improvmenets for atom components

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

@@ -4,8 +4,6 @@
 // Auto-generated by tape_atom_annotation_pass.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)) \
@@ -45,7 +43,7 @@ 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. */

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

@@ -440,6 +440,8 @@ 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 +715,14 @@ 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

@@ -31,13 +31,9 @@ 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 +122,7 @@ 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),
@@ -287,7 +283,10 @@ 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)),

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/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,
  
@@ -174,13 +146,10 @@ 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)),

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.lua

@@ -1,8 +1,6 @@
 -- duffle.lua
 --
 -- Shared primitives + domain tables for the tape-atom metaprograms.
--- Both `tape_atom_annotation_pass.lua` and `tape_atom.offset_gen.meta.lua`
--- `require("duffle")` for these.
 --
 -- 5.3-compatible Lua (no 5.4/5.5-only features):
 --   - no <close> / <toclose>
@@ -186,12 +184,25 @@ function M.write_file(path, content)
     f:close()
 end
 
+-- Cache of directories already verified to exist in this process. Each
+-- ensure_dir() call may otherwise spawn a `cmd.exe mkdir` (50-100ms
+-- per call on Windows) — calling it inside per-source loops added 1.5+
+-- seconds to the report pass. Cache makes ensure_dir idempotent within
+-- the process lifetime (safe across passes; the dir state doesn't change).
+local _ensured_dirs = {}
+
 function M.ensure_dir(path)
+    if _ensured_dirs[path] then return end
+    _ensured_dirs[path] = true
     local is_win = package.config:sub(1, 1) == "\\"
     os.execute(is_win and ('if not exist "' .. path .. '" mkdir "' .. path .. '"')
                   or ('mkdir -p "' .. path .. '" 2>/dev/null'))
 end
 
+-- Test helper: clear the cache (used by tests + between process runs).
+-- Not normally needed since Lua state is per-process.
+function M._reset_ensured_dirs() _ensured_dirs = {} end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Section 4: C-language scanner primitives
 -- ════════════════════════════════════════════════════════════════════════════
@@ -499,54 +510,92 @@ M.WAVE_CONTEXT_REGS = {
     ["R_OtBase"]     = { alias = "R_T6", size = 4, role = "base pointer (ordering table)" },
 }
 
-M.MACRO_EXPANSION = {
-    ["phase_init"]      = "init",
-    ["phase_bind"]      = "bind",
-    ["phase_setup"]     = "setup",
-    ["phase_work"]      = "work",
-    ["phase_commit"]    = "commit",
-    ["phase_terminate"] = "terminate",
-    ["REGION_PRIM_ARENA"]   = "prim_arena",
-    ["REGION_FACE_ARENA"]   = "face_arena",
-    ["REGION_VERTEX_ARENA"] = "vertex_arena",
-    ["REGION_OT_ARENA"]     = "ot_arena",
-    ["REGION_HEAP_3D"]      = "heap_3d_models",
-    ["REGION_CDROM_STREAM"] = "cdrom_stream",
-    ["REGION_VRAM"]         = "vram_heap",
-    ["CADENCE_FRAME"]       = "frame",
-    ["CADENCE_ONCE"]        = "once",
-    ["CADENCE_ONDEMAND"]    = "ondemand",
-}
-
-M.KNOWN_PHASES = {
-    ["init"]      = true, ["bind"]   = true, ["setup"]    = true,
-    ["work"]      = true, ["commit"] = true, ["terminate"] = true,
-}
-M.KNOWN_REGIONS = {
-    ["prim_arena"]     = true, ["face_arena"]     = true,
-    ["vertex_arena"]   = true, ["ot_arena"]       = true,
-    ["heap_3d_models"] = true, ["cdrom_stream"]   = true,
-    ["vram_heap"]      = true,
-}
-M.KNOWN_CADENCES = {
-    ["frame"]    = true, ["once"] = true, ["ondemand"] = true,
-}
-
+-- The annotation DSL has been reduced to a single annotation macro:
+--   atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...))
+-- All phase / region / cadence / async / resource / group tokens have
+-- been dropped. They may be reintroduced later as optional sub-calls
+-- of atom_info; for now, the parser only recognizes atom_info + its
+-- three sub-calls (atom_bind, atom_reads, atom_writes).
 M.TAPE_ATOM_MACROS = {
-    ["atom_annot"]     = { kind = "work",      binds = false },
-    ["atom_bind"]      = { kind = "bind",      binds = true  },
-    ["atom_setup"]     = { kind = "setup",     binds = false },
-    ["atom_commit"]    = { kind = "commit",    binds = false },
-    ["atom_init"]      = { kind = "init",      binds = false },
-    ["atom_terminate"] = { kind = "terminate", binds = false },
+    ["atom_info"] = { kind = "info", binds = false },
 }
 
-M.ATOM_PRAGMA_KINDS = {
-    ["resource"] = { kind = "string" },
-    ["region"]   = { kind = "ident",  allowed = M.KNOWN_REGIONS },
-    ["group"]    = { kind = "ident" },
-    ["cadence"]  = { kind = "ident",  allowed = M.KNOWN_CADENCES },
-    ["async"]    = { kind = "ident",  allowed = { ["true"] = true, ["false"] = true } },
+-- GTE pipeline-fill latency table (static-analysis Phase 1).
+--
+-- For each `gte_cmdw_*` macro in code/duffle/gte.h, the minimum number
+-- of consecutive COP2 "nop" words that MUST appear before any other
+-- COP2 read or non-nop instruction (so the GTE pipeline latency is
+-- fully retired).  Latencies are sourced from the doxygen comments
+-- in gte.h (e.g. `* @brief Rotate, Translate and Perspective Triple
+-- (23 cycles)` with body `Two nop words fill the COP2 pipeline
+-- latency`).
+--
+-- The check (`scripts/passes/static_analysis.lua ::
+-- check_gte_pipeline_fill`) walks each atom body, counts the
+-- consecutive nop words after every `gte_cmdw_*` invocation, and
+-- reports a finding if the count is below this minimum.  Aliases
+-- are dereferenced before lookup (gté_cmdw_rtps_alias ->
+-- gte_cmdw_rtps -> 2).
+--
+-- Values verified against PSX-SPX gte.txt (rtpt 23cy / 8cy per divide
+-- => 2 nops; nclip 8cy => 2 nops; avsz3/avsz4 14cy => 2 nops; op
+-- single-cycle atomic => 0 nops; mvmva 8cy matrix-vector => 2 nops).
+M.GTE_PIPELINE_LATENCY = {
+    -- Minimum number of consecutive `nop` words that must appear
+    -- IMMEDIATELY BEFORE a `gte_cmdw_<X>` invocation -- to retire
+    -- any preceding `lwc2` / `swc2` / pre-existing C2 state writes
+    -- before the GTE pipeline starts reading from V0/V1/V2 or
+    -- MAC0..3 / OTZ / IR0..3 at the command's issue cycle.
+    --
+    -- Values are from the doxygen comments in code/duffle/gte.h and
+    -- cross-checked against PSX-SPX `geometrytransformationenginegte.md`:
+    --
+    --   cmd      cycles  min pre-nops  rationale
+    --   rtps        14      2         8c per perspective divide + 6c for IR1..4 + mac write
+    --   rptt        22      2         3x rtps worth of pipeline depth
+    --   nclip        7      2         MAC0 write + 5c for sign
+    --   avsz3       14      2         14c to compute average + write OTZ
+    --   avsz4       16      2         avsz3 + 2c extra for avg over 4
+    --   mvmva        8      2         IR1..4 write + matrix work
+    --   op          5      0         output to MAC0 only (atomic 5c calc)
+    --
+    -- The `gte_rtpt()` / `gte_nclip()` / `gte_avsz3()` wrapper macros in
+    -- gte.h emit the pre-cmd nops internally (asm_words(nop, nop, ...)),
+    -- but THOSE WRAPPERS ARE NOT USED INSIDE ATOM BODIES in this
+    -- codebase. Every MipsAtom_(name) body uses raw `nop2,
+    -- gte_cmdw_<X>, ...` form instead -- that `nop2,` is the pre-fill
+    -- this check validates. So values here must reflect the source-level
+    -- convention, NOT the wrapper-internal pre-fill (which is invisible
+    -- at the source level).
+    --
+    -- Existing clean-atom bodies (cube_g4_face, floor_f3_face,
+    -- diag_gte) all emit `nop2,` before every `gte_cmdw_<X>` (which
+    -- matches values >= 2). The check passes them all.
+    --
+    -- Aliases are listed separately because source code may use either
+    -- the alias or the canonical name. The check looks up the EXACT
+    -- macro text, so both forms must be in the table.
+
+    -- Canonical macros (from code/duffle/gte.h)
+    ["gte_cmdw_rtps"]   = 2,
+    ["gte_cmdw_rtpt"]   = 2,
+    ["gte_cmdw_nclip"]  = 2,
+    ["gte_cmdw_op"]     = 0,
+    ["gte_cmdw_mvmva"]  = 2,
+    ["gte_cmdw_avsz3"]  = 2,
+    ["gte_cmdw_avsz4"]  = 2,
+
+    -- Aliases (must have the same value as their canonical target)
+    ["gte_cmdw_rotate_translate_perspective_single"] = 2,
+    ["gte_cmdw_rotate_translate_perspective_triple"] = 2,
+    ["gte_cmdw_avg_sort_z4"]                          = 2,
+
+    -- Outer product aliases (same canonical op, 0 pre-fill nops).
+    -- gte_cmdw_op          = canonical GTE-internal short form
+    -- gte_cmdw_outer_product = NOCASH / SDK-readable form
+    -- gte_cmdw_wedge         = geometric-algebra (exterior-product) form
+    ["gte_cmdw_outer_product"] = 0,
+    ["gte_cmdw_wedge"]        = 0,
 }
 
