eliminated most lag (runs in ms)

script running is slow need to fix.

.gitignore

@@ -15,3 +15,5 @@ toolchain/PSn00bSDK
 *.a
 .sentry-native
 .vscode/settings.json
+toolchain/lfs
+toolchain/lpeg

code/duffle/atom_dsl.h

@@ -2,207 +2,123 @@
  * atom_dsl.h
  * ============================================================================
  *
- * ATOM DSL — annotation layer for tape atoms (lottes_tape.h).
+ * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
  *
- * This header turns `__attribute__((annotate(...)))` and `_Pragma(...)` into
- * a small named DSL that the metaprogram can validate against.
- *
- *   The C compiler treats every macro below as a no-op:
- *     - atom_init / atom_terminate / atom_bind / atom_setup / atom_commit /
- *       atom_annot all expand to `__attribute__((annotate("..."))) MipsAtom_(name)`
- *       — accepted by GCC (with -Wno-attributes), absent at runtime.
- *     - atom_resource / atom_region / atom_group / atom_cadence / atom_async
- *       expand to `_Pragma("...")` — accepted by any C11 preprocessor.
- *
- *   The metaprogram (tape_atom_annotation_pass.lua) reads the source-as-written
+ *   WHAT THIS HEADER IS
+ *   -------------------
+ *   The metaprogram (scripts/passes/annotation.lua) reads 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
+ *     - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X),
+ *       atom_reads(...), atom_writes(...)) in any order. All optional.
+ *       (No phase token for now; phases may be reintroduced later.)
+ *     - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a
+ *       real Binds_* struct declaration.
+ *     - wave-context positions only reference the canonical 4-register
+ *       set: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
+ *     - atom word-counts in word_counts.metadata.h agree with the body's
+ *       actual .word count.
+ *
+ *   WHY A PURE MACRO (atom_info, atom_bind, atom_reads, atom_writes, atom_label)
+ *   -----------------------------------------------------------------
+ *   Each of these expands to a C comment or to nothing. The C preprocessor
+ *   strips them to whitespace. The metaprogram reads the literal token from
+ *   source-as-written, NOT from the preprocessed output. This means:
+ *     - the C compiler does no work for them (no __attribute__, no
+ *       _Pragma, no asm side-effects)
+ *     - they can never silently drift from the metaprogram's view
+ *       (the metaprogram re-reads the source on every build)
+ *     - the annotation is invisible to the linker, debugger, and IDE
  *
  * ============================================================================
  *
- *  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).
+ *      1. atom_info(...) is OPTIONAL. Most atoms have no annotation.
+ *         Atoms without atom_info are silently skipped by the metaprogram.
  *
- *      2. atom_init and atom_terminate take only the name.
+ *      2. If present, atom_info takes up to three sub-calls, all
+ *         order-independent within the arg list:
+ *           - atom_bind(Binds_X) (optional; only for rbind atoms)
+ *           - atom_reads(...) (optional; wave-context registers)
+ *           - atom_writes(...) (optional; wave-context registers)
  *
- *      3. atom_setup and atom_commit take name + reads.
+ *      3. atom_bind(Binds_X) pins the ABI-struct shape -- the metaprogram
+ *         cross-references Binds_X against the
+ *         `typedef struct Binds_X { ... } Binds_X;` declaration.
  *
- *      4. atom_bind takes name + Binds_* type + writes.
+ *      4. atom_reads(...) and atom_writes(...) args are wave-context
+ *         registers: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
+ *         Closed set. GTE / SP / DMA / I/O state is declared in source
+ *         comments, not in atom_reads/atom_writes.
  *
- *      5. atom_annot takes name + phase token + reads + writes.
- *         Phase tokens: phase_init / phase_bind / phase_setup / phase_work /
- *         phase_commit / phase_terminate.
+ *      5. atom_label(name) is an anchor -- the macro is empty in C; the
+ *         metaprogram records the marker at the current pos for offset
+ *         calculation.
  *
- *      6. 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
- *
- * ============================================================================
+ *      6. atom_offset(F, T) is resolved by gen/atom_offsets.h, generated
+ *         from the atom_label markers.
  */
 
- #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:
+ * WAVE-CONTEXT REGISTERS -- canonical register set for the tape wave model.
  *
  *   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.
+ * Closed set. If your atom needs to touch GTE / SP / DMA / other side state,
+ * declare it at the source level as you normally would -- but DO NOT put
+ * those registers in atom_reads/atom_writes.
  *
  * ============================================================================*/
 
 /* ============================================================================
- * REGION TOKENS — memory regions atoms may allocate from or write into.
+ * atom_reads(...) / atom_writes(...) -- wave-context register list
  *
- * 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)
+ *   atom_reads(R_PrimCursor, R_FaceCursor)
+ *         ->  (R_PrimCursor, R_FaceCursor)   // comma-evaluated, discarded
  *
  * The macro produces a comma-evaluated expression that the C compiler
- * silently discards (it'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.
+ * silently discards (it sits in an unused arg position -- the result is
+ * never bound). The Lua tool pattern-matches the "atom_reads(...)" /
+ * "atom_writes(...)" token to extract the list.
  *
- * You can have at most one 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
+ * You can have at most one atom_reads(...) and at most one atom_writes(...)
+ * in an atom_info(...) call. To declare multiple disjoint sets (rare), just
+ * declare the union -- the metaprogram doesn't track which reads need which
  * writes at this granularity.
  *
  * ============================================================================*/
@@ -212,227 +128,45 @@
 /* ============================================================================
  * 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_info -- single unified annotation. OPTIONAL. Most atoms have none.
+ *
+ *      MipsAtom_(cube_tri) atom_info(
+ *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor)
+ *      ){ ... };
+ *
+ *   Shape (sub-args order-independent; all optional):
+ *     - atom_bind(Binds_X):  at most one; pins the ABI-struct shape
+ *     - atom_reads(...):     at most one; comma-list of wave-context registers
+ *     - atom_writes(...):    at most one; comma-list of wave-context registers
+ *
+ *   No phase token for now. The metaprogram doesn't check ordering across
+ *   atoms -- phases (init / bind / setup / work / commit / terminate) will
+ *   be reintroduced when ordering checks are added.
+ *
+ *   The macro expands to a C comment (or to nothing). The C compiler does
+ *   no work. The metaprogram reads the source-as-written directly.
  *
  * ============================================================================*/
+#define atom_info(...)  /* atom_info(__VA_ARGS__) */
 
 /* ----------------------------------------------------------------------------
- * atom_init — entry into tape_runtime_main
+ * atom_bind(Binds_X) -- rbind sub-call of atom_info
  *
- *   atom_init(tape_main)
- *   internal MipsAtom_(tape_main) { ... };
+ *   MipsAtom_(rbind_cube_tri) atom_info(
+ *         atom_bind(Binds_CubeTri)
+ *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+ *   ){ ... };
  *
- * Implies: no reads, no writes (wave-context not established yet).
- * ----------------------------------------------------------------------------*/
-#define atom_init(name) __attribute__((annotate("atom_init")))
-
-/* ----------------------------------------------------------------------------
- * atom_terminate — exit from tape_runtime_main
- *
- *   atom_terminate(tape_exit)
- *   internal MipsAtom_(tape_exit) { ... };
- *
- * Implies: no reads, no writes (wave-context destroyed at this point).
- * ----------------------------------------------------------------------------*/
-#define atom_terminate(name) __attribute__((annotate("atom_terminate")))
-
-/* ----------------------------------------------------------------------------
- * atom_setup — pre-work atom: prepares engine state (e.g., set_gte_world)
- *
- *   atom_setup(set_gte_world, tape_regs(R_TapePtr))
- *   internal MipsAtom_(set_gte_world) { ... };
- *
- * Reads:   anything (the engine state you're reading)
- * Writes:  engine state (GTE / DMA / etc. — declared in source, not part of
- *          wave-context, so doesn't go in tape_regs)
- *
- * The metaprogram checks that setup is followed (in atomic order) by a work
- * atom in the same wave — there's no point in setting up state if no one
- * reads it.
- * ----------------------------------------------------------------------------*/
-#define atom_setup(name, reads) __attribute__((annotate("atom_setup")))
-
-/* ----------------------------------------------------------------------------
- * atom_commit — post-work atom: flushes wave-context back to C-side state
- *
- *   atom_commit(sync_prim_cursor, tape_regs(R_PrimCursor))
- *   internal MipsAtom_(sync_prim_cursor) { ... };
- *
- * Reads:   wave-context registers (the ones you sync back to C)
- * Writes:  C-side mirror (declared in source — not part of wave-context)
- *
- * The metaprogram checks that commit is preceded (in atomic order) by a
- * work atom that wrote the registers this commit is reading.
- * ----------------------------------------------------------------------------*/
-#define atom_commit(name, reads) __attribute__((annotate("atom_commit")))
-
-/* ----------------------------------------------------------------------------
- * atom_bind — rbind atom: read wave-context registers from tape pointer
- *
- *   atom_bind(rbind_cube_tri, Binds_CubeTri,
- *             tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase))
- *   internal MipsAtom_(rbind_cube_tri) { ... };
- *
- * The binds_struct MUST be a typedef'd type (declared via
+ * The Binds_X 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.
+ * atom_bind is a SUB-CALL of atom_info, not a standalone annotation macro.
+ *
+ * The macro expands to a C comment. The metaprogram reads source-as-written.
  * ----------------------------------------------------------------------------*/
-#define atom_bind(name, binds_struct, writes) __attribute__((annotate("atom_bind")))
-
-/* ----------------------------------------------------------------------------
- * atom_annot — generic work atom with explicit phase
- *
- *   atom_annot(cube_tri, phase_work,
- *              tape_regs(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
- *              tape_regs(R_PrimCursor, R_FaceCursor))
- *   internal MipsAtom_(cube_tri) { ... };
- *
- * Use this for the bulk of your atoms. For init/setup/commit/bind, prefer
- * the convenience macros above — they pin the phase for you.
- *
- * The phase arg is one of: phase_init / phase_bind / phase_setup /
- * phase_work / phase_commit / phase_terminate. Spelling mistakes are errors.
- * ----------------------------------------------------------------------------*/
-#define atom_annot(name, phase, reads, writes) __attribute__((annotate("atom_annot")))
-
-/* ============================================================================
- * RESOURCE / GROUP / CADENCE / REGION / ASYNC — optional atom metadata
- *
- *   These don't annotate the atom semantically (phase/reads/writes do that).
- *   They attach extra context that the metaprogram uses to catch:
- *     - same resource loaded twice in different ways
- *     - atoms that span multiple regions (likely bug — pick one)
- *     - frame-cadence atoms that should be once-cadence (perf / correctness)
- *     - ondemand atoms that aren't async (CDROM races)
- *
- *   You can use as many as apply to a given atom, in any order, immediately
- *   above the atom_*() macro.
- *
- * ============================================================================*/
-
-/* ----------------------------------------------------------------------------
- * atom_resource — name the logical resource the atom references
- *
- *   atom_resource(cube_tri, "model_ship_cube")
- *   atom_resource(load_track_faces, "track_lavender_field_0x42")
- *   atom_resource(play_engine_sfx, "sfx_engine_loop")
- *
- * Use any human-readable string. The metaprogram:
- *   - validates resource strings are non-empty and don't contain control chars
- *   - flags duplicates across atoms with the same name (two atoms claiming
- *     ownership of a resource is usually a refactor artifact or bug)
- *   - flags references to resources that no atom actually defines
- *
- * The arg is a STRING LITERAL, so it can't accidentally alias a variable.
- * ----------------------------------------------------------------------------*/
-#define atom_resource(name, res_id) //_Pragma("atom " #name " resource=" res_id)
-
-/* ----------------------------------------------------------------------------
- * atom_region — name the memory region the atom touches
- *
- *   atom_region(cube_tri,   REGION_PRIM_ARENA)
- *   atom_region(load_faces, REGION_HEAP_3D)
- *   atom_region(load_tex,   REGION_VRAM)
- *
- * Use REGION_* tokens above. The metaprogram enforces the closed set.
- *
- * Edge cases the metaprogram catches:
- *   - rbind atom that doesn't declare a SOURCE region (where is it loading from?)
- *   - work atom with no destination region (where is it pushing to?)
- *   - region that disagrees with the Binds_* struct layout (you said it's a
- *     prim_arena rbind but the struct has 4 faces in it — wait, that's wrong)
- * ----------------------------------------------------------------------------*/
-#define atom_region(name, region) //_Pragma("atom " #name " region=" #region)
-
-/* ----------------------------------------------------------------------------
- * atom_group — bundle atoms into a logical batch (track-load, sound-load, etc.)
- *
- *   atom_group(load_track_face_42,  GROUP_LOAD_FACES)
- *   atom_group(load_track_face_43,  GROUP_LOAD_FACES)
- *   atom_group(swap_face_42_43,     GROUP_VISIBILITY_SWAP)
- *
- * Use any token as the group id. The metaprogram:
- *   - validates all atoms in a group emit their waves in the same tb_group
- *     (no spawning other waves inside a group)
- *   - flags groups with only one member (probably a typo — meant to be a group?)
- *   - validates cross-group edges (no atom reads what another group writes,
- *     unless explicitly grouped together)
- *
- * Useful when:
- *   - subdivisible work (track-face batches, polygon subdivision) needs to
- *     confirm that all batches of one logical visible scene are emitted
- *     together
- *   - async loads (CDROM -> VRAM) need to be grouped so all batches complete
- *     before the swap
- *
- * Use GROUPS for sound effects to track which sound plays during which atom,
- * which is needed if the sound tool ever has to validate "this atom is the
- * trigger for an audio play".
- * ----------------------------------------------------------------------------*/
-#define atom_group(name, group_id) //_Pragma("atom " #name " group=" #group_id)
-
-/* ----------------------------------------------------------------------------
- * atom_cadence — declare execution frequency
- *
- *   atom_cadence(render_frame,     CADENCE_FRAME)       // every vsync
- *   atom_cadence(load_track_faces, CADENCE_ONDEMAND)    // on demand
- *   atom_cadence(init_heap,        CADENCE_ONCE)        // process lifetime
- *
- * Default (no atom_cadence call) is CADENCE_FRAME — most atoms run every
- * frame. Override explicitly when not.
- *
- * The metaprogram's checks:
- *   - CADENCE_ONCE atoms inside `if (frame == 0)` or `if (!initialized)` are
- *     tagged, validating that guards are required (or warning if missing)
- *   - CADENCE_FRAME atoms that mutate state outside the wave context get
- *     flagged (likely a bug — state should persist through commits)
- *   - CADENCE_ONDEMAND atoms must have atom_async — otherwise the trigger
- *     mechanism is undefined
- * ----------------------------------------------------------------------------*/
-#define atom_cadence(name, cadence) //_Pragma("atom " #name " cadence=" #cadence)
-
-/* ----------------------------------------------------------------------------
- * atom_async — declare whether the atom yields / interacts with CDROM DMA
- *
- *   atom_async(load_track_tex,  true)   // CDROM read yield
- *   atom_async(load_vram,       true)   // VRAM upload DMA
- *   atom_async(render_frame,    false)  // pure compute, no async
- *
- * The metaprogram requires this for CADENCE_ONDEMAND atoms. For
- * CADENCE_FRAME, it's optional but documents intent.
- *
- * Note: CDROM ATOMS in Psy-Q are typically implemented as a chain of
- * "async-init" atom followed by a "wait-for-completion" atom. Both atoms
- * should be marked async=true, and both should have the same resource/group
- * tag (so the metaprogram can verify they're paired).
- * ----------------------------------------------------------------------------*/
-#define atom_async(name, is_async) //_Pragma("atom " #name " async=" #is_async)
-
-/* ============================================================================
- * WORD-COUNT ANNOTATION FOR A #define MAC
- *
- *   tape_words(mac_yield, 1)
- *   #define mac_yield()  \
- *       load_word(R_AtomJmp, R_TapePtr, 0), \
- *       add_ui_self(R_TapePtr, 4), \
- *       jump_reg(R_AtomJmp), \
- *       nop
- *
- * The compiler accepts the unknown _Pragma. The Lua tool reads it and
- * cross-checks against WORD_COUNT(mac_yield, 1) in tape_atom.metadata.h.
- * If they disagree, build fails.
- *
- * Use sparingly — only on multi-word macros (single-word ones don't need
- * drift tracking; they're checked by the .word-count pass anyway).
- *
- * ============================================================================*/
-#define tape_words(name, n) //_Pragma(#name " tape_atom words=" #n)
+#define atom_bind(binds_struct)  /* atom_bind(binds_struct) */
 
