)
- * Writes:  engine state (GTE / DMA / etc. — declared in source, not part of
- *          wave-context, so doesn't go in tape_regs)
- *
- * The metaprogram checks that setup is followed (in atomic order) by a work
- * atom in the same wave — there's no point in setting up state if no one
- * reads it.
- * ----------------------------------------------------------------------------*/
-#define atom_setup(name, reads) __attribute__((annotate("atom_setup")))
-
-/* ----------------------------------------------------------------------------
- * atom_commit — post-work atom: flushes wave-context back to C-side state
- *
- *   atom_commit(sync_prim_cursor, tape_regs(R_PrimCursor))
- *   internal MipsAtom_(sync_prim_cursor) { ... };
- *
- * Reads:   wave-context registers (the ones you sync back to C)
- * Writes:  C-side mirror (declared in source — not part of wave-context)
- *
- * The metaprogram checks that commit is preceded (in atomic order) by a
- * work atom that wrote the registers this commit is reading.
- * ----------------------------------------------------------------------------*/
-#define atom_commit(name, reads) __attribute__((annotate("atom_commit")))
-
-/* ----------------------------------------------------------------------------
- * atom_bind — rbind atom: read wave-context registers from tape pointer
- *
- *   atom_bind(rbind_cube_tri, Binds_CubeTri,
- *             tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
- *   internal MipsAtom_(rbind_cube_tri) { ... };
- *
- * The binds_struct MUST be a typedef'd type (declared via
- * `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
- * The Lua tool cross-references this. Missing struct = error.
- *
- * Implicit: reads R_TapePtr, writes the four wave-context registers.
- * ----------------------------------------------------------------------------*/
-#define atom_bind(name, binds_struct, writes) __attribute__((annotate("atom_bind")))
-
-/* ----------------------------------------------------------------------------
- * atom_annot — generic work atom with explicit phase
- *
- *   atom_annot(cube_tri, phase_work,
- *              tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
- *              tape_regs(R_PrimCursor, R_FaceCursor))
- *   internal MipsAtom_(cube_tri) { ... };
- *
- * Use this for the bulk of your atoms. For init/setup/commit/bind, prefer
- * the convenience macros above — they pin the phase for you.
- *
- * The phase arg is one of: phase_init / phase_bind / phase_setup /
- * phase_work / phase_commit / phase_terminate. Spelling mistakes are errors.
- * ----------------------------------------------------------------------------*/
-#define atom_annot(name, phase, reads, writes) __attribute__((annotate("atom_annot")))
+#define atom_reg /* atom_reg: opt the preceding enum entry into the DWARF registry */
 
 /* ============================================================================
- * RESOURCE / GROUP / CADENCE / REGION / ASYNC — optional atom metadata
- *
- *   These don't annotate the atom semantically (phase/reads/writes do that).
- *   They attach extra context that the metaprogram uses to catch:
- *     - same resource loaded twice in different ways
- *     - atoms that span multiple regions (likely bug — pick one)
- *     - frame-cadence atoms that should be once-cadence (perf / correctness)
- *     - ondemand atoms that aren't async (CDROM races)
- *
- *   You can use as many as apply to a given atom, in any order, immediately
- *   above the atom_*() macro.
- *
+ *   atom_info :
+ *      MipsAtom_(cube_tri) atom_info(
+ *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor)
+ *      ){ ... };
+ * 
+ *  - atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *  - atom_reads(...):    comma-list of registers
+ *  - atom_writes(...):   comma-list of registers
  * ============================================================================*/
+#define atom_info(...)  /* atom_info(__VA_ARGS__) */
 
 /* ----------------------------------------------------------------------------
- * atom_resource — name the logical resource the atom references
+ * DEBUG SOURCE-STEP MARKERS
  *
- *   atom_resource(cube_tri, "model_ship_cube")
- *   atom_resource(load_track_faces, "track_lavender_field_0x42")
- *   atom_resource(play_engine_sfx, "sfx_engine_loop")
- *
- * Use any human-readable string. The metaprogram:
- *   - validates resource strings are non-empty and don't contain control chars
- *   - flags duplicates across atoms with the same name (two atoms claiming
- *     ownership of a resource is usually a refactor artifact or bug)
- *   - flags references to resources that no atom actually defines
- *
- * The arg is a STRING LITERAL, so it can't accidentally alias a variable.
+ * 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_resource(name, res_id) //_Pragma("atom " #name " resource=" res_id)
+#define atom_dbg_skip_over() /* atom_dbg_skip_over: skip the following atom or component source view */
 
 /* ----------------------------------------------------------------------------
- * atom_region — name the memory region the atom touches
+ * 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.
  *
- *   atom_region(cube_tri,   REGION_PRIM_ARENA)
- *   atom_region(load_faces, REGION_HEAP_3D)
- *   atom_region(load_tex,   REGION_VRAM)
- *
- * Use REGION_* tokens above. The metaprogram enforces the closed set.
- *
- * Edge cases the metaprogram catches:
- *   - rbind atom that doesn't declare a SOURCE region (where is it loading from?)
- *   - work atom with no destination region (where is it pushing to?)
- *   - region that disagrees with the Binds_* struct layout (you said it's a
- *     prim_arena rbind but the struct has 4 faces in it — wait, that's wrong)
+ * 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_region(name, region) //_Pragma("atom " #name " region=" #region)
+#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_group — bundle atoms into a logical batch (track-load, sound-load, etc.)
+ * atom_bind(Binds_X) -- rbind sub-call of atom_info
  *
- *   atom_group(load_track_face_42,  GROUP_LOAD_FACES)
- *   atom_group(load_track_face_43,  GROUP_LOAD_FACES)
- *   atom_group(swap_face_42_43,     GROUP_VISIBILITY_SWAP)
+ *   MipsAtom_(rbind_cube_tri) atom_info(
+ *         atom_bind(Binds_CubeTri)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *   ){ ... };
  *
- * Use any token as the group id. The metaprogram:
- *   - validates all atoms in a group emit their waves in the same tb_group
- *     (no spawning other waves inside a group)
- *   - flags groups with only one member (probably a typo — meant to be a group?)
- *   - validates cross-group edges (no atom reads what another group writes,
- *     unless explicitly grouped together)
- *
- * Useful when:
- *   - subdivisible work (track-face batches, polygon subdivision) needs to
- *     confirm that all batches of one logical visible scene are emitted
- *     together
- *   - async loads (CDROM -> VRAM) need to be grouped so all batches complete
- *     before the swap
- *
- * Use GROUPS for sound effects to track which sound plays during which atom,
- * which is needed if the sound tool ever has to validate "this atom is the
- * trigger for an audio play".
+ * The Binds_X MUST be a typedef'd type (declared via `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
  * ----------------------------------------------------------------------------*/
-#define atom_group(name, group_id) //_Pragma("atom " #name " group=" #group_id)
-
-/* ----------------------------------------------------------------------------
- * atom_cadence — declare execution frequency
- *
- *   atom_cadence(render_frame,     CADENCE_FRAME)       // every vsync
- *   atom_cadence(load_track_faces, CADENCE_ONDEMAND)    // on demand
- *   atom_cadence(init_heap,        CADENCE_ONCE)        // process lifetime
- *
- * Default (no atom_cadence call) is CADENCE_FRAME — most atoms run every
- * frame. Override explicitly when not.
- *
- * The metaprogram's checks:
- *   - CADENCE_ONCE atoms inside `if (frame == 0)` or `if (!initialized)` are
- *     tagged, validating that guards are required (or warning if missing)
- *   - CADENCE_FRAME atoms that mutate state outside the wave context get
- *     flagged (likely a bug — state should persist through commits)
- *   - CADENCE_ONDEMAND atoms must have atom_async — otherwise the trigger
- *     mechanism is undefined
- * ----------------------------------------------------------------------------*/
-#define atom_cadence(name, cadence) //_Pragma("atom " #name " cadence=" #cadence)
-
-/* ----------------------------------------------------------------------------
- * atom_async — declare whether the atom yields / interacts with CDROM DMA
- *
- *   atom_async(load_track_tex,  true)   // CDROM read yield
- *   atom_async(load_vram,       true)   // VRAM upload DMA
- *   atom_async(render_frame,    false)  // pure compute, no async
- *
- * The metaprogram requires this for CADENCE_ONDEMAND atoms. For
- * CADENCE_FRAME, it's optional but documents intent.
- *
- * Note: CDROM ATOMS in Psy-Q are typically implemented as a chain of
- * "async-init" atom followed by a "wait-for-completion" atom. Both atoms
- * should be marked async=true, and both should have the same resource/group
- * tag (so the metaprogram can verify they're paired).
- * ----------------------------------------------------------------------------*/
-#define atom_async(name, is_async) //_Pragma("atom " #name " async=" #is_async)
-
-/* ============================================================================
- * WORD-COUNT ANNOTATION FOR A #define MAC
- *
- *   tape_words(mac_yield, 1)
- *   #define mac_yield()  \
- *       load_word(R_AtomJmp, R_TapePtr, 0), \
- *       add_ui_self(R_TapePtr, 4), \
- *       jump_reg(R_AtomJmp), \
- *       nop
- *
- * The compiler accepts the unknown _Pragma. The Lua tool reads it and
- * cross-checks against WORD_COUNT(mac_yield, 1) in tape_atom.metadata.h.
- * If they disagree, build fails.
- *
- * Use sparingly — only on multi-word macros (single-word ones don't need
- * drift tracking; they're checked by the .word-count pass anyway).
- *
- * ============================================================================*/
-#define tape_words(name, n) //_Pragma(#name " tape_atom words=" #n)
+#define atom_bind(binds_struct)  /* atom_bind(binds_struct) */
 
 /* ============================================================================
  * atom_label / atom_offset — branch target machinery
@@ -443,25 +145,11 @@
  *
  *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
  *
- * The metaprogram generates gen/atom_offsets.h with one
- *   #define atom_offset__culling__bounds_chk  ((target - branch_pos - 1))
- * per atom_offset(F, T) call. The preprocessor then expands your call to
- * the right immediate value.
- *
- * If gen/atom_offsets.h is stale (or atom_label(name) is undefined),
- * `atom_offset__F__T` becomes an undefined macro and the C build fails.
- * This catches:
- *   - typo in atom_label (no anchor → metaprogram doesn't emit the macro)
- *   - .offsets.h not regenerated after body edits
- *   - body edit that broke the offset math (recompile + retest picks it up
- *     in CPU emulator)
+ * The metaprogram generates gen/atom_offsets.h with one #define with the offset value per atom_offset(F, T) call.
+ * The preprocessor then expands the call to the right immediate value.
  *
+ * If gen/atom_offsets.h is stale (or atom_label(name) is undefined), `atom_offset_F_T` becomes an undefined macro and the C build fails.
  * ============================================================================*/
-
 #define atom_offset(F, T) atom_offset_ ## F ## _ ## T
-/* atom_label is a pure annotation for the metaprogram's offset calculations.
- * The macro expands to a C comment, so the C preprocessor strips it to
- * whitespace — NO instruction word is emitted in the asm. The metaprogram
- * still recognises the literal `atom_label(name)` token in source and
- * records the marker at the current pos. */
+// atom_label is a pure annotation for the metaprogram's offset calculations.
 #define atom_label(name) /* atom_label anchor: name */

code/duffle/gen/duffle.macs.h

@@ -0,0 +1,125 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#define WORD_COUNT(name, count)  enum { words_##name = (count) };
+#endif
+
+#define mac_yield(...) \
+	load_word(R_AtomJmp, R_TapePtr, 0) \
+,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
+,	jump_reg( R_AtomJmp) \
+,	nop
+WORD_COUNT(mac_yield, 4)
+
+/* Words: 3; Loads 3 S2 indices from the face array */
+#define mac_load_tri_indices(...) \
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)) \
+,	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)) \
+,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
+WORD_COUNT(mac_load_tri_indices, 3)
+
+#define mac_gte_load_tri_verts(...) \
+	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
+,	add_u_self(R_AT, R_VertBase) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	gte_mv_to_data_r(R_V0, C2_VXY0) \
+,	gte_mv_to_data_r(R_V1, C2_VZ0) \
+,	shift_lleft(R_AT, R_T1, v3s2_byteoff) \
+,	add_u_self(R_AT, R_VertBase) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	gte_mv_to_data_r(R_V0, C2_VXY1) \
+,	gte_mv_to_data_r(R_V1, C2_VZ1) \
+,	shift_lleft(R_AT, R_T2, v3s2_byteoff) \
+,	add_u_self(R_AT, R_VertBase) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	gte_mv_to_data_r(R_V0, C2_VXY2) \
+,	gte_mv_to_data_r(R_V1, C2_VZ2)
+WORD_COUNT(mac_gte_load_tri_verts, 18)
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
+#define mac_insert_ot_tag_f3(...) \
+	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
+,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
+,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
+,	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (5 - 1) << 24 = 4 << 24 */ \
+,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
+,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
+,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
+,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
+,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
+,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
+WORD_COUNT(mac_insert_ot_tag_f3, 11)
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
+#define mac_insert_ot_tag_g4(...) \
+	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
+,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
+,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
+,	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (9 - 1) << 24 = 8 << 24 */ \
+,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
+,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
+,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
+,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
+,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
+,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
+WORD_COUNT(mac_insert_ot_tag_g4, 11)
+
+#define mac_pack_color_word(off, cmd, r, g, b) \
+	load_upper_i(R_AT, (cmd) << 8  | (b)) \
+,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
+,	store_word(  R_AT, R_PrimCursor, (off))
+WORD_COUNT(mac_pack_color_word, 3)
+
+#define mac_format_f3_color(r, g, b) \
+	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
+WORD_COUNT(mac_format_f3_color, 3)
+
+/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
+ * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
+#define mac_gte_store_f3_post_rtpt(...) \
+	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
+,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
+,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
+WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
+
+#define mac_format_g4_color(r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
+	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
+,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
+,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
+,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
+WORD_COUNT(mac_format_g4_color, 12)
+
+/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
+ * G4 triangle portion to p0/p1/p2.
+ * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
+ * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
+ * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
+ * three registers aligned with v0/v1/v2 you must store before RTPS).
+ * The macro name declares the pipeline position; check #6 (GTE state-
+ * machine validation) verifies the call site matches the declaration. */
+#define mac_gte_store_g4_p012_post_rtpt_pre_rtps(...) \
+	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)) \
+,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)) \
+,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
+WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
+
+/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
+ * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
+ * single-vertex result to SXY2; SXY0 still holds v0.screen from the
+ * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
+ * prevents).
+ */
+#define mac_gte_store_g4_p3_post_rtps(...) \
+	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
+WORD_COUNT(mac_gte_store_g4_p3_post_rtps, 1)
+

code/duffle/gen/duffle.offsets.h

@@ -1,4 +1,4 @@
-// Auto-generated by tape_atom_offset_gen.meta.lua — DO NOT EDIT
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
 // Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
 #pragma once
 

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
- *  PSYQ-style names.
+ *  They live in the opt-in `gp_vendor_sym.h` for users who prefer the PSYQ-style names.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -41,20 +39,15 @@
 /* ============================================================================
  *  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
- * tape-side macros that pin a register to hold the IO base and access
- * ports via offsets — `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.
- *
- * See lottes_tape.h `R_GpIoBase` + `mac_gp0_send_imm` for the
- * wave-context form that composes these primitives. */
+ * The 16-bit upper half `IO_BASE_ADDR_HI16` is the form used by tape-side macros that pin a register
+ * to hold the IO base and access ports via offsets:
+ *   `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
+ * Mirrors the `IO_BASE_ADDR equ 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s. */
 enum {
     IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
     IO_BASE_ADDR_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
@@ -79,14 +72,11 @@ 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; pre-baked 32-bit words are in §10.4.
- *  The layer-1 bitfield-layout constants live in the same enum block
- *  so the encoder in §10.4 can reference them by name. NO macro body
- *  past this point uses a raw shift or raw mask — every shift/width/mask
- *  is named here, named once. Mirrors the OPCODE_SHIFT / RS_SHIFT /
- *  REG_MASK convention from mips.h.
+ *  8-bit GP0 opcodes (the upper byte of a primitive's first word). These are the BYTE only.
+ *  The layer-1 bitfield-layout constants live in the same enum block so the encoder can reference them by name. 
+ *  NO macro body past this point uses a raw shift or raw mask.
+ *  Every shift/width/mask is named here, named once.
+ *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
  * ============================================================================ */
 enum {
     gp0_cmd_Nop             = 0x00,
@@ -120,8 +110,7 @@ enum {
     gp0_cmd_tile_8          = 0x68,
     gp0_cmd_tile_16         = 0x70,
 
-    /* State setters (not drawing primitives; set render context).
-     * Per PSX-SPX graphicsprocessingunitgpu.md §"GP0 Other Commands". */
+    /* State setters (not drawing primitives; set render context). */
     gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
     gp0_cmd_SetTextureWindow     = 0xE2,
     gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
@@ -130,9 +119,7 @@ enum {
     gp0_cmd_SetMaskBit           = 0xE6,
 
     /* bitfield shifts / widths / masks ----
-     *
-     * Generic GP0/GP1 command byte (upper 8 bits of every word sent
-     * to either port). Used by `enc_gp0_cmd(cmd)` and friends below. */
+     * Generic GP0/GP1 command byte (upper 8 bits of every word sent to either port). */
     gp0_cmd_shift = 24,
     gp0_cmd_width =  8,
     gp0_cmd_mask  = 0xFF,
@@ -151,21 +138,19 @@ 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. 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.
+ *  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.
  * ============================================================================ */
 
 /* ---- Layer 1.5: per-field encoders ---- */
-#define enc_gp0_cmd(cmd)       (((cmd) & gp0_cmd_mask)           << gp0_cmd_shift)
+#define enc_gp0_cmd(cmd)       (((cmd) & gp0_cmd_mask)         << gp0_cmd_shift)
 
-#define enc_gp0_color_cmd(cmd) (((cmd) & gp0_color_cmd_mask)     << gp0_color_cmd_shift)
-#define enc_gp0_color_r(r)     (((r)   & gp0_color_red_mask)     << gp0_color_red_shift)
-#define enc_gp0_color_g(g)     (((g)   & gp0_color_green_mask)   << gp0_color_green_shift)
-#define enc_gp0_color_b(b)     (((b)   & gp0_color_blue_mask)    << gp0_color_blue_shift)
+#define enc_gp0_color_cmd(cmd) (((cmd) & gp0_color_cmd_mask)   << gp0_color_cmd_shift)
+#define enc_gp0_color_r(r)     (((r)   & gp0_color_red_mask)   << gp0_color_red_shift)
+#define enc_gp0_color_g(g)     (((g)   & gp0_color_green_mask) << gp0_color_green_shift)
+#define enc_gp0_color_b(b)     (((b)   & gp0_color_blue_mask)  << gp0_color_blue_shift)
 
 /* ---- Layer 2: composite encoders ---- */
 #define enc_color_word(cmd, r, g, b) (enc_gp0_color_cmd(cmd) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
@@ -194,10 +179,9 @@ enum {
 /* ============================================================================
  *  GP1 command byte constants + Layer 1 (display-mode + range + draw-area bitfield shifts)
  *  ============================================================================
- *
- *  GP1 status bits are read from HW_GP1; ctrl writes use GP1 commands
- *  packed into 32-bit words (cmd byte in the upper 8 bits via
- *  `enc_gp0_cmd(cmd)` — never a raw shift).
+ *  GP1 status bits are read from HW_GP1;
+ *  ctrl writes use GP1 commands packed into 32-bit words 
+ *  (cmd byte in the upper 8 bits via `enc_gp0_cmd(cmd)`).
  * ============================================================================ */
 enum {
     gp1_cmd_Reset                  = 0x00,
@@ -210,10 +194,9 @@ enum {
     gp1_cmd_VerticalDisplayRange   = 0x07,
     gp1_cmd_DisplayMode            = 0x08,
     /* Note: GP1 only has commands 0x00..0x08. 
-		 * The state-setter commands (SetTextureWindow, * SetDrawArea*, 
-		 * SetDrawOffset, SetMaskBit) live in the GP0 enum as * 0xE1..0xE6. 
-		 * DrawArea word builders are below as GP0s * macros 
-		 * (since they emit GP0 commands). */
+		 * The state-setter commands (SetTextureWindow, * SetDrawArea*, SetDrawOffset, SetMaskBit) 
+		 * live in the GP0 enum as * 0xE1..0xE6.
+		 * DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
 
     /* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
      * Bit positions match the encoder shifts below; values are the
@@ -267,8 +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
- * both use the same X/Y 10-bit signed payload as GP1 DisplayRange. */
+ * GP0(0xE3) SetDrawArea top-left and GP0(0xE4) SetDrawArea bottom-right 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))
 
@@ -290,16 +272,7 @@ enum {
 /* ============================================================================
  *  Pre-baked GPU state words
  *  ============================================================================
- *
- *  Common command words for boot-time GPU init and standard
- *  display configurations. Each one is a pure compile-time integer
- *  constant ready to drop into a `.word` directive.
- *
- *  These are the equivalents of the `gp_HorizontalDisplayRange_3168_608`,
- *  `gp_VerticalDisplayRange_264_24`, `gp_DisplayMode_320x240_15bit_NTSC`,
- *  `gp_SetDrawMode_DrawAllowed`, `gp_DMA_*` `.equ`s from the pre-rewrite
- *  gp.h / graphics_hello/gp.s, rebuilt using the layer-cake encoders so
- *  no magic numbers appear in any body.
+ *  Common command words for boot-time GPU init and standard display configurations.
  * ============================================================================ */
 
 /* ---- Display enable (1-bit payload on DisplayEnable cmd) ---- */
@@ -339,7 +312,7 @@ enum {
 #define gp1_word_vertical_range_pal    enc_gp1_vrange_word(gp1_vrange_PAL_y1,  gp1_vrange_PAL_y2)
 
 /* ---- Draw-mode setting (TPage / draw-area allowance) ---- */
-/* The pre-baked "drawing enabled" word is the standard post-init state. */
+/* The "drawing enabled" word is the standard post-init state. */
 enum {
     gp0_DrawMode_DrawToDispBit = 10,
 };
@@ -372,17 +345,14 @@ enum {
 /* ============================================================================
  *  Primitive structs (8 polygon variants + tag)
  *  ============================================================================
+ *  Each struct follows the GPU-documented memory layout for the corresponding primitive command. 
+ *  The PolyTag is the OT-link header; the rest of the struct is the primitive's body.
  *
- *  Each struct follows the GPU-documented memory layout for the corresponding
- *  primitive command. The PolyTag is the OT-link header; the rest of the
- *  struct is the primitive's body.
- *
- *  The current working layouts match the existing demo (floor_tri uses
- *  Poly_F3; cube_tri uses Poly_G4). They are NOT necessarily byte-identical
- *  to the PSX-SPX reference layout — the demo layout uses color+vertex
- *  interleaving that doesn't match the standard PSX SDK file format. For
- *  PSX-SDK file compatibility, the textured variants (FT*, GT*) would need
- *  layout adjustments; out of scope for this track.
+ *  The current working layouts match the existing demo 
+ *  (floor_tri uses Poly_F3; cube_tri uses Poly_G4). 
+ *  They are NOT necessarily byte-identical to the PSX-SPX reference layout. 
+ *  The demo layout uses color+vertex interleaving that doesn't match the standard PSX SDK file format. 
+ *  For PSX-SDK file compatibility, the textured variants (FT*, GT*) would need layout adjustments.
  * ============================================================================ */
 
 /* ---------- RGB8 (3-byte packed color) ---------- */
@@ -391,12 +361,12 @@ typedef Struct_(RGB8) { B1 r; B1 g; B1 b; };
 
 /* ---------- PolyTag (the OT-link header; 1 word) ---------- */
 enum {
-    polytag_len_bits  =  8,
-    polytag_addr_bits = 24,
+    PolyTag_len_bits  =  8,
+    PolyTag_addr_bits = 24,
 };
 typedef Struct_(PolyTag) {
 	union {
-		U4 bf_addr_len;
+		U4 code;
 		struct {
 			U4 addr: 24;
 			U4 len:  8;
@@ -404,15 +374,14 @@ typedef Struct_(PolyTag) {
 	};
 };
 
-/* DSL cast convention: every cast uses `C_()`, every pointer qualifier
- * is `R_` (restrict) or `V_` (volatile). No raw C-style casts. RHS values
- * are assumed to be `U4` — caller passes a `U4` directly. */
+/* DSL cast convention: every cast uses `C_()`, every pointer qualifier is `R_` (restrict) or `V_` (volatile).
+ * No raw C-style casts. RHS values are assumed to be `U4` — caller passes a `U4` directly. */
 #define set_len(tag,v)  (C_(PolyTag_R,tag)->len  = u4_(v))
 #define set_addr(tag,v) (C_(PolyTag_R,tag)->addr = u4_(v))
-/* `set_code` is no longer in the new PolyTag design — the code byte lives
- *  in the primitive body (e.g. `((Poly_F3*)(p))->code`), not in the tag.
- *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the
- *  `set_poly_*` setters, which set both the tag's length and the code. */
+/* `set_code` is no longer in the new PolyTag design — the code byte lives in the primitive body
+ * (e.g. `((Poly_F3*)(p))->code`), not in the tag.
+ *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters,
+ *  which set both the tag's length and the code. */
 #define get_len(tag)  C_(U4,C_(PolyTag_R,tag)->len)
 #define get_addr(tag) C_(U4,C_(PolyTag_R,tag)->addr)
 
@@ -438,7 +407,7 @@ typedef Struct_(Poly_F4) {
     };
 };
 
-/* ---------- Poly_G3 (Gouraud Triangle; 6 words) ---------- */
+/* ---------- Poly_G3 (Gouraud Triangle; 7 words) ---------- */
 typedef Struct_(Poly_G3) {
     U4    tag; RGB8 c0; B1 code;
     V2_S2 p0;  RGB8 c1; B1 pad1;
@@ -446,7 +415,7 @@ typedef Struct_(Poly_G3) {
     V2_S2 p2;
 };
 
-/* ---------- Poly_G4 (Gouraud Quad; 5 words in the demo's interleaved layout) ---------- */
+/* ---------- Poly_G4 (Gouraud Quad; 9 words) ---------- */
 typedef Struct_(Poly_G4) {
     U4    tag; RGB8 c0; B1 code;
     V2_S2 p0;  RGB8 c1; B1 pad1;
@@ -468,7 +437,7 @@ typedef Struct_(Poly_FT3) {
     V2_S2 p2; U1 u2; U1 v2;
 };
 