 -- Expose the lpeg_ok flag so callers can detect the LPeg-back path.

scripts/passes/annotation.lua

@@ -47,12 +47,8 @@ local split_top_level_commas = duffle.split_top_level_commas
 
 -- Domain tables (single source of truth in duffle.lua).
 local WAVE_CONTEXT_REGS = duffle.WAVE_CONTEXT_REGS
-local MACRO_EXPANSION   = duffle.MACRO_EXPANSION
-local KNOWN_PHASES      = duffle.KNOWN_PHASES
 local TAPE_ATOM_MACROS  = duffle.TAPE_ATOM_MACROS
-local ATOM_PRAGMA_KINDS = duffle.ATOM_PRAGMA_KINDS
 
-local function valid_phase(p)          return KNOWN_PHASES[p]      or false end
 local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -109,32 +105,44 @@ end
 -- Parse TAPE_ATOM_ANNOT(...) calls
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Recognize a `atom_reads(...)` or `atom_writes(...)` register-list
--- call embedded inside an annotation arg list. Returns the kind
--- ("atom_reads" / "atom_writes") and the inner content, or nil if the
--- token isn't a recognized register-list form. Flattened via a
--- prefix lookup instead of a nested if/elseif chain.
+-- Recognize a `atom_bind(...)`, `atom_reads(...)`, or `atom_writes(...)`
+-- sub-call embedded inside an atom_info arg list. Returns the kind
+-- ("atom_bind" / "atom_reads" / "atom_writes") and the inner content,
+-- or nil if the token isn't a recognized sub-call form. Flattened via
+-- a prefix lookup instead of a nested if/elseif chain.
 local REGS_CALL_PREFIX = {
-	["atom_reads("]  = { kind = "atom_reads",  inner_offset = 12 },
 	["atom_writes("] = { kind = "atom_writes", inner_offset = 13 },
+	["atom_reads("]  = { kind = "atom_reads",  inner_offset = 12 },
+	["atom_bind("]   = { kind = "atom_bind",   inner_offset = 11, single_ident = true },
 }
 
 local function parse_regs_call(s)
 	if s:sub(-1) ~= ")" then return nil end
-	local spec = REGS_CALL_PREFIX[s:sub(1, 12)]  -- longest prefix first wins
+	-- Try longest prefix first so "atom_writes(" wins over "atom_reads("
+	-- when both 12-char prefixes would otherwise match. Lengths:
+	--   atom_writes(  = 12 chars, offset 13
+	--   atom_reads(   = 11 chars, offset 12
+	--   atom_bind(    = 10 chars, offset 11
+	local spec = REGS_CALL_PREFIX[s:sub(1, 12)]
+	if not spec then
+		spec = REGS_CALL_PREFIX[s:sub(1, 11)]
+	end
+	if not spec then
+		spec = REGS_CALL_PREFIX[s:sub(1, 10)]
+	end
 	if not spec then return nil end
-	-- The 12-char prefix "atom_reads(" also matches "atom_writes("
-	-- would be ambiguous; the table order above handles it.
-	-- (atom_reads prefix is 11 chars, atom_writes is 12; the 12-char
-	-- lookup matches atom_writes first.)
-	return spec.kind, s:sub(spec.inner_offset, -2)
+	local inner = s:sub(spec.inner_offset, -2)
+	if spec.single_ident then
+		-- atom_bind takes a single Binds_* type ident. Trim and pass through.
+		return spec.kind, trim(inner)
+	end
+	return spec.kind, inner
 end
 
 -- Resolve any phase_* / R_* alias macros in a register list.
+-- (Phase / region / cadence aliases have been dropped. Kept as an
+-- identity function so callers can stay uniform.)
 local function resolve_reg_aliases(regs)
-	for i, r in ipairs(regs) do
-		if MACRO_EXPANSION[r] then regs[i] = MACRO_EXPANSION[r] end
-	end
 	return regs
 end
 