 /* ============================================================================
  * atom_label / atom_offset — branch target machinery

code/duffle/gen/duffle.macs.h

@@ -1,11 +1,9 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #pragma once
 #endif
-// Auto-generated by tape_atom_annotation_pass.lua — DO NOT EDIT
+// 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_*)
-// + auto word-counts (so tape_atom.metadata.h stays manual-only
-//   for encoding macros).
 
 #ifndef WORD_COUNT
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
@@ -26,7 +24,7 @@ WORD_COUNT(mac_yield, 4)
 WORD_COUNT(mac_load_tri_indices, 3)
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-#define mac_load_tri_verts(...) \
+#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)) \
@@ -45,21 +43,21 @@ WORD_COUNT(mac_load_tri_indices, 3)
 ,	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_load_tri_verts, 18)
+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,bf_addr_len)) /* 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 */ \
+,	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,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 */
+,	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.
@@ -67,18 +65,18 @@ WORD_COUNT(mac_insert_ot_tag_f3, 11)
 #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,bf_addr_len)) /* 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 */ \
+,	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,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 */
+,	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, code, r, g, b) \
-	load_upper_i(R_AT, (code) << 8  | (b)) \
+#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)
@@ -104,8 +102,8 @@ 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
+ * 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-

code/duffle/gp.h

@@ -49,12 +49,8 @@
 /* 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. */
+ * 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` */
@@ -80,13 +76,12 @@ 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.
+ *  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 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.
+ *  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 +115,7 @@ 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 +124,7 @@ 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,
@@ -153,19 +145,19 @@ enum {
  *  ============================================================================
  *
  *  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.
+ *  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))
@@ -291,15 +283,7 @@ 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 +323,7 @@ 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,
 };
@@ -373,16 +357,14 @@ 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 +373,12 @@ 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 +386,14 @@ 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. */
+ *  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 +419,7 @@ 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 +427,7 @@ 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 +449,7 @@ 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 +462,7 @@ 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 +475,7 @@ 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 +490,7 @@ 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
@@ -647,7 +628,7 @@ enum {
  *
  *  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.
+ *  struct + enum level.
  * ============================================================================ */
 enum {
     tim_file_id_magic = 0x10,
@@ -660,7 +641,7 @@ 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 */
 };

code/duffle/gp_vendor_sym.h

@@ -27,7 +27,7 @@
  *  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
+ *  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.

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

@@ -21,12 +21,6 @@
  *      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

@@ -17,8 +17,7 @@ typedef Slice_MipsCode MipsAtom;
 #define MipsAtom_(sym) MipsCode tmpl(code,sym) [] align_(4) =
 
 // Bare form: file-scope declaration with hardcoded body.
-// Used for components with no args (e.g., ac_load_tri_indices) or
-// identifier-args (hardcoded register names).
+// 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 };
@@ -31,13 +30,9 @@ typedef Slice_MipsCode MipsAtom;
 //   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. The metaprogram puts them
+// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
 
-/* Register aliases (moved up from the Tape Drive region below so that
- * mac_yield's body and the Mips Atom Builder functions can reference
- * them. The C compiler processes the file top-to-bottom, so the enum
- * must be visible before any use.) */
+/* Register aliases */
 enum {
 	R_AtomJmp    = R_T9,
 	R_TapePtr    = R_T8,  /* The Instruction Stream Pointer */
@@ -126,7 +121,7 @@ MipsAtomComp_(ac_load_tri_indices) {
 };
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-MipsAtomComp_(ac_load_tri_verts) {
+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),
@@ -137,14 +132,14 @@ MipsAtomComp_(ac_load_tri_verts) {
 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,bf_addr_len)), // 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
+	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,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
+	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.
@@ -152,23 +147,21 @@ MipsAtomComp_(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,bf_addr_len)), // 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
+	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,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
+	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. */
-FI_ MipsAtom ac_pack_color_word(U4 off, U4 code, U1 r, U1 g, U1 b)
+ * 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, (code) << 8  | (b)),
+	load_upper_i(R_AT, (cmd) << 8  | (b)),
 	or_i_self(   R_AT, ((g)   << 8) | (r)),
 	store_word(  R_AT, R_PrimCursor, (off)),
 })
@@ -203,8 +196,8 @@ MipsAtomComp_Proc_(ac_format_g4_color, {
 
 /* 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-
@@ -287,7 +280,10 @@ internal MipsAtom_(mips_flush_icache) {
 typedef Struct_(Binds_SetGteWorld) {
 	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)),
@@ -322,9 +318,9 @@ 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/mips_vendor_sym.h

@@ -21,12 +21,6 @@
  *      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

@@ -1,4 +1,4 @@
-// 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).
 //

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/hello_gte.c

@@ -18,8 +18,8 @@
 #	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/gte_hello.offsets.h"
 #include "hello_gte.h"
 

code/gte_hello/hello_gte_tape.c

@@ -3,7 +3,7 @@
 #	include "duffle/gen/duffle.offsets.h"
 #	include "duffle/atom_dsl.h"
 #	include "duffle/lottes_tape.h"
-#	include "tape_atom.metadata.h"
+#	include "duffle/word_count.metadata.h"
 #	include "gen/gte_hello.offsets.h"
 #	include "hello_gte.h"
 #endif
@@ -22,88 +22,66 @@ typedef Struct_(Binds_CubeTri) {
 	V3_S2* VertBase;
 	U4*    OtBase;
 };
-internal MipsAtom_(rbind_cube_g4_face) {
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
 	/* Pop 4 arguments from the tape directly into the workspace registers */
 	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)), 
 	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)), 
 	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)), 
 	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)), 
 	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
-	// 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) ───────── */
+MipsAtom_(cube_g4_face) atom_info(
+		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,
-
-	/* ── 12. Insert into Ordering Table (length = 8 words for Poly_G4) ──── */
 	mac_insert_ot_tag_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 */
@@ -116,14 +94,11 @@ typedef Struct_(Binds_FloorTri) {
 	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) {
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri)
+	, 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)), 
@@ -133,16 +108,13 @@ MipsAtom_(rbind_floor_f3_face) {
 	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),
+MipsAtom_(floor_f3_face) atom_info(
+	, 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,
  
@@ -151,7 +123,6 @@ MipsAtom_(floor_f3_face) {
 	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(),
 	
@@ -174,13 +145,10 @@ atom_label(floor_f3_face_exit)
 };
 
 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)),      
 	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),

readme.md

@@ -6,19 +6,27 @@ 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.
+
 [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 +65,4 @@ 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)

scripts/audit_lua_nesting.lua

@@ -0,0 +1,227 @@
+--- 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 `dir /b *.lua` (Windows) or `ls *.lua` (Unix).
+	local function is_dir(p)
+		local f = io.open(p, "r")
+		if    f then f:close() return false end
+		return true
+	end
+	local function list_lua(dir)
+		local out = {}
+		local cmd
+		if package.config:sub(1, 1) == "\\" then
+			cmd = 'dir /b "' .. dir .. '\\*.lua" 2>nul'
+		else
+			cmd = 'ls -1 "' .. dir .. '"/*.lua 2>/dev/null'
+		end
+		local p = io.popen(cmd)
+		if p then
+			for line in p:lines() do
+				if line:match("%.lua$") then
+					out[#out + 1] = dir .. "/" .. line
+				end
+			end
+			p:close()
+		end
+		return out
+	end
+
+	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