-/* ---------- Poly_FT4 (Flat Textured Quad; placeholder layout) ---------- */
+/* ---------- Poly_FT4 (Flat Textured Quad) ---------- */
 typedef Struct_(Poly_FT4) {
     U4   tag;
     RGB8 color;
@@ -481,7 +450,7 @@ typedef Struct_(Poly_FT4) {
     V2_S2 p3; U1 u3; U1 v3;
 };
 
-/* ---------- Poly_GT3 (Gouraud Textured Triangle; placeholder layout) ---------- */
+/* ---------- Poly_GT3 (Gouraud Textured Triangle) ---------- */
 typedef Struct_(Poly_GT3) {
     U4    tag; RGB8 c0; B1 code;
     V2_S2 p0;  RGB8 c1; B1 pad1;
@@ -494,7 +463,7 @@ typedef Struct_(Poly_GT3) {
     V2_S2 tp2; U1 u2; U1 v2;
 };
 
-/* ---------- Poly_GT4 (Gouraud Textured Quad; placeholder layout) ---------- */
+/* ---------- Poly_GT4 (Gouraud Textured Quad) ---------- */
 typedef Struct_(Poly_GT4) {
     U4    tag; RGB8 c0; B1 code;
     V2_S2 p0;  RGB8 c1; B1 pad1;
@@ -509,7 +478,7 @@ typedef Struct_(Poly_GT4) {
     V2_S2 tp3; U1 u3; U1 v3;
 };
 
-/* ---------- Primitive setters (C-level, no emitted words) ----------
+/* ---------- Primitive setters (C-level) ----------
  * DSL cast convention: every cast via C_(), every pointer via R_/V_. */
 #define set_poly_f3(p)  set_len(p, 4),  C_(Poly_F3_R, p)->code = gp0_cmd_poly_f3
 #define set_poly_ft3(p) set_len(p, 7),  C_(Poly_FT3_R,p)->code = gp0_cmd_poly_ft3
@@ -530,7 +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)
@@ -594,7 +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)
@@ -627,7 +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)
@@ -645,9 +611,8 @@ enum {
  *         +0x06 U2 px_height
  *         +0x08 ... pixel data
  *
- *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that
- *  emits the necessary GP0 commands. Stoppped for now at the
- *  struct + enum level for this track.
+ *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that emits the necessary GP0 commands.
+ *  Stoppped for now at the struct + enum level.
  * ============================================================================ */
 enum {
     tim_file_id_magic = 0x10,
@@ -660,7 +625,7 @@ enum {
 };
 
 typedef Struct_(TIM_Header) {
-    U4 file_id;        /* always 0x10 = "TIM" magic */
+    U4 file_id;       /* always 0x10 = "TIM" magic */
     U4 version;       /* ignored; always 0 */
     U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
 };
@@ -678,35 +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
- *  Lottes tape model uses OT-DMA for primitive submission, so atom bodies
- *  write to main RAM (the OT/primitive buffer) and to GTE state — never
- *  directly to the GPU ports at 0x1F801810 / 0x1F801814. See
- *  `mac_format_f3_color`, `mac_insert_ot_tag`, `mac_gte_store_f3` in
- *  lottes_tape.h for the patterns atom bodies actually use.
+ *  No `mac_gp0_send` or related macros live in gp.h.
+ *  Rationale: the Lottes tape model uses OT-DMA for primitive submission, so atom bodies write to main RAM (the OT/primitive buffer)
+ *  and to GTE state — never directly to the GPU ports at 0x1F801810 / 0x1F801814.
+ *  See `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
- *  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_*`
- *  family — the encoder infrastructure is already in place:
+ *  If a feature need arises requires tape-side GPU port writes 
+ *  (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_*` family:
+ *    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.
- *         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.
- *
- *  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.
+ *  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
- *  primitive setters and OT operations. The duffle snake_case names are
- *  primary; this header is for users who prefer the PSYQ SDK function
- *  names from the legacy C API.
+ *  Provides the PSYQ-style CamelCase aliases for the canonical duffle GPU primitive setters and OT operations.
+ *  The duffle snake_case names are 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
- *  descriptive; no vendor alias is provided for them.
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (tape_atom.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries (the ones emitted by
- *  mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13
- *  (objdump byte-identical) holds.
+ *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and 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.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/gte.h

@@ -37,49 +37,6 @@
  *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word`
  *  constants from C. No GCC inline-assembly string syntax in the code body.
  *
- *  PHILOSOPHY
- *  ----------
- *  1. A 32-bit instruction word is composed from per-field encoders. Each
- *     encoder knows only its own bit range; the composite ORs them together.
- *     No magic numbers inside any encoder body. Every shift and mask is a
- *     named constant from the bitfield-layout enum below.
- *
- *  2. Pure (compile-time) instructions. Every GTE *command* (RTPS, RTPT,
- *     NCLIP, MVMVA, …) and every COP2 *transfer* (ctc2/cfc2) with a constant
- *     rs/rt/rd — are emitted as a single integer constant via
- *     `asm_inline(...)` from gcc_asm.h. The C compiler constant-folds
- *     these into `.word` directives in .rodata.
- *
- *  3. Runtime-base-register instructions (lwc2, swc2, lw, sw, …) cannot be
- *     a pure compile-time word because the `rs` field is chosen by the
- *     compiler at codegen. For these we use a "placeholder-pun" pattern:
- *     a fixed register number (R_T4 = $12) is baked into the rs field of
- *     the `.word` constant, and the macro declares a `"r"(arg)` input
- *     constraint plus a clobber on the same register. The compiler is
- *     therefore *forced* to bind `arg` to that exact register, and the
- *     constant is correct.
- *
- *  USAGE
- *  -----
- *      // Pure command sequence — all bits compile-time:
- *      asm volatile(
- *          asm_inline( gte_cmd_rtpt , gte_cmd_nclip , gte_cmd_avsz3 )
- *          asm_clobber( clbr_volatile_gprs )
- *      );
- *
- *      // Runtime-base-register load — caller picks the base GPR:
- *      register V3_S2* p_in_12 __asm__("$12") = verts[0].ptr;
- *      gte_load_v0(p_in_12, R_T4);   // R_T4 = 12 = $t4 = $12
- *
- *      // Three independent bases for an RTPT pipeline:
- *      register V3_S2* p0 gcc_reg(R_T4) = verts[0].ptr;
- *      register V3_S2* p1 gcc_reg(R_T5) = verts[1].ptr;
- *      register V3_S2* p2 gcc_reg(R_T6) = verts[2].ptr;
- *      gte_load_v0(p0, R_T4);
- *      gte_load_v1(p1, R_T5);
- *      gte_load_v2(p2, R_T6);
- *      gte_rtpt();
- *
  *  STYLE NOTES
  *  -----------
  *  - Per-field encoders are named `enc_gte_<field>(value)` and each one
@@ -94,8 +51,7 @@
  *
  *  SEE ALSO
  *  --------
- *  - gcc_asm.h: the `.word` emitter (`asm_inline`, `asm_clobber`, clobbers)
- *  - mips.h:    the MIPS encoder layer this builds on
+ *  - mips.h: The MIPS encoder layer this builds on.
  */
 
 /* C2 data registers */
@@ -222,8 +178,8 @@ enum {
  *                                    \_____ GTE_PAYLOAD _____/       \__ GTE_CMD __/
  *
  * Shifts/masks below are the *bit positions* and *bit widths* of each
- * configurable field, used by the ENC_GTE_CMD encoder. Mirrors the
- * OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
+ * configurable field, used by the ENC_GTE_CMD encoder. 
+ * Mirrors the OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
  */
 
 	gte_shift_sf  = 19,  gte_width_sf  = 1,  gte_mask_sf  = 0x1,
@@ -440,6 +396,8 @@ enum { _C2_TX_SUBS_ = 0
 #define gte_cmdw_rtpt   (gte_cmd_base | enc_gte_cmd(gte_cmd_rtpt ) | gte_cmdw_psyq_compat)
 #define gte_cmdw_nclip  (gte_cmd_base | enc_gte_cmd(gte_cmd_nclip))
 #define gte_cmdw_op     (gte_cmd_base | enc_gte_cmd(gte_cmd_op   ))
+#define gte_cmdw_outer_product gte_cmdw_op      /* "outer product" -- NOCASH/Sdk terminology */
+#define gte_cmdw_wedge        gte_cmdw_op      /* "wedge product"  -- geometric-algebra terminology */
 #define gte_cmdw_mvmva  (gte_cmd_base | enc_gte_cmd(gte_cmd_mvmva))
 
 #define gte_cmdw_rotate_translate_perspective_single gte_cmdw_rtps
@@ -713,14 +671,14 @@ enum {
 		asm_words(                          \
 			  load_word(R_T5, R_T4,  0)       \
 			, load_word(R_T6, R_T4,  4)       \
-			, gte_mt(   R_T5,        0)       \
-			, gte_mt(   R_T6,        1)       \
+			, gte_mv_to_data_r(   R_T5,  0)   \
+			, gte_mv_to_data_r(   R_T6,  1)   \
 			, load_word(R_T5, R_T4,  8)       \
 			, load_word(R_T6, R_T4, 12)       \
 			, load_word(R_T4, R_T4, 16)       \
-			, gte_mt(   R_T5,  2)             \
-			, gte_mt(   R_T6,  3)             \
-			, gte_mt(   R_T4,  4)             \
+			, gte_mv_to_data_r(   R_T5,  2)   \
+			, gte_mv_to_data_r(   R_T6,  3)   \
+			, gte_mv_to_data_r(   R_T4,  4)   \
 		)                                   \
 		, r_use(r0)                         \
 		asm_clobber: clbr_volatile_gprs, rlit(R_T4), rlit(R_T5), rlit(R_T6) \

code/duffle/gte_vendor_sym.h

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

code/duffle/lottes_tape.h

@@ -6,29 +6,42 @@
 # include "gte.h"
 # include "memory.h"
 #	include "atom_dsl.h"
+#	include "gen/duffle.macs.h"
+#	include "gen/duffle.offsets.h"
 #endif
 
 typedef U4 const MipsCode;
-#define MipsAtom_(sym) MipsCode tmpl(code,sym) [] align_(4) =
+typedef Slice_(MipsCode);
+typedef Slice_MipsCode MipsAtom;
 
-#pragma region Tape Drive
-/* ---------------------------------------------------------------------------
- *  TAPE DRIVE ABI & REGISTER ALIASES
- * ---------------------------------------------------------------------------
- *  We map the MIPS temporary registers to a persistent global workspace.
- *  The C compiler is completely unaware of these bindings.
- * ---------------------------------------------------------------------------*/
+#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+
+// Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
+//   MipsAtomComp_(ac_X) { body }
+// expands to:
+//   MipsCode ac_X[] align_(4) = { body };
+#define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+
+// Used for components with value-args (e.g., ac_format_f3_color).
+//   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+// expands to:
+//   FI_ MipsAtom ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+#define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
+
+// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
+
+/* Register aliases */
 enum {
-	R_AtomJmp    = R_T9,
-	R_TapePtr    = R_T8,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,  /* Input data cursor */
+	R_AtomJmp    = R_T9 atom_reg,  /* debug-visible; tape yield handshake scratch */
+	R_TapePtr    = R_T8 atom_reg,  /* The Instruction Stream Pointer */
+	R_InCursor   = R_T4,
 
-	R_PrimCursor = R_T7,  /* VRAM output cursor (primitive buffer) */
-	R_FaceCursor = R_T4,  /* Input data cursor (indices/faces) */
-	R_VertBase   = R_T5,  /* Base address of the vertex array */
-	R_OtBase     = R_T6,  /* Base address of the Ordering Table */
+	R_PrimCursor = R_T7 atom_reg atom_type(U4 *),    /* VRAM output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2 *), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2 *), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4 *),    /* Base address of the Ordering Table */
 
-/* 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
 
@@ -38,6 +51,11 @@ enum {
 #define R_OtBase_Code      R_T6_Code
 };
 
+#pragma region Tape Drive
+/* ---------------------------------------------------------------------------
+ *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
+ * ---------------------------------------------------------------------------*/
+
 /* The 'Exit' Atom */
 MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
@@ -69,13 +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_U4 tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_U4){ C_(U4*,tb->ptr), tb->used }; }
 FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Slice_U4){ 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
 
@@ -85,99 +103,115 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
  *  These do NOT yield. They are expanded inline inside Tape Atoms.
  * ---------------------------------------------------------------------------*/
 
-/* The 'Yield' sequence for Tape Atoms.
- * Loads the next pointer from the tape, advances the tape, and jumps. 
- * Cost: ~ 4 cycles */
-#define mac_yield()                               \
-	  load_word(R_AtomJmp, R_TapePtr, 0)            \
-	, add_ui_self(         R_TapePtr, S_(MipsCode)) \
-	, jump_reg( R_AtomJmp)                          \
-	, nop
+// The 'Yield' sequence for Tape Atoms (mac_yield). 
+MipsAtomComp_(ac_yield) {
+	load_word(R_AtomJmp, R_TapePtr, 0),
+	add_ui_self(         R_TapePtr, S_(MipsCode)),
+	jump_reg( R_AtomJmp), nop,
+};
 
 /* Words: 3; Loads 3 S2 indices from the face array */
-#define mac_load_tri_indices(rId_0, rId_1, rId_2) \
-	  load_half_u(rId_0, R_FaceCursor, 0 * S_(S2)) \
-	, load_half_u(rId_1, R_FaceCursor, 1 * S_(S2)) \
-	, load_half_u(rId_2, R_FaceCursor, 2 * S_(S2))
+MipsAtomComp_(ac_load_tri_indices) {
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+};
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-#define mac_load_tri_verts(rId_0, rId_1, rId_2) \
-	  shift_lleft(R_AT, rId_0, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0) \
-	, shift_lleft(R_AT, rId_1, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1) \
-	, shift_lleft(R_AT, rId_2, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2)
+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),
+	shift_lleft(R_AT, R_T2, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+};
 
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list */
-#define mac_insert_ot_tag(r_otz, prim_type) \
-	  shift_lleft( R_T1, r_otz, S_(U4)/2)                       /* T1 = r_otz * S_(U4) */ \
-	, add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-	, load_word(   R_AT, R_T1,         O_(PolyTag,bf_addr_len)) /* AT = old_ot_head */ \
-	, load_upper_i(R_V0, (S_(prim_type)/S_(U4) - S_(PolyTag)/S_(U4)) << polytag_len_bits) /* V0 = S_(prim_type without tag field) */ \
-	, mask_upper(  R_AT, R_AT,         S_(polytag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-	, or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-	, store_word(  R_AT, R_PrimCursor, O_(PolyTag,bf_addr_len)) /* prim->tag = packed(prim_length, old_addr) */ \
-	, shift_lleft( R_AT, R_PrimCursor, S_(polytag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-	, shift_lright(R_AT, R_AT,         S_(polytag_len_bits)) \
-	, store_word(  R_AT, R_T1,         O_(PolyTag,bf_addr_len)) /* OrderingTable[OTZ] = PrimCursor */
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
+MipsAtomComp_(ac_insert_ot_tag_f3) {
+	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
+	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
+	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (5 - 1) << 24 = 4 << 24
+	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
+	or_u(        R_AT, R_AT, R_V0),                            // Merge length
+	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
+	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
+	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
+	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
+};
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
+MipsAtomComp_(ac_insert_ot_tag_g4) {
+	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
+	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
+	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
+	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (9 - 1) << 24 = 8 << 24
+	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
+	or_u(        R_AT, R_AT, R_V0),                            // Merge length
+	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
+	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
+	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
+	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
+};
 
 /* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
- * byte offset. Internal helper used by the *_format_*_color macros.
- * Args: off = U4 byte offset, code = GP0 cmd byte (0 for c1/c2/c3 of
- * a Poly_G4), r/g/b = 8-bit RGB byte values. */
-#define mac_pack_color_word(off, code, r,g,b) \
-		load_upper_i(R_AT, (code) << 8  | (b))    \
-	, or_i_self(   R_AT, ((g)   << 8) | (r))    \
-	, store_word(  R_AT, R_PrimCursor, (off))
+ * byte offset. Internal helper used by the *_format_*_color macros. */
+FI_ MipsAtom ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
+MipsAtomComp_Proc_(ac_pack_color_word, {
+	load_upper_i(R_AT, (cmd) << 8  | (b)),
+	or_i_self(   R_AT, ((g)  << 8) | (r)),
+	store_word(  R_AT, R_PrimCursor, (off)),
+})
 
 /* Words: 3; Emits the F3 command+color word (cmd byte | BLUE | GREEN | RED)
- * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields).
- * Migrated from hello_gte_tape.c; takes RGB form per the Phase 3
- * convention. */
-#define mac_format_f3_color(r,g,b) mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r,g,b)
+ * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields). */
+FI_ MipsAtom ac_format_f3_color(U1 r, U1 g, U1 b)
+MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
 
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
- * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-#define mac_gte_store_f3_post_rtpt() \
-		gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
-	, gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
-	, gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
+ * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
+MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
+	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
+	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
+	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
+};
 
 /* Words: 12; Emits the four (code|color) words of a Poly_G4.
  * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-#define mac_format_g4_color(r0,g0,b0, r1,g1,b1, r2,g2,b2, r3,g3,b3)  \
-		mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
-	, mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
-	, mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
-	, mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
+FI_ MipsAtom ac_format_g4_color(
+	U1 r0, U1 g0, U1 b0,
+  U1 r1, U1 g1, U1 b1,
+  U1 r2, U1 g2, U1 b2,
+  U1 r3, U1 g3, U1 b3)
+MipsAtomComp_Proc_(ac_format_g4_color, {
+	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
+	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1),
+	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2),
+	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3),
+})
 
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
  * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen,
- * SXY2=v2.screen). MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
+ * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
+ * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
  * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
  * three registers aligned with v0/v1/v2 you must store before RTPS).
  * The macro name declares the pipeline position; check #6 (GTE state-
  * machine validation) verifies the call site matches the declaration. */
-#define mac_gte_store_g4_p012_post_rtpt_pre_rtps() \
-		gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0))  \
-	, gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1))  \
-	, gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
+MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
+	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
+	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
+	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
+};
 
 /* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
  * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
  * single-vertex result to SXY2; SXY0 still holds v0.screen from the
  * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
  * prevents).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration.
- *
- * History: this macro was named `mac_gte_store_g4_p3` until 2026-07-09
- * when it was discovered to be reading C2_SXY0 (which held v0.screen)
- * instead of C2_SXY2 (which holds v3.screen after RTPS). The rename
- * encodes the pipeline position in the name so the next bug of this
- * class is impossible. */
-#define mac_gte_store_g4_p3_post_rtps() gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
+ */
+MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
 
 #pragma endregion Macro Atom Components
 
@@ -186,7 +220,6 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
 
 typedef Struct_(FMipsAtom512) { U4 data[512]; U4 used; };
 
-typedef Slice_(MipsCode); typedef Slice_MipsCode MipsAtom; 
 // FArena Related
 typedef Relative_(FArena) Struct_(MipsAtomBuilder) { U4 start; U4 capacity; U4 used; };
 // Whatever the builder is writting to should most likely coresspond
@@ -201,9 +234,8 @@ 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_(code_yield), S_(code_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(code_yield));
+	mem_copy(ab->start, u4_(ac_yield), S_(ac_yield));
+	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
 }
 
 #define mipsatom_from_builder(ab) (MipsAtom){ab.start, ab.used}
@@ -219,33 +251,34 @@ enum {
 };
 
 /* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
- *
  * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
  *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
  *   2. $a0 = bios_flushcache (arg0)
  *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
- *   4. jalr $t0, $ra                    ; call BIOS(flushcache)
- *      nop                              ; branch delay slot
- *   5. lw $ra, 4($sp);  jr $ra          ; restore & return
+ *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
+ *      nop                             ; branch delay slot
+ *   5. lw $ra, 4($sp);  jr $ra         ; restore & return
  *   6. sp += 8
  */
 internal MipsAtom_(mips_flush_icache) {
-	  add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment) /* sp -= 8 */
-	, store_word(rret_addr, rstack_ptr, S_(U4))           /* sw  $ra,   4($sp)   */
-	, add_ui(rret_0, rdiscard, bios_flushcache)           /* addiu $a0, $0, 0x44 */
-	, add_ui(rtmp_0, rdiscard, bios_table_addr)           /* addiu $t0, $0, 0xA0 */
-	, jump_link(rtmp_0, rret_addr)                        /* jalr  $t0, $ra      */
-	, nop                                                 /* BD slot             */
-	, load_word(rret_addr, rstack_ptr, S_(U4))            /* lw   $ra, 4($sp)    */
-	, jump_reg(rret_addr)                                 /* jr   $ra            */
-	, add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment)  /* sp += 8 (BD) */
-	, mac_yield()
+	add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment), // sp -= 8
+	store_word(rret_addr, rstack_ptr, S_(U4)),           // sw  $ra,   4($sp) 
+	add_ui(rret_0, rdiscard, bios_flushcache),           // addiu $a0, $0, 0x44 
+	add_ui(rtmp_0, rdiscard, bios_table_addr),           // addiu $t0, $0, 0xA0 
+	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
+	load_word(rret_addr, rstack_ptr, S_(U4)),            // lw   $ra, 4($sp) 
+	jump_reg(rret_addr),                                 // jr   $ra 
+	add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment),  // sp += 8 (BD) 
+	mac_yield(),
 };
 
 typedef Struct_(Binds_SetGteWorld) {
-	U4 transform;
+	M3_S2* transform;
 };
-internal MipsAtom_(set_gte_world) {
+internal MipsAtom_(set_gte_world) atom_info(
+		atom_bind(Binds_SetGteWorld)
+	,	atom_reads(R_TapePtr)
+){
 	/* Pop matrix address from tape into R_T3 ($11) */
 	load_word(R_T3, R_TapePtr, O_(Binds_SetGteWorld,transform)),
 	add_ui_self(    R_TapePtr, S_(Binds_SetGteWorld)),
@@ -259,8 +292,6 @@ internal MipsAtom_(set_gte_world) {
 	mac_yield()
 };
 
-// TODO(Ed): I'm not sure yet if the bindings are redundant with the floortri atom yet.
-
 /* DIAGNOSTIC 1: Pure tape loop test */
 internal MipsAtom_(diag_yield) { mac_yield() };
 
@@ -282,9 +313,9 @@ internal MipsAtom_(diag_color) {
 	add_u_self(      R_T1,       R_T6),         
 	
 	load_word(   R_AT, R_T1, 0),        
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << polytag_len_bits),
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
 	store_word(  R_AT, R_T7, 0),       
-	shift_lleft(R_AT, R_T7, S_(polytag_len_bits)), shift_lright(R_AT, R_AT, S_(polytag_len_bits)),         
+	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
 	or_u_self(  R_AT, R_V0),          
 	store_word( R_AT, R_T1, 0),       
 

code/duffle/memory.h

@@ -73,6 +73,7 @@ typedef Slice_(B1);
 
 #define slice_iter(container, iter)     (T_((container).ptr) iter = (container).ptr; iter != slice_end(container); ++ iter)
 #define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = array_decl(type,__VA_ARGS__), .len = array_len( array_decl(type,__VA_ARGS__)) }
+#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = S_(array) }
 
 FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(s.ptr, s.len); }
 #define  slice_zero(s)        slice_zero_(slice_to_ut(s))