@@ -150,18 +158,19 @@ local function parse_regs_list(inner)
 end
 
 -- Parse a single token (from split_csv_top) into an arg entry.
--- Three forms: register-list call, bare identifier (with alias),
+-- Three forms: register-list call, bare identifier,
 -- "other" (preserved as text).
 local function parse_arg_token(s)
 	local kind, inner = parse_regs_call(s)
 	if kind then
+		if kind == "atom_bind" then
+			return { kind = kind, value = inner }  -- single ident, not a list
+		end
 		return { kind = kind, value = parse_regs_list(inner) }
 	end
 	local id = read_ident(s, 1)
 	if id and trim(s) == id then
-		local v = id
-		if MACRO_EXPANSION[v] then v = MACRO_EXPANSION[v] end
-		return { kind = "ident", value = v }
+		return { kind = "ident", value = id }
 	end
 	return { kind = "other", value = s }
 end
@@ -262,136 +271,6 @@ local function find_macro_word_annotations(source)
 	return out
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Parse `atom_<...>` Pragma / _Pragma annotations
--- ════════════════════════════════════════════════════════════════════════════
-
-local ATOM_ATTR_MACROS = {
-	["atom_resource"] = "resource",
-	["atom_region"]   = "region",
-	["atom_group"]    = "group",
-	["atom_cadence"]  = "cadence",
-	["atom_async"]    = "async",
-}
-
---- Parse macro form: `atom_<key>(atom_name, value, ...)`.
---- Returns (true, entry, str_end) on success, (false) on no match.
-local function try_parse_atom_attr_macro(source, i, line_of)
-	local ident, after = read_ident(source, i)
-	if not ident then return false end
-	local key = ATOM_ATTR_MACROS[ident]
-	if not key then return false end
-	local open = skip_ws_and_cmt(source, after)
-	if source:sub(open, open) ~= "(" then return false end
-
-	local body, body_end = read_parens(source, open)
-	local first, after_name = read_ident(body, 1)
-	if not first then return false end
-
-	local j = after_name
-	while j <= #body and is_space(body:sub(j, j)) do j = j + 1 end
-	if body:sub(j, j) ~= "," then return false end
-	j = j + 1
-	while j <= #body and is_space(body:sub(j, j)) do j = j + 1 end
-
-	local value
-	if body:sub(j, j) == '"' then
-		local k = j + 1
-		while k <= #body do
-			local c = body:sub(k, k)
-			if c == "\\" then
-				k = k + 2
-			elseif c == '"' then
-				break
-			else
-				k = k + 1
-			end
-		end
-		if body:sub(k, k) ~= '"' then return false end
-		value = body:sub(j + 1, k - 1)
-	else
-		local id2, after_id = read_ident(body, j)
-		if not id2 then return false end
-		value = id2
-		if MACRO_EXPANSION[value] then value = MACRO_EXPANSION[value] end
-	end
-
-	return true, {
-		line  = line_of(i),
-		name  = first,
-		attrs = { [key] = value },
-	}, body_end
-end
-
-local function find_atom_pragmas(source)
-	local line_of = duffle.LineIndex(source)
-	local out = {}
-	local len = #source
-	local i   = 1
-	while i <= len do
-		i = skip_ws_and_cmt(source, i); if i > len then break end
-		if source:sub(i, i) == "#" then
-			local j = i
-			while j <= len and source:sub(j, j) ~= "\n" do j = j + 1 end
-			i = j + 1
-		else
-			local got, entry, next_i = try_parse_atom_attr_macro(source, i, line_of)
-			if got then
-				out[#out + 1] = entry
-				i = next_i
-			else
-				local ident, after = read_ident(source, i)
-				if not ident then
-					i = i + 1
-				elseif ident == "_Pragma" then
-					local open = skip_ws_and_cmt(source, after)
-					if source:sub(open, open) == "(" then
-						local str, str_end = read_parens(source, open)
-						str = trim(str)
-						if str:sub(1, 1) == '"' and str:sub(-1) == '"' then
-							local inner = str:sub(2, -2)
-							local sp1 = find_byte(inner, " ", 1)
-							if sp1 and trim(inner:sub(1, sp1 - 1)) == "atom" then
-								local rest = trim(inner:sub(sp1 + 1))
-								local sp2 = find_byte(rest, " ", 1)
-								if sp2 then
-									local name = trim(rest:sub(1, sp2 - 1))
-									local attrs_str = trim(rest:sub(sp2 + 1))
-									local attrs = {}
-									local got_any = false
-									for _, pair in ipairs(split_ws(attrs_str)) do
-										local eq = find_byte(pair, "=", 1)
-										if eq then
-											local k = trim(pair:sub(1, eq - 1))
-											local v = trim(pair:sub(eq + 1))
-											if MACRO_EXPANSION[v] then v = MACRO_EXPANSION[v] end
-											attrs[k] = v
-											got_any = true
-										end
-									end
-									if got_any then
-										out[#out + 1] = {
-											line  = line_of(i),
-											name  = name,
-											attrs = attrs,
-										}
-									end
-								end
-							end
-						end
-						i = str_end
-					else
-						i = open + 1
-					end
-				else
-					i = after
-				end
-			end
-		end
-	end
-	return out
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Parse `typedef Struct_(Binds_X) { ... };` declarations
 -- ════════════════════════════════════════════════════════════════════════════
@@ -526,58 +405,35 @@ local function is_regs_arg(a)
 end
 
 --- Per-macro arg-shape handlers. Each takes (entry, args) and mutates
---- entry.{reads, writes, phase, binds, errors}. Replaces the 5-way
---- `if/elseif/elseif/elseif/elseif` chain inside find_atom_annotations.
+--- Per-atom_info sub-call dispatch. Each takes (entry, args) and mutates
+--- entry.{reads, writes, binds, errors}. The new annotation shape is:
+---
+---   MipsAtom_(name) atom_info(
+---        atom_bind(Binds_X)
+---      , atom_reads(...)
+---      , atom_writes(...)
+---   ) { ... };
+---
+--- All sub-calls are order-independent; each is dispatched on its
+--- `kind` (atom_bind / atom_reads / atom_writes) when parsed.
 local ANNOT_ARG_HANDLERS = {}
 
--- atom_bind(name, Binds_Struct, writes)
-function ANNOT_ARG_HANDLERS.bind(entry, args)
-	if #args >= 2 and args[2].kind == "ident" then
-		entry.binds = args[2].value
-	end
-	if #args >= 3 and is_regs_arg(args[3]) then
-		entry.writes = args[3].value
-	end
-end
-
--- atom_init(name) / atom_terminate(name): name only, no extra slots.
-ANNOT_ARG_HANDLERS.init      = function() end
-ANNOT_ARG_HANDLERS.terminate = function() end
-
--- Macro name -> handler key. Replaces the `macro_def.binds` check
--- plus the 4-way ident elseif chain.
-local MACRO_HANDLER_KEY = {
-	["atom_bind"]      = "bind",
-	["atom_annot"]     = "annot",
-	["atom_setup"]     = "reads_only",
-	["atom_commit"]    = "reads_only",
-	["atom_init"]      = "init",
-	["atom_terminate"] = "terminate",
-}
-
--- atom_setup(name, reads) / atom_commit(name, reads): reads from slot 2.
-function ANNOT_ARG_HANDLERS.reads_only(entry, args)
-	if #args >= 2 and is_regs_arg(args[2]) then
-		entry.reads = args[2].value
-	end
-end
-
--- atom_annot(name, phase, reads, writes)
-function ANNOT_ARG_HANDLERS.annot(entry, args)
-	if #args >= 2 and args[2].kind == "ident" then
-		entry.phase = MACRO_EXPANSION[args[2].value] or args[2].value
-	end
-	if #args >= 3 and is_regs_arg(args[3]) then
-		if args[3].kind == "atom_writes" then
-			entry.errors[#entry.errors + 1] = "reads slot has atom_writes — swap order?"
+-- atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...))
+function ANNOT_ARG_HANDLERS.info(entry, args)
+	for _, arg in ipairs(args) do
+		if arg.kind == "atom_bind" then
+			entry.binds = arg.value
+		elseif arg.kind == "atom_reads" then
+			entry.reads = arg.value
+		elseif arg.kind == "atom_writes" then
+			entry.writes = arg.value
+		elseif arg.kind == "ident" then
+			-- Reserved for future phase tokens. Currently ignored.
+			-- (Could be reintroduced as `phase_*` sub-calls of atom_info.)
+		else
+			entry.errors[#entry.errors + 1] = string.format(
+				"unexpected atom_info arg kind=%s value=%s", arg.kind, tostring(arg.value))
 		end
-		entry.reads = args[3].value
-	end
-	if #args >= 4 and is_regs_arg(args[4]) then
-		if args[4].kind == "atom_reads" then
-			entry.errors[#entry.errors + 1] = "writes slot has atom_reads — swap order?"
-		end
-		entry.writes = args[4].value
 	end
 end
 