@@ -317,32 +317,14 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-# ps1-meta orchestrator. Replaces generate-TapeAtomOffsets +
-# generate-TapeAtomAnnotations with a single invocation. Dispatches
-# the 6 passes (word-counts / components / annotation / offsets /
-# static-analysis / report) in dependency-topological order.
-
-function any-stale {
-	param([Parameter(Mandatory=$true)][string[]]$sources,
-	      [Parameter(Mandatory=$true)][string]$metadata,
-	      [Parameter(Mandatory=$true)][string]$out_root)
-	if (-not (test-path $out_root)) { return $true }
-	$out_mtime  = (get-item $out_root).LastWriteTimeUtc
-	$src_mtime  = ($sources  | ForEach-Object { (get-item $_).LastWriteTimeUtc } | Measure-Object -Maximum).Maximum
-	$meta_mtime = (get-item $metadata).LastWriteTimeUtc
-	return ($src_mtime -gt $out_mtime) -or ($meta_mtime -gt $out_mtime)
-}
-
-function ps1-meta {
-	param(
+function ps1-meta { param(
 		[Parameter(Mandatory=$true)][string[]]$sources,
 		[Parameter(Mandatory=$true)][string]$metadata,
 		[string]$out_root = (join-path $path_build 'gen'),
-		[string[]]$passes  = @('--all')
+		[string[]]$passes = @('--all')
 	)
 	$script = join-path $path_scripts 'ps1_meta.lua'
-	write-host "ps1-meta  $($sources.Count) source(s), passes=$($passes -join ',')" `
-		-ForegroundColor Magenta
+	write-host "ps1-meta  $($sources.Count) source(s), passes=$($passes -join ',')" ` -ForegroundColor Magenta
 	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root)
 	foreach ($s in $sources) { $arg_list += @('--source', $s) }
 	& luajit $script @arg_list
@@ -357,17 +339,11 @@ 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')
-
-	if (any-stale -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')) {
-		ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
-	} else {
-		write-host "ps1-meta  all $($atom_sources.Count) source(s) up-to-date" `
-			-ForegroundColor DarkGray
-	}
+	ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -412,18 +388,14 @@ 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 +1,113 @@
+--- 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 this script's path can't be parsed (shouldn't happen — dofile/debug.getinfo always populates source),
+--- fall back to a defensive walk: starting from this script's directory, walk UP until we find a parent that
+--- contains a `scripts/` directory. The first match is the repo root.
+--- @return string|nil
+local function find_repo_root()
+	if package.loaded[CACHE_KEY] then return package.loaded[CACHE_KEY] end
+
+	local source = debug.getinfo(1, "S").source
+	-- Strip the leading `@` (Lua's dofile marker) and the trailing `/duffle_paths.lua` filename.
+	-- What remains is the directory containing this script, i.e. `<repo>/scripts/` (with trailing slash or not).
+	local scripts_dir = source and source:match("^@?(.*)[/\\]duffle_paths%.lua$")
+	if scripts_dir then
+		-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
+		local root = scripts_dir:gsub("scripts[\\/]?$", "")
+		root = root:gsub("\\", "/")
+		if root == "" then root = "./" end
+		if not root:match("/$") then root = root .. "/" end
+		package.loaded[CACHE_KEY] = root
+		return root
+	end
+
+	-- Defensive fallback: walk UP from this script's directory until we find a parent that contains `scripts/`.
+	-- In practice this branch never fires — debug.getinfo always returns a source for dofile()'d chunks.
+	local lfs = pcall(require, "lfs") and require("lfs") or nil
+	if lfs then
+		local dir = source and source:match("^@?(.*[/\\])") or "./"
+		dir = dir:gsub("\\", "/")
+		while dir and dir ~= "" do
+			local candidate_scripts = dir .. "scripts"
+			if lfs.attributes(candidate_scripts, "mode") == "directory" then
+				dir = dir:gsub("/$", "")
+				package.loaded[CACHE_KEY] = dir .. "/"
+				return dir .. "/"
+			end
+			local parent = dir:match("^(.*)/[^/]+/$")
+			if not parent then break end
+			dir = parent .. "/"
+		end
+	end
+
+	return nil
+end
+
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
+--- `package.cpath` (for `lpeg.dll`).
+---
+--- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$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/passes/offsets.lua

@@ -1,287 +1,215 @@
--- passes/offsets.lua
---
--- Generate <module>/gen/<basename>.offsets.h with branch offset
--- immediates for every atom_offset(F, T) reference in atom bodies.
--- Ported from scripts/tape_atom.offset_gen.meta.lua:148-389.
---
--- The branch offset regression we just fixed in commit 98e27c2 must
--- NOT return. The fix was in duffle.lua's split_top_level_commas +
--- tape_atom_annotation_pass.lua's compute_component_word_count.
--- word_count_eval.count_token_words preserves the fix.
---
--- THIS MODULE ALSO REQUIRES the recent fix to duffle.lua's
--- split_top_level_commas (the second-half of the 98e27c2 fix):
--- top-level comments must be appended to the previous token, not
--- stripped, so the emit path preserves `// trailing comment` text
--- for convert_line_comments_to_block to convert to `/* */`.
---
--- Coding standard: tabs (1/level), EmmyLua annotations, no regex.
+--- 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
 -- ════════════════════════════════════════════════════════════════════════════
 
-local script_path = arg and arg[0] or "?"
-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;" .. script_dir .. "?.lua;" .. package.path
-
-local duffle              = require("duffle")
-local trim                = duffle.trim
-local read_ident          = duffle.read_ident
-local is_space            = duffle.is_space
-local is_alpha            = duffle.is_alpha
-local is_alnum            = duffle.is_alnum
-local skip_ws_and_cmt     = duffle.skip_ws_and_cmt
-local skip_str_or_cmt     = duffle.skip_str_or_cmt
-local split_top_level_commas = duffle.split_top_level_commas
-local read_parens         = duffle.read_parens
-local read_braces         = duffle.read_braces
-local write_file          = duffle.write_file
-local dirname             = duffle.dirname
-local basename_no_ext     = duffle.basename_no_ext
-local ensure_dir          = duffle.ensure_dir
-
-local word_count_eval     = require("word_count_eval")
-local count_token_words   = word_count_eval.count_token_words
+-- 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
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Local helpers (ported from offset_gen.meta.lua lines 67-93)
+-- Constants
 -- ════════════════════════════════════════════════════════════════════════════
 
-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
+-- Marker-call identifiers inside atom bodies.
+local LABEL_MARKER    = "atom_label"
+local OFFSET_MARKER   = "atom_offset"
 
-local function to_upper(s) return s:upper() end
+-- Offset macro/enum naming prefixes (the emitted header uses these).
+local OFFSET_MACRO_PREFIX = "_atom_offset_"
+local OFFSET_ENUM_PREFIX  = "atom_offset_"
 
-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
+-- Column width for the `#define _atom_offset_F_T = N` alignment.
+local OFFSET_MACRO_COL = 44
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Marker-call helpers (ported from offset_gen.meta.lua lines 148-205)
+-- Type declarations
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Extract comma-separated identifier args from a parenthesized group
---- after a function-like macro call.
+--- @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 = skip_ws_and_cmt(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 = read_parens(token, arg_start)
+	local inner, after_paren = duffle.read_parens(token, arg_start)
+	-- scan: <marker>(<args>)
 
-	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)
+	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)
-			n = after
+			pos = after
 		else
-			n = n + 1
+			pos = pos + 1
 		end
-		n = skip_ws_and_cmt(inner, n)
-		if n <= len and inner:sub(n, n) == "," then n = n + 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
 
---- Scan a single token for atom_label/atom_offset markers, walking through
---- balanced groups transparently (so nested calls are found).
+-- (internal) Record a `atom_label(name)` marker — `at_pos` is the branch-free word position within the atom body.
+-- @param labels table<string, integer>
+-- @param args string[]
+-- @param at_pos integer
+local function record_label_marker(labels, args, at_pos)
+	if #args >= 1 then labels[args[1]] = at_pos end
+end
+
+-- (internal) Record a `atom_offset(tag, target)` marker.
+-- @param branches table[]  -- list of {pos=, target=, tag=}
+-- @param args string[]
+-- @param at_pos integer
+local function record_offset_marker(branches, args, at_pos)
+	if #args >= 2 then
+		table.insert(branches, { pos = at_pos, target = args[2], tag = args[1] })
+	end
+end
+
+--- Scan a single token for atom_label/atom_offset markers, walking through balanced groups transparently (so nested calls are found).
+--- @param token string
+--- @param at_pos integer    -- the branch-free word position of this token in the body
+--- @param labels table<string, integer>
+--- @param branches table[]
 local function scan_for_atom_markers(token, at_pos, labels, branches)
-	local 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 pos     = 1
+	local tok_len = #token
+	while pos <= tok_len do
+		pos = duffle.skip_ws_and_cmt(token, pos)
+		if pos > tok_len then break end
+		local ch = token:sub(pos, pos)
+		if duffle.is_alpha(ch) then
+			local ident, after = duffle.read_ident(token, pos)
+			if ident == LABEL_MARKER then
 				local args, after_paren = extract_ident_args(token, after)
-				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
+				record_label_marker(labels, args, at_pos)
+				pos = after_paren or after
+			elseif ident == OFFSET_MARKER then
 				local args, after_paren = extract_ident_args(token, after)
-				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
+				record_offset_marker(branches, args, at_pos)
+				pos = after_paren or after
 			else
-				i = after
+				pos = after
 			end
 		else
-			local nx = skip_str_or_cmt(token, i)
-			if nx > i then i = nx else i = i + 1 end
+			local nx = duffle.skip_str_or_cmt(token, pos)
+			pos = (nx > pos) and nx or (pos + 1)
 		end
 	end
 end
 
---- 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_space(c) then
-			i = i + 1
-		elseif c == "/" then
-			-- comment — skip past it (delegated to duffle.skip_str_or_cmt)
-			local nx = skip_str_or_cmt(tok, i)
-			if nx > i then i = nx else i = i + 1 end
-		else
-			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 or (i + 1)
-		end
-	end
-	return 0
+-- (internal) Count words emitted by the rest of `tok` after a marker call
+-- (the marker call itself emits 0 words, but the source pattern may bundle the marker with the next instruction on the same line,
+-- separated by no top-level comma).
+-- Returns the word count contributed by that rest.
+-- @param tok string
+-- @param word_counts table
+-- @return integer
+local function count_marker_rest(tok, word_counts)
+	-- duffle.find_marker_call_end returns the position PAST the closing `)` of the marker call
+	-- (or nil if `tok` isn't a marker call). Canonical impl in duffle.lua is faster than the
+	-- file-local copy that used to live here (byte-indexed, no `tok:sub` per char).
+	local marker_end = duffle.find_marker_call_end(tok)
+	if not marker_end or marker_end >= #tok then return 0 end
+	local rest = duffle.trim(tok:sub(marker_end))
+	if rest == "" then return 0 end
+	return count_token_words(rest, word_counts)
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Atom scanner (ported from offset_gen.meta.lua lines 245-321)
--- ════════════════════════════════════════════════════════════════════════════
-
---- Skip C qualifier keywords (static, const, etc.) and return the position
---- past the last qualifier.
-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
+-- (internal) Is this token a marker call (`atom_label` or `atom_offset`)?
+-- @param tok string
+-- @return boolean
+local function is_marker_token(tok)
+	local leading_ident = duffle.read_ident(tok, 1)
+	return leading_ident == LABEL_MARKER or leading_ident == OFFSET_MARKER
 end
 