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). Combine with
- *              _u or _s (single-letter modifier + type combined): add_ui,
- *              add_si. Examples: add_ui, add_si, and_i, or_i, xor_i,
- *              load_upper_i. and_i is sign-agnostic (andi zero-extends).
- *              load_upper_i is a unique verb; _i is the immediate marker,
- *              not a modifier+type combination.
- *
- *    _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.
+ *    _i: Immediate value (16-bit constant operand). 
+ *        Combine with _u or _s (single-letter modifier + type combined): add_ui, add_si.
+ *        Examples: add_ui, add_si, and_i, or_i, xor_i, load_upper_i. and_i is sign-agnostic (andi zero-extends).
+ *        load_upper_i is a unique verb; _i is the immediate marker, not a modifier+type combination.
+ *    _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 single letter `l` (logical) or `a` (arithmetic). Direction
- *    is the word `left` or `right`. Combined: `_lleft`, `_lright`,
- *    `_aright`. Examples: shift_lleft( rd, rt, shamt) (= sll)
- *                         shift_lright(rd, rt, shamt) (= srl)
- *                         shift_aright(rd, rt, shamt) (= sra)
- *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical
- *    to logical-left, so use shift_lleft for that case)
+ *    The shift macros use `shift_<modifier><direction>`.
+ *    Modifier is the single letter `l` (logical) or `a` (arithmetic).
+ *    Direction is the word `left` or `right`. Combined: `_lleft`, `_lright`, `_aright`.
+ *    Examples: shift_lleft( rd, rt, shamt) (= sll)
+ *              shift_lright(rd, rt, shamt) (= srl)
+ *              shift_aright(rd, rt, shamt) (= sra)
+ *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical to logical-left, so use shift_lleft for that case)
  *
  *  --- Jump/Call family ---
- *    Simple jumps keep the original short names: jump (j), jump_reg
- *    (jr), jump_link (jalr rs, rd). The jump-and-link-to variants
- *    (jal, jalr rs with default $ra) get the `call_` verb instead:
+ *    Simple jumps keep the original short names: jump (j), jump_reg (jr), jump_link (jalr rs, rd).
+ *    The jump-and-link-to variants (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
- *              disambiguation (e.g., GTE data register vs control
- *              register). NOT used in plain R-type arithmetic (the
- *              R-type is implicit). Examples: gte_mv_to_data_r,
- *              gte_mv_to_ctrl_r.
+ *    _r: Register marker — used only when the register type needs disambiguation (e.g., GTE data register vs control register).
+ *        NOT used in plain R-type arithmetic (the R-type is implicit). Examples: gte_mv_to_data_r, gte_mv_to_ctrl_r.
+ *    _self:     Destination equals one source operand.
+ *               Examples: add_ui_self (I-type, to self), add_u_self (R-type, to self).
+ *    _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.
- *              Examples: add_ui_self (I-type, to self),
- *                        add_u_self (R-type, to self).
- *
- *    _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)
+ *    _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
- *  header. They live in the opt-in `mips_vendor_sym.h` for users who
- *  prefer the textbook MIPS assembly mnemonics.
+ *  Vendor mnemonics (sll, srl, sra, jr, j, jal, jalr) are NOT in this header.
+ *  They live in the opt-in `mips_vendor_sym.h` for users who 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
- *  preprocessor can stringify the integer (e.g. for asm clobber lists and
- *  register-variable declarations via `rgcc(R_X)`). The enum value is
- *  bound to the `#define` so the two forms cannot drift apart.
+ *  Every R_* enum below has a parallel R_*_Code `#define` so that the preprocessor can stringify the integer
+ *  (e.g. for asm clobber lists and register-variable declarations via `rgcc(R_X)`).
+ *  The enum value is bound to the `#define` so the two forms cannot drift apart.
  *
- *  Only registers that get stringified need a `_Code` form; the rest are
- *  plain enum values. If you need to add a new one, follow the pattern:
+ *  Only registers that get stringified need a `_Code` form; the rest are plain enum values.
+ *  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
- *  sites and let `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify
- *  cases — never write the bare number `12`.
+ *  User code should always reference the enum form (`R_T4`) at arithmetic sites and let
+ *  `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify 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)
- *   store_word(rt, base, off)  → sw   rt, off(base)
- *   add_ui(rt, rs, imm)        → addiu rt, rs, imm
- *   shift_lleft(rd, rt, shamt)   → sll   rd, rt, shamt
- *   shift_lright(rd, rt, shamt)  → srl   rd, rt, shamt
- *   shift_aright(rd, rt, shamt)  → sra   rd, rt, shamt
- *   jump_reg(rs)               → jr    rs
- *   jump_link(rs, rd)          → jalr  rs        (link in rd, default $ra)
- *   nop                        → sll   $0, $0, 0
+ *   load_word(rt, base, off)    → lw   rt, off(base)
+ *   store_word(rt, base, off)   → sw   rt, off(base)
+ *   add_ui(rt, rs, imm)         → addiu rt, rs, imm
+ *   shift_lleft(rd, rt, shamt)  → sll   rd, rt, shamt
+ *   shift_lright(rd, rt, shamt) → srl   rd, rt, shamt
+ *   shift_aright(rd, rt, shamt) → sra   rd, rt, shamt
+ *   jump_reg(rs)                → jr    rs
+ *   jump_link(rs, rd)           → jalr  rs        (link in rd, default $ra)
+ *   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
- * _Generic-based signed integer-arithmetic helpers for U1/U2/U4. Those
- * live in a different conceptual layer (generic arithmetic on DSL
- * types) and would collide with the instruction encoders here. The
- * `#undef` below lets the gas-style names below win; if a file needs
- * both, the dsl.h versions can be reached via their long forms
+ * NOTE: dsl.h defines `add_s`/`sub_s`/`mut_s`/`gt_s`/etc. as _Generic-based signed integer-arithmetic helpers for U1/U2/U4.
+ * Those live in a different conceptual layer (generic arithmetic on DSL types) and would collide with the instruction encoders here.
+ * The `#undef` below lets the gas-style names below win; if a file needs both, the dsl.h versions can be reached via their long forms
  * (e.g. `def_signed_op`-style or the underlying `add_s1/s2/s4`). */
 #undef add_s
 #undef sub_s
@@ -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
- * and always emits 2 .words. Use this when:
+ * Granular companion to `load_imm`: skips the compile-time range checks and always emits 2 .words. Use this when:
  *   - you know `imm` is > 0xFFFF (otherwise you're wasting a word), OR
  *   - `imm` is not a compile-time constant and you want predictable
  *     2-word emission without the `__builtin_constant_p` branches.
  *
  * The lo16 strategy is still chosen at expansion time on the lo half:
- *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui
- *                              already cleared bits 15..0)
- *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the
- *                              intended bit pattern)
+ *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui already cleared bits 15..0)
+ *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the intended bit pattern)
  *
- * For situations where you need to bypass even this choice (e.g. to
- * force a specific encoding for a known discontiguous high/low pair),
+ * For situations where you need to bypass even this choice 
+ * (e.g. to force a specific encoding for a known discontiguous high/low pair),
  * see `load_imm_2w_ori_forced` and `load_imm_2w_addi_forced` below.
- *
  * Statement-level (not expression-level): emits its own `asm volatile(...)`.
  */
 #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
- * signed downstream) and you want a smaller effective instruction
- * (the assembler/MIPS hardware will sign-extend the imm16). */
+ * Use when you know sign-extension is fine (e.g. lo16 is treated as signed downstream) 
+ * and you want a smaller effective instruction (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
- * 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  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:
+ * Dispatches at compile time on the immediate's range, picking the 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 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:
  *      load_imm(R_T4, 0x12345678);          // emits 2 .words
  *
- * Falls back to a 2-word form if `imm` is not a compile-time constant,
- * but that path is unusual (load_imm is most useful with literal
- * addresses and magic numbers). */
+ * Falls back to a 2-word form if `imm` is not a compile-time constant, but that path is unusual
+ * (load_imm is most useful with literal 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
- * "memory" barrier. The register ids are passed through `rlit` so
- * the R_*_Code `#define`s are stringified into "$N" at expansion time. */
+ * GPRs that the kernel treats as volatile (v0/v1/t0/t1/ra) plus the "memory" barrier.
+ * The register ids are passed through `rlit` so the R_*_Code `#define`s are stringified into "$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
- *  canonical duffle macros in mips.h. The duffle names are primary; this
- *  header is for users who prefer the textbook mnemonics.
+ *  Provides the textbook MIPS assembly mnemonics as thin aliases to the canonical duffle macros in mips.h.
+ *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/mips_vendor_sym.h"   // after mips.h
  *
@@ -21,12 +20,6 @@
  *      jal  -> call_addr         (jump-and-link to immediate address)
  *      jalr -> call_reg          (jump-and-link to register, default $ra)
  *      (for the 2-arg `jalr rs, rd`, use `jump_link(rs, rd)` directly)
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (tape_atom.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries. Verification: V2 (objdump
- *  byte-identical) holds.
- *
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/word_count.metadata.h

@@ -1,10 +1,15 @@
-// tape_atom.metadata.h
+// word_count.metadata.h
 // Single source of truth for instruction-word counts.
 // Used by C (to define compile-time constants) AND Python (to count positions).
 //
 // Format: WORD_COUNT(MACRO_NAME, COUNT)
 // One line per macro that appears in your atom sources.
 //
+// This file is encoding-macros-only. 
+// The auto-generated component macros (mac_X) live in duffle/gen/<dir>.macs.h (included separately by the unity build).
+// The unity build should include THIS file and the .macs.h file in the same TU, with both wrapped 
+// (or the include guard order handled) to avoid WORD_COUNT redeclaration.
+//
 // To regenerate: hand-count the instructions in each macro definition.
 // (You'll only need to do this once per macro — they don't change often.)
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
@@ -47,15 +52,5 @@ WORD_COUNT(gte_cmdw_nclip,                   1)
 WORD_COUNT(gte_avg_sort_z3,                  1)
 WORD_COUNT(sub_u,                            1)
 WORD_COUNT(nop2,                             2)
-WORD_COUNT(mac_load_tri_indices,             3)
-WORD_COUNT(mac_load_tri_verts,              18)
-WORD_COUNT(mac_format_f3_color,              3)
-WORD_COUNT(mac_gte_store_f3_post_rtpt,       3)
-WORD_COUNT(mac_pack_color_word,              3)
-WORD_COUNT(mac_format_g4_color,             12)
-WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
-WORD_COUNT(mac_gte_store_g4_p3_post_rtps,    1)
-WORD_COUNT(mac_insert_ot_tag,               11)
-WORD_COUNT(mac_yield,                        4)
 
 #undef WORD_COUNT

code/graphics_hello_psyq/hello_gpu.c

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

code/graphics_hello_psyq/hello_gpu.h

@@ -5,8 +5,8 @@
 #	include "duffle/gp.h"
 #endif
 
-typedef def_struct(DrawEnv_Packed) { U4 tag; U4 code[15]; };
-typedef def_struct(DrawEnv) {
+typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };
+typedef Struct_(DrawEnv) {
 	Rect_S2 clip_area;
 	A2_S2   drawing_offset;
 	Rect_S2 texture_window;
@@ -17,7 +17,7 @@ typedef def_struct(DrawEnv) {
 	RGB8    initial_bg_color;
 	DrawEnv_Packed dr_env; // reserved
 };
-typedef def_struct(DisplayEnv) {
+typedef Struct_(DisplayEnv) {
 	Rect_S2 display_area;
 	Rect_S2 screen;
 	B1      vinterlace;
@@ -25,9 +25,9 @@ typedef def_struct(DisplayEnv) {
 	B1      pad0;
 	B1      pad1;
 };
-typedef def_farray(DrawEnv,    2);
-typedef def_farray(DisplayEnv, 2);
-typedef def_struct(DoubleBuffer) {
+typedef Array_(DrawEnv,    2);
+typedef Array_(DisplayEnv, 2);
+typedef Struct_(DoubleBuffer) {
 	A2_DrawEnv    draw;
 	A2_DisplayEnv display;
 };
@@ -58,7 +58,7 @@ U4 vsync(U4 mode)     __asm__("VSync");
 
 void draw_orderingtbl(U4* buf) __asm__("DrawOTag");
 
-typedef def_struct(PolyTag) {
+typedef Struct_(PolyTag) {
 	U4   addr: 24;
 	U4   len:  8;
 	RGB8 color;
@@ -106,7 +106,7 @@ typedef def_struct(PolyTag) {
 // #define setLineF4(p)	set_len(p, 6),  set_code(p, 0x4c),(p)->pad = 0x55555555
 // #define setLineG4(p)	set_len(p, 9),  set_code(p, 0x5c),(p)->pad = 0x55555555, (p)->p2 = 0, (p)->p3 = 0
 
-typedef def_struct(Poly_F3) {
+typedef Struct_(Poly_F3) {
 	U4   tag;
 	RGB8 color;
 	B1   code;
@@ -120,14 +120,14 @@ typedef def_struct(Poly_F3) {
 	};
 };
 
-typedef def_struct(Poly_G3) {
+typedef Struct_(Poly_G3) {
 	U4    tag; RGB8 c0; B1 code;
 	V2_S2 p0;  RGB8 c1; B1 pad1;
 	V2_S2 p1;  RGB8 c2; B1 pad2;
 	V2_S2 p2;
 };
 
-typedef def_struct(Poly_F4) {
+typedef Struct_(Poly_F4) {
 	U4   tag;
 	RGB8 color;
 	B1   code;
@@ -142,7 +142,7 @@ typedef def_struct(Poly_F4) {
 	};
 };
 
-typedef def_struct(Poly_G4) {
+typedef Struct_(Poly_G4) {
 	U4    tag; RGB8 c0; B1 code;
 	V2_S2 p0;  RGB8 c1; B1 pad1;
 	V2_S2 p1;  RGB8 c2; B1 pad2;
@@ -150,7 +150,7 @@ typedef def_struct(Poly_G4) {
 	V2_S2 p3;
 };
 
-typedef def_struct(Tile) {
+typedef Struct_(Tile) {
 	U4      tag;
 	RGB8    color;
 	B1      code;
@@ -169,7 +169,7 @@ M3_S2* m3s2_scale      (M3_S2* mat, V3_S4* vec) __asm__("ScaleMatrix");
 // Rotation, Translation, Perspective
 
 S4 rtp_v3s2_raw(V3_S2* vec, S4* xy, S4* pp, S4* flag) __asm__("RotTransPers");
-FI_ S4 rtp_v3s2(V3_S2* vec, V2_S2* xy, A2_S2* pp, S4* flag) { return rtp_v3s2_raw(vec, cast(S4*R_, & xy->x), cast(S4*R_, pp), r_(flag)); }
+FI_ S4 rtp_v3s2(V3_S2* vec, V2_S2* xy, A2_S2* pp, S4* flag) { return rtp_v3s2_raw(vec, C_(S4*R_, & xy->x), C_(S4*R_, pp), r_(flag)); }
 
 S4 rtp_avg_nclip_a3_v3s2_raw(V3_S2* v0, V3_S2* v1, V3_S2* v2, S4* xy1, S4* xy2, S4* xy3, S4* pp, S4* otz, S4* flag) __asm__("RotAverageNclip3");
 FI_  S4 rtp_avg_nclip_a3_v3s2(
@@ -179,8 +179,8 @@ FI_  S4 rtp_avg_nclip_a3_v3s2(
 ){
 	return rtp_avg_nclip_a3_v3s2_raw(
 		v0, v1, v2, 
-		cast(S4*R_, xy0), cast(S4*R_, xy1), cast(S4*R_, xy2),
-		cast(S4*R_, pp),  cast(S4*R_, otz), cast(S4*R_, flag)
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
 	);
 }
 
@@ -192,8 +192,8 @@ FI_ S4 rtp_avg_nclip_a4_v3s2(
 ){
 	return rtp_avg_nclip_a4_v3s2_raw(
 		v0, v1, v2, v3,
-		cast(S4*R_, xy0), cast(S4*R_, xy1), cast(S4*R_, xy2), cast(S4*R_, xy3),
-		cast(S4*R_, pp),  cast(S4*R_, otz), cast(S4*R_, flag)
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2), C_(S4*R_, xy3),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
 	);
 }
 

code/gte_hello/gen/gte_hello.offsets.h

@@ -1,4 +1,4 @@
-// Auto-generated by tape_atom_offset_gen.meta.lua — DO NOT EDIT
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
 // Source: C:\projects\Pikuma\ps1\code\gte_hello\hello_gte_tape.c
 #pragma once
 

code/gte_hello/hello_gte.c

@@ -8,18 +8,21 @@
 #include "duffle/dsl.h"
 #include "duffle/memory.h"
 #include "duffle/math.h"
+
 #include "duffle/gcc_asm.h"
 #include "duffle/mips.h"
 #include "duffle/gp.h"
 #include "duffle/gte.h"
 
-#	include "duffle/gen/lottes_tape.offsets.h"
+#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/duffle.offsets.h"
 #include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
+#include "duffle/word_count.metadata.h"
 
-#	include "tape_atom.metadata.h"
-#	include "gen/hello_gte_tape.offsets.h"
+#	include "gen/gte_hello.offsets.h"
 #include "hello_gte.h"
+
 #include "hello_gte_tape.c"
 
 typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
@@ -256,9 +259,9 @@ 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));
-		TapeBuilder tb = tb_make_old(&tape_arena); tb_scope(& tb) {
-			tb_emit(& tb, code_rbind_cube_g4_face);
+		LP_ U4 mem_temp_tape[512];
+		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
 				tb_data(& tb, prim_cursor);
 				tb_data(& tb, u4_(smem.cube.faces));
 				tb_data(& tb, u4_(smem.cube.verts));
@@ -266,10 +269,10 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 
 			for (U4 i = 0; i < Cube_num_faces; i++) {
 				// Two triangles per quad face: (x,y,z) and (x,z,w)
-				tb_emit(& tb, code_cube_g4_face);
+				tb_emit(& tb, cube_g4_face);
 			}
 
-			tb_emit(& tb, code_sync_primitive_arena);
+			tb_emit(& tb, sync_primitive_arena);
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}
@@ -337,9 +340,6 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
 		m3s2_translation(& smem.tform_world, & smem.floor.pos);
 		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
-		// TODO(Ed): This can either be in the tape or here...
-		// gte_matrix_set_rotation   (& smem.tform_world);
-		// gte_matrix_set_translation(& smem.tform_world);
 
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
@@ -350,21 +350,20 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		// Prepare the tape. (Push protocol to tape)
 		LP_ U4 mem_temp_tape[512];
 		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
-			// TODO(Ed): This is bugged.
-			tb_emit(& tb, code_set_gte_world);
+			tb_emit(& tb, 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.
-			tb_emit(& tb, code_sync_primitive_arena);
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}

code/gte_hello/hello_gte_tape.c

@@ -1,109 +1,82 @@
 #ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/lottes_tape.h"
+#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/duffle.offsets.h"
 #	include "duffle/atom_dsl.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/word_count.metadata.h"
+#	include "gen/gte_hello.offsets.h"
 #	include "hello_gte.h"
-#	include "tape_atom.metadata.h"
-#	include "gen/hello_gte_tape.offsets.h"
 #endif
 
 #pragma region MACs (Mips Atom components)
-/* The macros mac_format_f3_color and mac_gte_store_f3 moved to
- * lottes_tape.h during the Phase 3 gp.h overhaul. Both are now RGB-form
- * (mac_format_f3_color takes _r, _g, _b byte values rather than raw
- * 16-bit half-words). */
+
+
 
 #pragma endregion MACs
 
 #pragma region Baked Atoms
 
 typedef Struct_(Binds_CubeTri) {
-	U4 PrimCursor;
-	U4 FaceCursor;
-	U4 VertBase;
-	U4 OtBase;
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
 };
-internal MipsAtom_(rbind_cube_g4_face) {
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
 	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)), 
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)), 
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)), 
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)), 
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
 	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
-	// Note(Ed): This entire thing is argument shuffle?
-	// TODO(Ed): Eliminate
 	mac_yield()
 };
 
-/* ============================================================================
- *  cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
- * ============================================================================
- *
- *  Reads 4 indices from R_FaceCur (V4_S2 = 8 bytes), loads 4 vertices into
- *  the GTE, runs the PsyQ RotAverageNclip4 sequence, and renders a Poly_G4.
- */
-	atom_region  (cube_g4_face, REGION_PRIM_ARENA)
-	atom_group   (cube_g4_face, GROUP_RENDER_PRIMS)
-	atom_cadence (cube_g4_face, CADENCE_FRAME)
-	atom_annot(cube_g4_face, phase_work,
-			atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-			atom_writes(R_PrimCursor, R_FaceCursor))
-internal 
-MipsAtom_(cube_g4_face) {
-	/* ── 1. Load 4 face indices from R_FaceCur (V4_S2 = 8 bytes) ───────── */
+ // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
 	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
 	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
 	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
 	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
 
-	/* ── 2. Load V0, V1, V2 into GTE (parallel to mac_load_tri_verts) ── */
-	mac_load_tri_verts(R_T0, R_T1, R_T2),
-
-	/* ── 3. RTPT — transforms V0/V1/V2 → SXY0/SXY1/SXY2 + SZ1/SZ2/SZ3 ─── */
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple,
-
-	/* ── 4. NCLIP — backface culling on SXY0/SXY1/SXY2 (p0,p1,p2) ──────── */
-	/*     MUST be done BEFORE V3-RTPS overwrites SXY0 with p3.             */
 	nop2, gte_cmdw_nclip,
 
-	/* ── 5. Cull check: skip format/insert if MAC0 ≤ 0 (backface) ───────── */
 	nop2,  gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,                                                          /* COP2 stall */
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
-	nop,                                                          /* BD slot    */
+	nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
 
-	/* ── 6. Format c0..c3 (color+code words) BEFORE V3-RTPS ─────────────── */
 	store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
 	mac_format_g4_color(
 		/* c0 magenta */ 0xFF, 0x00, 0xFF,
 		/* c1 yellow  */ 0xFF, 0xFF, 0x00,
 		/* c2 cyan    */ 0x00, 0xFF, 0xFF,
 		/* c3 green   */ 0x00, 0xFF, 0x00),
-
-	/* ── 7. Store p0..p2 BEFORE V3-RTPS overwrites SXY0 ─────────────────── */
 	mac_gte_store_g4_p012_post_rtpt_pre_rtps(),
 
-	/* ── 8. Load V3 = verts[face->w] into V0 ─────────────────────────────── */
 	shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
 	load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
 	gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
 
-	/* ── 9. RTPS — transforms V0 (now V3) → SXY0 (p3) + SZ3 ─────────────── */
 	nop2, gte_cmdw_rotate_translate_perspective_single,
 	mac_gte_store_g4_p3_post_rtps(),
 
-	/* ── 10. AVSZ4 — average Z from SZ0/SZ1/SZ2/SZ3 ─────────────────────── */
 	nop2, gte_cmdw_avg_sort_z4,
 	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
 
-	/* ── 11. Bounds check OTZ < OrderingTbl_Len ─────────────────────────── */
 	add_ui(      R_AT, R_0,  OrderingTbl_Len),
 	set_lt_u(    R_AT, R_T1, R_AT),
 	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+	mac_insert_ot_tag_g4(),
 
-	/* ── 12. Insert into Ordering Table (length = 8 words for Poly_G4) ──── */
-	mac_insert_ot_tag(R_T1, Poly_G4),
-
-	/* ── 13. Advance cursors & yield (both branch targets land here) ────── */
 atom_label(cube_g4_face_exit)
 	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
 	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
@@ -111,78 +84,69 @@ atom_label(cube_g4_face_exit)
 };
 
 typedef Struct_(Binds_FloorTri) {
-	U4 PrimCursor;
-	U4 FaceCursor;
-	U4 VertBase;
-	U4 OtBase;
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
 };
-	atom_region(rbind_floor_f3_face, REGION_PRIM_ARENA)
-	atom_group(rbind_floor_f3_face, GROUP_RENDER_FLOOR)
-	atom_cadence(rbind_floor_f3_face, CADENCE_FRAME)
-	atom_annot(rbind_floor_f3_face, phase_bind
-		, atom_reads()
-		, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
-internal 
-MipsAtom_(rbind_floor_f3_face) {
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
 	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)), 
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)), 
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)), 
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)), 
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
 	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
 	mac_yield()
 };
 
-	atom_region( floor_f3_face, REGION_PRIM_ARENA)
-	atom_group(  floor_f3_face, GROUP_RENDER_FLOOR)
-	atom_cadence(floor_f3_face, CADENCE_FRAME)
-	atom_annot(  floor_f3_face, phase_work,
-		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-		atom_writes(R_PrimCursor, R_FaceCursor)) 
-internal 
-MipsAtom_(floor_f3_face) {
-	mac_load_tri_indices(R_T0, R_T1, R_T2),
-	mac_load_tri_verts(  R_T0, R_T1, R_T2),
+internal
+atom_dbg_skip_over()
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(  R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple,
 	nop2, gte_cmdw_nclip,
- 
+
 	/* Culling (Branch forward if Backface) */
 	nop2, gte_mv_from_data_r(R_T0, C2_MAC0),
 	nop,
 	branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop,
 	/* Format Primitive */
-	// mac_format_f3_color(0x20FF, 0xFFFF),  // works
 	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
 	mac_gte_store_f3_post_rtpt(),
-	
+
 	/* Calculate Depth */
-	nop2, gte_avg_sort_z3,            
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),      
+	nop2, gte_avg_sort_z3,
+	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
 	/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
-	add_ui(      R_AT, R_0,  OrderingTbl_Len),   
-	set_lt_u(    R_AT, R_T1, R_AT), 
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop, 
+	add_ui(      R_AT, R_0,  OrderingTbl_Len),
+	set_lt_u(    R_AT, R_T1, R_AT),
+	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
 	/* Insert into Ordering Table Linked List */
-	mac_insert_ot_tag(R_T1, Poly_F3),
+	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()
 };
 
 typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
-	atom_region( sync_primitive_arena, REGION_PRIM_ARENA)
-	atom_group(  sync_primitive_arena, GROUP_RENDER_FLOOR)
-	atom_cadence(sync_primitive_arena, CADENCE_FRAME)
-	atom_annot(  sync_primitive_arena, phase_work,
-		atom_reads( R_TapePtr, R_PrimCursor),
-		atom_writes(R_TapePtr)) 
-internal MipsAtom_(sync_primitive_arena) {
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
 	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
-	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),      
+	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

@@ -6,19 +6,34 @@ A rest from the usual.
 
 ## Dependencies
 
-I will be programming from a Windows 11 machine:
+I will be programming from a Windows 11 machine (may eventually try this on the Steam Deck...):
 ![system_info](./docs/assets/system_info.png)
 
-```ps1
-# not really used yet for scripts (may never)
-scoop install lua
-```
-
 [armips](https://github.com/Kingcom/armips)
 
 * Supports doing bare-metal assembly for the ps1
 * `scoop install armips` or just clone and build..
 
+[luajit-2.1](https://github.com/LuaJIT/LuaJIT.git)
+
+```
+scoop install luajit
+```
+
+* Used for lua scripts
+* Particularly, ps1_meta.lua which is a staged metaprogram pass for the custom C11 Assembly DSL used in this codebase.
+
+[lpeg](https://github.com/roberto-ieru/LPeg.git)
+
+* Lua is slow (even jitted) so this helps.
+
+[lfs (LuaFileSystem)](https://github.com/lunarmodules/luafilesystem)
+
+* Native directory enumeration + `mkdir` for the build scripts.
+* Used by `passes/word_count_eval.lua :: scan_dir` (native walk vs. `dir /b /s` subprocess,
+  ~2ms vs. ~56ms) and by `duffle.lua :: ensure_dir` + `to_absolute_path` (avoids
+  `cmd.exe mkdir` + `cd` shell spawns, ~50ms each).
+
 [pscx-redux](https://github.com/grumpycoders/pcsx-redux/): A collection of tools, research, hardware design, and libraries aiming at development and reverse engineering on the PlayStation 1.
 