@@ -589,16 +445,17 @@ local function new_annot_entry(line, ident, name, kind)
 		name   = name,
 		kind   = kind,
 		binds  = nil,
-		phase  = nil,
 		reads  = {},
 		writes = {},
 		errors = {},
 	}
 end
 
---- Find every TAPE_ATOM_* macro call in source and convert it to a
---- normalized annotation entry. Dispatches per-macro arg-shape via
---- ANNOT_ARG_HANDLERS (lookup table; no nested if/elseif chain).
+--- Find every MipsAtom_(name) declaration in source, then look for an
+--- immediately-following atom_info(...) call. If present, parse its
+--- sub-calls into a normalized annotation entry linked to the MipsAtom_
+--- name. If no atom_info follows, emit NO annotation entry (atoms
+--- without annotations are valid in the new minimal shape).
 local function find_atom_annotations(source)
 	local line_of = duffle.LineIndex(source)
 	local annots = {}
@@ -607,9 +464,6 @@ local function find_atom_annotations(source)
 	while i <= len do
 		i = skip_ws_and_cmt(source, i); if i > len then break end
 		-- Skip preprocessor directives (lines starting with #).
-		-- Without this guard, `#define atom_init(name) ...` macro
-		-- definitions get misinterpreted as annotation calls with
-		-- the literal placeholder "name" as the atom name.
 		if source:sub(i, i) == "#" then
 			local j = i
 			while j <= len and source:sub(j, j) ~= "\n" do j = j + 1 end
@@ -620,30 +474,45 @@ local function find_atom_annotations(source)
 		local ident, after = read_ident(source, i)
 		if not ident then
 			i = i + 1
-		elseif TAPE_ATOM_MACROS[ident] then
+		elseif ident == "MipsAtom_" then
 			local open = skip_ws_and_cmt(source, after)
 			if source:sub(open, open) ~= "(" then
 				i = open + 1
-			else
-				local inner, after_paren = read_parens(source, open)
-				local args = parse_atom_annot_args(inner)
-				local macro_def = TAPE_ATOM_MACROS[ident]
+				goto continue
+			end
+			local inner, after_paren = read_parens(source, open)
+			local a = 1
+			while a <= #inner and is_space(inner:sub(a, a)) do a = a + 1 end
+			local b = a
+			while b <= #inner and is_alnum(inner:sub(b, b)) do b = b + 1 end
+			local name = inner:sub(a, b - 1)
 
-				if #args < 1 then
-					annots[#annots + 1] = {
-						line  = line_of(i),
-						macro = ident,
-						kind  = macro_def.kind,
-						error = "missing atom name (first arg)",
-					}
-				else
-					local entry = new_annot_entry(line_of(i), ident, args[1].value, macro_def.kind)
-					local handler = ANNOT_ARG_HANDLERS[MACRO_HANDLER_KEY[ident]]
-					if handler then handler(entry, args) end
+			-- Look for atom_info(...) right after MipsAtom_(name).
+			local lookahead = skip_ws_and_cmt(source, after_paren)
+			local look_ident, look_after = read_ident(source, lookahead)
+			if look_ident == "atom_info" then
+				local info_open = skip_ws_and_cmt(source, look_after)
+				if source:sub(info_open, info_open) == "(" then
+					local info_inner, info_after = read_parens(source, info_open)
+					local args = parse_atom_annot_args(info_inner)
+					local entry = new_annot_entry(line_of(lookahead), "atom_info", name, "info")
+					ANNOT_ARG_HANDLERS.info(entry, args)
 					annots[#annots + 1] = entry
+					i = info_after
+				else
+					i = info_open + 1
 				end
+			else
+				-- No atom_info follows this MipsAtom_. Valid in new shape.
 				i = after_paren
 			end
+
+			-- Skip past the body { ... } if present.
+			local brace = scan_to_char(source, "{", i)
+			if brace then
+				local _, after_brace = read_braces(source, brace)
+				i = after_brace
+			end
 		else
 			i = after
 		end
@@ -661,7 +530,6 @@ local function validate(ctx, src)
 
 	local annots      = find_atom_annotations(source)
 	local macros     = find_macro_word_annotations(source)
-	local pragmas    = find_atom_pragmas(source)
 	local binds      = find_binds_structs(source)
 	local atoms      = find_atom_names(source)
 
@@ -692,37 +560,26 @@ local function validate(ctx, src)
 		end
 	end
 
-	-- 2. Every atom must have exactly one annotation (no orphans, no duplicates).
+	-- 2. Every atom may have AT MOST ONE annotation (no duplicates).
+	--    (Atoms with ZERO annotations are valid in the new minimal shape.)
 	local count_per_atom = {}
 	for _, a in ipairs(annots) do
 		if a.name and not a.error then
 			count_per_atom[a.name] = (count_per_atom[a.name] or 0) + 1
 		end
 	end
-	for _, atom in ipairs(atoms) do
-		local n = count_per_atom[atom.name] or 0
-		if n == 0 then
-			warnings[#warnings + 1] = {
-				line = atom.line,
-				msg  = string.format("MipsAtom_(%s) has no TAPE_ATOM_* annotation", atom.name),
-			}
-		elseif n > 1 then
+	for name, n in pairs(count_per_atom) do
+		if n > 1 then
 			errors[#errors + 1] = {
-				line = atom.line,
-				msg  = string.format("MipsAtom_(%s) has %d annotations (expected 1)", atom.name, n),
+				line = atom_index[name] and atom_index[name].line or 0,
+				msg  = string.format("MipsAtom_(%s) has %d annotations (expected at most 1)", name, n),
 			}
 		end
 	end
 
-	-- 3. Phase validity.
-	for _, a in ipairs(annots) do
-		if a.name and not a.error and a.phase and not valid_phase(a.phase) then
-			errors[#errors + 1] = {
-				line = a.line,
-				msg  = string.format("'%s' has unknown phase '%s' (expected one of init/bind/setup/work/commit/terminate)", a.name, a.phase),
-			}
-		end
-	end
+	-- 3. (Phase validity check DROPPED. Phases were removed from the
+	--    annotation DSL. They may be reintroduced later as sub-calls of
+	--    atom_info, at which point ordering checks will go here.)
 
 	-- 4. BIND atoms must reference a real Binds_* struct.
 	for _, a in ipairs(annots) do
@@ -764,16 +621,14 @@ local function validate(ctx, src)
 		end
 	end
 
-	-- 6. WORK reads should be a subset of BIND writes (the wave contract).
+	-- 6. INFO reads should be wave-context registers (or R_TapePtr for rbind).
 	for _, a in ipairs(annots) do
-		if a.kind == "work" then
-			for _, r in ipairs(a.reads) do
-				if not is_wave_context_reg(r) and r ~= "R_TapePtr" then
-					warnings[#warnings + 1] = {
-						line = a.line,
-						msg  = string.format("work atom '%s' reads '%s' which is not a known wave-context register", a.name, r),
-					}
-				end
+		for _, r in ipairs(a.reads) do
+			if not is_wave_context_reg(r) and r ~= "R_TapePtr" then
+				warnings[#warnings + 1] = {
+					line = a.line,
+					msg  = string.format("atom '%s' reads '%s' which is not a known wave-context register", a.name, r),
+				}
 			end
 		end
 	end
@@ -805,51 +660,19 @@ local function validate(ctx, src)
 		check_macro_drift(m, ctx.shared.word_counts[m.name])
 	end
 
-	-- 8. atom_<...> _Pragma validation: resource/region/group/cadence/async
-	for _, p in ipairs(pragmas) do
-		if not atom_index[p.name] then
-			errors[#errors + 1] = {
-				line = p.line,
-				msg  = string.format("pragma references unknown atom '%s'", p.name),
-			}
-		end
+	-- 8. (atom_<...> _Pragma validation DROPPED. The pragma macros
+	--    atom_resource / atom_region / atom_group / atom_cadence /
+	--    atom_async were removed from atom_dsl.h. They may be
+	--    reintroduced later as sub-calls of atom_info.)
 