---- Find every MipsAtom_(name) { ... } in a source.
-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
-		-- Find the brace after the parens.
-		local brace_pos = duffle.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 = duffle.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
-
--- ════════════════════════════════════════════════════════════════════════════
--- Per-atom body scan (ported from offset_gen.meta.lua lines 207-239)
--- ════════════════════════════════════════════════════════════════════════════
-
 --- Scan an atom body for labels + branches, count total words.
 --- Returns (labels, branches, total_words).
-local function scan_atom_body(body, word_counts)
+--- @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 _, 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
+	for _, t in ipairs(body_tokens) do
+		local tok = t.tok
+		if is_marker_token(tok) then
 			-- Marker call: record at the current pos, do NOT advance pos.
-			-- 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 = count_token_words(rest, word_counts)
-					pos = pos + rest_words
-				end
-			end
+			pos = pos + count_marker_rest(tok, word_counts)
 		else
 			local words = count_token_words(tok, word_counts)
 			scan_for_atom_markers(tok, pos, labels, branches)
@@ -292,10 +220,14 @@ local function scan_atom_body(body, word_counts)
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Offset computation + header generation (ported lines 327-383)
+-- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Compute branch offsets as (target_word - branch_word - 1).
+-- Compute branch offsets as `target_word - branch_word - 1`
+-- (the standard MIPS branch-immediate encoding).
+-- @param labels table<string, integer>
+-- @param branches table[]
+-- @return BranchOffset[]
 local function compute_offsets(labels, branches)
 	local results = {}
 	for _, br in ipairs(branches) do
@@ -303,17 +235,62 @@ local function compute_offsets(labels, branches)
 		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})
+		results[#results + 1] = { target = br.target, tag = br.tag, offset = target - br.pos - 1 }
 	end
 	return results
 end
 
---- Generate the per-source .offsets.h header.
+-- 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 = basename_no_ext(source_path)
+	local basename = duffle.basename_no_ext(source_path)
 
 	local lines = {}
-	local function add(s) table.insert(lines, s) end
+	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)
@@ -323,28 +300,7 @@ local function generate_header(source_path, atoms_data)
 	add("")
 	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
+		emit_atom_offsets(add, atom)
 	end
 	add("#pragma endregion " .. basename)
 	add("")
@@ -352,13 +308,58 @@ local function generate_header(source_path, atoms_data)
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- M.run — orchestrator entry
+-- M — module exports
 -- ════════════════════════════════════════════════════════════════════════════
 
---- @class M
-
 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)
@@ -367,29 +368,13 @@ function M.run(ctx)
 	local warnings = {}
 
 	for _, src in ipairs(ctx.sources) do
-		local atoms = find_atoms(src.text)
-		if #atoms > 0 then
-			local atoms_data = {}
-			for _, atom in ipairs(atoms) do
-				local labels, branches, total = scan_atom_body(atom.body, ctx.shared.word_counts)
-				local offsets = compute_offsets(labels, branches)
-				table.insert(atoms_data, {
-					name        = atom.name,
-					total_words = total,
-					offsets     = offsets,
-				})
-			end
-
-			local out_path = src.dir .. "/gen/" .. basename_no_ext(src.dir) .. ".offsets.h"
-			if not ctx.dry_run then
-				ensure_dir(dirname(out_path))
-				write_file(out_path, generate_header(src.path, atoms_data))
-			end
-			table.insert(outputs, { offsets_h = out_path })
+		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
+return M

scripts/passes/report.lua

@@ -1,124 +1,311 @@
--- passes/report.lua
---
--- Render the per-project summary (build/gen/annotation_validation.txt)
--- + the per-source annotation reports (build/gen/<basename>.annotations.txt).
--- Aggregates errors + warnings from upstream annotation pass results.
---
--- The annotation pass stashes its per-source results in ctx.flags._annot_results
--- (set by passes/annotation.lua). This report pass renders them.
---
--- Coding standard: tabs (1/level), EmmyLua annotations, no regex.
+--- 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
 -- ════════════════════════════════════════════════════════════════════════════
 
-local script_path = arg and arg[0] or "?"
-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
+-- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
+-- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
+-- Bootstrap: see `ps1_meta.lua` for the rationale.
+-- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works
+-- both standalone and when require'd from the orchestrator.
+-- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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
-local script_dir = last_sep == 0 and "./" or script_path:sub(1, last_sep)
-package.path   = script_dir .. "../?.lua;" .. script_dir .. "../?/init.lua;" .. script_dir .. "?.lua;" .. package.path
 
-local duffle        = require("duffle")
-local ensure_dir    = duffle.ensure_dir
-local write_file    = duffle.write_file
+-- (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
 
--- ════════════════════════════════════════════════════════════════════════════
--- Per-source annotation report (ported from tape_atom_annotation_pass.lua:1411-1486)
--- ════════════════════════════════════════════════════════════════════════════
+-- (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
 
-local function render_source_report(source_path, result)
+-- (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
+
+--- 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("========================================================")
-	add("ANNOTATION PASS — " .. source_path)
-	add("========================================================")
-	add("")
-	add(string.format("Atoms: %d   Annotations: %d   Pragmas: %d   Binds structs: %d   Macro decls: %d",
-		#result.atoms, #result.annots,
-		(result.pragmas and #result.pragmas or 0),
-		#result.binds, #result.macros))
+	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("")
 
-	add("── Atoms ────────────────────────────────────────────────")
-	for _, a in ipairs(result.atoms) do
-		add(string.format("  MipsAtom_(%s)   line %d", a.name, a.line))
-	end
+	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("")
 
-	add("── Annotations ──────────────────────────────────────────")
-	for _, a in ipairs(result.annots) do
-		if a.error then
-			add(string.format("  ✗ line %d  %s  [ERROR: %s]", a.line, a.macro or "?", a.error))
-		else
-			local line = string.format("  %s  line %d  %s  phase=%s",
-				a.kind == "work" and "●" or (a.kind == "bind" and "◆" or "○"),
-				a.line, a.name, a.phase or a.kind)
-			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
-			add(line)
-		end
-	end
-	add("")
-
-	add("── Binds_* structs ──────────────────────────────────────")
-	for _, b in ipairs(result.binds) do
-		add(string.format("  %s   line %d   %d bytes", b.name, b.line, b.bytes))
-		for _, f in ipairs(b.fields) do
-			add(string.format("    +%2d: %s", f.offset, f.name))
-		end
-	end
-	add("")
-
-	add("── Macro word-count declarations ─────────────────────────")
-	for _, m in ipairs(result.macros) do
-		add(string.format("  %s  line %d  words=%d", m.name, m.line, m.words))
-	end
-	add("")
-
-	add("── Atom pragmas (resource / region / group / cadence / async) ─")
-	if not result.pragmas or #result.pragmas == 0 then add("  (none)") end
-	for _, p in ipairs(result.pragmas or {}) do
-		local kvs = {}
-		for k, v in pairs(p.attrs) do kvs[#kvs + 1] = k .. "=" .. v end
-		table.sort(kvs)
-		add(string.format("  ◇ line %d  %s  {%s}", p.line, p.name, table.concat(kvs, ", ")))
-	end
-	add("")
-
-	add("── Errors ──────────────────────────────────────────────")
-	if #result.errors == 0 then add("  (none)") end
-	for _, e in ipairs(result.errors) do
-		add(string.format("  ✗ line %d  %s", e.line, e.msg))
-	end
-	add("")
-
-	add("── Warnings ────────────────────────────────────────────")
-	if #result.warnings == 0 then add("  (none)") end
-	for _, w in ipairs(result.warnings) do
-		add(string.format("  ⚠ line %d  %s", w.line, w.msg))
-	end
-	add("")
+	render_module_atoms_section(add, results)
+	render_module_annots_section(add, results)
+	render_module_binds_section(add, results)
+	render_module_macros_section(add, results)
+	render_module_errors_section(add, results, total_errors)
+	render_module_warnings_section(add, results, total_warnings)
 
 	return table.concat(lines, "\n") .. "\n"
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Per-project summary (ported from tape_atom_annotation_pass.lua:1488-1528)
+-- 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
@@ -128,9 +315,9 @@ local function render_project_report(all_results)
 		total_warnings = total_warnings + #r.warnings
 	end
 
-	add("========================================================")
+	add(RULE_THICK)
 	add("ANNOTATION VALIDATION — project summary")
-	add("========================================================")
+	add(RULE_THICK)
 	add("")
 	add(string.format("Atoms:        %d", total_atoms))
 	add(string.format("Annotations:  %d", total_annots))
@@ -145,7 +332,8 @@ local function render_project_report(all_results)
 		add("Per-source error counts:")
 		for _, r in ipairs(all_results) do
 			if #r.errors > 0 then
-				add(string.format("  %s : %d error(s)", r.source, #r.errors))
+				local src_name = source_basename(r.source)
+				add(string.format("  %s : %d error(s)", src_name, #r.errors))
 			end
 		end
 		add("")
@@ -155,13 +343,59 @@ local function render_project_report(all_results)
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- M.run — orchestrator entry
+-- Orchestration helpers
 -- ════════════════════════════════════════════════════════════════════════════
 
---- @class M
+-- (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)
@@ -169,39 +403,41 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- The annotation pass stashes per-source results in ctx.flags._annot_results.
-	-- Render each as build/gen/<basename>.annotations.txt and aggregate into
-	-- build/gen/annotation_validation.txt.
-	local annot_results = (ctx.flags and ctx.flags._annot_results) or {}
+	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)
 
-	-- Render per-source reports.
-	for _, entry in ipairs(annot_results) do
-		local src = entry.source
-		local result = entry.result
-		local out_path = ctx.out_root .. "/" .. src.basename .. ".annotations.txt"
-		if not ctx.dry_run then
-			ensure_dir(ctx.out_root)
-			write_file(out_path, render_source_report(src.path, result))
+	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
-		table.insert(outputs, { annotations_txt = out_path })
 	end
 
-	-- Render project summary.
-	if not ctx.dry_run then
-		-- The project report references each source by its absolute path.
-		-- Augment the entries with a .source field for the per-source error counts.
-		local all_results = {}
-		for _, entry in ipairs(annot_results) do
-			entry.result.source = entry.source.path
-			table.insert(all_results, entry.result)
-		end
-		ensure_dir(ctx.out_root)
+	if not ctx.dry_run and #all_results_for_summary > 0 then
 		local summary_path = ctx.out_root .. "/annotation_validation.txt"
-		write_file(summary_path, render_project_report(all_results))
-		table.insert(outputs, { summary_txt = summary_path })
+		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
+return M