 * Used as the runtime sandbox emulated the ps1
@@ -57,3 +72,4 @@ scoop install lua
 ![polys!](./docs/assets/pcsx-redux.main_2025-08-03_20-45-35.png)  
 ![hello_psyq!](./docs/assets/pcsx-redux_2025-08-05_23-01-19.png)
 ![cube!](./docs/assets/pcsx-redux_2025-10-11_03-04-01.png)
+![cube and floor!](./docs/assets/pcsx-redux_2026-07-10_22-47-02.png)

scratch/gte_debug.lua

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

scripts/audit_lua_nesting.lua

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

scripts/build_psyq.ps1

@@ -287,7 +287,7 @@ function build-graphis_hello {
 
 	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
 	$module_asm_crt = join-path $path_build         'crt0.o'
-	# assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
 
 	$src_asm    = join-path $path_module 'hello_gpu.s'
 	$module_asm = join-path $path_build  'hello_gpu.o'
@@ -317,100 +317,22 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function generate-TapeAtomOffsets {param([Parameter(Mandatory=$true)] [string[]]$sources, [Parameter(Mandatory=$true)] [string]$metadata)
-	$gen_atom_offsets_script = join-path $path_scripts 'tape_atom.offset_gen.meta.lua'
-
-    $any_stale = $false
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $out      = join-path $gen_dir "$basename.offsets.h"
-
-        if (-not (test-path $out)) { $any_stale = $true; break }
-        $src_mtime  = (get-item $src).LastWriteTimeUtc
-        $out_mtime  = (get-item $out).LastWriteTimeUtc
-        $meta_mtime = (get-item $metadata).LastWriteTimeUtc
-        if (($src_mtime -gt $out_mtime) -or ($meta_mtime -gt $out_mtime)) {
-            $any_stale = $true
-            break
-        }
-    }
-
-    if (-not $any_stale) {
-        write-host "AtomOffsets  all $($sources.Count) source(s) up-to-date" -ForegroundColor DarkGray
-        return
-    }
-
-    write-host "AtomOffsets  $($sources.Count) source(s)" -ForegroundColor Magenta
-    & luajit $gen_atom_offsets_script $metadata @sources
-    if ($LASTEXITCODE -ne 0) {
-        write-error "Atom offset generation failed. Aborting."
-        exit 1
-    }
-}
-
-function generate-TapeAtomAnnotations {param([Parameter(Mandatory=$true)] [string[]]$sources, [Parameter(Mandatory=$true)] [string]$metadata)
-	# Sibling to generate-TapeAtomOffsets. Validates TAPE_ATOM_* / TAPE_WORDS
-	# annotations against the metadata manifest. Emits gen/<basename>.errors.h
-	# containing #error directives for the C build to fail on annotation drift.
-	$gen_atom_annot_script = join-path $path_scripts 'tape_atom_annotation_pass.lua'
-
-    $any_stale = $false
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $out_txt  = join-path $gen_dir "$basename.annotations.txt"
-        $out_err  = join-path $gen_dir "$basename.errors.h"
-
-        if (-not (test-path $out_txt) -or -not (test-path $out_err)) { $any_stale = $true; break }
-        $src_mtime     = (get-item $src).LastWriteTimeUtc
-        $out_txt_mtime = (get-item $out_txt).LastWriteTimeUtc
-        $out_err_mtime = (get-item $out_err).LastWriteTimeUtc
-        $out_mtime     = if ($out_txt_mtime -gt $out_err_mtime) { $out_txt_mtime } else { $out_err_mtime }
-        $meta_mtime    = (get-item $metadata).LastWriteTimeUtc
-        if (($src_mtime -gt $out_mtime) -or ($meta_mtime -gt $out_mtime)) {
-            $any_stale = $true
-            break
-        }
-    }
-
-    if (-not $any_stale) {
-        write-host "AtomAnnotations all $($sources.Count) source(s) up-to-date" -ForegroundColor DarkGray
-        return
-    }
-
-    write-host "AtomAnnotations $($sources.Count) source(s)" -ForegroundColor Magenta
-    & luajit $gen_atom_annot_script $metadata @sources
-    if ($LASTEXITCODE -ne 0) {
-        write-error "Atom annotation generation failed. Aborting."
-        exit 1
-    }
-
-    # If any source produced annotation errors, surface them now and halt the
-    # build. The errors.h files are also #include'd via -include below, so
-    # the C build would fail at preprocessing time anyway — failing here gives
-    # a more readable error in the build log.
-    $err_count = 0
-    foreach ($src in $sources) {
-        $basename = [System.IO.Path]::GetFileNameWithoutExtension($src)
-        $dir      = split-path -Path $src -Parent
-        $gen_dir  = join-path $dir 'gen'
-        $ann_txt  = join-path $gen_dir "$basename.annotations.txt"
-        $err_h    = join-path $gen_dir "$basename.errors.h"
-        if ((test-path $ann_txt) -and (test-path $err_h)) {
-            $txt = get-content $ann_txt -raw
-            if ($txt -match 'Errors:\s+([1-9]\d*)') {
-                $err_count += [int]$Matches[1]
-                write-warning "Annotation errors in $src — see $ann_txt"
-            }
-        }
-    }
-    if ($err_count -gt 0) {
-        write-error "Annotation pass failed: $err_count error(s) across $($sources.Count) source(s). Aborting."
-        exit 1
-    }
+function ps1-meta { param(
+		[Parameter(Mandatory=$true)][string[]]$sources,
+		[Parameter(Mandatory=$true)][string]$metadata,
+		[string]$out_root = (join-path $path_build 'gen'),
+		[string[]]$passes = @('--pre-link'),
+		[string[]]$extra_args = @()
+	)
+	$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', $out_root) + @($extra_args)
+	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 {
@@ -418,13 +340,11 @@ function build-gte_hello {
 
 	$path_module        = join-path $path_code   'gte_hello'
 	$path_duffle        = join-path $path_code   'duffle'
-	$path_atom_metadata = join-path $path_module 'tape_atom.metadata.h'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
 
 	$source_dirs  = @($path_duffle, $path_module)
 	$atom_sources = Get-SourceFiles -paths $source_dirs -extensions @('.h', '.c')
-
-	generate-TapeAtomAnnotations -sources $atom_sources -metadata $path_atom_metadata
-	generate-TapeAtomOffsets     -sources $atom_sources -metadata $path_atom_metadata
+	ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -433,7 +353,7 @@ function build-gte_hello {
 
 	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
 	$module_asm_crt = join-path $path_build         'crt0.o'
-	# assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
 
 	# $src_asm    = join-path $path_module 'hello_gte.s'
 	# $module_asm = join-path $path_build  'hello_gte.o'
@@ -458,25 +378,94 @@ function build-gte_hello {
 	$link_args = @()
 	$link_args += $f_debug
 	# $link_args += $f_optimize_size
-	link-modules @($module_asm_crt, $module_c) $elf $link_args
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
 	make-binary $elf $exe
+
+	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
+	ps1-meta -sources $atom_sources -metadata $path_atom_metadata `
+		-out_root (join-path $path_build 'gen') `
+		-passes @('--post-link') `
+		-extra_args @('--elf', $elf)
+
+	# F' + G' splice: collapse 9 objcopy subprocess invocations into 3.
+	# - 1 call: 3x --update-section for F' (line / aranges / rnglists)
+	# - 1 call: 3x --update-section for G' (info / abbrev / str)
+	# - 1 call: 2x --add-section for G' (loc / loclists — these don't exist in the source ELF)
+	# - 1 call: 1x --set-section-flags (.rodata / .data enable code flag)
+	# = 4 objcopy calls (was 9; saved 5 spawns).
+	$dwarfLineBin     = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_line.bin'
+	$dwarfArangesBin  = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_aranges.bin'
+	$dwarfRnglistsBin = join-path (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
+
+		# Single objcopy call: 3x --update-section for F' (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;
+		}
+
+		# G' 5-section splice: 3 update-section (info / abbrev / str) + 2 add-section (loc / loclists).
+		$dwarfInfoBin     = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_info.bin'
+		$dwarfAbbrevBin   = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_abbrev.bin'
+		$dwarfStrBin      = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_str.bin'
+		$dwarfLocBin      = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_loc.bin'
+		$dwarfLoclistsBin = join-path (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 shipping 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
 
 
 # NO idea if this works yet...
-function Send-ToEmulator { param(
-    [string]$exePath
-)
+function Send-ToEmulator { param( [string]$exePath )
     $uri = "http://localhost:8080/api/v1/load-exec"
 
     # Absolute path is safest for the emulator web server
     $absolutePath = [System.IO.Path]::GetFullPath($exePath)
 
     # Create JSON payload pointing to your compiled .ps-exe
-    $body = @{
-        filename = $absolutePath
-    } | ConvertTo-Json
+    $body = @{ filename = $absolutePath } | ConvertTo-Json
 
     Write-Host "Pushing hot-reload to PCSX-Redux..." -ForegroundColor Magenta
     try {

scripts/duffle_paths.lua

@@ -0,0 +1,91 @@
+--- duffle_paths.lua — Single-line bootstrap helper for the tape-atom Lua scripts.
+---
+--- Each entry script (ps1_meta.lua + the 7 passes/*.lua files) starts with one of:
+---   ```lua
+---   -- Entry script (ps1_meta.lua — `arg[0]` is set):
+---   local duffle = dofile((arg[0]:match("(.*[/\\])") or "./") .. "duffle_paths.lua")
+---
+---   -- Pass module (debug.getinfo path resolution; works both standalone and when require'd):
+---   local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+---   local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+---   ```
+---
+--- That small bootstrap: (a) locates this helper via `arg[0]` / `debug.getinfo`,
+--- (b) loads it (which sets `package.path` + `package.cpath` via cached `git rev-parse`),
+--- (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 = {}
+
+-- Cache key for the repo root. Stored in `package.loaded` (process-global) so all 8 entry scripts + passes scripts share one resolution.
+local CACHE_KEY = "__duffle_repo_root__"
+
+--- Resolve the repo root from this script's own path. Zero shell spawn.
+--- `duffle_paths.lua` always lives at `<repo>/scripts/duffle_paths.lua`, so the repo root is the
+--- parent of the directory containing this script. We derive it directly from `debug.getinfo(1, "S").source`
+--- (returns `@<path>` for the currently-running chunk).
+---
+--- Replaces the prior `io.popen("git rev-parse --show-toplevel")` approach, which cost ~100-180ms per
+--- LuaJIT process on Windows due to git's CLI startup. The path-derive approach costs <1ms.
+---
+--- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source),
+--- return nil and let `M.setup()` fail loud.
+--- @return string|nil
+local function find_repo_root()
+	if package.loaded[CACHE_KEY] then return package.loaded[CACHE_KEY] end
+
+	local source = debug.getinfo(1, "S").source
+	-- Strip the leading `@` (Lua's dofile marker) and the trailing `/duffle_paths.lua` filename.
+	-- What remains is the directory containing this script, i.e. `<repo>/scripts/`.
+	local scripts_dir = source and source:match("^@?(.*)[/\\]duffle_paths%.lua$")
+	if not scripts_dir then return nil end
+
+	-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
+	local root = scripts_dir:gsub("scripts[\\/]?$", "")
+	root = root:gsub("\\", "/")
+	if root == "" then root = "./" end
+	if not root:match("/$") then root = root .. "/" end
+	package.loaded[CACHE_KEY] = root
+	return root
+end
+
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
+--- `package.cpath` (for `lpeg.dll`).
+---
+--- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$PATH` mods. 
+--- It just sets `package.path` and `package.cpath` (the standard Lua way to register module search dirs). 
+--- lpeg is built by `update_deps.ps1` to `toolchain/lpeg/`, 
+--- which we wire into `package.cpath` here (so `require("lpeg")` from `duffle.lua` resolves without any global state).
+function M.setup()
+	local repo_root = find_repo_root()
+	if not repo_root then
+		io.stderr:write("[duffle_paths] git rev-parse failed -- not in a git repo?\n")
+		os.exit(2)
+	end
+
+	local scripts_dir = repo_root .. "scripts/"
+	local passes_dir  = repo_root .. "scripts/passes/"
+	package.path = scripts_dir .. "?.lua;"
+		.. scripts_dir .. "?/init.lua;"
+		.. passes_dir  .. "?.lua;"
+		.. passes_dir  .. "?/init.lua;"
+		.. package.path
+
+	-- lpeg: built by `update_deps.ps1` to `toolchain/lpeg/lpeg.dll`.
+	-- lfs: compiled from pcsx-redux's vendored luafilesystem source to `toolchain/lfs/lfs.dll`.
+	-- Wire both directories into cpath so `require("lpeg")` and `require("lfs")` resolve.
+	local lpeg_dir = repo_root .. "toolchain/lpeg/"
+	local lfs_dir  = repo_root .. "toolchain/lfs/"
+	package.cpath = lpeg_dir .. "?.dll;"
+		.. lfs_dir  .. "?.dll;"
+		.. package.cpath
+end
+
+-- Run the setup as a side effect.
+M.setup()
+
+-- Now that package.path includes scripts/, `require("duffle")` resolves. Return the duffle module
+-- so callers can do `local duffle = dofile(...duffle_paths.lua)` in one line.
+return require("duffle")

scripts/elf_dwarf.lua

@@ -0,0 +1,925 @@
+--- 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).
+-- Track A Task 10: 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.
+--- - We strip the `code_` prefix to match the previous `read_nm` output. 
+--- - `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>` since the `code_` prefix was removed from the MipsAtom_ macro).
+				-- 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
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Parse a FORMAT_VERSION <expected_version> atoms-meta file (sourcemap or provenance).
+--- Shared by M.parse_source_map_file + M.parse_provenance_file.
+--- The two callers differ only in how they parse WORD lines; that's `extract_word(line)`.
+--- Returns the standard `{name -> {total, words}}` shape.
+--- Returns `{}` on format-version mismatch (and logs to stderr).
+--- @param path             string
+--- @param expected_version integer
+--- @param extract_word     fun(line: string): table|nil  -- caller-supplied per-line parser
+--- @return table<string, table>
+function M.parse_atom_records(path, expected_version, extract_word)
+	local out = {}
+	local cur_name, cur_words = nil, {}
+	for raw in io.lines(path) do
+		local line = raw
+		if line:match("^#") then
+			local ver = line:match("^# FORMAT_VERSION%s+(%d+)")
+			if ver and tonumber(ver) ~= expected_version then
+				io.stderr:write(string.format(
+					"[elf_dwarf.parse_atom_records] version mismatch (got %s, expected %d) in %s\n",
+					ver, expected_version, path))
+				return {}
+			end
+			-- skip other comments
+		elseif line:sub(1, 4) == "ATOM" then
+			-- ATOM <name> "<abs-source-path>" <total>
+			local _, _, name = line:find("ATOM%s+(%S+)%s+\"[^\"]*\"%s+(%d+)")
+			if name then
+				cur_name  = name
+				cur_words = {}
+				out[name] = { total = 0, words = cur_words }
+			end
+		elseif line == "ENDATOM" then
+			-- Update the recorded total from the entries count
+			-- (matches the `lines[1] = lines[1]:gsub(" 0$", " " .. total)` patch in atoms_source_map.lua:170).
+			if cur_name and out[cur_name] then
+				out[cur_name].total = #cur_words
+			end
+			cur_name, cur_words = nil, {}
+		elseif line:sub(1, 4) == "WORD" and cur_name then
+			local field = extract_word(line)
+			if field then
+				cur_words[#cur_words + 1] = field
+			end
+		end
+	end
+	return out
+end
+
+--- Parse a FORMAT_VERSION <expected_version> `*.atoms.sourcemap.txt` file.
+--- Returns `{name -> {total = N, words = {{pos, line}, ...}}}`.
+--- Returns `{}` on format-version mismatch (and logs to stderr).
+---
+--- **Wire format** (emitted by `passes/atoms_source_map.lua`):
+--- ```
+--- # FORMAT_VERSION <n>
+--- ATOM <name> "<abs-source-path>" <total>
+--- WORD <n> LINE <line> TEXT <text...>
+--- ...
+--- ENDATOM
+--- ```
+---
+--- **Conventions:** the in-memory shape uses `{pos, line, text}`
+--- (`atoms_source_map.lua:142`); the `.txt` file uses `WORD <n>` so the parser maps `n` → `pos` field name.
+--- @param sm_path Path
+--- @param expected_version integer  -- expected FORMAT_VERSION line
+--- @return table<string, table>
+function M.parse_source_map_file(sm_path, expected_version)
+	return M.parse_atom_records(sm_path, expected_version, function(line)
+		local _, n, _, src_line = line:find("WORD%s+(%d+)%s+LINE%s+(%d+)")
+		if n and src_line then
+			return { pos = tonumber(n), line = tonumber(src_line) }
+		end
+	end)
+end
+
+--- Parse a FORMAT_VERSION <expected_version> `*.atoms.provenance.txt` file.
+--- Returns `{name -> {total = N, words = {{pos, call_file, call_line, comp_name, comp_file, comp_line}, ...}}}`.
+--- Returns `{}` on format-version mismatch (and logs to stderr).
+---
+--- **Wire format** (emitted by `passes/atoms_source_map.lua`):
+--- ```
+--- # FORMAT_VERSION <n>
+--- ATOM <name> "<abs-source-path>" <total>
+--- WORD <n>  CALL <src-file>:<src-line>  RAW
+--- WORD <n>  CALL <src-file>:<src-line>  MACRO <comp_name> "<comp-file>:<comp-line>"
+--- ...
+--- ENDATOM
+--- ```
+---
+--- **Used by** `passes/dwarf_injection.lua` to:
+---   - group consecutive MACRO rows into component invocations (one `DW_TAG_inlined_subroutine` each)
+---   - emit abstract `DW_TAG_subprogram` per unique component name
+---   - extend `.debug_line` so stepping into a `mac_X(...)` lands on the component's source line.
+--- @param prov_path string  -- path to *.atoms.provenance.txt
+--- @param expected_version integer  -- expected FORMAT_VERSION line
+--- @return table<string, table>
+function M.parse_provenance_file(prov_path, expected_version)
+	return M.parse_atom_records(prov_path, expected_version, function(line)
+		-- Two accepted shapes:
+		--   WORD <n>  CALL <call-file>:<call-line>  RAW
+		--   WORD <n>  CALL <call-file>:<call-line>  MACRO <comp_name> "<comp-file>:<comp-line>"
+		local pos, call_file, call_line, comp_name, comp_file, comp_line =
+			line:match('WORD%s+(%d+)%s+CALL%s+(.-):(%d+)%s+MACRO%s+(%S+)%s+"([^"]*):(%d+)"')
+		if pos then
+			return {
+				pos       = tonumber(pos),
+				call_file = call_file,
+				call_line = tonumber(call_line),
+				comp_name = comp_name,
+				comp_file = comp_file,
+				comp_line = tonumber(comp_line),
+			}
+		end
+		-- RAW row.
+		local raw_pos, raw_file, raw_line = line:match('WORD%s+(%d+)%s+CALL%s+(.-):(%d+)%s+RAW')
+		if raw_pos then
+			return {
+				pos       = tonumber(raw_pos),
+				call_file = raw_file,
+				call_line = tonumber(raw_line),
+				comp_name = nil,
+				comp_file = nil,
+				comp_line = nil,
+			}
+		end
+	end)
+end
+
+return M

scripts/gdb/gdb_tape_atoms.gdb

@@ -0,0 +1,105 @@
+# scripts/gdb/gdb_tape_atoms.gdb
+#
+# Wrapper for the tape-atom step-debug helpers.
+# The 9 user commands are defined here as STUBS (degraded-state messages). 
+# The real implementations + the per-atom data tables are emitted by `passes/atoms_source_map.lua` 
+# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/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

@@ -0,0 +1,692 @@
+--- passes/annotation.lua — Atom-annotation DSL validator.
+---
+--- Validates `MipsAtom_(name) atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...)) { ... }` declarations in source files.
+--- Also reads: `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`)
+---
+--- Source scanning: done ONCE upstream by `duffle.scan_source()` (ps1_meta.lua pre-scans each source and stashes the result in `src.scan`).
+---
+--- 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`
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, 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 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 write_file     = duffle.write_file
+local ensure_dir     = duffle.ensure_dir
+
+-- The annotation pass now consults the source-derived registries built by scan_source:
+--   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
+--   * pipe_ctx.type_name_registry      — for atom_dbg_reg_default(<T>, ...) and atom_reg_types(<T>, ...) type-identity checks
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+--- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+
+--- @class PassCtx
+--- @field sources            SourceFile[]
+--- @field metadata_path      string
+--- @field shared             table
+--- @field shared.word_counts table<string, integer>
+--- @field out_root           string
+--- @field project_root       string
+--- @field upstream           table<string, table>
+--- @field flags              table
+--- @field flags._annot_results table[] -- stashed by annotation pass; consumed by report.lua
+--- @field dry_run            boolean
+--- @field verbose            boolean
+
+--- @class PassResult
+--- @field outputs  table[]
+--- @field errors   table[]
+--- @field warnings table[]
+
+--- @class AtomAnnotation
+--- @field line    integer       -- source line of the atom_info call
+--- @field macro   string        -- the macro name (always "atom_info" in the new shape)
+--- @field name    string        -- the atom name
+--- @field kind    string        -- always "info"
+--- @field binds   string|nil    -- Binds_X name if any
+--- @field reads   string[]      -- R_* names (read targets)
+--- @field writes  string[]      -- R_* names (write targets)
+--- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
+
+--- @class 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 msg  string  -- finding message
+
+--- @class Findings
+--- @field errors   Finding[]
+--- @field warnings Finding[]
+--- @field info     Finding[]
+
+--- @class PipeCtx
+--- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
+--- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
+--- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
+--- @field types          table<string, RegTypeDefault> -- from scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
+--- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field seen_field     table<string, integer>        -- Binds_* -> count of fields (set/checked by check_binds_no_duplicate_fields)
+--- @field _scan          SourceScan                    -- full scan payload (typed-view sub-calls live here) 
+
+--- @class AnnotatedResult
+--- @field atoms    AtomEntry[]
+--- @field annots   AtomAnnotation[]
+--- @field macros   MacroEntry[]
+--- @field binds    BindsEntry[]
+--- @field errors   Finding[]
+--- @field warnings Finding[]
+--- @field info     Finding[]
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-check functions (the CHECK_RULES table's payload)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- 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
+
+--- Migration warning emitted alongside the new registry-membership check.
+---
+--- R_TapePtr / R_AtomJmp / R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase
+--- are the wave-context aliases opted in via `#define atom_reg` in lottes_tape.h (Task 21).
+--- Any source referencing an R_X that's NOT in the registry will trip the new check; a single pass-level info entry
+--- (emitted only when at least one such rejection lands in this source) tells users where to look.
+---
+--- This check is a stop-gap until users migrate off raw C-ABI register names.
+--- @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.
+-- No source walking; no parsing. The scan was done once upstream.
+
+--- Validate one source against its pre-scanned SourceScan payload.
+--- @param ctx PassCtx
+--- @param src SourceFile
+--- @return AnnotatedResult
+local function validate(ctx, src)
+	local scan = src.scan
+
+	-- Project the pre-scanned atoms to the AtomEntry shape this pass needs.
+	local atoms = {}
+	for _, a in ipairs(scan.atoms) do
+		if a.kind == "atom" then
+			atoms[#atoms + 1] = { line = a.line, name = a.raw_name }
+		end
+	end
+
+	-- Project the pre-scanned atom_infos to AtomAnnotation shape.
+	local annots = {}
+	for _, info in ipairs(scan.atom_infos) do
+		annots[#annots + 1] = {
+			line   = info.info_line,
+			macro  = "atom_info",
+			name   = info.atom_name,
+			kind   = "info",
+			binds  = info.binds,
+			reads  = info.reads or {},
+			writes = info.writes or {},
+			errors = info.errors,
+		}
+	end
+
+	-- Build pipe_ctx (Fleury: expose structure). Pre-compute everything the per-check functions need.
+	-- Single source of truth for atom / binds / annotation-count lookups.
+	-- pipe_ctx.types / pipe_ctx.atom_views / pipe_ctx.seen_defaults are projected from the scan payload so per_source check rules can iterate.
+	local seen_defaults = {}
+	for reg, _ in pairs(scan.types or {}) do
+		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
+	end
+	local atom_infos_list = {}
+	for _, ai in ipairs(scan.atom_infos or {}) do
+		atom_infos_list[#atom_infos_list + 1] = ai
+	end
+
+	local pipe_ctx = {
+		atom_index               = {},
+		binds_index              = {},
+		annot_counts             = {},
+		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 {},
+		-- Project the source-derived registries from the scan payload so per_source checks consult them instead of the deleted
+		-- SEMANTIC_DEFAULT_REGS / KNOWN_REG_DEFAULT_TYPES / etc.
+		register_alias_registry  = scan.register_alias_registry or {},
+		type_name_registry       = scan.type_name_registry or {},
+	}
+	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
+	for _, a in ipairs(annots)     do
+		if a.name then
+			pipe_ctx.annot_counts[a.name] = (pipe_ctx.annot_counts[a.name] or 0) + 1
+		end
+	end
+
+	-- Findings live in a single struct with three lists (errors / warnings / info).
+	-- Each check writes to the list appropriate for its severity.
+	local findings = { errors = {}, warnings = {}, info = {} }
+
+	-- 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 a.errors then
+			for _, msg in ipairs(a.errors) do
+				findings.errors[#findings.errors + 1] = {
+					line = a.line,
+					msg  = string.format("'%s': %s", a.name, msg),
+				}
+			end
+		end
+	end
+
+	-- THE per-annotation pipeline. ONE loop. CHECK_RULES dispatches per_annot rules.
+	for _, a in ipairs(annots) do
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_annot then rule.per_annot(a, pipe_ctx, findings) end
+		end
+	end
+
+	-- Post-loop rules (one-shot checks that need full-corpus aggregation in pipe_ctx).
+	for _, rule in ipairs(CHECK_RULES) do
+		if rule.post then rule.post(pipe_ctx, findings) end
+	end
+
+	-- Per-skip-marker rules. 
+	-- Each raw marker recorded by scan_source (in scan.skip_over.markers) is validated independently; 
+	-- the check emits at most one error per marker.
+	-- 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
+
+	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
+	local wc = ctx.shared.word_counts
+	for _, m in ipairs(scan.macros) do
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_macro then rule.per_macro(m, wc, findings) end
+		end
+	end
+
+	-- Per-source rules (reg defaults, atom_view layout, compute-register type overrides, Binds_* field uniqueness).
+	-- Each per_source rule sees the full scan payload via pipe_ctx.
+	for _, rule in ipairs(CHECK_RULES) do
+		if rule.per_source then rule.per_source(src, pipe_ctx, findings) end
+	end
+
+	-- Information summary (always emitted).
+	findings.info[#findings.info + 1] = {
+		line = 0,
+		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)",
+			#atoms, #annots, #scan.macros, #scan.binds),
+	}
+
+	return {
+		atoms    = atoms,
+		annots   = annots,
+		macros   = scan.macros,
+		binds    = scan.binds,
+		errors   = findings.errors,
+		warnings = findings.warnings,
+		info     = findings.info,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-DIRECTORY (per-module) output: errors.h + annotations.txt
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
+--- Empty directories (no errors, no atoms) produce no file.
+local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
+	if ctx.dry_run then return nil end
+	if atoms_count == 0 and #errors == 0 then
+		return nil
+	end
+	local out_path = ctx.out_root .. "/" .. dir_basename .. ".errors.h"
+	local lines    = {
+		"// Auto-generated by ps1_meta.lua (passes/annotation.lua) — DO NOT EDIT",
+		string.format("// Module: %s   Sources: %d", dir_basename, #sources),
+		"#pragma once",
+		"",
+	}
+	if #errors == 0 then
+		lines[#lines + 1] = "// annotation pass OK"
+	else
+		for _, e in ipairs(errors) do
+			local src_tag = ""
+			if e.source then
+				local src_name = e.source:match("([^/\\]+)$") or e.source
+				src_tag = src_name .. ": "
+			end
+			lines[#lines + 1] = string.format('#error "%s%s (line %d)"', src_tag, e.msg, e.line)
+		end
+	end
+	ensure_dir(ctx.out_root)
+	write_file(out_path, table.concat(lines, "\n") .. "\n")
+	return out_path
+end
+
+--- Stash aggregated per-module results for the report pass to consume.
+local function emit_module_annotations_stub(ctx, dir, dir_basename, atoms_count)
+	ctx.flags                = ctx.flags                or {}
+	ctx.flags._annot_results = ctx.flags._annot_results or {}
+	ctx.flags._annot_results[#ctx.flags._annot_results + 1] = {
+		dir          = dir,
+		dir_basename = dir_basename,
+		atoms_count  = atoms_count,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M.run — orchestrator entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class M
+
+local M = {}
+
+-- Expose `validate` for downstream passes (e.g. report.lua) that need to re-render the per-source results into a per-MODULE report.
+M.validate = validate
+
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	-- Per-DIRECTORY (per-module) aggregation. Group sources by `src.dir`, validate every source in the dir, then emit ONE errors.h per dir.
+	-- `ctx.by_dir` is pre-computed in build_ctx (shared across all passes).
+	local by_dir = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+
+	for dir, dir_sources in pairs(by_dir) do
+		local dir_basename = dir:match("([^/\\]+)$") 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)
+			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 }
+			end
+			for _, w in ipairs(result.warnings) do
+				dir_warnings[#dir_warnings + 1] = { line = w.line, msg = w.msg }
+				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
+			end
+		end
+
+		local err_path = emit_module_errors_h(ctx, dir_basename, dir_atoms, dir_errors, dir_sources)
+		if err_path then
+			table.insert(outputs, { errors_h = err_path })
+		end
+
+		emit_module_annotations_stub(ctx, dir, dir_basename, dir_atoms)
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/atoms_source_map.lua