-		for k, v in pairs(p.attrs) do
-			local spec = ATOM_PRAGMA_KINDS[k]
-			if not spec then
-				errors[#errors + 1] = {
-					line = p.line,
-					msg  = string.format("'%s' has unknown pragma key '%s' (allowed: resource/region/group/cadence/async)", p.name, k),
-				}
-			elseif spec.allowed and not spec.allowed[v] then
-				local allowed = {}
-				for kk in pairs(spec.allowed) do allowed[#allowed + 1] = kk end
-				table.sort(allowed)
-				local allowed_str = table.concat(allowed, ", ")
-				errors[#errors + 1] = {
-					line = p.line,
-					msg  = string.format("'%s' pragma %s=%s but '%s' is not allowed (allowed: %s)", p.name, k, v, v, allowed_str),
-				}
-			end
-		end
-	end
-
-	-- 9. CADENCE_ONDEMAND requires async=true. Flattened as a guard
-	--    (single condition, no nested if).
-	for _, p in ipairs(pragmas) do
-		if p.attrs.cadence == "ondemand" and p.attrs.async ~= "true" then
-			errors[#errors + 1] = {
-				line = p.line,
-				msg  = string.format("'%s' is CADENCE_ONDEMAND but does not declare atom_async(true)", p.name),
-			}
-		end
-	end
+	-- 9. (CADENCE_ONDEMAND requires async check DROPPED. Same reason
+	--    as #8.)
 
 	-- 10. Information summary.
 	info[#info + 1] = {
 		line = 0,
-		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d pragma(s), %d macro-word-decl(s), %d binds struct(s)",
-			#atoms, #annots, #pragmas, #macros, #binds),
+		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)",
+			#atoms, #annots, #macros, #binds),
 	}
 
 	return {
@@ -940,4 +763,4 @@ function M.run(ctx)
 	return { outputs = outputs, errors = errors, warnings = warnings }
 end
 
-return M
+return M

scripts/passes/components.lua

@@ -459,9 +459,7 @@ local function compute_component_word_count(c, components, wc)
 					-- It's a `mac_X(...)` call. Recurse.
 					n = n + rec(comp_name)
 				elseif comp_name and wc and wc[comp_name] then
-					-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2,
-					-- nop2 = 2). Trust the metadata — tape_atom.metadata.h is the
-					-- single source of truth for word counts.
+					-- Encoding macro or pseudo-instruction (e.g. mask_upper = 2, nop2 = 2).
 					n = n + wc[comp_name]
 				else
 					-- Unrecognized token. Fall back to 1 word.
@@ -628,12 +626,8 @@ local function emit_component_macros_h(ctx, src, components)
 		"// Auto-generated by tape_atom_annotation_pass.lua — DO NOT EDIT",
 		"// Source: " .. to_absolute_path(src.path),
 		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
-		"// + auto word-counts (so tape_atom.metadata.h stays manual-only",
-		"//   for encoding macros).",
 		"",
 		-- Self-contained: define WORD_COUNT if not already defined.
-		-- The metadata file (tape_atom.metadata.h) defines it as
-		--     enum { words_##name = (count) };
 		-- We use the same definition here so the auto-generated
 		-- entries below expand to compile-time constants whether
 		-- the metadata file is included first or not.
@@ -694,4 +688,4 @@ function M.run(ctx)
 	return { outputs = outputs, errors = errors, warnings = warnings }
 end
 
-return M
+return M

scripts/passes/offsets.lua

@@ -2,7 +2,6 @@
 --
 -- 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 +
@@ -392,4 +391,4 @@ function M.run(ctx)
 	return { outputs = outputs, errors = errors, warnings = warnings }
 end
 
-return M
+return M

scripts/passes/report.lua

@@ -38,9 +38,8 @@ local function render_source_report(source_path, result)
 	add("ANNOTATION PASS — " .. source_path)
 	add("========================================================")
 	add("")
-	add(string.format("Atoms: %d   Annotations: %d   Pragmas: %d   Binds structs: %d   Macro decls: %d",
+	add(string.format("Atoms: %d   Annotations: %d   Binds structs: %d   Macro decls: %d",
 		#result.atoms, #result.annots,
-		(result.pragmas and #result.pragmas or 0),
 		#result.binds, #result.macros))
 	add("")
 
@@ -55,9 +54,7 @@ local function render_source_report(source_path, result)
 		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)
+			local line = string.format("  ●  line %d  %s", a.line, a.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
@@ -81,16 +78,6 @@ local function render_source_report(source_path, result)
 	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
@@ -175,12 +162,13 @@ function M.run(ctx)
 	local annot_results = (ctx.flags and ctx.flags._annot_results) or {}
 
 	-- Render per-source reports.
+	-- Hoist ensure_dir out of the loop (cache + hoisting = single mkdir).
+	if not ctx.dry_run then ensure_dir(ctx.out_root) end
 	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))
 		end
 		table.insert(outputs, { annotations_txt = out_path })
@@ -190,12 +178,12 @@ function M.run(ctx)
 	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.
+		-- ensure_dir was already hoisted above; cache makes this a no-op.
 		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)
 		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 })
@@ -204,4 +192,4 @@ function M.run(ctx)
 	return { outputs = outputs, errors = errors, warnings = warnings }
 end
 