scripts/passes/scan_source.lua

@@ -0,0 +1,484 @@
+--- passes/scan_source.lua — Source pre-scan pass (the "mega entity" pass).
+---
+--- Single source-walk pass that produces the fat `SourceScan` payload consumed by all downstream passes. Walks each `ctx.sources` entry once, 
+--- extracting every construct type the metaprograms need:
+---
+---   MipsAtom_              (kind = "atom", with optional atom_info inner)
+---   MipsAtomComp_          (kind = "comp_bare")
+---   MipsAtomComp_Proc_     (kind = "comp_proc", body inside last {})
+---   MipsCode code_<name>   (kind = "raw_atom", offsets pass only)
+---   typedef Struct_(Binds_X) { fields }
+---   #pragma mac_X tape_atom words=N   +   _Pragma("...")
+---
+--- The result is attached to each `src.scan` so downstream passes can read from `src.scan.atoms` / `src.scan.binds` / etc. without re-walking the source.
+--- This is the first pass in the dep graph (no deps).
+--- Every other pass that reads source structure depends on this one — see `ps1_meta.lua :: PASSES`.
+---
+--- **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")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class SourceScan
+--- @field atoms      AtomEntry[]     -- MipsAtom_ + MipsAtomComp_ + MipsAtomComp_Proc_
+--- @field raw_atoms  AtomEntry[]     -- MipsCode code_<name> { body } (offsets pass only)
+--- @field binds      BindsEntry[]    -- typedef Struct_(Binds_X) { fields } (fields pre-parsed)
+--- @field atom_infos AtomInfoEntry[] -- MipsAtom_(name) atom_info(...) (sub-calls pre-parsed)
+--- @field macros     MacroEntry[]    -- #pragma mac_X tape_atom words=N  +  _Pragma("...")
+--- @field line_of    fun(pos: integer): integer  -- shared LineIndex closure
+
+--- @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 (set by this pass)
+
+--- @class PassCtx
+--- @field sources            SourceFile[]
+--- @field metadata_path      string
+--- @field shared             table
+--- @field out_root           string
+--- @field project_root       string
+--- @field upstream           table<string, table>
+--- @field flags              table
+--- @field dry_run            boolean
+--- @field verbose            boolean
+
+--- @class PassResult
+--- @field outputs  table[]
+--- @field errors   table[]
+--- @field warnings table[]
+
+--- @class AtomEntry
+--- @field line       integer
+--- @field name       string         -- atom name (for components: without ac_ prefix)
+--- @field body       string         -- brace-delimited body (without the braces)
+--- @field body_off   integer        -- char offset of body[1] in source
+--- @field kind       string         -- "atom" | "comp_bare" | "comp_proc" | "raw_atom"
+--- @field raw_name   string         -- un-stripped name (for components: with ac_ prefix)
+--- @field ident_pos  integer        -- position of the MipsAtom_/MipsAtomComp_ ident start
+--- @field after_paren integer       -- position past the closing paren
+--- @field args       string|nil     -- populated by components pass (backward lookup)
+--- @field comment    string|nil     -- populated by components pass (backward lookup)
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Local helpers
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- C qualifier keywords that may precede a MipsAtom_ / MipsCode declaration.
+-- (typedef is NOT a qualifier here — it's a separate construct (`typedef Struct_(Binds_X) { ... };`)
+-- and must be read as an ident so the typedef check below can match it.)
+local QUALIFIER_KEYWORDS = {
+	["static"]   = true, ["const"] = true, ["volatile"] = true, ["extern"] = true,
+	["register"] = true, ["auto"]  = true, ["inline"]   = true,
+	["internal"] = true, ["LP_"]   = true, ["global"]   = true, ["gkknown"] = true,
+}
+
+-- Parse the U4 fields from a Binds_X body. Returns (fields, byte_count).
+local function scan_binds_fields(body)
+	local fields   = {}
+	local byte_off = 0
+	local body_pos = 1
+	while body_pos <= #body do
+		body_pos = duffle.skip_ws_and_cmt(body, body_pos)
+		if body_pos > #body then break end
+		local type_ident, type_end = duffle.read_ident(body, body_pos)
+		if not type_ident then
+			body_pos = body_pos + 1
+		elseif type_ident == "U4" then
+			local field_ident, field_end = duffle.read_ident(body, duffle.skip_ws_and_cmt(body, type_end))
+			if field_ident then
+				fields[#fields + 1] = { name = field_ident, offset = byte_off }
+				byte_off = byte_off + 4
+			end
+			body_pos = field_end or (type_end + 1)
+		else
+			body_pos = type_end + 1
+		end
+	end
+	return fields, byte_off
+end
+
+-- Parse the register list from inside `atom_reads(...)` or `atom_writes(...)`.
+local function scan_reg_list(sub_inner)
+	local regs = {}
+	local sub_inner_pos = 1
+	while sub_inner_pos <= #sub_inner do
+		sub_inner_pos = duffle.skip_ws_and_cmt(sub_inner, sub_inner_pos)
+		if sub_inner_pos > #sub_inner then break end
+		local reg_ident, reg_end = duffle.read_ident(sub_inner, sub_inner_pos)
+		if reg_ident then
+			regs[#regs + 1] = duffle.trim(reg_ident)
+			sub_inner_pos = reg_end
+		else
+			sub_inner_pos = sub_inner_pos + 1
+		end
+		if sub_inner_pos > #sub_inner then break end
+		if sub_inner:sub(sub_inner_pos, sub_inner_pos) == "," then sub_inner_pos = sub_inner_pos + 1 end
+	end
+	return regs
+end
+
+-- Parse the sub-calls inside `atom_info(atom_bind(...), atom_reads(...), atom_writes(...))`.
+-- Returns (binds, reads, writes).
+local function scan_atom_info_subcalls(info_inner)
+	local binds, reads, writes = nil, nil, nil
+	local sub_pos = 1
+	while sub_pos <= #info_inner do
+		sub_pos = duffle.skip_ws_and_cmt(info_inner, sub_pos)
+		if sub_pos > #info_inner then break end
+		local sub_ident, sub_end = duffle.read_ident(info_inner, sub_pos)
+		if not sub_ident then
+			sub_pos = sub_pos + 1
+		elseif sub_ident == "atom_bind" then
+			local sub_open = duffle.skip_ws_and_cmt(info_inner, sub_end)
+			if info_inner:sub(sub_open, sub_open) == "(" then
+				local sub_inner, sub_after2 = duffle.read_parens(info_inner, sub_open)
+				-- scan: atom_bind(<Binds_X>)
+				binds = duffle.trim(sub_inner)
+				sub_pos = sub_after2
+			else
+				sub_pos = sub_open + 1
+			end
+		elseif sub_ident == "atom_reads" or sub_ident == "atom_writes" then
+			local kind = sub_ident
+			local sub_open = duffle.skip_ws_and_cmt(info_inner, sub_end)
+			if info_inner:sub(sub_open, sub_open) == "(" then
+				local sub_inner, sub_after2 = duffle.read_parens(info_inner, sub_open)
+				-- scan: atom_reads(<regs>)  OR  atom_writes(<regs>)
+				local regs = scan_reg_list(sub_inner)
+				if kind == "atom_reads" then reads = regs else writes = regs end
+				sub_pos = sub_after2
+			else
+				sub_pos = sub_open + 1
+			end
+		else
+			sub_pos = sub_end
+		end
+	end
+	return binds, reads, writes
+end
+
+-- Skip C qualifier keywords and return the position past the last one.
+local function scan_skip_qualifiers(source, pos)
+	while true do
+		pos = duffle.skip_ws_and_cmt(source, pos)
+		local ident, after = duffle.read_ident(source, pos)
+		if not ident then return pos end
+		if QUALIFIER_KEYWORDS[ident] then pos = after else return pos end
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- The single source walker
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Single-pass source scan. Walks the source ONCE and extracts every construct type the metaprogram passes need.
+--- Returns a fat SourceScan table. Each pass filters from this payload instead of re-walking the source.
+--- @param source string
+--- @return table  -- SourceScan { atoms, raw_atoms, binds, atom_infos, macros, line_of }
+local function scan_source(source)
+	local line_of = duffle.LineIndex(source)
+	local atoms       = {}
+	local raw_atoms   = {}
+	local binds       = {}
+	local atom_infos  = {}
+	local macros      = {}
+	local pos         = 1
+	local src_len     = #source
+
+	while pos <= src_len do
+		pos = duffle.skip_ws_and_cmt(source, pos)
+		if pos > src_len then break end
+
+		-- Skip preprocessor directives (#define / #include / #pragma / etc).
+		-- _Pragma is an operator (not a directive) — it doesn't start with #.
+		local pp_pos = duffle.skip_preprocessor_line(source, pos)
+		if pp_pos then pos = pp_pos; goto continue end
+
+		-- Skip C qualifiers (static, const, etc.) that may precede a declaration.
+		pos = scan_skip_qualifiers(source, pos)
+		if pos > src_len then break end
+
+		local ident, ident_end = duffle.read_ident(source, pos)
+		-- scan: <ident>
+		if not ident then pos = pos + 1; goto continue end
+
+		-- ── MipsAtom_ / MipsAtomComp_ / MipsAtomComp_Proc_ ──
+		if ident == "MipsAtom_" or ident == "MipsAtomComp_" or ident == "MipsAtomComp_Proc_" then
+			local is_atom = ident == "MipsAtom_"
+			local is_comp = ident == "MipsAtomComp_"
+			local is_proc = ident == "MipsAtomComp_Proc_"
+			local kind = is_atom and "atom" or (is_comp and "comp_bare" or "comp_proc")
+			local open_paren = duffle.skip_ws_and_cmt(source, ident_end)
+			if source:sub(open_paren, open_paren) ~= "(" then pos = open_paren + 1; goto continue end
+
+			local inner, after_paren = duffle.read_parens(source, open_paren)
+			-- scan: <ident>(<args>)
+
+			if is_proc then
+				-- MipsAtomComp_Proc_(name, { body }) — body is inside the LAST { } in args.
+				local last_brace_pos
+				for search_pos = #inner, 1, -1 do
+					if inner:sub(search_pos, search_pos) == "{" then last_brace_pos = search_pos; break end
+				end
+				if last_brace_pos then
+					local depth = 1
+					local inner_pos = last_brace_pos + 1
+					while inner_pos <= #inner and depth > 0 do
+						local c = inner:byte(inner_pos)
+						if c == 123 then depth = depth + 1; inner_pos = inner_pos + 1
+						elseif c == 125 then depth = depth - 1; if depth == 0 then break end; inner_pos = inner_pos + 1
+						elseif c == 40 then local _, a = duffle.read_parens(inner, inner_pos); inner_pos = a
+						elseif c == 91 then local _, a = duffle.read_brackets(inner, inner_pos); inner_pos = a
+						elseif c == 34 or c == 39 then inner_pos = duffle.skip_str_or_cmt(inner, inner_pos) + 1
+						else inner_pos = inner_pos + 1 end
+					end
+					if depth == 0 then
+						-- scan: <ident>(<name>, { <body> })
+						local name_match = inner:match("^%s*([%w_]+)")
+						local raw_name = name_match or "?"
+						-- Strip "ac_" prefix for component names (components pass convention).
+						local name = raw_name
+						if #raw_name > 3 and raw_name:sub(1, 3) == "ac_" then
+							name = raw_name:sub(4)
+						end
+						local body = inner:sub(last_brace_pos + 1, inner_pos - 1)
+						local body_off = open_paren + 1 + last_brace_pos
+						atoms[#atoms + 1] = {
+							line = line_of(pos), name = name, body = body, body_off = body_off + 1,
+							kind = kind, raw_name = raw_name,
+							ident_pos = pos, after_paren = after_paren,
+							args = nil, comment = nil,
+						}
+					end
+				end
+				pos = after_paren
+			else
+				-- MipsAtom_(name) { body }  OR  MipsAtomComp_(name) { body }
+				local name_start = 1
+				while name_start <= #inner and inner:sub(name_start, name_start):match("[%s]") do name_start = name_start + 1 end
+				local name_end = name_start
+				while name_end <= #inner and inner:sub(name_end, name_end):match("[%w_]") do name_end = name_end + 1 end
+				local raw_name = inner:sub(name_start, name_end - 1)
+				-- scan: <ident>(<name>)
+				if raw_name ~= "" then
+					local brace = duffle.scan_to_char(source, "{", after_paren)
+					-- scan: <ident>(<name>) {
+					if brace then
+						local body, after_brace = duffle.read_braces(source, brace)
+						-- scan: <ident>(<name>) { <body> }
+						-- Strip "ac_" prefix for component names (components pass convention).
+						local disp_name = raw_name
+						if is_comp and #raw_name > 3 and raw_name:sub(1, 3) == "ac_" then
+							disp_name = raw_name:sub(4)
+						end
+						atoms[#atoms + 1] = {
+							line = line_of(pos), name = disp_name, body = body, body_off = brace + 1,
+							kind = kind, raw_name = raw_name,
+							ident_pos = pos, after_paren = after_paren,
+							args = nil, comment = nil,
+						}
+						pos = after_brace
+					else
+						pos = open_paren + 1
+					end
+				else
+					pos = open_paren + 1
+				end
+			end
+
+			-- For MipsAtom_ entries: check if atom_info(...) follows.
+			if is_atom then
+				local lookahead = duffle.skip_ws_and_cmt(source, after_paren)
+				local look_ident, look_end = duffle.read_ident(source, lookahead)
+				-- scan: MipsAtom_(<name>) <look_ident>
+				if look_ident == "atom_info" then
+					local info_open = duffle.skip_ws_and_cmt(source, look_end)
+					if source:sub(info_open, info_open) == "(" then
+						local info_inner, info_after = duffle.read_parens(source, info_open)
+						-- scan: MipsAtom_(<name>) atom_info(<binds>, <reads>, <writes>)
+						-- Find the atom name from the just-parsed atom entry (last one added).
+						local last_atom = atoms[#atoms]
+						local atom_name = last_atom and last_atom.raw_name or "?"
+						local ai_binds, ai_reads, ai_writes = scan_atom_info_subcalls(info_inner)
+						atom_infos[#atom_infos + 1] = {
+							atom_name = atom_name, binds = ai_binds,
+							reads = ai_reads or {}, writes = ai_writes or {},
+							info_line = line_of(lookahead),
+						}
+						-- Don't advance pos past info_after — the body { ... } still needs to be skipped
+						-- by the brace scan below. But if there's no body (forward decl), advance.
+						local body_brace = duffle.scan_to_char(source, "{", info_after)
+						if body_brace then
+							local _, after_body = duffle.read_braces(source, body_brace)
+							pos = after_body
+						else
+							pos = info_after
+						end
+					end
+				end
+			end
+			goto continue
+		end
+
+		-- ── MipsCode code_<name> { body } (raw atom form — offsets pass only) ──
+		if ident == "MipsCode" then
+			local next_pos = duffle.skip_ws_and_cmt(source, ident_end)
+			local next_ident, next_after = duffle.read_ident(source, next_pos)
+			-- scan: MipsCode <next_ident>
+			if next_ident and #next_ident > 5 and next_ident:sub(1, 5) == "code_" then
+				local atom_name = next_ident:sub(6)
+				-- scan: MipsCode code_<name>
+				local brace_pos = duffle.scan_to_char(source, "{", next_after)
+				-- scan: MipsCode code_<name> {
+				if brace_pos then
+					local body, after_brace = duffle.read_braces(source, brace_pos)
+					-- scan: MipsCode code_<name> { <body> }
+					raw_atoms[#raw_atoms + 1] = {
+						line = line_of(pos), name = atom_name, body = body, body_off = brace_pos + 1,
+						kind = "raw_atom", raw_name = atom_name,
+					}
+					pos = after_brace
+					goto continue
+				end
+			end
+			pos = ident_end
+			goto continue
+		end
+
+		-- ── typedef Struct_(Binds_X) { fields } ──
+		if ident == "typedef" then
+			local after_typedef = duffle.skip_ws_and_cmt(source, ident_end)
+			local id2, id2_end = duffle.read_ident(source, after_typedef)
+			-- scan: typedef <id2>
+			if id2 == "Struct_" then
+				local open_paren = duffle.skip_ws_and_cmt(source, id2_end)
+				if source:sub(open_paren, open_paren) == "(" then
+					local inner, after_paren = duffle.read_parens(source, open_paren)
+					-- scan: typedef Struct_(<name>)
+					local name = duffle.trim(inner)
+					local brace = duffle.scan_to_char(source, "{", after_paren)
+					-- scan: typedef Struct_(<name>) {
+					if brace then
+						local body, after_brace = duffle.read_braces(source, brace)
+						-- scan: typedef Struct_(<name>) { <fields> }
+						if name:sub(1, 6) == "Binds_" then
+							local fields, byte_off = scan_binds_fields(body)
+							binds[#binds + 1] = { line = line_of(pos), name = name, fields = fields, bytes = byte_off }
+						end
+						pos = after_brace
+						goto continue
+					end
+					pos = open_paren + 1
+					goto continue
+				end
+				pos = id2_end or (after_typedef + 1)
+				goto continue
+			end
+			pos = ident_end
+			goto continue
+		end
+
+		-- ── _Pragma("mac_X tape_atom words=N") (operator form) ──
+		if ident == "_Pragma" then
+			local open_paren = duffle.skip_ws_and_cmt(source, ident_end)
+			if source:sub(open_paren, open_paren) == "(" then
+				local str, str_end = duffle.read_parens(source, open_paren)
+				-- scan: _Pragma(<string>)
+				str = duffle.trim(str)
+				if str:sub(1, 1) == '"' and str:sub(-1) == '"' then
+					local inner = str:sub(2, -2)
+					local space = duffle.find_byte(inner, 32, 1)
+					if space then
+						local name = inner:sub(1, space - 1)
+						local rest = inner:sub(space + 1)
+						local eq = duffle.find_byte(rest, 61, 1)
+						if eq then
+							local key = duffle.trim(rest:sub(1, eq - 1))
+							local val = duffle.trim(rest:sub(eq + 1))
+							if key == "tape_atom words" or key == "words" then
+								macros[#macros + 1] = { line = line_of(pos), name = name, words = tonumber(val) or 0 }
+							end
+						end
+					end
+				end
+				pos = str_end
+				goto continue
+			end
+			pos = open_paren + 1
+			goto continue
+		end
+
+		-- ── #pragma mac_X tape_atom words=N (directive form) ──
+		-- (preprocessor skip above handles # lines, but pragma is an ident here
+		--  only if it appeared without a leading # — which happens when the
+		--  preprocessor skip didn't fire because the # was on a previous line.
+		--  The annotation pass handles this via its own skip_preprocessor_line,
+		--  but scan_source handles it here by checking the ident.)
+		if ident == "pragma" then
+			-- This shouldn't normally fire — #pragma lines are skipped by
+			-- skip_preprocessor_line above. If we get here, it's a _Pragma
+			-- variant or a non-#-prefixed pragma. Just advance.
+			pos = ident_end
+			goto continue
+		end
+
+		-- ── Unrecognized ident — advance past it ──
+		pos = ident_end
+
+		::continue::
+	end
+
+	return {
+		atoms      = atoms,
+		raw_atoms  = raw_atoms,
+		binds      = binds,
+		atom_infos = atom_infos,
+		macros     = macros,
+		line_of    = line_of,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class M
+
+local M = {}
+
+--- Walk each source once and attach the fat SourceScan payload to `src.scan`.
+--- No output files; this is a pure in-memory pre-processing pass.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	for _, src in ipairs(ctx.sources) do
+		src.scan = scan_source(src.text)
+		-- Pre-tokenize each atom body once (plex: single source of truth).
+		-- Downstream passes (offsets, word-counts, components, static-analysis) read from
+		-- `atom.body_tokens` instead of calling `split_top_level_commas` / `tokenize_body` independently.
+		-- The tokens are memoized in duffle.lua's cache, so re-access is O(1).
+		for _, atom in ipairs(src.scan.atoms) do
+			atom.body_tokens = duffle.tokenize_body(atom.body)
+		end
+		for _, atom in ipairs(src.scan.raw_atoms or {}) do
+			atom.body_tokens = duffle.tokenize_body(atom.body)
+		end
+	end
+	return { outputs = {}, errors = {}, warnings = {} }
+end
+
+return M