@@ -0,0 +1,749 @@
+--- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
+---
+--- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
+--- for `MipsAtom_(name)` (kind="atom"), `MipsAtomComp_` / `MipsAtomComp_Proc_` (kind="comp_*"),
+--- and `MipsCode code_<name>` (kind="raw_atom") declarations. 
+--- Walks each atom's pre-tokenized body (`{{tok=string, rel=integer}, ...}` from `duffle.tokenize_body`), 
+--- counts per-token word contributions via `ctx.shared.word_counts`, and emits one
+--- `WORD N LINE L TEXT T` line per `.word` to `<out_root>/<basename>.atoms.sourcemap.txt`.
+---
+--- **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 calls (`atom_label(...)`, `atom_offset(...)`) emit 0 `.word`s. 
+--- They share the same walking convention as `passes/offsets.lua :: scan_atom_body`: 
+--- Markers do NOT advance the word-offset counter, but if a marker is bundled on the same token with a trailing instruction 
+--- (e.g. `atom_label(foo) load_half_u(...)`), the trailing instruction's word count is added. This matches `offsets.lua :: count_marker_rest`.
+---
+--- **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")
+local word_count_eval  = require("word_count_eval")
+local count_token_words = word_count_eval.count_token_words
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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
+
+-- Marker-call identifiers (mirrors offsets.lua:33-34).
+local LABEL_MARKER  = "atom_label"
+local OFFSET_MARKER = "atom_offset"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class AtomSourceMapCtx
+--- @field sources            table[] -- SourceScan payload per source (from `ctx.sources`)
+--- @field shared             table   -- `ctx.shared`
+--- @field shared.word_counts table   -- macro name -> word count (populated by word-counts + components passes)
+--- @field out_root           string  -- output root (e.g. "build/gen")
+--- @field dry_run            boolean -- if true, compute but don't write
+--- @field flags              table   -- `ctx.flags`; reads `flags.gdb_runtime` + `flags.elf_path`
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Helpers
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Provenance emission
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Component-macro invocation prefix (mirrors components.lua's MAC_PREFIX).
+local MAC_PREFIX     = "mac_"
+local MAC_PREFIX_LEN = 4
+
+--- Strip the `mac_` prefix from a token's leading identifier.
+--- Returns nil if the identifier doesn't start with `mac_`
+--- (so non-component tokens like `load_half_u`, `nop2`, `gte_cmdw_*` fall through cleanly).
+--- @param tok string
+--- @return string|nil
+local function strip_mac_prefix_from_token(tok)
+	local leading = duffle.read_ident(tok, 1)
+	if not leading then return nil end
+	if leading:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+		return leading:sub(MAC_PREFIX_LEN + 1)
+	end
+	return nil
+end
+
+--- Fetch the per-word body lines for a `mac_X(...)` invocation.
+--- Walks the component's pre-tokenized body in lockstep with `count_token_words` and attributes each emitted `.word`
+--- to a source line via `idx.line_of(...)`. 
+--- Atom labels (`atom_label(...)`) emit 0 `.word`s and are skipped.
+--- @param bare string|nil          -- the bare component name (e.g. `gte_load_tri_verts`)
+--- @param comp_body_index table
+--- @param wc table
+--- @return table|nil                -- list of source lines, 1-based by word position
+local function fetch_body_lines(bare, comp_body_index, wc)
+	if not (bare and comp_body_index) then return nil end
+	local idx = comp_body_index[bare]
+	if not (idx and idx.body_tokens and idx.line_of) then return nil end
+	local lines = {}
+	for _, bt in ipairs(idx.body_tokens) do
+		local bt_tok = duffle.trim(bt.tok or "")
+		if bt_tok ~= "" then
+			local leading = duffle.read_ident(bt_tok, 1)
+			local bt_words
+			if leading == "atom_label" or leading == "atom_offset" then
+				bt_words = 0
+			else
+				bt_words = count_token_words(bt_tok, wc)
+			end
+			if bt_words > 0 then
+				local body_line = idx.line_of(idx.body_off + bt.rel)
+				for _ = 1, bt_words do lines[#lines + 1] = body_line end
+			end
+		end
+	end
+	return lines
+end
+
+--- Unified per-word entry walker. `mode` is "sourcemap" (3 fields) or "provenance" (8 fields including component + body-line lookup).
+--- Returns (entries, total_words). Markers contribute 0 entries.
+--- @param atom table
+--- @param src  table
+--- @param wc   table
+--- @param mode string           -- "sourcemap" | "provenance"
+--- @param comp table|nil        -- shared.components map (provenance only)
+--- @param comp_body_index table|nil  -- per-source body index (provenance only)
+--- @return table[], integer
+local function compute_word_entries(atom, src, wc, mode, comp, comp_body_index)
+	local entries = {}
+	local pos     = 0
+	for _, t in ipairs(atom.body_tokens) do
+		local tok = t.tok
+		local rel = t.rel
+
+		local words
+		if duffle.is_marker_token(tok) then
+			words = duffle.count_marker_rest(tok, wc, count_token_words)
+		else
+			words = count_token_words(tok, wc)
+		end
+
+		-- Provenance-only: resolve component + body_lines (one fetch per token).
+		local comp_name, comp_line, comp_path, comp_kind
+		local body_lines
+		if mode == "provenance" then
+			local bare = strip_mac_prefix_from_token(tok)
+			if bare and comp and comp[bare] then
+				comp_name = bare
+				comp_line = comp[bare].line
+				comp_path = comp[bare].path
+				comp_kind = comp[bare].kind
+			end
+			if comp_name then body_lines = fetch_body_lines(bare, comp_body_index, wc) end
+		end
+
+		if words > 0 then
+			local line = src.scan.line_of(atom.body_off + rel)
+			local text = duffle.trim(tok):gsub("[\t\r\n]+", " ")
+			for i = 1, words do
+				local entry
+				if mode == "provenance" then
+					entry = {
+						pos       = pos,
+						line      = line,
+						text      = text,
+						comp_name = comp_name,
+						comp_line = comp_line,
+						comp_path = comp_path,
+						comp_kind = comp_kind,
+						body_line = body_lines and body_lines[i],
+					}
+				else  -- "sourcemap" (default)
+					entry = { pos = pos, line = line, text = text }
+				end
+				entries[#entries + 1] = entry
+				pos = pos + 1
+			end
+		end
+	end
+	return entries, pos
+end
+
+--- Render one atom's provenance stanza. Format:
+---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>"  [BODY <line>]`  (for component words)
+---   `WORD N  CALL <src-path>:<src-line>  RAW`                                                  (for direct instructions)
+--- `BODY <line>` is the source line of THIS specific word within the macro body
+--- (lottes_tape.h:N where N is the per-word body line).
+--- Absent for RAW rows and for component rows whose component declaration could not be indexed (older pass combinations / external macros).
+--- Downstream consumers (dwarf_injection, tests) fall back to DefLine / comp_line when BODY is absent.
+--- Returns (lines, total_words).
+--- @param src    table
+--- @param atom   table
+--- @param wc     table
+--- @param comp   table  -- shared.components map
+--- @param comp_body_index table  -- per-source component body index: bare_name -> {body_off, body_tokens, line_of}
+--- @return string[], integer
+local function emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
+	local lines    = {}
+	local rel_path = src.path:gsub("\\", "/")
+	local entries, total = compute_word_entries(atom, src, wc, "provenance", comp, comp_body_index)
+
+	-- ATOM header line with placeholder total (patched after we know it).
+	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
+
+	for _, pe in ipairs(entries) do
+		if pe.comp_name then
+			local body_suffix = ""
+			if pe.body_line then
+				body_suffix = "  BODY " .. tostring(pe.body_line)
+			end
+			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d"%s',
+				pe.pos, rel_path, pe.line, pe.comp_name, pe.comp_path, pe.comp_line, body_suffix)
+		else
+			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", pe.pos, rel_path, pe.line)
+		end
+	end
+
+	-- Patch the placeholder total in the ATOM header line.
+	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
+	lines[#lines + 1] = "ENDATOM"
+	return lines, total
+end
+
+--- Build a per-source component body index keyed by the bare component name (e.g. `gte_load_tri_verts`).
+--- Each entry holds the data we need to map each emitted `.word` to its actual source line within the macro body:
+---   body_off    -- byte offset of the `{` (start of body) in the component's source file.
+---   body_tokens -- list of {tok, rel} pairs; `rel` is the byte offset within the body.
+---   line_of     -- closure resolving byte offsets in the component's source file to lines.
+--- Only `comp_bare` + `comp_proc` declarations contribute (a macro invocation can only resolve to one of those).
+--- First declaration wins (subsequent redeclarations would collide; today's sources declare each component exactly once).
+--- Render the full provenance file content for one source (one `.atoms.provenance.txt` per source).
+--- @param src table
+--- @param wc table
+--- @param comp table  -- shared.components map
+--- @param comp_body_index table  -- cross-source component body index (built once in M.run; may be empty)
+--- @return string
+local function render_provenance(src, wc, comp, comp_body_index)
+	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."
+
+	-- The cross-source component body index is passed in from M.run (one global lookup shared across every source's provenance file).
+	-- A per-source lookup would miss every component whose declaration is in another source (e.g. `gte_load_tri_verts` is declared in `lottes_tape.h` but invoked from `hello_gte_tape.c`).
+
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		local stanza = emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		local stanza = emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line 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, wc)
+	local lines              = {}
+	local rel_path           = src.path:gsub("\\", "/")
+	local entries, total     = compute_word_entries(atom, src, wc)
+
+	-- ATOM header line with placeholder total (patched after we know it).
+	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
+	for _, we in ipairs(entries) do
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			we.pos, we.line, we.text)
+	end
+
+	-- Patch the placeholder total in the ATOM header line.
+	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, wc)
+	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"
+
+	for   _, atom in ipairs(src.scan.atoms or {}) do
+		local stanza = emit_atom_stanza(src, atom, wc)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for   _, atom in ipairs(src.scan.raw_atoms or {}) do
+		local stanza = emit_atom_stanza(src, atom, wc)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line 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 $var = "..."` literal.
+--- gdb uses C-style escaping; we escape `\` and `"` (newlines were flattened earlier).
+--- @param s string
+--- @return string
+local function gdb_escape(s)
+	return (s:gsub("\\", "\\\\"):gsub('"', '\\"'))
+end
+
+--- Build the list of atoms with addresses + word entries.
+--- Shared helper for the gdb-runtime file emission.
+--- @param ctx PassCtx
+--- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
+local function build_atom_table(ctx)
+	local wc    = (ctx.shared and ctx.shared.word_counts) or {}
+	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
+
+	local matched = {}
+	for _, src in ipairs(ctx.sources) do
+		if src.scan then
+			local file_base = src.path:match("([^/\\]+)$") or src.path
+			for _, atom in ipairs(src.scan.atoms or {}) do
+				if atom.kind == nil or atom.kind == "atom" then
+					local name    = atom.raw_name or atom.name
+					local info    = addrs[name]
+					if info then
+						local entries, total = compute_word_entries(atom, src, wc)
+						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
+				end
+			end
+			for _, atom in ipairs(src.scan.raw_atoms or {}) do
+				local name = atom.name
+				local info = addrs[name]
+				if info then
+					local entries, total = compute_word_entries(atom, src, wc)
+					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
+			end
+		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] = ""
+
+	-- ── show_c2 ──
+	-- GTE data regs (COP2). pcsx-redux's gdb stub doesn't expose COP2 (only 72 regs: 32 GPR + COP0 + FPR).
+	--   curl http://localhost:8080/api/v1/lua/gte
+	-- We keep the command definition as a stub that points the user at the plugin.
+	lines[#lines + 1] = "define show_c2"
+	lines[#lines + 1] = '  echo "[gdb_tape_atoms] show_c2: gdb stub does not expose COP2 in this build."'
+	lines[#lines + 1] = '  echo "[gdb_tape_atoms] Use scripts/pcsx_debug_helper.zip + curl http://localhost:8080/api/v1/lua/gte"'
+	lines[#lines + 1] = '  echo "[gdb_tape_atoms] (or pcsx-redux Debug > Registers window for a native view)"'
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document show_c2"
+	lines[#lines + 1] = "  Stub. The gdb stub in this pcsx-redux build does not expose COP2 regs."
+	lines[#lines + 1] = "  For GTE data + control state, use the pcsx_debug_helper Lua plugin or the"
+	lines[#lines + 1] = "  pcsx-redux Debug > Registers window."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── show_c2ctl ──
+	lines[#lines + 1] = "define show_c2ctl"
+	lines[#lines + 1] = '  echo "[gdb_tape_atoms] show_c2ctl: see show_c2 for the same workaround."'
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document show_c2ctl"
+	lines[#lines + 1] = "  Stub. Same workaround as show_c2."
+	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"
+	if not ctx.dry_run then
+		duffle.ensure_dir(duffle.dirname(out_path))
+		duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
+	end
+	io.stderr:write(string.format(
+		"[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+--- Build the cross-source component body index used by `render_provenance` to attribute each emitted `.word` to its actual line within the macro body.
+---
+--- Components are declared in one source (the header that contains `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`) 
+--- but invoked from many source files (every atom body that calls `mac_X(...)`).
+--- The body_offset + body_tokens + line_of live with the declaration source, so a per-source index would miss invocations from other sources.
+---
+--- The cross-source index is keyed by the bare component name (`gte_load_tri_verts`, NOT `ac_gte_load_tri_verts`)
+--- `strip_mac_prefix_from_token` strips the `mac_` prefix from call-site identifiers and yields that exact bare name;
+--- matching it here keeps the lookup aligned with the `ctx.shared.components` map's keying convention.
+--- First declaration wins (subsequent redeclarations would collide; today's sources declare each component exactly once).
+--- @param ctx PassCtx
+--- @return table<string, table>  -- {[comp_name] = {body_off, body_tokens, line_of}}
+local function build_cross_source_component_body_index(ctx)
+	local index = {}
+	for _, src in ipairs(ctx.sources or {}) do
+		if src.scan and src.scan.atoms then
+			local line_of = src.scan.line_of
+			for _, atom in ipairs(src.scan.atoms) do
+				if atom.kind == "comp_bare" or atom.kind == "comp_proc" then
+					-- Prefer `atom.name` (stripped of `ac_` prefix); fall back to `raw_name`
+					-- only if the stripped name is absent (defensive — current scan-source always sets both).
+					local name = atom.name or atom.raw_name
+					if name and not index[name] then
+						index[name] = {
+							body_off    = atom.body_off,
+							body_tokens = atom.body_tokens,
+							line_of     = line_of,
+						}
+					end
+				end
+			end
+		end
+	end
+	return index
+end
+
+--- 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 = {}
+
+	-- word-counts + components passes must have populated shared.word_counts.
+	-- If absent, the orchestrator wired the deps wrong — fail loud.
+	local wc = (ctx.shared and ctx.shared.word_counts) or {}
+	if not wc or not next(wc) then
+		warnings[#warnings + 1] = {
+			line = 0,
+			msg  = "atoms_source_map: ctx.shared.word_counts is empty; the word-counts + components passes may not have populated it. Check the PASSES dep edges.",
+		}
+	end
+
+	-- shared.components map is populated by `passes/components.lua`.
+	-- Used to attribute each emitted `.word` to either a component macro or the enclosing atom body.
+	-- If absent, all words fall through as RAW (correct behavior — provenance is additive).
+	local comp = (ctx.shared and ctx.shared.components) or {}
+
+	-- Cross-source component body index.
+	-- Built ONCE so every source's provenance writer can resolve `mac_X(...)` invocations back to the macro's body tokens (regardless of which source declared the component).
+	-- Per-source copies were insufficient — the atom file (`hello_gte_tape.c`) does not contain the `MipsAtomComp_(...)` declarations, 
+	-- so the body data would be missing for every component invocation the atom file emitted.
+	local comp_body_index = build_cross_source_component_body_index(ctx)
+
+	-- Always emit the canonical text form (per-source).
+	for _, src in ipairs(ctx.sources) do
+		if src.scan then
+			local n_atoms     = src.scan.atoms and #src.scan.atoms or 0
+			local n_raw_atoms = src.scan.raw_atoms and #src.scan.raw_atoms or 0
+			if n_atoms + n_raw_atoms > 0 then
+				local basename = duffle.basename_no_ext(src.path)
+
+				-- (1) atoms.sourcemap.txt — per-.word line map (unchanged contract).
+				local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
+				local sourcemap_body = render_source_map(src, wc)
+
+				-- (2) atoms.provenance.txt — per-.word provenance with `mac_X(...)` component resolution back to the component's definition file:line.
+				--     Consumed by `passes/dwarf_injection.lua` to synthesize `DW_TAG_inlined_subroutine` instances for source-level Step Into on component invocations.
+				local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
+				local prov_body = render_provenance(src, wc, comp, comp_body_index)
+
+				if not ctx.dry_run then
+					duffle.ensure_dir(duffle.dirname(sourcemap_path))
+					duffle.write_file_lf(sourcemap_path, sourcemap_body)
+					duffle.write_file_lf(prov_path,    prov_body)
+				end
+
+				outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
+				outputs[#outputs + 1] = { kind = "report", path = prov_path }
+			end
+		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