-return M
+return M

scripts/passes/static_analysis.lua

@@ -1,7 +1,635 @@
 -- passes/static_analysis.lua
 --
--- [FUTURE] Per-atom static-analysis checks. Stub for now; the upcoming
--- static_analysis_atoms_20260708 track will deliver the 5 (+1) checks.
+-- Per-atom static-analysis checks for the tape-atom build pipeline.
+-- Currently ships Phase 1 checks (GTE pipeline-fill + mac_yield
+-- uniformity). Phases 2/3 (ABI handoff discipline, GPU port-store
+-- shape, per-atom cycle budget) extend this file.
+--
+-- Workspace boundary: same conventions as annotation.lua
+--   - primitives from duffle.lua (read_parens, read_braces, scan_to_char, LineIndex)
+--   - LPeg not needed: pure hand-rolled string scanning
+--   - 5.3-compatible (no <close>, no continue keyword)
+--   - no :match/:gmatch
+--   - tab indent, EmmyLua @class/@param annotations
+--
+-- The orchestrator (ps1_meta.lua) wires this module in via the
+-- PASSES table:
+--   ["static-analysis"] = {
+--       module = "passes.static_analysis",
+--       kind   = "validation",      -- errors stop the build
+--       deps   = {"word-counts", "components"},
+--       out    = { { kind = "report",
+--                    path_template = "<out_root>/<basename>.static_analysis.txt" } },
+--   }
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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 read_ident         = duffle.read_ident
+local skip_ws_and_cmt    = duffle.skip_ws_and_cmt
+local read_parens        = duffle.read_parens
+local read_braces        = duffle.read_braces
+local read_brackets      = duffle.read_brackets
+local scan_to_char       = duffle.scan_to_char
+local split_top_level_commas = duffle.split_top_level_commas
+local trim               = duffle.trim
+local ensure_dir         = duffle.ensure_dir
+local write_file         = duffle.write_file
+local basename_no_ext    = duffle.basename_no_ext
+
+-- Latency table lives in duffle.lua (shared between this pass + future
+-- per-atom cycle-budget pass). Lazily read on first use.
+local GTE_PIPELINE_LATENCY = duffle.GTE_PIPELINE_LATENCY
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Source walkers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Walk source-as-written, return a list of `{line, name, body,
+--- body_off, kind}` for every:
+---   `MipsAtom_(name) { body };`               -> kind = "atom"     (baked atom)
+---   `MipsAtomComp_(name) { body };`           -> kind = "comp_bare" (static-array component)
+---   `MipsAtomComp_Proc_(name, { body })`     -> kind = "comp_proc" (procedural component)
+---
+--- All three forms are recognized because the user explicitly uses
+--- both bare components (e.g. `ac_gte_store_f3_post_rtpt`) and
+--- procedural components (e.g. `ac_format_f3_color(r, g, b)`) inside
+--- atom bodies. The two component forms generate macro equivalents
+--- (in gen/duffle.macs.h) that atoms call via `mac_*` -- so the parent
+--- atom body is what needs the GTE pipeline-fill + mac_yield checks.
+--- The component bodies themselves don't need mac_yield (control
+--- transfer is the parent atom's job) but they DO need pre-fill nops
+--- before any gte_cmdw_X they contain.
+---
+--- Comments / strings inside `name` and `body` are tolerated; `body`
+--- is the raw brace inner text (with surrounding whitespace, no
+--- leading/trailing `{` `}`). `body_off` is the character offset of
+--- `body[1]` in `source_text`, used to compute per-token line numbers
+--- later.
+local function find_atom_bodies(source_text)
+	local line_of  = duffle.LineIndex(source_text)
+	local out      = {}
+	local len      = #source_text
+	local i        = 1
+	while i <= len do
+		i = skip_ws_and_cmt(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_"
+		    or ident == "MipsAtomComp_"
+		    or ident == "MipsAtomComp_Proc_" then
+			-- Determine the kind from the exact ident (3 distinct macros,
+			-- each with its own kind).
+			local kind
+			if     ident == "MipsAtom_"             then kind = "atom"
+			elseif ident == "MipsAtomComp_"         then kind = "comp_bare"
+			else                                         kind = "comp_proc"
+			end
+
+			local open = skip_ws_and_cmt(source_text, after)
+			if source_text:sub(open, open) ~= "(" then
+				i = open + 1
+			else
+				local inner, after_paren = read_parens(source_text, open)
+
+				if kind == "comp_proc" then
+					-- MipsAtomComp_Proc_(sym, { body })
+					-- The body is inside the LAST `{ ... }` in the args
+					-- (the macro takes 2 args: sym name, then body in {}).
+					-- Find the last `{` in `inner`, then the matching `}`.
+					local last_open
+					for k = #inner, 1, -1 do
+						if inner:sub(k, k) == "{" then last_open = k; break end
+					end
+					if not last_open then
+						i = open + 1
+					else
+						-- Walk forward to find matching `}` honoring balanced
+						-- ()/[] and strings. We could call duffle.read_braces
+						-- from last_open+1, but read_braces expects to start at
+						-- the brace itself. Inline the walk for clarity.
+						local depth = 1
+						local j = last_open + 1
+						while j <= #inner and depth > 0 do
+							local c = inner:byte(j)
+							if c == 123 then
+								depth = depth + 1; j = j + 1
+							elseif c == 125 then
+								depth = depth - 1
+								if depth == 0 then break end
+								j = j + 1
+							elseif c == 40 then
+								local _, a = read_parens(inner, j); j = a
+							elseif c == 91 then
+								local _, a = read_brackets(inner, j); j = a
+							elseif c == 34 or c == 39 then
+								j = duffle.skip_str_or_cmt(inner, j) + 1
+							else
+								j = j + 1
+							end
+						end
+						if depth ~= 0 then
+							-- unmatched; bail
+							i = open + 1
+						else
+							-- First ident in `inner` is the comp name.
+							local name_match = inner:match("^%s*([%w_]+)")
+							local name = name_match or "?"
+							local body     = inner:sub(last_open + 1, j - 1)
+							-- body_off in full source: position right after the
+							-- LAST `{` in `inner`, which sits at `open+1+last_open`
+							-- (open+1 = just inside the outer paren, +last_open
+							-- = at the `{`).
+							local body_off = open + 1 + last_open
+							out[#out + 1] = {
+								line     = line_of(i),
+								name     = name,
+								body     = body,
+								body_off = body_off + 1,
+								kind     = kind,
+							}
+							i = after_paren
+						end
+					end
+				else
+					-- MipsAtom_(sym) { body };  OR
+					-- MipsAtomComp_(sym) { body };
+					-- name is the first arg, body is the FIRST { ... } after
+					-- the paren.
+					local a = 1
+					while a <= #inner and inner:sub(a, a):match("[%s]") do a = a + 1 end
+					local b = a
+					while b <= #inner and inner:sub(b, b):match("[%w_]") do b = b + 1 end
+					local name = inner:sub(a, b - 1)
+					if name == "" then
+						i = open + 1
+					else
+						local brace = scan_to_char(source_text, "{", after_paren)
+						if brace then
+							local body, after_brace = read_braces(source_text, brace)
+							local body_off = brace + 1
+							out[#out + 1] = {
+								line     = line_of(i),
+								name     = name,
+								body     = body,
+								body_off = body_off,
+								kind     = kind,
+							}
+							i = after_brace
+						else
+							i = open + 1
+						end
+					end
+				end
+			end
+		else
+			i = after
+		end
+	end
+	return out
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Body tokenizer (top-level comma splitter + per-token classification)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Build a map: `body_relative_char_offset` -> `body_relative_line`.
+--- Used by the checks to convert per-token offsets in the body to line
+--- numbers relative to the start of `body`. The atom's source-line of
+--- the body-start is added by the caller.
+---
+--- Simple line-counting: count `\n` chars from offset 1 up to the
+--- offset; that count + 1 is the line number (1-based).