scripts/passes/word_count_eval.lua

@@ -0,0 +1,213 @@
+--- 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.
+-- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
+-- Bootstrap: see `ps1_meta.lua` for the rationale.
+-- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
+-- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Windows separator chars — used to convert `dir` output (which uses `\`) into POSIX paths (which our scripts expect).
+local PATH_SEP_BACKSLASH = "\\"
+local PATH_SEP_FORWARD   = "/"
+
+-- Fallback glob command (subprocess). Used when `lfs` (LuaFileSystem) is not available.
+-- Scoped to `code\` to avoid walking `.git/`, `toolchain/`, `build/`, etc.
+local DIR_GLOB_CMD = 'dir /b /s "%s\\code\\%s" 2>nul'
+
+-- Try to load lfs (LuaFileSystem). If available, scan_dir uses native directory enumeration (~2ms)
+-- instead of spawning `dir /b /s` as a subprocess (~56ms). Built by update_deps.ps1 into toolchain/lfs/lfs.dll.
+local lfs = pcall(require, "lfs") and require("lfs") or nil
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @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                                           │
+-- └────────────────────────────────────────────────────────────────────┘
+
+--- Recursively scan a directory for files matching a glob suffix.
+--- No regex per the no_regex constraint — uses plain byte matching via `dir /b /s` on Windows.
+---
+--- The `.macs.h` files produced by the components pass always live at `<project_root>/<module>/gen/`.
+--- We can shortcut the `dir /b /s` walk by listing modules first (one `dir /b /ad`), then walking each `<module>/gen/`
+--- (one `dir /b` per module, no recursion). 
+--- For projects with 2 modules and 0 .macs.h files, this drops the cost from ~52ms 
+--- (full recursive walk of the entire project tree) to ~5ms.
+---
+--- @param dir string        -- directory to scan (absolute or relative)
+--- @param suffix string     -- file pattern, e.g. "*.macs.h"
+--- @return string[]
+-- Cache the scan_dir result per (dir, suffix) in package.loaded. 
+-- Each `io.popen` call on Windows is ~50-100ms of subprocess overhead, so caching the result saves a fixed cost on every build.
+-- The cache persists for the lifetime of the Lua process (cleared when ps1_meta.lua exits).
+-- If a build removes/creates .macs.h files mid-process, the caller can invalidate by calling `M._invalidate_scan_cache()`.
+local SCAN_CACHE_KEY = "__word_count_eval_scan_cache__"
+
+--- Scan `code/` for files matching `suffix` (e.g. `*.macs.h`).
+--- Uses `lfs` (LuaFileSystem) when available — native directory enumeration at ~2ms.
+--- Falls back to `dir /b /s` subprocess (~56ms) when `lfs` is not compiled.
+---
+--- @param dir string        -- project root directory
+--- @param suffix string     -- file pattern, e.g. "*.macs.h"
+--- @return string[]
+function M.scan_dir(dir, suffix)
+	local key = dir .. "\0" .. suffix
+
+	local cache = package.loaded[SCAN_CACHE_KEY]
+	if    cache and cache[key] then return cache[key] end
+
+	local results = {}
+
+	if lfs then
+		-- Native walk: list code/<module>/gen/ for matching files. Zero subprocess spawns.
+		local code_dir = dir .. "/code"
+		if lfs.attributes(code_dir, "mode") == "directory" then
+			for mod_name in lfs.dir(code_dir) do
+				if mod_name ~= "." and mod_name ~= ".." then
+					local gen_path = code_dir .. "/" .. mod_name .. "/gen"
+					if lfs.attributes(gen_path, "mode") == "directory" then
+						for fname in lfs.dir(gen_path) do
+							if fname:match("%.macs%.h$") then
+								results[#results + 1] = gen_path .. "/" .. fname
+							end
+						end
+					end
+				end
+			end
+		end
+	else
+		-- Fallback: single `dir /b /s` subprocess scoped to code\.
+		local pipe = io.popen(DIR_GLOB_CMD:format(dir, suffix))
+		if pipe then
+			for raw_line in pipe:lines() do
+				results[#results + 1] = raw_line:gsub(PATH_SEP_BACKSLASH, PATH_SEP_FORWARD)
+			end
+			pipe:close()
+		end
+	end
+
+	-- Cache the result (including empty results).
+	cache      = cache or {}
+	cache[key] = results
+	package.loaded[SCAN_CACHE_KEY] = cache
+
+	return results
+end
+
+--- Invalidate the scan cache (call after creating new .macs.h files in the same Lua process — usually not needed).
+function M._invalidate_scan_cache() package.loaded[SCAN_CACHE_KEY] = nil end
+
+-- ┌────────────────────────────────────────────────────────────────────┐
+-- │ Pass entry: M.run(ctx) — "word-counts" pass                        │
+-- └────────────────────────────────────────────────────────────────────┘
+
+--- Load metadata.h + scan for existing *.macs.h files into ctx.shared.word_counts.
+--- 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/ps1_meta.lua