@@ -0,0 +1,651 @@
+--- 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 per-source backward lookups
+--- for the function-args string (from the preceding `FI_ MipsAtom ac_X(...)` function declaration)
+--- 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 + `WORD_COUNT(mac_X, N)`
+--- 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 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 word_count_eval = require("word_count_eval")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Atom component declaration identifiers.
+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_LEN   = 3
+local MAC_PREFIX      = "mac_" -- prefix on generated macros; the rest is the atom name
+local MAC_PREFIX_LEN  = 4
+
+-- ASCII byte values used in tokenization.
+local BYTE_NEWLINE    = 10
+local BYTE_SLASH      = 47
+
+-- Source dir basename used as the output `.macs.h` filename.
+local GEN_SUBDIR      = "gen"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+--- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+
+--- @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<string, integer> -- populated by word-counts + components
+--- @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
+--- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+
+--- @class Component
+--- @field name    string     -- atom name (without `ac_` prefix)
+--- @field body    string     -- brace-delimited body (without the braces)
+--- @field args    string|nil -- function-args string (function form only)
+--- @field line    integer    -- source line of the declaration
+--- @field comment string|nil -- preceding `/* */` or `//` comment block (signature doc)
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Local helpers (file I/O + path normalization)
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Back-walk helpers (composed into the 2 entry points below: find_function_args_for + preceding_comment_block)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- 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 with possible qualifiers between).
+---
+--- @param source string
+--- @param name string
+--- @param before_pos integer
+--- @return string|nil
+local function find_function_args_for(source, name, before_pos)
+	-- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
+	local name_open  = name .. "("
+	local last_idx   = nil
+	local scan_pos   = 1
+	while true do
+		-- Pass `before_pos + 1` so string.find only returns positions < before_pos + 1
+		-- (string.find's 4th arg `plain` is true; we use the 3rd arg `init` for the upper bound).
+		local found = source:find(name_open, scan_pos, true)
+		if not found or found >= before_pos then break end
+		last_idx = found
+		scan_pos = found + #name_open
+	end
+	if not last_idx then return nil end
+
+	-- Verify the preceding context ends with "MipsAtom" (with possible qualifiers between).
+	local before   = source:sub(1, last_idx - 1)
+	local trimmed  = duffle.trim(before)
+	if trimmed:sub(-#MIPS_ATOM) ~= MIPS_ATOM then
+		-- Preceding context is not a function declaration.
+		return nil
+	end
+
+	local open_paren = last_idx + #name  -- position of "("
+	-- scan: MipsAtom ac_X(
+	local inner      = duffle.read_parens(source, open_paren)
+	-- scan: MipsAtom ac_X(<args>)
+	if not inner then return nil end
+	return inner
+end
+
+--- 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
+--- @return string
+local function preceding_comment_block(source, pos)
+	local scan_pos = pos
+	local pieces   = {}
+	while true do
+		-- Skip whitespace (space/tab/newline/CR) backward from `scan_pos`,
+		-- returning the position of the first non-whitespace char.
+		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
+
+		local is_block_close = non_ws >= 2 and source:sub(non_ws - 1, non_ws) == "*/"
+		local is_line_end    = source:sub(non_ws, non_ws) == "\n" or source:sub(non_ws, non_ws) == "\r"
+
+		if is_block_close then
+			-- Find the opening `/*` for a block comment whose `*/` ends at `non_ws`.
+			-- Walk back from `non_ws` over `/*` candidates.
+			local prefix  = source:sub(1, non_ws - 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
+			if not open_at then break end
+			-- Walk back from `open_at` over leading spaces + tabs to include the indentation before the `/*`.
+			local block_start = open_at
+			while block_start > 1 do
+				local ch = source:sub(block_start - 1, block_start - 1)
+				if ch == " " or ch == "\t" then
+					block_start = block_start - 1
+				else
+					break
+				end
+			end
+			table.insert(pieces, 1, source:sub(block_start, non_ws))
+			scan_pos = block_start
+		elseif is_line_end then
+			-- Walk back from `non_ws` to the start of the source line (the 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
+				table.insert(pieces, 1, line)
+				scan_pos = line_start - 1
+			else
+				break
+			end
+		else
+			break
+		end
+	end
+	if #pieces == 0 then return "" end
+	return table.concat(pieces, "\n")
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Argument-name extraction
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Extract just the parameter NAMES from a function-args string (stripping type annotations). E.g.,
+---   `"U4 off, U4 code, U1 r, U1 g, U1 b"`  ->  `{"off", "code", "r", "g", "b"}`
+---   `"U4 *ptr"`                            ->  `{"ptr"}`
+---   `""`                                   ->  nil
+--- @param args_str string|nil
+--- @return string[]|nil
+local function extract_arg_names(args_str)
+	if not args_str or args_str == "" then return nil end
+	local names  = {}
+	local tokens = duffle.split_top_level_commas(args_str)
+	for _, tok in ipairs(tokens) do
+		local trimmed = duffle.trim(tok)
+		if trimmed ~= "" then
+			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
+			-- then walk back over the identifier chars (alnum + `_`).
+			-- Plex: inlined the 2 single-caller helpers (no 2-caller rule met).
+			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(ch) or ch == "_" then
+					ident_start = ident_start - 1
+				else
+					break
+				end
+			end
+			ident_start = ident_start + 1
+			local name = trimmed:sub(ident_start, ident_end)
+			if name ~= "" then names[#names + 1] = name end
+		end
+	end
+	if #names == 0 then return nil end
+	return names
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Component projection (read from pre-scanned SourceScan)
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
+-- 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 instead of calling duffle.tokenize_body again.
+-- @param source string  -- the full source text (needed for backward lookups)
+-- @param scan  table   -- SourceScan from duffle.scan_source
+-- @return Component[]
+local function project_components(source, scan)
+	local out = {}
+	for _, a in ipairs(scan.atoms) do
+		if a.kind == "comp_bare" or a.kind == "comp_proc" then
+			local args    = find_function_args_for(source, a.raw_name, a.ident_pos)
+			local comment = preceding_comment_block(source, a.ident_pos)
+			out[#out + 1] = {
+				line        = a.line,
+				name        = a.name,
+				body        = a.body,
+				body_tokens = a.body_tokens,
+				args        = args,
+				comment     = comment,
+				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
+			}
+		end
+	end
+	return out
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Line-comment → block-comment conversion
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Convert `//` line comments to `/* */` block comments in a token.
+-- C macros use `\` line-continuations; a `//` comment before `\` would consume the continuation,
+-- breaking the macro. We convert `//` to `/* */` so the multi-line macro structure is preserved.
+--
+-- Skips `//` sequences that are inside string or character literals
+-- (a rough heuristic — sufficient for component bodies which don't have those constructs).
+--
+--- @param s string
+--- @return string
+local function convert_line_comments_to_block(s)
+	local result = s
+	local pos    = 1
+	local len    = #result
+	while pos <= len do
+		local is_double_slash = result:byte(pos) == BYTE_SLASH
+			and pos + 1 <= len and result:byte(pos + 1) == BYTE_SLASH
+		if not is_double_slash then
+			pos = pos + 1
+		else
+			-- Find end of line.
+			local eol = pos
+			while eol <= len and result:byte(eol) ~= BYTE_NEWLINE do
+				eol = eol + 1
+			end
+			local before  = result:sub(1, pos - 1)
+			local comment = result:sub(pos + 2, eol - 1)  -- skip the `//`
+			local after
+			if eol <= len and result:byte(eol) == BYTE_NEWLINE then
+				after = " */" .. result:sub(eol)  -- keep the newline
+			else
+				after = " */"
+			end
+			result = before .. "/*" .. comment .. after
+			pos = #before + 2 + #comment + 3  -- skip past converted comment
+		end
+	end
+	return result
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Word-count computation (memoized recursive lookup)
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
+-- Returns the ident unchanged if it doesn't start with the prefix 
+-- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
+-- @param ident string|nil
+-- @return string|nil
+local function strip_mac_prefix(ident)
+	if not ident then return nil end
+	if ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+		return ident:sub(MAC_PREFIX_LEN + 1)
+	end
+	return ident
+end
+
+-- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
+-- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
+-- @param name string -- the component name (without `mac_`)
+-- @param comp_by_name table<string, Component>
+-- @param wc table<string, integer>
+-- @param cache table<string, integer>
+-- @return integer
+local function word_count_rec(name, comp_by_name, wc, cache)
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1  -- mark in-progress (cycle detection)
+	local cc = comp_by_name[name]
+	local n
+	if cc then
+		n = 0
+		local tokens = cc.body_tokens
+		for _, t in ipairs(tokens) do
+			local trimmed = t.tok
+			if trimmed ~= "" then
+				local lookup = strip_mac_prefix(duffle.read_ident(trimmed, 1))
+				if lookup and comp_by_name[lookup] then
+					-- It's a `mac_X(...)` call. Recurse.
+					n = n + word_count_rec(lookup, comp_by_name, wc, cache)
+				elseif lookup and wc and wc[lookup] then
+					-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2, nop2 = 2).
+					n = n + wc[lookup]
+				else
+					-- Unrecognized token. Fall back to 1 word.
+					n = n + 1
+				end
+			end
+		end
+	else
+		-- Not a known component: assume 1 word (regular instruction).
+		n = 1
+	end
+	cache[name] = n
+	return n
+end
+
+--- Compute word counts for every component in `components` in a single pass.
+--- The name-lookup table + memoization cache are built ONCE (per source) instead of per-component,
+--- so the cache survives across siblings and a component's recursive `mac_Y(...)` 
+--- references hit memoized values instead of re-walking the body.
+--- Cycle detection (A -> B -> A) is preserved via the in-progress `-1` sentinel in `cache`.
+--- @param components Component[]
+--- @param wc table<string, integer>
+--- @return table<string, integer>  -- map of component name (without `mac_`) -> word count
+local function count_all_components(components, wc)
+	local comp_by_name = {}
+	for _, cc in ipairs(components) do comp_by_name[cc.name] = cc end
+	local cache        = {}
+	local counts       = {}
+	for _, c in ipairs(components) do
+		counts[c.name] = word_count_rec(c.name, comp_by_name, wc, cache)
+	end
+	return counts
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-component emit logic
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Split a (possibly multi-line) comment into per-line entries.
+--- Hand-rolled (no regex patterns used).
+--- @param s string
+--- @return string[]
+local function split_comment_lines(s)
+	local out  = {}
+	local pos  = 1
+	local s_len = #s
+	while pos <= s_len do
+		local nl = s:find("\n", pos, true)
+		if not nl then
+			out[#out + 1] = s:sub(pos)
+			break
+		end
+		out[#out + 1] = s:sub(pos, nl - 1)
+		pos = nl + 1
+	end
+	return out
+end
+
+--- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
+--- @param args_str string|nil
+--- @return string
+local function signature_from_args(args_str)
+	local arg_names = extract_arg_names(args_str)
+	if    arg_names and #arg_names > 0 then
+		return table.concat(arg_names, ", ")
+	end
+	return "..."
+end
+
+--- Strip the trailing `" \"` (space + backslash) line continuation from the last body line.
+--- The last 2 chars are always that pair.
+local function strip_trailing_continuation(lines)
+	local last = lines[#lines]
+	if    last:sub(-2) == " \\" then
+		lines[#lines] = last:sub(1, -3)
+	end
+end
+
+--- Emit the `#define mac_X(sig) \<newline>\t<tok1> \<newline>,\t<tok2> ...` block.
+--- Converts `//` line comments to `/* */` block comments in each token so they don't break the C macro `\` line continuations.
+local function emit_macro_body(lines, c, sig, tokens)
+	for tok_idx = 1, #tokens do
+		tokens[tok_idx] = convert_line_comments_to_block(tokens[tok_idx])
+	end
+	lines[#lines + 1] = "#define mac_" .. c.name .. "(" .. sig .. ") \\"
+	lines[#lines + 1] = "\t" .. tokens[1] .. " \\"
+	for tok_idx = 2, #tokens do
+		lines[#lines + 1] = ",\t" .. tokens[tok_idx] .. " \\"
+	end
+	strip_trailing_continuation(lines)
+end
+
+--- Build the list of lines for one component
+--- (signature comment, `#define mac_X(...)` line with backslash-continued tokens, then `WORD_COUNT(mac_X, N)` entry).
+--- @param c Component
+--- @param components Component[]
+--- @param wc table<string, integer>
+--- @return string[]  -- list of lines for this component
+local function build_component_lines(c, counts)
+	local lines = {}
+
+	if c.comment and c.comment ~= "" then
+		for _, line in ipairs(split_comment_lines(c.comment)) do
+			lines[#lines + 1] = line
+		end
+	end
+
+	local tokens = duffle.split_top_level_commas(c.body)
+	for i = 1, #tokens do tokens[i] = duffle.trim(tokens[i]) end
+	local sig    = signature_from_args(c.args)
+	-- Direct lookup against the per-source precomputed `counts` table (built once by count_all_components).
+	local n      = counts[c.name]
+
+	if n > 0 then
+		emit_macro_body(lines, c, sig, tokens)
+	end
+
+	-- Emit the WORD_COUNT(mac_<X>, N) entry.
+	lines[#lines + 1] = "WORD_COUNT(mac_" .. c.name .. ", " .. n .. ")"
+	lines[#lines + 1] = ""
+
+	return lines
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-source emit logic
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
+-- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
+-- @param src SourceFile
+-- @return string[]
+local function header_boilerplate(src)
+	return {
+		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
+		-- The build does manual unity includes (the user controls include order), so the pragma is only active for IDE/tooling.
+		"#ifdef INTELLISENSE_DIRECTIVES",
+		"#pragma once",
+		"#endif",
+		"// Auto-generated by ps1_meta.lua — DO NOT EDIT",
+		"// Source: " .. duffle.to_absolute_path(src.path),
+		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
+		"",
+		-- Self-contained: define WORD_COUNT if not already defined.
+		-- We use the same definition here so the auto-generated entries below expand
+		-- to compile-time constants whether the metadata file is included first or not.
+		"#ifndef WORD_COUNT",
+		"#define WORD_COUNT(name, count)  enum { words_##name = (count) };",
+		"#endif",
+		"",
+	}
+end
+
+-- Compute the output path for one source's `.macs.h` file.
+-- The pre-rework convention uses the *directory* basename
+-- (not the source file basename) e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+-- This matches what the C codebase #includes.
+-- @param src SourceFile
+-- @return string  -- the output directory
+-- @return string  -- the full output path
+local function compute_macs_h_path(src)
+	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
+	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+	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 preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
+--- Honors `ctx.dry_run`: prints the intended path but does not write the file.
+--- @param ctx PassCtx
+--- @param src SourceFile
+--- @param components Component[]
+--- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @return string|nil  -- path to the written file (nil if no components)
+local function emit_component_macros_h(ctx, src, components, counts)
+	if #components == 0 then return nil end
+	local out_dir, out_path = compute_macs_h_path(src)
+	local lines             = header_boilerplate(src)
+
+	for _, c in ipairs(components) do
+		for _, l in ipairs(build_component_lines(c, counts)) do
+			lines[#lines + 1] = l
+		end
+	end
+
+	local content = table.concat(lines, "\n") .. "\n"
+	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))
+	return out_path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Pass entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- (internal) Extend `ctx.shared.word_counts` with this source's component macros so offsets sees them without re-reading the file.
+-- @param ctx PassCtx
+-- @param components Component[]
+-- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+local function update_shared_word_counts(ctx, components, counts)
+	local wc = ctx.shared.word_counts
+	for _, c in ipairs(components) do
+		wc["mac_" .. c.name] = counts[c.name]
+	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) Extend `ctx.shared.components` with this source's components-by-name map so downstream passes
+--- (atoms_source_map, dwarf_injection) can resolve `mac_X(...)` invocations back to their component definition file:line.
+--- provenance emission uses this to attribute each emitted `.word` to either a component macro or the enclosing atom body.
+-- @param ctx PassCtx
+-- @param src SourceFile
+-- @param components Component[]
+local function update_shared_components(ctx, src, components)
+	ctx.shared.components         = ctx.shared.components or {}
+	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 strips the `mac_` prefix from the call site identifier before lookup.
+		ctx.shared.components[c.name] = {
+			name = c.name,
+			line = c.line,
+			path = rel_path,
+			kind = c.kind or "comp_bare",
+		}
+	end
+end
+
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	-- Initialize shared component map.
+	-- The atoms_source_map and dwarf_injection passes consume `ctx.shared.components` to resolve `mac_X(...)` 
+	-- invocations back to the component's definition file:line.
+	ctx.shared.components = ctx.shared.components or {}
+
+	for _, src in ipairs(ctx.sources) 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).
+			local counts = count_all_components(components, ctx.shared.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)
+				-- share component definitions with downstream passes.
+				-- `mac_X(...)` invocations in atom bodies resolve back to (path, line) via this map.
+				update_shared_components(ctx, src, components)
+			end
+		end
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/offsets.lua

@@ -0,0 +1,361 @@
+--- passes/offsets.lua — Branch-offset generator.
+---
+--- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
+--- for `MipsAtom_(name)` and `MipsCode code_<name>` declarations, computes the word offset
+--- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
+--- `<dir_basename>.offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+---
+--- The offset is `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding: branch_offset = relative_pc_in_words - 1).
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
+-- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works
+-- both standalone and when require'd from the orchestrator.
+-- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+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_"
+
+-- Column width for the `#define _atom_offset_F_T = N` alignment.
+local OFFSET_MACRO_COL = 44
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+--- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+
+--- @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 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
+--- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+
+--- @class BranchOffset
+--- @field tag    string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field target string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field pos    integer  -- the branch's word position within the atom body
+--- @field offset integer  -- computed `target_word - branch_word - 1`
+
+--- @class AtomData
+--- @field name        string         -- atom name
+--- @field total_words integer        -- total word count of the atom body
+--- @field offsets     BranchOffset[] -- per-branch offset list
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-token marker-call helpers (atom_label / atom_offset inside bodies)
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Extract comma-separated identifier args from a parenthesized group after a function-like macro call.
+-- Returns (args, after_paren) where `after_paren` is the position just past the closing `)`, or nil if `token` did not start with `(`.
+-- @param token string
+-- @param after_ident integer
+-- @return string[], integer|nil
+local function extract_ident_args(token, after_ident)
+	local arg_start = duffle.skip_ws_and_cmt(token, after_ident)
+	if token:sub(arg_start, arg_start) ~= "(" then return {}, nil end
+	local inner, after_paren = duffle.read_parens(token, arg_start)
+	-- scan: <marker>(<args>)
+
+	local args      = {}
+	local pos       = 1
+	local inner_len = #inner
+	while pos <= inner_len do
+		pos = duffle.skip_ws_and_cmt(inner, pos)
+		if pos > inner_len then break end
+		local ident, after = duffle.read_ident(inner, pos)
+		if ident and ident ~= "" then
+			table.insert(args, ident)
+			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 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
+
+-- MARKER_TO_HANDLER — data-driven marker dispatch (the plex pattern).
+-- Maps the marker ident to its recorder function. Each handler takes (out_table, args, at_pos).
+-- Adding a new marker type = 1 row + 1 recorder function.
+local MARKER_TO_HANDLER = {
+	[LABEL_MARKER]  = record_label_marker,
+	[OFFSET_MARKER] = record_offset_marker,
+}
+
+--- 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)
+			local handler = MARKER_TO_HANDLER[ident]
+			if handler then
+				local args, after_paren = extract_ident_args(token, after)
+				-- Marker found — dispatch to its recorder. markers share labels and branches as
+				-- out-tables; the recorder picks which one(s) to write to based on its semantics.
+				-- (record_label_marker writes to labels; record_offset_marker writes to branches.)
+				handler(ident == LABEL_MARKER and labels or 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
+
+--- 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 duffle.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 + duffle.count_marker_rest(tok, word_counts, count_token_words)
+		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` (the standard MIPS branch-immediate encoding).
+-- @param labels table<string, integer>
+-- @param branches table[]
+-- @return BranchOffset[]
+local function compute_offsets(labels, branches)
+	local results = {}
+	for _, br in ipairs(branches) do
+		local target = labels[br.target]
+		if not target then
+			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.pos .. ")")
+		end
+		results[#results + 1] = { target = br.target, tag = br.tag, offset = target - br.pos - 1 }
+	end
+	return results
+end
+
+-- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
+-- @param s string
+-- @param w integer
+-- @return string
+local function pad_right(s, w)
+	return s .. string.rep(" ", math.max(0, w - #s))
+end
+
+-- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
+-- @param r BranchOffset
+-- @return table
+local function make_offset_const(r)
+	return {
+		macro_name = OFFSET_MACRO_PREFIX .. r.tag .. "_" .. r.target,
+		enum_name  = OFFSET_ENUM_PREFIX  .. r.tag .. "_" .. r.target,
+		value      = r.offset,
+	}
+end
+
+-- (internal) Emit one atom's offset constants + enum into the lines buffer.
+-- @param add fun(s: string)
+-- @param atom AtomData
+local function emit_atom_offsets(add, atom)
+	if #atom.offsets == 0 then return end
+	add("// --- atom: " .. atom.name .. " (" .. atom.total_words .. " words) ---")
+	add("")
+	local consts = {}
+	for _, r in ipairs(atom.offsets) do
+		consts[#consts + 1] = make_offset_const(r)
+	end
+	for _, c in ipairs(consts) do
+		add("#define " .. pad_right(c.macro_name, OFFSET_MACRO_COL) .. "  " .. c.value)
+	end
+	add("")
+	add("enum {")
+	for _, c in ipairs(consts) do
+		add("    " .. c.enum_name .. " = " .. c.macro_name .. ",")
+	end
+	add("};")
+	add("")
+end
+
+-- Generate the per-source .offsets.h header.
+-- @param source_path string
+-- @param atoms_data AtomData[]
+-- @return string
+local function generate_header(source_path, atoms_data)
+	local basename = duffle.basename_no_ext(source_path)
+
+	local lines = {}
+	local function add(s) lines[#lines + 1] = s end
+
+	add("// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT")
+	add("// Source: " .. source_path)
+	add("#pragma once")
+	add("")
+	add("#pragma region " .. basename)
+	add("")
+	add("")
+	for _, atom in ipairs(atoms_data) do
+		emit_atom_offsets(add, atom)
+	end
+	add("#pragma endregion " .. basename)
+	add("")
+	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.
+-- MipsAtom_ entries have kind="atom"; MipsCode code_<name> entries have kind="raw_atom".
+-- `body_tokens` is set by scan-source on every `scan.atoms[i]` / `scan.raw_atoms[i]`; we carry it forward
+-- so `scan_atom_body` reads from the precomputed table directly (no per-atom tokenize_body fallback).
+-- @param scan table  -- SourceScan from duffle.scan_source
+-- @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 labels, branches, total = scan_atom_body(atom.body_tokens, ctx.shared.word_counts)
+		atoms_data[#atoms_data + 1] = {
+			name        = atom.name,
+			total_words = total,
+			offsets     = compute_offsets(labels, branches),
+		}
+	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.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 every `atom_offset(F, T)` reference in the source's atoms.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	for _, src in ipairs(ctx.sources) do
+		local out_path = process_source(ctx, src)
+		if out_path then
+			outputs[#outputs + 1] = { offsets_h = out_path }
+		end
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/report.lua

@@ -0,0 +1,475 @@
+--- passes/report.lua — Per-MODULE annotation report renderer +
+--- project-wide summary writer.
+---
+--- Two output files per build:
+---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
+---   - `build/gen/annotation_validation.txt` — the project summary.
+---
+--- The annotation pass stashes per-MODULE summary entries in `ctx.flags._annot_results` (set by `passes/annotation.lua`). 
+--- This pass re-validates each source via `annotation.validate()` to get the detailed per-source results needed for the report. 
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
+-- Bootstrap: see `ps1_meta.lua` for the rationale.
+-- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works
+-- both standalone and when require'd from the orchestrator.
+-- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Section separators used in the rendered text reports.
+-- The thin rules are hand-tuned to align with the per-section content width; do not change without also checking the section renderers below.
+local RULE_THICK = "========================================================"
+local SECTION_HEADER_ATOMS     = "── Atoms ────────────────────────────────────────────────"
+local SECTION_HEADER_ANNOTS    = "── Annotations ──────────────────────────────────────────"
+local SECTION_HEADER_BINDS     = "── Binds_* structs ──────────────────────────────────────"
+local SECTION_HEADER_MACROS    = "── Macro word-count declarations ─────────────────────────"
+local SECTION_HEADER_ERRORS    = "── Errors ──────────────────────────────────────────────"
+local SECTION_HEADER_WARNINGS  = "── Warnings ────────────────────────────────────────────"
+
+-- Lua pattern that captures the basename (last path segment) of a
+-- forward- or back-slash separated path.
+local BASENAME_PATTERN = "([^/\\]+)$"
+
+-- Debug flag name — set to truthy in `_G` to enable verbose logging.
+local DEBUG_FLAG = "_DEBUG_REPORT"
+
+-- Pass identifier for log messages.
+local PASS_NAME = "report"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+
+--- @class PassCtx
+--- @field sources            SourceFile[]         -- all source files in the build
+--- @field metadata_path      string               -- path to word_count.metadata.h
+--- @field shared             table                -- cross-pass shared state
+--- @field out_root           string               -- output root (e.g. "build/gen")
+--- @field project_root       string               -- project root (e.g. "code/")
+--- @field upstream           table<string, table> -- per-pass upstream outputs
+--- @field flags              table                -- CLI flags + per-pass stash
+--- @field flags._annot_results ModuleEntry[]      -- stashed by annotation pass
+--- @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
+--- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+
+-- Shapes produced by `passes/annotation.lua`'s `M.validate()`.
+
+--- @class AtomEntry
+--- @field name string   -- atom name (e.g. "cube_g4_face")
+--- @field line integer  -- source line of the atom declaration
+
+--- @class AnnotEntry
+--- @field line    integer  -- source line
+--- @field macro   string   -- the macro name (e.g. "atom_reads")
+--- @field name    string   -- the atom name (if a `name(...)` was given)
+--- @field kind    string   -- "atom_info" | "atom_bind" | ...
+--- @field binds   string|nil  -- Binds_X name if any
+--- @field reads   string[]    -- R_* names (read targets)
+--- @field writes  string[]    -- R_* names (write targets)
+--- @field error   string|nil  -- error message if annotation was malformed
+
+--- @class BindsField
+--- @field name   string   -- field name
+--- @field offset integer  -- byte offset within the Binds_X struct
+
+--- @class BindsStruct
+--- @field name   string         -- struct name (e.g. "Binds_Floor")
+--- @field line   integer        -- source line of the typedef
+--- @field bytes  integer        -- total byte size
+--- @field fields BindsField[]   -- the field list
+
+--- @class MacroEntry
+--- @field name  string   -- macro name (e.g. "WORD_COUNT(my_macro, 4)")
+--- @field line  integer  -- source line
+--- @field words integer  -- declared word count
+
+--- @class Finding
+--- @field line integer  -- source line
+--- @field msg  string   -- finding message
+
+--- @class AnnotationResult
+--- @field source   string         -- set by this pass; original source path
+--- @field atoms    AtomEntry[]    -- atom declarations in this source
+--- @field annots   AnnotEntry[]   -- annotation entries
+--- @field macros   MacroEntry[]   -- macro word-count declarations
+--- @field binds    BindsStruct[]  -- Binds_* struct declarations
+--- @field errors   Finding[]      -- errors from validation
+--- @field warnings Finding[]      -- warnings from validation
+--- @field info     table          -- info summary (not rendered here)
+
+--- @class ModuleEntry
+--- @field dir          string   -- absolute directory path
+--- @field dir_basename string   -- basename (e.g. "duffle", "gte_hello")
+--- @field atoms_count  integer  -- pre-counted atoms for filtering
+
+--- @class ModuleReport
+--- @field dir       string             -- module directory
+--- @field sources   SourceFile[]       -- sources in this module
+--- @field results   AnnotationResult[] -- per-source validate() results
+
+--- @class ProjectReport
+--- @field results AnnotationResult[]  -- all per-source results
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-MODULE annotation report (aggregated across all sources in a dir)
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
+-- @param path string
+-- @return string
+local function source_basename(path)
+	return path:match(BASENAME_PATTERN) or path
+end
+
+-- (internal) Format a single annotation entry as one rendered line.
+-- @param a AnnotEntry
+-- @param src_name 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.reads > 0          then line = line .. "  reads={" .. table.concat(a.reads, ",") .. "}" 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.
+-- @param results AnnotationResult[]
+-- @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
+	for _, r in ipairs(results) do
+		total_atoms    = total_atoms    + #r.atoms
+		total_annots   = total_annots   + #r.annots
+		total_binds    = total_binds    + #r.binds
+		total_macros   = total_macros   + #r.macros
+		total_errors   = total_errors   + #r.errors
+		total_warnings = total_warnings + #r.warnings
+	end
+	return total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings
+end
+
+-- (internal) Section renderer: per-source atom declarations.
+local function render_module_atoms_section(add, results)
+	add(SECTION_HEADER_ATOMS)
+	for _, r in ipairs(results) do
+		local src_name = source_basename(r.source)
+		for _, a in ipairs(r.atoms) do
+			add(string.format("  MipsAtom_(%s)   line %d   [%s]", a.name, a.line, src_name))
+		end
+	end
+	add("")
+end
+
+-- (internal) Section renderer: per-source annotation entries.
+local function render_module_annots_section(add, results)
+	add(SECTION_HEADER_ANNOTS)
+	for _, r in ipairs(results) do
+		local src_name = source_basename(r.source)
+		for _, a in ipairs(r.annots) do
+			add(format_annot_line(a, src_name))
+		end
+	end
+	add("")
+end
+
+-- (internal) Section renderer: per-source Binds_* struct declarations.
+local function render_module_binds_section(add, results)
+	add(SECTION_HEADER_BINDS)
+	for _, r in ipairs(results) do
+		local src_name = source_basename(r.source)
+		for _, b in ipairs(r.binds) do
+			add(string.format("  %s   line %d   %d bytes   [%s]", b.name, b.line, b.bytes, src_name))
+			for _, f in ipairs(b.fields) do
+				add(string.format("    +%2d: %s", f.offset, f.name))
+			end
+		end
+	end
+	add("")
+end
+
+-- (internal) Section renderer: per-source macro word-count declarations.
+local function render_module_macros_section(add, results)
+	add(SECTION_HEADER_MACROS)
+	for _, r in ipairs(results) do
+		local src_name = source_basename(r.source)
+		for _, m in ipairs(r.macros) do
+			add(string.format("  %s  line %d  words=%d  [%s]", m.name, m.line, m.words, src_name))
+		end
+	end
+	add("")
+end
+
+-- (internal) Section renderer: per-source errors (one-line + "(none)" if empty).
+local function render_module_errors_section(add, results, total_errors)
+	add(SECTION_HEADER_ERRORS)
+	if total_errors == 0 then
+		add("  (none)")
+	else
+		for _, r in ipairs(results) do
+			local src_name = source_basename(r.source)
+			for _, e in ipairs(r.errors) do
+				add(string.format("  ✗ line %d  %s  [%s]", e.line, e.msg, src_name))
+			end
+		end
+	end
+	add("")
+end
+
+-- (internal) Section renderer: per-source warnings (one-line + "(none)" if empty).
+local function render_module_warnings_section(add, results, total_warnings)
+	add(SECTION_HEADER_WARNINGS)
+	if total_warnings == 0 then
+		add("  (none)")
+	else
+		for _, r in ipairs(results) do
+			local src_name = source_basename(r.source)
+			for _, w in ipairs(r.warnings) do
+				add(string.format("  ⚠ line %d  %s  [%s]", w.line, w.msg, src_name))
+			end
+		end
+	end
+	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
+--- @param results AnnotationResult[]  -- per-source validate() results
+--- @return string                     -- the rendered report text
+local function render_module_report(dir, sources, results)
+	local lines = {}
+	local function add(s) lines[#lines + 1] = s end
+
+	add(RULE_THICK)
+	add("ANNOTATION PASS — module " .. source_basename(dir))
+	add(RULE_THICK)
+	add(string.format("Sources: %d", #sources))
+	for _, s in ipairs(sources) do add("  " .. s.path) end
+	add("")
+
+	local total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings = tally_module_totals(results)
+	add(string.format("Atoms: %d   Annotations: %d   Binds structs: %d   Macro decls: %d",
+		total_atoms, total_annots, total_binds, total_macros))
+	add("")
+
+	-- Bundle the totals so the section renderers don't need separate parameter lists.
+	-- Errors/warnings sections need their total count to decide "(none)" vs iterate.
+	-- Sections without totals (atoms/annots/binds/macros) ignore this arg.
+	local totals = {
+		atoms    = total_atoms,  annots = total_annots,  binds  = total_binds,
+		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
+		add(section.header)
+		section.render(add, results, totals)
+		add("")
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-project summary
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render the per-project summary (`build/gen/annotation_validation.txt`).
+--- Aggregates totals across all sources; lists per-source error counts if any source has errors.
+--- @param all_results AnnotationResult[]
+--- @return string
+local function render_project_report(all_results)
+	local lines = {}
+	local function add(s) lines[#lines + 1] = s end
+
+	local total_atoms, total_annots, total_macros, total_binds = 0, 0, 0, 0
+	local total_errors, total_warnings = 0, 0
+	for _, r in ipairs(all_results) do
+		total_atoms    = total_atoms    + #r.atoms
+		total_annots   = total_annots   + #r.annots
+		total_macros   = total_macros   + #r.macros
+		total_binds    = total_binds    + #r.binds
+		total_errors   = total_errors   + #r.errors
+		total_warnings = total_warnings + #r.warnings
+	end
+
+	add(RULE_THICK)
+	add("ANNOTATION VALIDATION — project summary")
+	add(RULE_THICK)
+	add("")
+	add(string.format("Atoms:        %d", total_atoms))
+	add(string.format("Annotations:  %d", total_annots))
+	add(string.format("Macros:       %d", total_macros))
+	add(string.format("Binds:        %d", total_binds))
+	add("")
+	add(string.format("Errors:       %d", total_errors))
+	add(string.format("Warnings:     %d", total_warnings))
+	add("")
+
+	if total_errors > 0 then
+		add("Per-source error counts:")
+		for _, r in ipairs(all_results) do
+			if #r.errors > 0 then
+				local src_name = source_basename(r.source)
+				add(string.format("  %s : %d error(s)", src_name, #r.errors))
+			end
+		end
+		add("")
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Orchestration helpers
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- (internal) Pull per-source validate() results from the annotation pass's stash.
+-- 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.
+-- 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 result then
+			result.source = src.path                                    -- defensive (annotation tags it too; this guards against cache misses from earlier iterations)
+			module_results[#module_results + 1] = result
+			all_results[#all_results + 1] = result
+		end
+	end
+	return module_results, all_results
+end
+
+-- (internal) Does this module's results contain anything worth emitting?
+-- @param module_results AnnotationResult[]
+-- @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
+		or #r.macros > 0 or #r.errors > 0 or #r.warnings > 0 then
+			return true
+		end
+	end
+	return false
+end
+
+-- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
+-- @param fmt string
+local function debug_log(fmt, ...)
+	if _G[DEBUG_FLAG] then
+		io.stderr:write(string.format("[%s] " .. fmt, PASS_NAME, ...))
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+--- Run the report pass. 
+--- Renders one `<dir_basename>.annotations.txt` per source-directory that has content, plus the project-wide `annotation_validation.txt` summary.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local module_entries = (ctx.flags and ctx.flags._annot_results) or {}
+	local by_dir         = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+
+	if not ctx.dry_run then duffle.ensure_dir(ctx.out_root) end
+
+	local all_results_for_summary = {}
+	for _, entry in ipairs(module_entries) 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 {}))
+
+		if entry.atoms_count > 0 or #(by_dir[entry.dir] or {}) > 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"
+				if not ctx.dry_run then
+					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")
+			end
+		end
+	end
+
+	if not ctx.dry_run and #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 }
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/word_count_eval.lua