+local function build_body_line_index(body)
+	local index = {}
+	local len = #body
+	local newline_count = 0
+	for i = 1, len do
+		if i > 1 then
+			index[i] = newline_count + 1   -- line of `i` relative to body
+		end
+		if body:byte(i) == 10 then        -- '\n'
+			newline_count = newline_count + 1
+		end
+	end
+	-- Offsets beyond the body still resolve to the final line
+	index[len + 1] = newline_count + 1
+	return index
+end
+
+--- Count of COP2-nop words contributed by a single top-level token.
+--   `nop`                  -> 1
+--   `nop2`                 -> 2 (i.e. `nop, nop` baked into one asm arg)
+--   `nop,` / `nop2,`        -> same as above; strip trailing comma defensively
+--   anything else          -> 0
+--
+-- (Branch-delay-slot nops like `branch_*(..., nop)` are tokenized
+-- separately by split_top_level_commas: the branch arg ends before
+-- the trailing comma, and `nop` becomes its own token. So no special
+-- handling is needed here.)
+local function nop_word_count(token)
+	local s = trim(token)
+	-- strip trailing comma(s) (defensive against raw text via, but our
+	-- tokenize_body already strips them; this is a safety net)
+	s = s:gsub(",$", "")
+	s = trim(s)
+	if s == "nop"  then return 1 end
+	if s == "nop2" then return 2 end
+	return 0
+end
+
+--- Tokenize the body inner-text into a flat list of `(token, body_rel_offset)`
+--- pairs (nested parens/braces/brackets are honored; comments and strings
+--- are skipped). `body_rel_offset` is the char offset within `body` of the
+--- start of the token — callers add it to the atom's `body_off` to get
+--- an absolute source position for line tracking.
+local function tokenize_body(body)
+	local out = {}
+	local len     = #body
+	local rel     = 1
+	while rel <= len do
+		-- Find next non-whitespace, non-comment start
+		local ws_end = skip_ws_and_cmt(body, rel)
+		if ws_end > rel then
+			rel = ws_end
+		end
+		if rel > len then break end
+
+		-- Find comma/newline/semicolon after this token. Read balanced
+		-- groups so commas inside parens/braces/brackets aren't treated
+		-- as separators. Comments / strings are skipped.
+		local i = rel
+		while i <= len do
+			local c = body:byte(i)
+			if c == 44 then break end               -- ','
+			if c == 10 then break end               -- '\n'
+			if c == 59 then break end               -- ';'
+			if c == 40 then                         -- '('
+				local _, a = read_parens(body, i); i = a
+			elseif c == 123 then                     -- '{'
+				local _, a = read_braces(body, i); i = a
+			elseif c == 91 then                      -- '['
+				local _, a = read_brackets(body, i); i = a
+			elseif c == 34 or c == 39 then           -- '"' or '\''
+				i = duffle.skip_str_or_cmt(body, i) + 1
+			else
+				i = i + 1
+			end
+		end
+		-- Extract token [rel .. i-1]
+		local tok = trim(body:sub(rel, i - 1))
+		if tok ~= "" then
+			out[#out + 1] = { tok = tok, rel = rel }
+		end
+		-- Move past the separator
+		if i <= len then
+			i = i + 1
+			-- Also skip whitespace before next token
+			local w = skip_ws_and_cmt(body, i)
+			if w > i then i = w end
+		end
+		rel = i
+	end
+	return out
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Check #1: GTE pipeline-fill
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Walk the token list. Whenever we hit a `gte_cmdw_<X>` token, count
+--- consecutive nop words starting at the next token. If count < the
+--- minimum declared in `GTE_PIPELINE_LATENCY[X]`, record a finding.
+---
+--- Aliases (`gte_cmdw_rotate_translate_perspective_single` etc.) are
+--- resolved against the lookup table directly; if a macro name is not
+--- in the table, emit a soft warning (the user might have added a new
+--- gte_cmdw_* but not updated duffle.lua).
+local function check_gte_pipeline_fill(atoms, findings, line_of)
+	-- Walk the token list. Whenever we hit a `gte_cmdw_<X>` token,
+	-- count consecutive `nop` words IMMEDIATELY PRECEDING it (the
+	-- source-level `nop2, gte_cmdw_X` idiom provides the pre-pipeline
+	-- fill that gte.h's wrapper functions provide internally). If
+	-- count < `GTE_PIPELINE_LATENCY[X]`, record a finding.
+	--
+	-- We count nops going backwards from the cmdw token, stopping at
+	-- the first non-nop token. Tokens like `mem_share` or `port_write`
+	-- (any non-nop) break the count. `gte_mv_to_data_r` (writes to
+	-- C2_DR registers) are non-nops in this sense -- they count as
+	-- "previous GTE state" but don't themselves count as pipeline
+	-- fill.
+	for _, a in ipairs(atoms) do
+		local tokens = tokenize_body(a.body)
+		local line_in_body = build_body_line_index(a.body)
+		local tn = #tokens
+		local ti = 1
+		while ti <= tn do
+			local tok = tokens[ti].tok
+			local cmdw_full = tok:match("^(gte_cmdw_[%w_]+)%s*[,%)]")
+			              or tok:match("^(gte_cmdw_[%w_]+)%s*$")
+			if cmdw_full then
+				local variant = cmdw_full:match("^gte_cmdw_(.+)$")
+				local need = GTE_PIPELINE_LATENCY[cmdw_full]
+				if need == nil then
+					-- alias or new gte_cmdw_<X> not yet in latency table
+					local line = a.line + line_in_body[tokens[ti].rel]
+					findings[#findings + 1] = {
+						atom  = a.name,
+						line  = line,
+						check = "gte_pipeline_fill",
+						kind  = "warning",
+						msg   = string.format(
+							"%s at line %d uses `gte_cmdw_%s` but that macro is not in duffle.GTE_PIPELINE_LATENCY -- add a min_nops entry",
+							a.name, line, variant),
+					}
+					ti = ti + 1
+				elseif need > 0 then
+					-- Count consecutive nops immediately BEFORE the cmdw
+					-- token. We walk tokens[ti - n] backwards, accumulating
+					-- nop_word_count, stopping at the first non-nop.
+					local have = 0
+					local where_ti = ti - 1
+					while where_ti >= 1 do
+						local n = nop_word_count(tokens[where_ti].tok)
+						if n == 0 then break end
+						have = have + n
+						where_ti = where_ti - 1
+					end
+					if have < need then
+						local line = a.line + line_in_body[tokens[ti].rel]
+						findings[#findings + 1] = {
+							atom  = a.name,
+							line  = line,
+							check = "gte_pipeline_fill",
+							kind  = "error",
+							msg   = string.format(
+								"%s at line %d needs %d nop word%s immediately BEFORE `gte_cmdw_%s`; only %d found",
+								a.name, line, need, need == 1 and "" or "s", variant, have),
+						}
+					end
+					ti = ti + 1
+				else
+					ti = ti + 1
+				end
+			else
+				ti = ti + 1
+			end
+		end
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Check #2: mac_yield uniformity
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Every atom body must contain exactly one `mac_yield()` call and it
+--- must be the LAST top-level token in the body (so the tape runtime
+--- can pick up cleanly at the next atom's bound registers).
+---
+--- Empty bodies are not currently flagged — runtime infrastructure
+--- atoms like `MipsAtom_(yield) { mac_yield() }` and `MipsAtom_(tape_exit)
+--- { jump_reg(rret_addr), nop }` are valid as-is; mac_yield at the end
+--- is the contract.
+local function check_mac_yield_uniformity(atoms, findings)
+	-- Per-kind semantics:
+	--   MipsAtom_          (baked atom): exactly 1 mac_yield at the end of
+	--                          the body. Control transfer is the atom's job.
+	--   MipsAtomComp_      (bare static-array component): ZERO mac_yield.
+	--                          The component is invoked from inside an atom
+	--                          body; the parent atom does the yield.
+	--   MipsAtomComp_Proc_ (procedural component): ZERO mac_yield.
+	--                          Same reasoning -- it's a function returning
+	--                          a MipsAtom slice, invoked from a parent atom.
+	--
+	-- The GTE pipeline-fill check applies to all 3 kinds (see
+	-- check_gte_pipeline_fill). Only the mac_yield rule branches on kind.
+	for _, a in ipairs(atoms) do
+		local tokens = tokenize_body(a.body)
+		local line_in_body = build_body_line_index(a.body)
+
+		local count = 0