@@ -1,84 +1,117 @@
--- ps1_meta.lua
---
--- Orchestrator entry point for the tape-atom metaprogram pipeline.
--- Dispatches to pass modules under scripts/passes/, resolving dependencies
--- topologically. Single CLI surface (`--<pass>` flags + auto-dep + --dry-run).
---
--- Coding standard: tabs (1/level), EmmyLua annotations, no regex,
--- Lua 5.3 compatible.
-
+--- ps1_meta.lua — Orchestrator entry point for the tape-atom metaprogram pipeline.
+---
+--- Dispatches to pass modules under `scripts/passes/`, resolving
+--- dependencies topologically (Kahn's algorithm + cycle detection).
+--- Single CLI surface (`--<pass>` flags + auto-dep expansion + --dry-run).
+---
+--- **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 payload. 
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- Lua 5.3 compatible.
+---
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
-local script_path = arg and arg[0] or "?"
-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
+-- Bootstrap: load `duffle_paths.lua` via `arg[0]` (this script's own path).
+-- 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 duffle = dofile((arg[0]:match("(.*[/\\])") or "./") .. "duffle_paths.lua")
 
-local duffle = require("duffle")
-local dirname          = duffle.dirname
-local basename_no_ext  = duffle.basename_no_ext
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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 desc     string         -- human description (used by --help + ASCII graph)
---- @field out      PassOutput[]   -- output paths (used by --dry-run + report)
+--- @field module string       -- module name passed to require()
+--- @field kind   string       -- "shared" | "header-output" | "validation" | "report"
+--- @field deps   string[]     -- names of upstream passes
+--- @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
---- @field text     string
---- @field dir      string
---- @field basename string
+--- @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[]
---- @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 dry_run            boolean
---- @field verbose            boolean
+--- @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  table[]
---- @field errors   table[]
---- @field warnings table[]
+--- @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
---- @field verbose       boolean
+--- @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
 -- ════════════════════════════════════════════════════════════════════════════
 
 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 = "word_count_eval",
+		module = "passes.word_count_eval",
 		kind   = "shared",
 		deps   = {},
 		desc   = "Build the shared metadata table (metadata.h + .macs.h)",
@@ -87,14 +120,14 @@ local PASSES = {
 	components = {
 		module = "passes.components",
 		kind   = "header-output",
-		deps   = {"word-counts"},
+		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   = {"word-counts"},
+		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" },
@@ -104,14 +137,14 @@ local PASSES = {
 	offsets = {
 		module = "passes.offsets",
 		kind   = "header-output",
-		deps   = {"word-counts", "components"},
+		deps   = {"scan-source", "word-counts", "components"},
 		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   = {"word-counts", "components"},
+		deps   = {"scan-source", "word-counts", "components"},
 		desc   = "[FUTURE] GTE pipeline-fill, mac_yield uniformity, etc.",
 		out    = { { kind = "report", path_template = "<out_root>/<basename>.static_analysis.txt" } },
 	},
@@ -140,27 +173,26 @@ local PASS_FLAG_TO_NAME = {
 	["--offsets"]         = "offsets",
 	["--static-analysis"] = "static-analysis",
 	["--report"]          = "report",
-	["--all"]             = "__all__",
+	["--scan-source"]     = "scan-source",
+	["--all"]             = ALL_PASSES_SENTINEL,
 }
 
 local ALL_PASS_NAMES = {
-	"word-counts", "components", "annotation",
+	"scan-source", "word-counts", "components", "annotation",
 	"offsets", "static-analysis", "report",
 }
 
---- Append every pass name to args.requested_set. Used by --all and
---- by the "default to --all if no pass flags were given" fallback.
+--- Append every pass name to args.requested_set. Used by --all and by the "default to --all if no pass flags were given" fallback.
+--- @param args ParsedArgs
 local function request_all_passes(args)
 	for _, n in ipairs(ALL_PASS_NAMES) do
 		args.requested_set[#args.requested_set + 1] = n
 	end
 end
 
--- Per-flag handlers. Each handler takes (args, argv, i) and returns
--- the new i (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.
+-- 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 = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -203,53 +235,28 @@ EXAMPLE:
 ]])
 end
 
--- Per-flag handlers. Each takes (args, argv, i) and returns the new i
--- (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.
+-- 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).
+-- 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["--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
 
-FLAG_HANDLERS["--verbose"] = function(args)
-	args.verbose = true
-end
-
-FLAG_HANDLERS["--source"] = function(args, argv, i)
-	args.sources[#args.sources + 1] = argv[i + 1]
-	return i + 1
-end
-
-FLAG_HANDLERS["--metadata"] = function(args, argv, i)
-	args.metadata = argv[i + 1]
-	return i + 1
-end
-
-FLAG_HANDLERS["--out-root"] = function(args, argv, i)
-	args.out_root = argv[i + 1]
-	return i + 1
-end
-
-FLAG_HANDLERS["--project-root"] = function(args, argv, i)
-	args.project_root = argv[i + 1]
-	return i + 1
-end
-
--- Pass-flag handler. Reads the closed-set table, expands --all,
--- appends to requested_set. Single-statement, no nesting.
-FLAG_HANDLERS["__pass__"] = function(args, a)
+-- Pass-flag handler. Reads the closed-set table, expands --all, appends to requested_set. Single-statement, no nesting.
+FLAG_HANDLERS[PASS_FLAG_DISPATCH_KEY] = function(args, a)
 	local name = PASS_FLAG_TO_NAME[a]
-	if name == "__all__" then
+	if name == ALL_PASSES_SENTINEL then
 		request_all_passes(args)
 		return
 	end
@@ -265,49 +272,47 @@ local function parse_args(argv)
 		requested_set = {},
 		sources       = {},
 		metadata      = nil,
-		out_root      = "build/gen",
+		out_root      = DEFAULT_OUT_ROOT,
 		project_root  = nil,
 		dry_run       = false,
 		verbose       = false,
 	}
 
-	local i = 1
-	while i <= #argv do
-		local a = argv[i]
+	local pos = 1
+	while pos <= #argv do
+		local a       = argv[pos]
 		local handler = FLAG_HANDLERS[a]
 		if handler then
-			i = handler(args, argv, i) or i
+			pos = handler(args, argv, pos) or pos
 		elseif PASS_FLAG_TO_NAME[a] then
-			FLAG_HANDLERS["__pass__"](args, a)
+			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(2)
+			os.exit(EXIT_INTERNAL_ERROR)
 		end
-		i = i + 1
+		pos = pos + 1
 	end
 
 	-- Default: --all if no explicit pass flags.
-	if #args.requested_set == 0 then
-		request_all_passes(args)
-	end
+	if #args.requested_set == 0 then request_all_passes(args) end
 
 	-- Defaults: project_root = dirname(metadata).
 	if args.metadata and not args.project_root then
-		local d = dirname(args.metadata)
-		if #d > 0 and (d:sub(-1) == "/" or d:sub(-1) == "\\") 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 = dirname(d)
+		args.project_root = duffle.dirname(d)
 	end
 
 	if not args.metadata then
 		io.stderr:write("ps1_meta: --metadata PATH is required\n")
-		os.exit(2)
+		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(2)
+		os.exit(EXIT_INTERNAL_ERROR)
 	end
 
 	return args
@@ -325,20 +330,23 @@ end
 local function build_ctx(args)
 	local sources = {}
 	for _, path in ipairs(args.sources) do
-		local f = io.open(path, "r")
+		local  f = io.open(path, "r")
 		if not f then
 			io.stderr:write("ps1_meta: cannot open --source " .. path .. "\n")
-			os.exit(2)
+			os.exit(EXIT_INTERNAL_ERROR)
 		end
 		local text = f:read("*a")
 		f:close()
 
-		local dir      = dirname(path)
-		local basename = basename_no_ext(path)
+		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,
@@ -347,8 +355,14 @@ local function build_ctx(args)
 		}
 	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      = {},
@@ -364,14 +378,12 @@ 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.
+--- Compute the dep-closure of `requested_set`: include every pass name transitively required by the requested set.
 ---
 --- @param passes table<string, PassDescriptor>
 --- @param requested_set string[]
---- @return string[]  -- execution order
-local function topo_sort(passes, requested_set)
-	-- Step 1: dep-closure.
+--- @return table<string, boolean>  -- set of pass names needed (including transitive deps)
+local function dep_closure(passes, requested_set)
 	local needed = {}
 	for _, name in ipairs(requested_set) do needed[name] = true end
 	local changed = true
@@ -390,55 +402,94 @@ local function topo_sort(passes, requested_set)
 			end
 		end
 	end
+	return needed
+end
 
-	-- Step 2: Kahn's algorithm.
+--- Count entries in a hash table (Lua's `#t` doesn't work for hash tables).
+--- @param t table
+--- @return integer
+local function count_entries(t)
+	local n = 0
+	for _ in pairs(t) do n = n + 1 end
+	return n
+end
+
+--- Compute in-degrees for the Kahn sort: for each pass in `needed`, the number of its deps that are also in `needed`.
+---
+--- @param passes table<string, PassDescriptor>
+--- @param needed table<string, boolean>
+--- @return table<string, integer>
+local function compute_in_degrees(passes, 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] or 0) + 1
+				in_degree[name] = in_degree[name] + 1
 			end
 		end
 	end
+	return in_degree
+end
 
-	-- Seed with passes that have no unmet deps. Sort for determinism.
+--- Seed the Kahn ready queue with passes whose in-degree is 0, sorted alphabetically for deterministic execution order.
+---
+--- @param in_degree table<string, integer>
+--- @return string[]
+local function seed_ready_queue(in_degree)
 	local ready = {}
 	for name, deg in pairs(in_degree) do
 		if deg == 0 then ready[#ready + 1] = name end
 	end
 	table.sort(ready)
+	return ready
+end
 
-	local order = {}
-	while #ready > 0 do
-		local n = table.remove(ready, 1)
-		order[#order + 1] = n
-		for name, _ in pairs(needed) do
-			if name ~= n then
-				for _, dep in ipairs(passes[name].deps) do
-					if dep == n then
-						in_degree[name] = in_degree[name] - 1
-						if in_degree[name] == 0 then
-							ready[#ready + 1] = name
-							table.sort(ready)
-						end
+-- (internal) Pop the next ready pass, decrement the in-degree of every remaining pass that depended on it
+-- (inserting newly-zero-degree passes back into the ready queue), and append to `order`. Keeps `ready` sorted.
+-- @param passes table<string, PassDescriptor>
+-- @param needed table<string, boolean>
+-- @param in_degree table<string, integer>
+-- @param ready string[]
+-- @param order string[]
+local function process_next_ready(passes, needed, in_degree, ready, order)
+	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.
-	--
-	-- Note: `#needed` returns 0 for hash tables (needed is a set,
-	-- not a sequence), so we count entries explicitly.
-	local needed_count = 0
-	for _ in pairs(needed) do needed_count = needed_count + 1 end
-	if #order ~= needed_count then
+--- 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
+local function topo_sort(passes, requested_set)
+	local needed    = dep_closure(passes, requested_set)
+	local in_degree = compute_in_degrees(passes, needed)
+	local ready     = seed_ready_queue(in_degree)
+
+	local order = {}
+	while #ready > 0 do
+		process_next_ready(passes, needed, in_degree, ready, order)
+	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 emspty order list, 
+	-- leaving the orchestrator to dispatch nothing.
+	if #order ~= count_entries(needed) then
 		for name, deg in pairs(in_degree) do
 			if deg > 0 then
 				error("dependency cycle detected involving pass '" .. name .. "'")
@@ -454,8 +505,7 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- 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.
+--- Falls back to the simpler "Resolved dependency order" list only if graph width exceeds terminal width.
 ---
 --- @param passes table<string, PassDescriptor>
 --- @param requested string[]  -- originally-requested passes (subset of closed)
@@ -466,50 +516,61 @@ local function render_dep_graph(passes, requested, closed)
 	local function add(s) lines[#lines + 1] = s end
 
 	add("[ps1_meta] Resolved dependency order (closed under deps):")
-	for i, name in ipairs(closed) do
+	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]",
-			i, name, deps_str, p.kind))
+			pass_idx, name, deps_str, p.kind))
 	end
 	add("")
 