@@ -0,0 +1,175 @@
+--- word_count_eval.lua — Word-counting logic for the tape-atom metaprogram pipeline.
+---
+--- Three responsibilities:
+---   1. **Public utilities** (used by `passes/components.lua`, `passes/offsets.lua`, `passes/annotation.lua`):
+---        - `M.count_token_words(token, wc)` — words emitted by one token
+---        - `M.scan_dir(dir, suffix)` — glob walk for *.macs.h
+---   2. **Pass entry** `M.run(ctx)` — loads metadata.h + *.macs.h into `ctx.shared.word_counts` for downstream passes.
+---   3. **Internal helpers** for the body scanner.
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
+-- Bootstrap: see `ps1_meta.lua` for the rationale.
+-- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
+-- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Required native extension: lfs (LuaFileSystem). Built by `update_deps.ps1` to
+-- `toolchain/lfs/lfs.dll` and wired into package.cpath by `scripts/duffle_paths.lua`.
+-- If lfs is missing, `require` throws — fail loud per the build-tool convention.
+local lfs = require("lfs")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class WordCounts
+--- @field [string] integer  -- macro name -> word count
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+
+--- @class PassCtx
+--- @field sources            SourceFile[]  -- all source files in the build
+--- @field metadata_path      string        -- path to word_count.metadata.h
+--- @field shared             table         -- cross-pass shared state
+--- @field shared.word_counts WordCounts    -- 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
+--- @field outputs  table[]  -- {kind=, path=} entries describing emit files
+--- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- ┌────────────────────────────────────────────────────────────────────┐
+-- │ Shared utility: count_token_words                                  │
+-- └────────────────────────────────────────────────────────────────────┘
+
+--- Count words emitted by a single comma-separated token inside an atom body.
+--- For most tokens (regular MIPS instructions) this returns 1.
+--- For `mac_X(...)` calls, this returns the resolved word count from `wc` (recursively if needed). For `nop2` etc., returns wc[name].
+--- For unknown macros, returns 1 and (optionally) warns.
+--- @param token string        -- a single token from split_top_level_commas
+--- @param wc WordCounts       -- the shared word-count table
+--- @return integer
+function M.count_token_words(token, wc)
+	local s = duffle.trim(token)
+	if    s == "" then return 0 end
+	local name, after = duffle.read_ident(s, 1)
+	if not name then return 1 end
+	if wc[name] then return wc[name] end
+	local paren_pos = duffle.skip_ws_and_cmt(s, after)
+	if s:sub(paren_pos, paren_pos) == "(" then
+		io.stderr:write("  warning: unknown macro '" .. name .. "', assuming 1 word\n")
+	end
+	return 1
+end
+
+-- ┌────────────────────────────────────────────────────────────────────┐
+-- │ Shared utility: scan_dir                                           │
+-- └────────────────────────────────────────────────────────────────────┘
+
+-- Cache the scan_dir result per (dir, suffix) in package.loaded.
+-- 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`).
+--- Native directory enumeration via lfs (~2ms). Zero subprocess spawns.
+--- @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 = {}
+	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
+
+	-- 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.
+--- Loading the .macs.h files is idempotent: entries from later (current-build) .macs.h files override metadata.h entries of the same name.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local wc = {}
+
+	-- 1. Load metadata.h (the encoding-macro source of truth).
+	local meta_counts = duffle.load_word_counts(ctx.metadata_path)
+	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
+	return { outputs = {}, errors = {}, warnings = {} }
+end
+
+return M

scripts/pcsx_debug_helper/autoexec.lua

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

scripts/ps1_meta.lua

@@ -0,0 +1,962 @@
+--- ps1_meta.lua — Orchestrator entry point for the tape-atom metaprogram.
+---
+--- Dispatches to pass modules under `scripts/passes/`, resolving dependencies topologically (Kahn's algorithm + cycle detection).
+---
+--- **Architecture**:
+---   - **PASSES table** — declarative dep graph (data, not code).
+---   - **FLAG_HANDLERS table** — per-flag CLI dispatchers (handler-map pattern; replaces an 8-way if/elseif chain).
+---   - **parse_args** → **build_ctx** (just opens + reads source files; no inline scanning) → **topo_sort** → **dispatch_passes**.
+---   - The first pass in the dep graph is `scan-source` (see `passes/scan_source.lua`).
+---     It calls `duffle.scan_source` once per source to produce the fat `SourceScan` payload, which is attached to each `src.scan`. 
+---     Every other pass that reads source structure depends on `scan-source` and consumes `src.scan` as a read-only. 
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+---
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Bootstrap: load `duffle_paths.lua` via this script's own path.
+-- Use `arg[0]` when this file is the entry script (`arg[0]` ends in
+-- "ps1_meta.lua"); fall back to `debug.getinfo(1, "S").source` when this
+-- file is being dofile()'d or require()'d (in which case `arg[0]` is the
+-- *caller's* path, not ours).
+--
+-- That single statement: (a) sets `package.path` + `package.cpath`
+-- (via cached `git rev-parse`), (b) at the bottom returns `require("duffle")`.
+-- So the dofile's return value is the duffle module.
+local _is_entry_script = arg and arg[0] and arg[0]:match("ps1_meta%.lua$") ~= nil
+local _bootstrap_src
+if _is_entry_script then
+	_bootstrap_src = arg[0]
+else
+	-- debug.getinfo(1, "S").source returns "@<path>" for the current chunk;
+	-- strip the leading "@" so the directory match works in both cases.
+	_bootstrap_src = debug.getinfo(1, "S").source:sub(2)
+end
+local duffle = dofile((_bootstrap_src:match("(.*[/\\])") or "./") .. "duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Exit codes (per the --help text and the post-build summary convention).
+local EXIT_OK                  = 0
+local EXIT_VALIDATION_ERRORS   = 1
+local EXIT_INTERNAL_ERROR      = 2
+
+-- Default --out-root value if not provided.
+local DEFAULT_OUT_ROOT         = "build/gen"
+
+-- Sentinel for "all passes" in `PASS_FLAG_TO_NAME`. Distinguishes `--all` from the per-pass flags (which map to individual pass names).
+local ALL_PASSES_SENTINEL      = "__all__"
+
+-- Sentinel key for the pass-flag dispatcher in `FLAG_HANDLERS`. 
+-- The actual pass names are looked up via `PASS_FLAG_TO_NAME`, not direct dispatch, so this key never matches a real flag.
+local PASS_FLAG_DISPATCH_KEY   = "__pass__"
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class PassDescriptor
+--- @field module string       -- module name passed to require()
+--- @field kind   string       -- "shared" | "header-output" | "validation" | "report"
+--- @field deps   string[]     -- names of upstream passes
+--- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of
+---                             -- (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
+--- @field desc   string       -- human description (used by --help + ASCII graph)
+--- @field out    PassOutput[] -- output paths (used by --dry-run + report)
+
+--- @class PassOutput
+--- @field kind          string  -- "header" | "report"
+--- @field path_template string  -- e.g. "<source_dir>/gen/<basename>.macs.h"
+
+--- @class SourceFile
+--- @field path     string  -- absolute path to the source file
+--- @field text     string  -- the full source text
+--- @field dir      string  -- the directory containing the source
+--- @field basename string  -- filename without extension
+
+--- @class PassCtx
+--- @field sources            SourceFile[]           -- all source files in the build
+--- @field metadata_path      string                 -- path to word_count.metadata.h
+--- @field shared             table                  -- cross-pass shared state
+--- @field shared.word_counts table<string, integer> -- populated by word-counts 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 output accumulator
+--- @field flags              table                  -- CLI flags + per-pass stash
+--- @field dry_run            boolean                -- if true, compute but don't write
+--- @field verbose            boolean                -- if true, log diagnostic info
+
+--- @class PassOutputEntry
+--- @field [string] string  -- dynamic shape; key is the output kind
+	-- (e.g. "macs_h", "offsets_h", "errors_h", "annotations_txt", "static_analysis_txt", "summary_txt"), value is the path
+
+--- @class Finding
+--- @field line integer  -- source line (or 0 for pass-level)
+--- @field msg  string   -- finding message
+
+--- @class PassResult
+--- @field outputs  PassOutputEntry[] -- emitted file paths
+--- @field errors   Finding[]         -- build-stops (per-pass kind policy)
+--- @field warnings Finding[]         -- informational
+
+--- @class ParsedArgs
+--- @field requested_set string[]  -- pass names to run (explicit --all expanded)
+--- @field sources       string[]  -- --source values
+--- @field metadata      string    -- --metadata value
+--- @field out_root      string    -- --out-root value (default "build/gen")
+--- @field project_root  string    -- --project-root value (default dirname(metadata))
+--- @field dry_run       boolean   -- if true, compute but don't write
+--- @field verbose       boolean   -- if true, log diagnostic info
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- PASSES table (data, not code) — the orchestrator's dep graph
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Build-phase groups: each PASSES row may declare membership in one or more
+-- named groups via `groups = { ... }`. The CLI flags --pre-link and
+-- --post-link request the *roots* of their group; topo_sort then closes
+-- transitive dependencies from those roots, and dispatch_passes runs every
+-- pass in the resulting closure without phase-filtering.
+--
+-- A row without a `groups` entry is dependency-only: it runs only when a
+-- transitive dep requests it, but it remains directly requestable through
+-- its explicit CLI flag (e.g. --atoms-source-map, --scan-source).
+
+local PASSES = {
+	["scan-source"] = {
+		module = "passes.scan_source",
+		kind   = "shared", deps = {},
+		desc   = "Walk each source once; produce the fat SourceScan payload for downstream passes",
+		out    = {},
+	},
+	["word-counts"] = {
+		module = "passes.word_count_eval",
+		kind   = "shared", deps = {},
+		desc   = "Build the shared metadata table (metadata.h + .macs.h)",
+		out    = {},
+	},
+	components = {
+		module = "passes.components",
+		kind   = "header-output",
+		deps   = {"scan-source", "word-counts"},
+		desc   = "Emit mac_X macros from MipsAtomComp_ declarations",
+		out    = { { kind = "header", path_template = "<source_dir>/gen/<basename>.macs.h" } },
+	},
+	annotation = {
+		module = "passes.annotation",
+		kind   = "validation",
+		deps   = {"scan-source", "word-counts"},
+		desc   = "Validate atom DSL usage; emit errors.h + annotations.txt",
+		out    = {
+			{ kind = "report", path_template = "<out_root>/<basename>.errors.h" },
+			{ kind = "report", path_template = "<out_root>/<basename>.annotations.txt" },
+		},
+	},
+	offsets = {
+		module = "passes.offsets",
+		kind   = "header-output",
+		deps   = {"scan-source", "word-counts", "components"},
+		groups = { "pre-link" },
+		desc   = "Compute branch offsets for atom_label / atom_offset",
+		out    = { { kind = "header", path_template = "<source_dir>/gen/<basename>.offsets.h" } },
+	},
+	["static-analysis"] = {
+		module = "passes.static_analysis",
+		kind   = "validation",
+		deps   = {"scan-source", "word-counts", "components"},
+		desc   = "Static analysis: GTE pipeline-fill, mac_yield uniformity, ABI handoff, GPU port-store shape, per-atom cycle budget, type consistency",
+		out    = { { kind = "report", path_template = "<out_root>/<basename>.static_analysis.txt" } },
+	},
+	["atoms-source-map"] = {
+		module = "passes.atoms_source_map",
+		kind   = "header-output",
+		deps   = {"word-counts", "components"},
+		desc   = "Emit gen/<basename>.atoms.sourcemap.txt (per-.word C source line map for gdb debugging) AND gen/<basename>.atoms.provenance.txt (per-.word provenance; each word tagged with its call-site file:line and, when emitted by a mac_X(...) component invocation, the component's definition file:line). Consumed by passes/dwarf_injection.lua to synthesize DW_TAG_inlined_subroutine instances for source-level Step Into on component invocations.",
+		out    = {
+			{ kind = "report", path_template = "<out_root>/<basename>.atoms.sourcemap.txt"   },
+			{ kind = "report", path_template = "<out_root>/<basename>.atoms.provenance.txt" },
+		},
+	},
+	["dwarf-injection"] = {
+		module = "passes.dwarf_injection",
+		kind   = "shared",
+		deps   = {"scan-source", "atoms-source-map"},
+		groups = { "post-link" },
+		desc   = "Inject per-atom .debug_line + .debug_aranges (F') + per-atom .debug_info subprogram + per-wave-context-reg .debug_info variables (G') into the ELF (post-link; writes 7 section .bin blobs plus one deterministic .gdbinit sidecar). (rbind composite) reads ctx.sources[i].scan to find atom_bind(Binds_X) atoms + their Binds_X struct fields; emits per-Binds_X DW_TAG_structure_type DIEs + per-rbind-atom DW_TAG_variable 'bind_args' DIEs with piece-chain DW_OP_bregN/DW_OP_piece location expressions.",
+		out    = {
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_line.bin"     },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_aranges.bin"  },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_rnglists.bin" },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_abbrev.bin"   },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_info.bin"     },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_str.bin"      },
+			{ kind = "report", path_template = "<out_root>/<basename>.dwarf_loc.bin"      },
+			{ kind = "report", path_template = "<out_root>/<basename>.gdbinit"            },
+		},
+	},
+	report = {
+		module = "passes.report",
+		kind   = "report",
+		deps   = {"annotation", "static-analysis"},
+		groups = { "pre-link" },
+		desc   = "Render the per-project summary",
+		out    = { { kind = "report", path_template = "<out_root>/annotation_validation.txt" } },
+	},
+}
+
+-- ────────────────────────────────────────────────────────────────────────────
+-- Phase-root selection: derive the sorted set of roots belonging to a named
+-- build-phase group, then append them to `args.requested_set`. topo_sort
+-- closes the transitive deps from there; dispatch_passes runs every resolved
+-- pass without phase-filtering.
+-- ────────────────────────────────────────────────────────────────────────────
+
+--- @param group_name string  -- the build-phase group ("pre-link" | "post-link")
+--- @return string[]          -- sorted root pass names belonging to that group
+local function roots_for_group(group_name)
+	local names = {}
+	for name, pass in pairs(PASSES) do
+		if pass.groups then
+			for _, g in ipairs(pass.groups) do
+				if g == group_name then
+					names[#names + 1] = name
+					break
+				end
+			end
+		end
+	end
+	table.sort(names)
+	return names
+end
+
+--- Append every root belonging to `group_name` to `args.requested_set`.
+--- Errors loudly if no PASSES row declares the group, so a typo'd or
+--- future-removed group name cannot silently fall through to pre-link
+--- (or any other default) and dispatch nothing.
+--- @param args ParsedArgs
+--- @param group_name string
+local function request_roots_for_group(args, group_name)
+	local roots = roots_for_group(group_name)
+	if #roots == 0 then
+		error(string.format(
+			"ps1_meta: build-phase group %q has zero roots in PASSES; "
+			.. "check PASSES rows for a `groups = { %q }` field",
+			group_name, group_name))
+	end
+	for _, name in ipairs(roots) do
+		args.requested_set[#args.requested_set + 1] = name
+	end
+end
+
+-- Pass-kind taxonomy: which kinds stop the build on errors?
+local PASS_KIND_STOP_ON_ERROR = {
+	["shared"]        = false,
+	["header-output"] = true,
+	["validation"]    = true,
+	["report"]        = false,
+}
+
+-- Closed set of CLI flags -> pass names.
+-- Per-pass flags (e.g. --word-counts) live here; phase flags (--pre-link,
+-- --post-link, --all) live in FLAG_HANDLERS because they own side effects
+-- or invoke group-derivation logic. --dwarf-injection is *also* a per-pass
+-- opt-in flag, but its selection + opt-in state are both owned by the
+-- explicit FLAG_HANDLERS entry below (it sets args.flags.dwarf_injection
+-- and appends "dwarf-injection" to requested_set), so it is intentionally
+-- absent from this table.
+local PASS_FLAG_TO_NAME = {
+	["--word-counts"]       = "word-counts",
+	["--components"]        = "components",
+	["--validate"]          = "annotation",
+	["--offsets"]           = "offsets",
+	["--static-analysis"]   = "static-analysis",
+	["--atoms-source-map"]  = "atoms-source-map",
+	["--report"]            = "report",
+	["--scan-source"]       = "scan-source",
+	["--all"]               = ALL_PASSES_SENTINEL,
+}
+
+--- Append every pass name to args.requested_set. Names are derived from
+--- PASSES (no parallel name list); used by --all and by any caller that
+--- wants the full closure.
+--- @param args ParsedArgs
+local function request_all_passes(args)
+	local names = {}
+	for name in pairs(PASSES) do names[#names + 1] = name end
+	table.sort(names)
+	for _, n in ipairs(names) do
+		args.requested_set[#args.requested_set + 1] = n
+	end
+end
+
+-- Per-flag handlers. Each handler takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it). 
+-- Returning nil + os.exit() handles termination flags (--help). 
+-- This replaces the 8-way `if/elseif/elseif...` chain that nested 4 levels deep and made the dispatch logic hard to scan.
+local FLAG_HANDLERS = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- CLI parsing
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Print the CLI usage to stdout and exit 0.
+local function print_help()
+	io.write([[
+ps1_meta.lua - Tape-atom metaprogram orchestrator
+
+USAGE:
+  ps1_meta.lua [PASS_FLAGS] [COMMON_FLAGS]
+
+PASS_FLAGS:
+  Pick a phase or one-or-more individual passes:
+    --pre-link            [phase; default] Run the pre-link group + transitive deps.
+                         The root set is data-driven from each PASSES row's
+                         `groups` field; no parallel name list is maintained.
+    --post-link           [phase] Run the post-link group + transitive deps.
+                         Requires --elf. Sets --gdb-runtime and --dwarf-injection
+                         opt-in flags as well.
+    --all                 Select every row of the PASSES table. Pass-local opt-in
+                         guards remain active, so --dwarf-injection still requires
+                         --elf and --gdb-runtime still requires a runtime emission.
+  Or pick any subset:
+    --scan-source         Scan sources into the fat SourceScan payload
+    --word-counts         Load metadata.h + scan for existing .macs.h
+    --components          Generate <module>/gen/<basename>.macs.h
+    --validate            Run atom annotation DSL validation
+    --offsets             Generate <module>/gen/<basename>.offsets.h
+    --atoms-source-map    Generate <basename>.atoms.sourcemap.txt per source
+    --dwarf-injection     [opt-in] Select the post-link dwarf-injection pass + set the
+                         opt-in flag. Requires --elf.
+    --static-analysis     Static analysis: GTE pipeline-fill, mac_yield, ABI handoff, cycle budget
+    --report              Render per-project summary
+
+COMMON_FLAGS:
+  --source FILE         Source file to process (repeatable)
+  --metadata PATH       Path to metadata.h (required)
+  --out-root DIR        Output root for reports (default: build/gen)
+  --project-root DIR    Project root for .macs.h scan (default: dirname(metadata))
+  --gdb-runtime         Also emit <out_root>/gdb_tape_atoms_runtime.gdb (post-link, requires --elf)
+  --elf PATH            Path to linked .elf (for --gdb-runtime / --dwarf-injection)
+  --dry-run             Print dep order + ASCII graph; exit 0 without running
+  --verbose             Print per-pass debug output
+  --help                Show this help and exit
+
+EXIT CODES:
+  0  All requested passes succeeded
+  1  Validation errors found
+  2  Metaprogram internal error
+
+EXAMPLE:
+  ps1_meta.lua --pre-link  --metadata metadata.h --source code/foo.c --source code/bar.c
+  ps1_meta.lua --post-link --metadata metadata.h --source code/foo.c --source code/bar.c --elf build/hello_gte.elf
+  ps1_meta.lua --all       --metadata metadata.h --source code/foo.c --source code/bar.c
+]])
+end
+
+-- Per-flag handlers. Each takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it).
+-- Termination flags like --help call os.exit() instead.
+-- This replaces the 8-way `if/elseif/elseif...` chain that nested 4 levels deep and made the dispatch logic hard to scan.
+--
+-- Populated AFTER print_help so the --help handler can reference it as an upvalue (Lua resolves locals at closure-call time, 
+-- but if the closure is defined before the local, it falls back to _G). 
+FLAG_HANDLERS["--help"] = function(args)
+	print_help()
+	os.exit(0)
+end
+
+FLAG_HANDLERS["--dry-run"]      = function(args) args.dry_run = true end
+FLAG_HANDLERS["--verbose"]      = function(args) args.verbose = true end
+FLAG_HANDLERS["--source"]       = function(args, argv, arg_idx) args.sources[#args.sources + 1] = argv[arg_idx + 1]; return arg_idx + 1 end
+FLAG_HANDLERS["--metadata"]     = function(args, argv, arg_idx) args.metadata                   = argv[arg_idx + 1]; return arg_idx + 1 end
+FLAG_HANDLERS["--out-root"]     = function(args, argv, arg_idx) args.out_root                   = argv[arg_idx + 1]; return arg_idx + 1 end
+FLAG_HANDLERS["--project-root"] = function(args, argv, arg_idx) args.project_root               = argv[arg_idx + 1]; return arg_idx + 1 end
+
+-- Per-pass stash flags. Read by `passes/atoms_source_map.lua` to opt into the post-link gdb-runtime emission.
+-- Same shape as the existing per-flag handlers. mutates `args.flags` (which propagates into `ctx.flags`).
+FLAG_HANDLERS["--gdb-runtime"]  = function(args) args.flags                = args.flags or {}; args.flags.gdb_runtime = true end
+FLAG_HANDLERS["--elf"]          = function(args, argv, arg_idx) args.flags = args.flags or {}; args.flags.elf_path    = argv[arg_idx + 1]; return arg_idx + 1 end
+-- Enable DWARF injection (default OFF). Opts in to the post-link pass and
+-- sets the flag in one shot — the explicit handler below owns both
+-- selection and opt-in state, so --dwarf-injection is intentionally absent
+-- from PASS_FLAG_TO_NAME.
+FLAG_HANDLERS["--dwarf-injection"] = function(args)
+	args.flags                 = args.flags or {}
+	args.flags.dwarf_injection = true
+	args.requested_set[#args.requested_set + 1] = "dwarf-injection"
+end
+-- Build-phase flags: --pre-link and --post-link request the roots of their
+-- declared groups (see roots_for_group). topo_sort closes transitive deps
+-- from those roots; dispatch_passes runs every pass in the resolved
+-- closure without phase-filtering.
+FLAG_HANDLERS["--pre-link"] = function(args)
+	request_roots_for_group(args, "pre-link")
+end
+-- Batch post-link phase: gdb-runtime + dwarf-injection in one luajit cold
+-- start. Sets the same opt-in flags as --gdb-runtime + --dwarf-injection
+-- and selects the post-link build-phase group.
+-- --elf is required; parse_args enforces it after all flags are parsed.
+FLAG_HANDLERS["--post-link"]      = function(args)
+	args.flags = args.flags or {}
+	args.flags.gdb_runtime     = true
+	args.flags.dwarf_injection = true
+	request_roots_for_group(args, "post-link")
+end
+
+-- G' (atom locals) is now consolidated into --dwarf-injection; no separate flag.
+
+-- Pass-flag handler. Reads the closed-set table, expands --all, appends to requested_set.
+FLAG_HANDLERS[PASS_FLAG_DISPATCH_KEY] = function(args, a)
+	local name = PASS_FLAG_TO_NAME[a]
+	if name == ALL_PASSES_SENTINEL then
+		request_all_passes(args)
+		return
+	end
+	args.requested_set[#args.requested_set + 1] = name
+end
+
+--- Parse argv into a structured table. Validates against a closed enum.
+--- @param argv string[]
+--- @return ParsedArgs
+local function parse_args(argv)
+	local args = {
+		requested_set = {},
+		sources       = {},
+		metadata      = nil,
+		out_root      = DEFAULT_OUT_ROOT,
+		project_root  = nil,
+		dry_run       = false,
+		verbose       = false,
+	}
+
+	local pos = 1
+	while pos <= #argv do
+		local a       = argv[pos]
+		local handler = FLAG_HANDLERS[a]
+		if handler then
+			pos = handler(args, argv, pos) or pos
+		elseif PASS_FLAG_TO_NAME[a] then
+			FLAG_HANDLERS[PASS_FLAG_DISPATCH_KEY](args, a)
+		else
+			io.stderr:write("ps1_meta: unknown flag '" .. a .. "'\n")
+			io.stderr:write("Run with --help for usage.\n")
+			os.exit(EXIT_INTERNAL_ERROR)
+		end
+		pos = pos + 1
+	end
+
+	-- Default: --pre-link if no explicit pass flags were given. The first
+	-- invocation of a build is always pre-link, so this avoids silently
+	-- also invoking post-link work in builds without an ELF artifact.
+	if #args.requested_set == 0 then request_roots_for_group(args, "pre-link") end
+
+	-- Defaults: project_root = dirname(metadata).
+	if args.metadata and not args.project_root then
+		local d = duffle.dirname(args.metadata)
+		if   #d > 0 and (d:sub(-1) == "/" or d:sub(-1) == "\\") then
+			d = d:sub(1, -2)
+		end
+		args.project_root = duffle.dirname(d)
+	end
+
+	if not args.metadata then
+		io.stderr:write("ps1_meta: --metadata PATH is required\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+	if #args.sources == 0 then
+		io.stderr:write("ps1_meta: at least one --source FILE is required\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	-- Post-link opt-ins (--gdb-runtime, --dwarf-injection) write output that
+	-- depends on the linked ELF. Without --elf the metaprogram can't satisfy
+	-- those requests, so refuse loud and early. This covers the explicit
+	-- --post-link batch, --dwarf-injection by itself, and --gdb-runtime by
+	-- itself.
+	local flags      = args.flags or {}
+	local elf_path   = flags.elf_path
+	local has_elf    = type(elf_path) == "string" and #elf_path > 0
+	local post_links = flags.gdb_runtime or flags.dwarf_injection
+	if post_links and not has_elf then
+		io.stderr:write("ps1_meta: --elf PATH is required for post-link output\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	return args
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Build ctx from parsed args
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Build the PassCtx from parsed args. Reads each source file once at startup;
+--- passes consume `src.text`, not the path (path is preserved for error reporting).
+--- @param args ParsedArgs
+--- @return PassCtx
+local function build_ctx(args)
+	local sources = {}
+	for _, path in ipairs(args.sources) do
+		-- lfs handles path metadata and directories, not file-content streams.
+		-- Keep this io.open local so this entry point preserves its tailored diagnostic and exit path below.
+		local  f = io.open(path, "r")
+		if not f then
+			io.stderr:write("ps1_meta: cannot open --source " .. path .. "\n")
+			os.exit(EXIT_INTERNAL_ERROR)
+		end
+		local text = f:read("*a")
+		f:close()
+
+		local dir      = duffle.dirname(path)
+		local basename = duffle.basename_no_ext(path)
+		if #dir > 0 and (dir:sub(-1) == "/" or dir:sub(-1) == "\\") then
+			dir = dir:sub(1, -2)
+		end
+
+		-- src.scan is populated by the "scan-source" pass (the first pass in the dep graph).
+		-- build_ctx just opens + reads the files; the scan itself happens in the pass module, not inline in the orchestrator.
+		sources[#sources + 1] = {
+			path     = path,
+			text     = text,
+			dir      = dir,
+			basename = basename,
+		}
+	end
+
+	-- Pre-compute the per-directory grouping once (Fleury: expose structure).
+	-- Three passes (annotation, report, static-analysis) call group_sources_by_dir with the same ctx.sources;
+	-- computing it here and stashing on ctx.by_dir eliminates 2 redundant calls.
+	local by_dir = duffle.group_sources_by_dir(sources)
+
+	return {
+		sources       = sources,
+		by_dir        = by_dir,
+		metadata_path = args.metadata,
+		shared        = {},
+		upstream      = {},
+		out_root      = args.out_root,
+		project_root  = args.project_root,
+		flags         = args.flags or {},
+		dry_run       = args.dry_run,
+		verbose       = args.verbose,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Topological sort (Kahn's algorithm + cycle detection)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Topologically sort the requested pass set, augmented with all transitive deps.
+--- Detects cycles and errors out with details.
+--- @param passes table<string, PassDescriptor>
+--- @param requested_set string[]
+--- @return string[]  -- execution order
+---
+--- Implementation note: the 4 algorithm phases (dep-closure, in-degree, ready-queue, sort) are inlined as 3 small blocks within this function. 
+--- Each was a 1-caller helper; the 2-caller rule doesn't apply, so inlining produces a single readable function
+--- (plex: small patterns → shared, but only when shared; here they're not).
+local function topo_sort(passes, requested_set)
+	-- Phase 1: dep-closure. Include every pass name transitively required by `requested_set`.
+	local needed = {}
+	for _, name in ipairs(requested_set) do needed[name] = true end
+	local changed = true
+	while changed do
+		changed = false
+		for name, _ in pairs(needed) do
+			local pass = passes[name]
+			if not pass then
+				error("unknown pass '" .. name .. "' requested")
+			end
+			for _, dep in ipairs(pass.deps) do
+				if not needed[dep] then
+					needed[dep] = true
+					changed = true
+				end
+			end
+		end
+	end
+
+	-- Phase 2: in-degrees for Kahn's algorithm. For each pass in `needed`, the number of its deps that are also in `needed`.
+	local in_degree = {}
+	for name, _ in pairs(needed) do in_degree[name] = 0 end
+	for name, _ in pairs(needed) do
+		for _, dep in ipairs(passes[name].deps) do
+			if needed[dep] then
+				in_degree[name] = in_degree[name] + 1
+			end
+		end
+	end
+
+	-- Phase 3: seed the ready queue with passes whose in-degree is 0, sorted alphabetically for deterministic order.
+	local ready = {}
+	for name, deg in pairs(in_degree) do
+		if deg == 0 then ready[#ready + 1] = name end
+	end
+	table.sort(ready)
+
+	-- Phase 4: drain the ready queue. For each popped pass, decrement the in-degree of every remaining pass that depended on it.
+	-- Newly-zero-degree passes are inserted back into the ready queue (kept sorted).
+	local order = {}
+	while #ready > 0 do
+		local just_finished = table.remove(ready, 1)
+		order[#order + 1] = just_finished
+		for name, _ in pairs(needed) do
+			if name ~= just_finished then
+				for _, dep in ipairs(passes[name].deps) do
+					if dep == just_finished then
+						in_degree[name] = in_degree[name] - 1
+						if in_degree[name] == 0 then
+							ready[#ready + 1] = name
+							table.sort(ready)
+						end
+					end
+				end
+			end
+		end
+	end
+
+	-- Cycle detection: if `order` doesn't include all needed passes, some are stuck with in_degree > 0
+	-- (the cycle closed on itself before Kahn could process them).
+	-- Without this check, a fully-closed cycle (e.g. A -> B -> A) would silently return an empty order list, leaving the orchestrator to dispatch nothing.
+	local needed_count = 0
+	for _ in pairs(needed) do needed_count = needed_count + 1 end  -- count hash entries; Lua's #t doesn't work
+	if #order ~= needed_count then
+		for name, deg in pairs(in_degree) do
+			if deg > 0 then
+				error("dependency cycle detected involving pass '" .. name .. "'")
+			end
+		end
+	end
+
+	return order
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- ASCII dep graph renderer (Decision 6 in the spec)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render the dep graph as ASCII art. Output width capped at 78 columns.
+--- Falls back to the simpler "Resolved dependency order" list only if graph width exceeds terminal width.
+--- @param passes table<string, PassDescriptor>
+--- @param closed string[]     -- dep-closed execution order
+--- @return string
+local function render_dep_graph(passes, closed)
+	local lines = {}
+	local function add(s) lines[#lines + 1] = s end
+
+	add("[ps1_meta] Resolved dependency order (closed under deps):")
+	for pass_idx, name in ipairs(closed) do
+		local p        = passes[name]
+		local deps_str = (#p.deps == 0) and "(no deps)" or
+			"(deps: " .. table.concat(p.deps, ", ") .. ")"
+		add(string.format("  %d. %-22s %-45s [%s]",
+			pass_idx, name, deps_str, p.kind))
+	end
+	add("")
+
+	-- Data-driven ASCII graph built from the actual PASSES table.
+	-- Kahn layers determine the row of each box; each pass becomes a 4-row
+	-- box (top border, name, kind+output-count, bottom border). Boxes in
+	-- the same layer are rendered side-by-side; layers are connected by a
+	-- 'v' marker row whose 'v' chars are centered under each box, indicating
+	-- the downward 'feeds into' direction.
+	--
+	-- Layout invariants enforced here:
+	--   * MAX_GRAPH_WIDTH = 78 cols: no emitted line exceeds this. The
+	--     "simplest" way to stay under the budget is to limit each sub-row
+	--     to MAX_BOXES_PER_ROW = 3 boxes; for the canonical 9-row PASSES
+	--     table the largest layer has 3 boxes, so no wrap engages today.
+	--     If a future layer grows past 3 boxes, the layer is split into
+	--     adjacent sub-rows (each ending in its own 'v'-marker row).
+	--   * No silent truncation: per-layer box width is computed from the
+	--     layer's widest content (max(name length, kind-suffix length))
+	--     plus a 1-char leading + 1-char trailing padding + 2 wall chars.
+	--     Long names widen the box; they are NEVER truncated.
+	--   * Collision-safety: every PASSES row has a unique name, kind, and
+	--     out-list, so two distinct passes cannot produce visually identical
+	--     boxes.
+	add("[ps1_meta] Pass graph (read top-to-bottom; edges = 'feeds into'):")
+	add("")
+
+	-- Compute Kahn layer per pass: layer L = max(deps' layer) + 1, layer 0
+	-- for deps-less passes. Repeated sweeps until every pass in `closed` is
+	-- assigned (handles forward refs that resolve on the second pass).
+	local pass_layer, max_layer = {}, 0
+	local sorted_closed = {}
+	for _, name in ipairs(closed) do sorted_closed[#sorted_closed + 1] = name end
+	table.sort(sorted_closed)
+	local function assign_layers()
+		local assigned_count = 0
+		for _, name in ipairs(sorted_closed) do
+			if pass_layer[name] == nil then
+				local p = passes[name]
+				local max_dep, ready = -1, true
+				for _, dep in ipairs(p.deps) do
+					if pass_layer[dep] == nil then ready = false; break end
+					if pass_layer[dep] > max_dep then max_dep = pass_layer[dep] end
+				end
+				if ready then
+					pass_layer[name] = max_dep + 1
+					if pass_layer[name] > max_layer then max_layer = pass_layer[name] end
+					assigned_count = assigned_count + 1
+				end
+			end
+		end
+		return assigned_count
+	end
+	while assign_layers() > 0 do end
+	-- Defensive invariant: any unresolved pass is a bug. topo_sort already
+	-- errors on cycles before this point, so reaching here means a logic
+	-- error in the renderer (or a synthetic call that bypassed topo_sort).
+	-- Surface the failure loudly with the offending names; never silently
+	-- place unresolved passes at layer 0 (which would corrupt the graph).
+	local unresolved = {}
+	for _, name in ipairs(sorted_closed) do
+		if pass_layer[name] == nil then unresolved[#unresolved + 1] = name end
+	end
+	if #unresolved > 0 then
+		error("render_dep_graph: unresolved Kahn layer for pass(es): "
+			.. table.concat(unresolved, ", ")
+			.. "; topo_sort should have caught this earlier")
+	end
+
+	-- Bucket passes by layer; sort each bucket alphabetically for stability.
+	local layers = {}
+	for i = 0, max_layer do layers[i] = {} end
+	for _, name in ipairs(sorted_closed) do
+		layers[pass_layer[name]][#layers[pass_layer[name]] + 1] = name
+	end
+	for i = 0, max_layer do table.sort(layers[i]) end
+
+	-- Layout constants. Boxes in the same layer share the same width
+	-- (computed as the layer's widest content + padding + walls).
+	local MAX_GRAPH_WIDTH   = 78
+	local MAX_BOXES_PER_ROW = 3
+	local GAP               = 3
+
+	local function pad_right(s, width)
+		if #s >= width then return s:sub(1, width) end
+		return s .. string.rep(" ", width - #s)
+	end
+
+	-- Compute the box width (in chars, including both walls) for a layer.
+	-- The interior is the wider of (a) the longest pass name + 1 leading
+	-- space and (b) the longest "<kind> <N>" suffix + 1 leading space.
+	-- Then add 2 for the wall chars.
+	--
+	-- Why +1 (not +2): the +1 formula means the layer's max-content row
+	-- has 0 padding before the right wall (the `|` is immediately after
+	-- the content). This is the collision-safe widening policy the task
+	-- requires — long names widen the box and never get trailing padding.
+	-- Shorter passes in the same layer get trailing padding to fill the
+	-- interior to the layer's uniform width; they never get truncated.
+	local function box_w_for_layer(bucket)
+		local max_content = 0
+		for _, name in ipairs(bucket) do
+			local p        = passes[name]
+			local out_n    = #(p.out or {})
+			local kind_suf = string.format("%s %d>", p.kind, out_n)
+			if #name     > max_content then max_content = #name     end
+			if #kind_suf > max_content then max_content = #kind_suf end
+		end
+		-- Interior = 1 leading space + max_content + 0 trailing (the
+		-- trailing `|` IS the right boundary); walls = 2.
+		return max_content + 3
+	end
+
+	-- Render one pass as a 4-row box at the given box_w. The interior is
+	-- always padded to fit exactly; no string is ever truncated.
+	local function render_box(name, box_w)
+		local p        = passes[name]
+		local out_n    = #(p.out or {})
+		local kind_suf = string.format("%s %d>", p.kind, out_n)
+		local interior = box_w - 2
+		local border   = "+" .. string.rep("-", interior) .. "+"
+		return {
+			border,
+			"|" .. pad_right(" " .. name,     interior) .. "|",
+			"|" .. pad_right(" " .. kind_suf, interior) .. "|",
+			border,
+		}
+	end
+
+	-- Join a single row (1..4) across all boxes in a sub-row, with GAP
+	-- spaces between adjacent boxes.
+	local function join_row(box_rows, row_idx)
+		local parts = {}
+		for i, b in ipairs(box_rows) do
+			parts[#parts + 1] = b[row_idx]
+			if i < #box_rows then parts[#parts + 1] = string.rep(" ", GAP) end
+		end
+		return table.concat(parts)
+	end
+
+	-- 'v' marker row beneath a sub-row: one 'v' centered under each box.
+	local function v_marker_row(box_rows)
+		local total = 0
+		local centers = {}
+		for i, b in ipairs(box_rows) do
+			local w = #b[1]  -- box width = length of the top border row
+			local center = total + math.floor(w / 2)
+			centers[#centers + 1] = center
+			total = total + w + GAP
+		end
+		-- total now includes a trailing GAP we don't want; trim it.
+		total = total - GAP
+		local s = string.rep(" ", total)
+		for _, c in ipairs(centers) do
+			s = s:sub(1, c) .. "v" .. s:sub(c + 2)
+		end
+		return s
+	end
+
+	-- Render a single sub-row (a contiguous chunk of a layer's bucket).
+	-- Emits the 4 box rows + an empty line + the 'v' marker row + an
+	-- empty line, EXCEPT the very last sub-row of the very last layer
+	-- omits the trailing 'v' marker (nothing flows below it).
+	local function render_subrow(bucket_chunk, is_last_subrow, is_last_layer)
+		local box_w = box_w_for_layer(bucket_chunk)
+		local boxes = {}
+		for _, name in ipairs(bucket_chunk) do boxes[#boxes + 1] = render_box(name, box_w) end
+		add(join_row(boxes, 1))  -- top borders
+		add(join_row(boxes, 2))  -- names
+		add(join_row(boxes, 3))  -- kind + output-count
+		add(join_row(boxes, 4))  -- bottom borders
+		-- 'v' marker row beneath this sub-row connects downward to the
+		-- next sub-row of the same layer (if any) OR to the next layer.
+		-- Skip the trailing 'v' only on the very last sub-row of the
+		-- final layer, where nothing flows below it.
+		if not is_last_subrow or not is_last_layer then
+			add("")
+			add(v_marker_row(boxes))
+			add("")
+		end
+	end
+
+	for layer_idx = 0, max_layer do
+		local bucket   = layers[layer_idx]
+		local is_last  = (layer_idx == max_layer)
+		-- Split the layer into sub-rows of at most MAX_BOXES_PER_ROW boxes.
+		-- With a 20-char box width (the canonical case: name "static-analysis"
+		-- is 15 chars, suffix "header-output 2>" is 16 chars) and GAP=3,
+		-- 3 boxes per sub-row = 3*20 + 2*3 = 66 cols + a 'v' row of 66 cols;
+		-- well within MAX_GRAPH_WIDTH. A 4th box would push to 4*20 + 3*3 = 89,
+		-- which is why MAX_BOXES_PER_ROW = 3 (wrap when >3).
+		local chunk_size = math.min(MAX_BOXES_PER_ROW, #bucket)
+		if chunk_size < 1 then chunk_size = 1 end
+		for chunk_start = 1, #bucket, chunk_size do
+			local chunk_end = math.min(chunk_start + chunk_size - 1, #bucket)
+			local chunk = {}
+			for i = chunk_start, chunk_end do chunk[#chunk + 1] = bucket[i] end
+			local is_last_subrow = (chunk_end == #bucket)
+			render_subrow(chunk, is_last_subrow, is_last)
+		end
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Main orchestrator
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- (internal) Push a pass's outputs + warnings into `ctx.upstream[name]` for downstream passes to consume.
+-- @param ctx PassCtx
+-- @param pass_name string
+-- @param result PassResult
+local function accumulate_pass_result(ctx, pass_name, result)
+	ctx.upstream[pass_name] = ctx.upstream[pass_name] or {}
+	for _, out in ipairs(result.outputs or {}) do
+		table.insert(ctx.upstream[pass_name], out)
+	end
+	for _, warn in ipairs(result.warnings or {}) do
+		table.insert(ctx.upstream[pass_name], warn)
+	end
+end
+
+-- (internal) If the pass's kind is in PASS_KIND_STOP_ON_ERROR and it reported errors, write each error to stderr.
+-- Returns true if any validation errors were reported.
+-- @param pass_name string
+-- @param pass PassDescriptor
+-- @param result PassResult
+-- @return boolean
+local function report_validation_errors(pass_name, pass, result)
+	local   has_errors = result.errors and #result.errors > 0
+	if not (has_errors and PASS_KIND_STOP_ON_ERROR[pass.kind]) then
+		return false
+	end
+	for _, e in ipairs(result.errors) do
+		io.stderr:write(string.format("[%s] line %d: %s\n",
+			pass_name, e.line or 0, e.msg or ""))
+	end
+	return true
+end
+
+-- (internal) Run each pass in `order` in topological sequence.
+-- @param ctx PassCtx
+-- @param order string[]
+-- @return boolean  -- true if any validation errors were reported
+local function dispatch_passes(ctx, order)
+	ctx.shared = {}
+	local had_errors = false
+	for _, pass_name in ipairs(order) do
+		local pass   = PASSES[pass_name]
+		io.stderr:write(string.format("[ps1_meta] %-22s running\n", pass_name))
+		local mod    = require(pass.module)
+		local result = mod.run(ctx)
+
+		accumulate_pass_result(ctx, pass_name, result)
+		if report_validation_errors(pass_name, pass, result) then
+			had_errors = true
+		end
+	end
+	return had_errors
+end
+
+--- Main entry point. Runs the requested passes in dep-topological order.
+--- @param argv string[]
+local function main(argv)
+	local ok, err = pcall(function()
+		local args = parse_args(argv)
+		local ctx  = build_ctx(args)
+
+		local requested = args.requested_set
+		local closed    = topo_sort(PASSES, requested)
+
+		-- --dry-run: print dep order + ASCII graph, exit OK.
+		if args.dry_run then
+			io.write(render_dep_graph(PASSES, closed))
+			os.exit(EXIT_OK)
+		end
+
+		local had_errors = dispatch_passes(ctx, closed)
+		if had_errors then os.exit(EXIT_VALIDATION_ERRORS) end
+	end)
+
+	if not ok then
+		io.stderr:write("[ps1_meta] internal error: " .. tostring(err) .. "\n")
+		os.exit(EXIT_INTERNAL_ERROR)
+	end
+
+	os.exit(EXIT_OK)
+end
+
+-- Module export for in-process consumers (tests that dofile this script).
+-- The closure above, `render_dep_graph`, and the canonical `PASSES` table
+-- are exposed so a test can render the graph for synthetic PASSES tables
+-- without spawning a subprocess. The conditional `main(...)` call below
+-- only fires when this file is invoked as the entry script (arg[0] ends
+-- in "ps1_meta.lua"); in dofile() mode (test's arg[0] does not match),
+-- main() is skipped and the chunk returns `_M` to the caller.
+local _M = {
+	render_dep_graph = render_dep_graph,
+	PASSES           = PASSES,
+}
+
+if arg and arg[0] and arg[0]:match("ps1_meta%.lua$") then
+	main({...})
+end
+
+return _M