+		local last_idx = 0
+		for i, t in ipairs(tokens) do
+			local tok = t.tok
+			-- Match `mac_yield(...)` or just `mac_yield`. The bareword
+			-- variant is rare in modern style but tolerated.
+			if tok:match("^mac_yield%s*%(") or tok == "mac_yield" then
+				count = count + 1
+				last_idx = i
+			end
+		end
+		local function line_for(idx)
+			return a.line + line_in_body[tokens[idx].rel]
+		end
+
+		if a.kind == "atom" then
+			-- Baked atom: exactly 1 yield at the end.
+			if count == 0 then
+				findings[#findings + 1] = {
+					atom  = a.name,
+					line  = a.line,
+					check = "mac_yield_uniformity",
+					kind  = "warning",
+					msg   = string.format(
+						"%s at line %d has no `mac_yield()`; every atom must hand control to the next via mac_yield at end",
+						a.name, a.line),
+				}
+			elseif count > 1 then
+				findings[#findings + 1] = {
+					atom  = a.name,
+					line  = line_for(last_idx),
+					check = "mac_yield_uniformity",
+					kind  = "warning",
+					msg   = string.format(
+						"%s at line %d has %d `mac_yield()` calls; exactly 1 is allowed",
+						a.name, line_for(last_idx), count),
+				}
+			elseif last_idx < #tokens then
+				-- 1 call, but not the last token. We DON'T fail if the
+				-- post-token is just `nop` or `nop2` or a branch with `, nop`
+				-- delay slot -- it's the standard "yield, then BD nop" idiom.
+				local post_non_nop = false
+				for j = last_idx + 1, #tokens do
+					local t = tokens[j].tok
+					if t ~= "" and t ~= "nop" and t ~= "nop2"
+					   and not t:match("%,%s*nop%)%s*$") then
+						post_non_nop = true
+						break
+					end
+				end
+				if post_non_nop then
+					findings[#findings + 1] = {
+						atom  = a.name,
+						line  = line_for(last_idx),
+						check = "mac_yield_uniformity",
+						kind  = "warning",
+						msg   = string.format(
+							"%s at line %d has `mac_yield()` at token %d/%d; the yield must be the LAST non-nop token in the body",
+							a.name, line_for(last_idx), last_idx, #tokens),
+					}
+				end
+			end
+		else
+			-- Component (comp_bare or comp_proc): ZERO yields. The parent
+			-- atom does the yield. A yield inside a component would either
+			-- be dead code (bare) or prematurely terminate the function
+			-- (proc). Both are bugs.
+			if count > 0 then
+				findings[#findings + 1] = {
+					atom  = a.name,
+					line  = line_for(last_idx),
+					check = "mac_yield_uniformity",
+					kind  = "warning",
+					msg   = string.format(
+						"%s at line %d is a %s component but has %d `mac_yield()` call(s); components must not yield (the parent atom does)",
+						a.name, line_for(last_idx), a.kind, count),
+				}
+			end
+		end
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-source validation
+-- ════════════════════════════════════════════════════════════════════════════
+
+local function validate(ctx, src)
+	local source = src.text
+	local atoms  = find_atom_bodies(source)
+
+	local findings = {}
+	check_gte_pipeline_fill(atoms, findings)
+	check_mac_yield_uniformity(atoms, findings)
+
+	-- Phase 2 (ABI handoff / GPU port-store shape) and Phase 3 (cycle
+	-- budget) hooks go here when those tracks are reactivated.
+	-- check_abi_handoff(atoms, findings)
+	-- check_gpu_portstore_shape(atoms, findings)
+	-- emit per-atom cycle counts via INSTRUCTION_LATENCY table
+
+	local errors   = {}
+	local warnings = {}
+	local info     = {}
+	for _, f in ipairs(findings) do
+		-- Per-finding severity is set by the check via `f.kind`
+		-- ("error" or "warning"). A `gte_pipeline_fill` finding can be
+		-- either severity (errors for missing nops; warnings for unknown
+		-- cmdw macros not in the latency table). Bin by `kind`, not by
+		-- check name.
+		if f.kind == "error" then
+			errors[#errors + 1] = { line = f.line, msg = f.msg }
+		else
+			warnings[#warnings + 1] = { line = f.line, msg = f.msg }
+		end
+	end
+	info[#info + 1] = {
+		line = 0,
+		msg  = string.format("scanned: %d atom bodies; %d findings", #atoms, #findings),
+	}
+
+	return {
+		atoms    = atoms,
+		findings = findings,
+		errors   = errors,
+		warnings = warnings,
+		info     = info,
+	}
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Per-source output: build/gen/<basename>.static_analysis.txt
+-- ════════════════════════════════════════════════════════════════════════════
+
+local function emit_static_analysis_txt(ctx, src, result)
+	local out_path = ctx.out_root .. "/" .. src.basename .. ".static_analysis.txt"
+	if ctx.dry_run then return out_path end
+	ensure_dir(ctx.out_root)
+
+	local lines = {}
+	local function add(s) lines[#lines + 1] = s end
+
+	add("========================================================")
+	add("STATIC ANALYSIS PASS -- " .. src.path)
+	add("========================================================")
+	add("")
+	-- Tally atoms by kind for the header summary
+	local n_atoms, n_bare, n_proc = 0, 0, 0
+	for _, a in ipairs(result.atoms) do
+		n_atoms = n_atoms + 1
+		if a.kind == "comp_bare" then n_bare = n_bare + 1
+		elseif a.kind == "comp_proc" then n_proc = n_proc + 1
+		end
+	end
+	local header_atoms = string.format("Atoms: %d", n_atoms)
+	if n_bare > 0 or n_proc > 0 then
+		header_atoms = header_atoms .. string.format("  (atoms: %d, comp_bare: %d, comp_proc: %d)",
+			n_atoms - n_bare - n_proc, n_bare, n_proc)
+	end
+	add(string.format("%s   Findings: %d   Errors: %d   Warnings: %d",
+		header_atoms, #result.findings, #result.errors, #result.warnings))
+	add("")
+
+	-- Group findings by atom for readability
+	local by_atom = {}
+	for _, f in ipairs(result.findings) do
+		by_atom[f.atom] = by_atom[f.atom] or {}
+		by_atom[f.atom][#by_atom[f.atom] + 1] = f
+	end
+
+		if next(by_atom) == nil then
+			add("  (no findings -- every atom passed all checks)")
+		else
+		add("── Findings by atom ─────────────────────────────────────")
+		for _, a in ipairs(result.atoms) do
+			local fs = by_atom[a.name]
+			if fs then
+				add(string.format("  %s   line %d", a.name, a.line))
+				for _, f in ipairs(fs) do
+					add(string.format("      [%s] %s", f.check, f.msg))
+				end
+			end
+		end
+	end
+
+	add("")
+	add("── Errors ──────────────────────────────────────────────")
+	if #result.errors == 0 then add("  (none)") end
+	for _, e in ipairs(result.errors) do
+		add(string.format("  X 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("")
+	add("── Info ────────────────────────────────────────────────")
+	for _, i_ in ipairs(result.info) do
+		add(string.format("  %s", i_.msg))
+	end
+
+	write_file(out_path, table.concat(lines, "\n") .. "\n")
+	return out_path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M.run — orchestrator entry
+-- ════════════════════════════════════════════════════════════════════════════
 
 --- @class M
 
@@ -10,7 +638,25 @@ local M = {}
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
-	return { outputs = {}, errors = {}, warnings = {} }
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	for _, src in ipairs(ctx.sources) do
+		local result = validate(ctx, src)
+		local out_path = emit_static_analysis_txt(ctx, src, result)
+		if out_path then
+			table.insert(outputs, { static_analysis_txt = out_path })
+		end
+		for _, e in ipairs(result.errors) do
+			errors[#errors + 1] = { line = e.line, msg = e.msg }
+		end
+		for _, w in ipairs(result.warnings) do
+			warnings[#warnings + 1] = { line = w.line, msg = w.msg }
+		end
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
 end
 
-return M
+return M

scripts/ps1_meta.lua

@@ -570,4 +570,4 @@ local function main(argv)
 	os.exit(0)
 end
 
-main({...})
+main({...})

scripts/word_count_eval.lua

@@ -81,10 +81,6 @@ local M = {}
 --- (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
@@ -109,10 +105,6 @@ end
 --- 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[]
@@ -138,10 +130,6 @@ end
 --- 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
@@ -177,9 +165,6 @@ end
 --- 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)