-	add("[ps1_meta] Pass graph (read top-to-bottom):")
+	-- Data-driven ASCII graph built from the actual PASSES table.
+	-- Shows the source -> scan_source -> pass chain. Each pass is
+	-- shown once; edges are "feeds into" arrows based on deps.
+	add("[ps1_meta] Pass graph (read top-to-bottom; edges = 'feeds into'):")
 	add("")
-	add("                       metadata.h")
-	add("                           |")
-	add("                           v")
-	add("   +-----------+   +-----------------+   +-----------------+")
-	add("   |   word-   |-->|    components   |-->|     offsets     |")
-	add("   |   counts  |   +-----------------+   +-----------------+")
-	add("   |   (load)  |            |                      ^")
-	add("   +-----------+            |                      |")
-	add("        |                   v                      |")
-	add("        |  code/<module>/gen/<basename>.macs.h     |")
-	add("        |     (header - co-located for #include)   |")
-	add("        |                                            |")
-	add("        |           +-----------------+              |")
-	add("        +---------->|    annotation   |--------------+")
-	add("        |           +-----------------+              |")
-	add("        |                   |                        |")
-	add("        |                   v                        |")
-	add("        |       build/gen/<basename>.errors.h        |")
-	add("        |       build/gen/<basename>.annotations.txt |")
-	add("        |         (report - NOT #included)            |")
-	add("        |                                            |")
-	add("        |           +-----------------+              |")
-	add("        +---------->| static-analysis |--------------+")
-	add("                    +-----------------+")
-	add("                            |")
-	add("                            v")
-	add("                    +---------------+")
-	add("                    |    report     |")
-	add("                    +---------------+")
-	add("                            |")
-	add("                            v")
-	add("                    build/gen/annotation_validation.txt")
-	add("                      (project summary)")
+
+	-- Compute which passes feed which other passes (reverse of deps).
+	local feeds = {}  -- feeds[X] = list of passes that X feeds into
+	for _, name in ipairs(closed) do feeds[name] = {} end
+	for name, p in pairs(passes) do
+		for _, dep in ipairs(p.deps) do
+			if feeds[dep] then feeds[dep][#feeds[dep] + 1] = name end
+		end
+	end
+
+	-- Layout: source -> scan_source -> word-counts -> {components, annotation, offsets, static-analysis} -> report
+	-- Outputs are listed under each pass.
+	local outputs_for = function(name)
+		local p = passes[name]
+		if not p or not p.out or #p.out == 0 then return "" end
+		local outs = {}
+		for _, o in ipairs(p.out) do outs[#outs + 1] = o.path_template end
+		return table.concat(outs, ", ")
+	end
+
+	add("   +-----------+   +-------------------+    +-----------------+")
+	add("   |  source   |-->|  scan_source      |--->|   word-counts   |")
+	add("   |  files    |   | (scan_source.lua) |    |   (load)        |")
+	add("   +-----------+   +-------------------+    +-----------------+")
+	add("                          (single walk)         |")
+	add("                                                |")
+	add("        +-------------------+-------------------+-----------+")
+	add("        v                   v                   v           v")
+	add("   +--------------+   +--------------+   +--------------+   +---------------+")
+	add("   | components   |   | annotation   |   |   offsets    |   |static-analysis|")
+	add("   +--------------+   +--------------+   +--------------+   +---------------+")
+	add("   |<src>/gen/    |   |build/gen/    |   |<src>/gen/    |   |build/gen/     |")
+	add("   |<base>.macs.h |   |<base>.errors |   |<base>.offsets|   |<base>.static  |")
+	add("   |   (header)   |   |  .h          |   |  .h          |   |  _analysis    |")
+	add("   +------+-------+   |  +annot.txt  |   |  (header)    |   |  .txt         |")
+	add("          |           +------+-------+   +--------------+   +------+--------+")
+	add("          v                v                                       v")
+	add("   +------+----------------+  +------+-------+                     |")
+	add("   |offsets|static-analysis|  |report|       |<--------------------+")
+	add("   |       |               |  +------+-------+")
+	add("   +-------+---------------+")
 
 	return table.concat(lines, "\n") .. "\n"
 end
@@ -518,6 +579,59 @@ 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]
+		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)
@@ -525,49 +639,25 @@ local function main(argv)
 		local args = parse_args(argv)
 		local ctx  = build_ctx(args)
 
-		-- 1. Compute requested set + dep-closed set.
 		local requested = args.requested_set
 		local closed    = topo_sort(PASSES, requested)
 
-		-- 2. --dry-run: print dep order + ASCII graph, exit 0.
+		-- --dry-run: print dep order + ASCII graph, exit OK.
 		if args.dry_run then
 			io.write(render_dep_graph(PASSES, requested, closed))
-			os.exit(0)
+			os.exit(EXIT_OK)
 		end
 
-		-- 3. Run passes in topological order.
-		ctx.shared = {}
-		for _, pass_name in ipairs(closed) do
-			local pass  = PASSES[pass_name]
-			local mod   = require(pass.module)
-			local result = mod.run(ctx)
-
-			-- Collect outputs + warnings into ctx.upstream.
-			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
-
-			-- Stop on errors (unless this is a non-stopping pass like report).
-			if (result.errors and #result.errors > 0) and PASS_KIND_STOP_ON_ERROR[pass.kind] then
-				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
-				os.exit(1)
-			end
-		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(2)
+		os.exit(EXIT_INTERNAL_ERROR)
 	end
 
-	os.exit(0)
+	os.exit(EXIT_OK)
 end
 
-main({...})
+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

scripts/word_count_eval.lua

@@ -1,246 +0,0 @@
--- word_count_eval.lua
---
--- Word-counting logic for the tape-atom metaprogram pipeline.
--- Used by:
---   - passes/components.lua (compute_component_word_count)
---   - passes/offsets.lua    (scan_atom_body)
---   - passes/annotation.lua (TAPE_WORDS <-> WORD_COUNT drift check)
---
--- This module ALSO exposes M.run(ctx) — the "word-counts" pass entry in
--- the PASSES table — which loads metadata.h + scans for existing
--- *.macs.h files into ctx.shared.word_counts.
---
--- Coding standard: tabs (1/level), EmmyLua annotations, no regex,
--- Lua 5.3 compatible.
-
--- ════════════════════════════════════════════════════════════════════════════
--- Module-scope requires + package.path setup
--- ════════════════════════════════════════════════════════════════════════════
-
-local script_path = arg and arg[0] or "?"
-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
-
-local duffle = require("duffle")
-local trim              = duffle.trim
-local read_ident        = duffle.read_ident
-local skip_ws_and_cmt   = duffle.skip_ws_and_cmt
-local load_word_counts  = duffle.load_word_counts
-local split_top_level_commas = duffle.split_top_level_commas
-local is_space          = duffle.is_space
-
--- ════════════════════════════════════════════════════════════════════════════
--- Type declarations
--- ════════════════════════════════════════════════════════════════════════════
-
---- @class WordCounts
---- @field [string] integer  -- macro name -> word count
-
---- @class PassCtx
---- @field sources            SourceFile[]
---- @field metadata_path      string
---- @field shared             table
---- @field shared.word_counts WordCounts
---- @field out_root           string
---- @field project_root       string
---- @field upstream           table<string, table>
---- @field flags              table
---- @field dry_run            boolean
---- @field verbose            boolean
-
---- @class PassResult
---- @field outputs  table[]
---- @field errors   table[]
---- @field warnings table[]
-
---- @class SourceFile
---- @field path     string
---- @field text     string
---- @field dir      string
---- @field basename string
-
--- ════════════════════════════════════════════════════════════════════════════
--- 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.
----
---- PORT NOTE: taken verbatim from tape_atom.offset_gen.meta.lua:130-141
---- (`word_count_of_token`). Behavior is identical to preserve the
---- branch-offset fix from commit 98e27c2.
----
---- @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 = 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
-
--- ┌────────────────────────────────────────────────────────────────────┐
--- │ Shared utility: scan_dir                                           │
--- └────────────────────────────────────────────────────────────────────┘
-
---- Recursively scan a directory for files matching a glob suffix.
---- No regex per the no_regex constraint — uses plain byte matching
---- via `dir /b /s` on Windows.
----
---- PORT NOTE: taken from tape_atom.offset_gen.meta.lua:432-443
---- (`scan_dir`). Adapted: removed the hardcoded project_root derivation;
---- the caller passes `dir` explicitly.
----
---- @param dir string        -- directory to scan (absolute or relative)
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
-function M.scan_dir(dir, suffix)
-	local results = {}
-	local p = io.popen('dir /b /s "' .. dir .. '\\' .. suffix .. '" 2>nul')
-	if not p then return results end
-	for raw_line in p:lines() do
-		local path = raw_line:gsub("\\", "/")
-		results[#results + 1] = path
-	end
-	p:close()
-	return results
-end
-
--- ┌────────────────────────────────────────────────────────────────────┐
--- │ Shared utility: count_body_words                                   │
--- └────────────────────────────────────────────────────────────────────┘
-
---- Count words emitted by an entire atom body (a brace-delimited block).
---- Splits by top-level commas; for each token, delegates to count_token_words.
---- Handles `atom_label(name)` / `atom_offset(tag, name)` markers (record at
---- current pos, do NOT advance pos; if the marker call bundles an instruction
---- after it, count that instruction too).
----
---- PORT NOTE: taken verbatim from tape_atom.offset_gen.meta.lua:207-239
---- (`scan_atom_body`). Behavior is identical to preserve the branch-offset
---- fix from commit 98e27c2.
----
---- @param body string         -- brace-delimited atom body (without braces)
---- @param wc WordCounts       -- the shared word-count table
---- @return integer            -- total words
-function M.count_body_words(body, wc)
-	local pos = 0
-	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 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.
-			local marker_end = M.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 = M.count_token_words(rest, wc)
-					pos = pos + rest_words
-				end
-			end
-		else
-			pos = pos + M.count_token_words(tok, wc)
-		end
-	end
-	return pos
-end
-
---- Find the end position (just past the closing ')') of the first
---- atom_label/atom_offset call in `tok`. Returns 0 if no such call.
---- Internal helper for count_body_words.
----
---- PORT NOTE: taken from tape_atom.offset_gen.meta.lua:181-205
---- (`find_marker_call_end`).
----
---- @param tok string
---- @return integer  -- 0 if no marker call found
-function M.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_space(c) then
-			i = i + 1
-		elseif c == "/" then
-			-- comment — skip past it (delegated to duffle.skip_str_or_cmt)
-			local nx = duffle.skip_str_or_cmt(tok, i)
-			if nx > i then i = nx else i = i + 1 end
-		else
-			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 = duffle.read_parens(tok, j)
-					return end_paren - 1
-				end
-				return 0
-			end
-			i = after or (i + 1)
-		end
-	end
-	return 0
-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 = 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(load_word_counts, macs_path)
-		if ok and type(mc) == "table" then
-			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