scripts/tape_atom.offset_gen.meta.lua

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

scripts/update_deps.ps1

@@ -4,25 +4,24 @@ $path_code      = join-path $path_root 'code'
 $path_scripts   = join-path $path_root 'scripts'
 $path_toolchain = join-path $path_root 'toolchain'
 
+# Halt on any error (instead of PowerShell's default `Continue`).
+$ErrorActionPreference = 'Stop'
+
 $misc = join-path $PSScriptRoot 'helpers/misc.ps1'
 . $misc
 
-# TODO(Ed): Review usage of these deps
-# I orgiinally cloned them when starting to get to the C runtime usage of the course
-# However, based on the heavy reliance of the PSX.Dev extension I might fallback; also
-# The gdb server doesn't need the full repo and were only using the src/mips 
-# which has a standalone repo (nuggets)
-# armips may not be used at all but I'm not sure...
-
 $url_armips     = 'https://github.com/Kingcom/armips.git'
 $url_pcsx_redux = 'https://github.com/grumpycoders/pcsx-redux.git'
 $url_psyq_iwyu  = 'https://github.com/johnbaumann/psyq_include_what_you_use.git'
+$url_lpeg       = 'https://github.com/roberto-ieru/LPeg.git'
 
 $path_armips     = join-path $path_toolchain 'armips'
 $path_pcsx_redux = join-path $path_toolchain 'pcsx-redux'
 $path_psyq_iwyu  = join-path $path_toolchain 'psyq_iwyu'
+$path_lpeg       = join-path $path_toolchain 'lpeg'
 
 clone-gitrepo $path_armips     $url_armips
+clone-gitrepo $path_lpeg       $url_lpeg
 clone-gitrepo $path_pcsx_redux $url_pcsx_redux
 clone-gitrepo $path_psyq_iwyu  $url_psyq_iwyu
 
@@ -37,3 +36,98 @@ pop-location
 # $path_pcsx_redux_binaries   = join-path $path_pcsx_redux_vsprojects 'x64/Release'
 
 # $psyq_obj_parser = join-path $path_pcsx_redux_binaries 'psyq-obj-parser.exe'
+
+# ════════════════════════════════════════════════════════════════════════════
+# PCSX-Redux — built via MSBuild (VS2022)
+#
+# Requires: Visual Studio 2022 with the C++ desktop workload.
+# The .vcxproj files target platform toolset v145, but VS2022 ships v143;
+# we pass /p:PlatformToolset=v143 to retarget at build time (no file edits).
+# NuGet packages (glfw, luajit.native, libFFmpeg-lite, x64sentry) are
+# restored automatically by MSBuild on first build.
+#
+# Output: toolchain\pcsx-redux\vsprojects\x64\Debug\pcsx-redux.exe
+# ════════════════════════════════════════════════════════════════════════════
+
+# Locate MSBuild from the VS2022 install (no hardcoded path — uses vswhere).
+$vswhere = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
+if (-not (Test-Path $vswhere)) {
+	write-error "vswhere not found at '$vswhere'. Install Visual Studio 2022 with the C++ desktop workload."
+	exit 1
+}
+$msbuild_exe = & $vswhere -latest -products * -requires Microsoft.Component.MSBuild -find "MSBuild\**\Bin\MSBuild.exe" 2>$null | Select-Object -First 1
+if (-not $msbuild_exe) {
+	write-error "MSBuild not found via vswhere. Install Visual Studio 2022 with the C++ desktop workload."
+	exit 1
+}
+
+$path_pcsx_sln = join-path $path_pcsx_redux 'vsprojects\pcsx-redux.sln'
+&	$msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
+
+# Locate luajit via scoop. `luajit.exe` is on PATH via scoop's shim;
+# we use `scoop prefix` to find the install root for the include dir
+# (needed to compile lpeg against luajit's headers).
+# If scoop or luajit is missing, fail fast with an actionable message.
+$luajit_prefix = & scoop prefix luajit 2>$null
+if (-not $luajit_prefix -or -not (Test-Path (Join-Path $luajit_prefix 'bin/luajit.exe'))) {
+	write-error "luajit not found via 'scoop prefix luajit'. Install via: scoop install luajit"
+	exit 1
+}
+
+# Discover the luajit include dir by globbing `include/luajit-*`.
+# This avoids hardcoding a specific version (e.g. `luajit-2.1`).
+$luajit_include_root = Join-Path $luajit_prefix 'include'
+$lua_inc_dir = Get-ChildItem -Path $luajit_include_root -Directory -Filter 'luajit-*' -ErrorAction SilentlyContinue |
+	Select-Object -First 1 -ExpandProperty FullName
+if (-not $lua_inc_dir) {
+	write-error "No 'luajit-*' include dir found under '$luajit_include_root'. The scoop luajit install may be broken."
+	exit 1
+}
+
+# Generate lpeg.dll by compiling the 6 source files directly.
+# `gcc` is on PATH (scoop's shim puts it there).
+# The source files: lpcap.c lpcode.c lpcset.c lpprint.c lptree.c lpvm.c
+# (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')
+$lpeg_compile_args = @(
+	'-O2', '-shared',
+	"-I$lua_inc_dir",
+	"-L$luajit_lib_dir",
+	'-o', 'lpeg.dll'
+) + $lpeg_sources + @('-lluajit-5.1')
+push-location $path_lpeg
+&	gcc @lpeg_compile_args
+pop-location
+
+# ════════════════════════════════════════════════════════════════════════════
+# lfs (LuaFileSystem) — compiled from pcsx-redux's vendored luafilesystem source.
+# Used by word_count_eval.lua :: scan_dir for native directory enumeration (~2ms)
+# instead of spawning `dir /b /s` as a subprocess (~56ms).
+# Source: toolchain/pcsx-redux/third_party/luafilesystem/src/lfs.c
+# Output: toolchain/lfs/lfs.dll
+# ════════════════════════════════════════════════════════════════════════════
+
+$path_lfs = join-path $path_toolchain 'lfs'
+verify-path $path_lfs
+$lfs_src = join-path $path_pcsx_redux 'third_party\luafilesystem\src\lfs.c'
+$lfs_dll = join-path $path_lfs 'lfs.dll'
+$lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
+& gcc -O2 -shared "-I$lua_inc_dir" -o $lfs_dll $lfs_src $lfs_dll_import
+
+# ════════════════════════════════════════════════════════════════════════════
+# OpenBIOS — built from the PCSX-Redux source tree via make + mipsel-none-elf
+#
+# OpenBIOS is an open-source PS1 BIOS implementation (no retail BIOS dump needed). 
+# It builds with the MIPS cross-toolchain (`mipsel-none-elf-gcc`, on PATH via the `mips` toolchain installer) 
+# + `make` (on PATH via scoop).
+#
+# Output: toolchain\pcsx-redux\src\mips\openbios\openbios.bin
+# ════════════════════════════════════════════════════════════════════════════
+
+$path_openbios = join-path $path_pcsx_redux 'src\mips\openbios'
+push-location $path_openbios
+&	make clean
+&	make
+pop-location