req c11,

Still don't have this thing in its final state for  the curse but its close.

code/duffle/atom_dsl.h

@@ -11,7 +11,7 @@
  * Pure macro anntation.
  * ---------------
  * Don't want to constraint the macro usage to some attribute placment constraint, etc, don't want ot dela with the compiler.
- * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip_over each expand to a C comment or to nothing
+ * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip each expand to a C comment or to nothing
  * (C preprocessor strips them to whitespace).
  *
  * ============================================================================
@@ -90,13 +90,18 @@
 #define atom_info(...)  /* atom_info(__VA_ARGS__) */
 
 /* ----------------------------------------------------------------------------
- * DEBUG SOURCE-STEP MARKERS
+ * DEBUG SOURCE-STEP MARKER
  *
- * Place atom_dbg_skip_over() before a MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_.
+ * Place `atom_dbg_skip` (BARE) before a MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_.
  * The following declaration kind determines whether the marker selects a whole atom or a component inline view.
  * The source scanner associates the marker with that declaration; placement diagnostics are handled by the annotation pass.
+ *
+ * Example:
+ *     atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+ *     atom_dbg_skip MipsAtomComp_(ac_yield) { ... };
+ *     atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { ... });
  * ----------------------------------------------------------------------------*/
-#define atom_dbg_skip_over() /* atom_dbg_skip_over: skip the following atom or component source view */
+#define atom_dbg_skip /* atom_dbg_skip: skip the following atom or component source view */
 
 /* ----------------------------------------------------------------------------
  * Typed-view annotations (Registry for DWARF RR_<R_X> chain resolution)
@@ -117,7 +122,7 @@
  *       The preferred correlation mechanism; atom_ctx is the escape hatch for non-natural cases.
  *
  * All three expand to C comments
- * (the bare-token convention matching `atom_reg` and `atom_dbg_skip_over`).
+ * (the bare-token convention matching `atom_reg` and `atom_dbg_skip`).
  * The Lua scanner reads the bare tokens in source-as-written; the C preprocessor strips them.
  * ----------------------------------------------------------------------------*/
 #define atom_type(T)        /* atom_type: associate <T> with the preceding enum entry (enum site) or this register (atom-info site) */

code/duffle/dsl.h

@@ -218,3 +218,6 @@ IA_ void assert(U8 cond) { if(cond){return;} else{debug_trap(); ms_exit_process(
 #endif
 #pragma endregion Debug
 #endif
+
+#define GCC_OPTIMIZATION_DISABLE _Pragma("GCC push_options") _Pragma("GCC optimize(\"O0\")")
+#define GCC_OPTIMIZATION_ENABLE  _Pragma("GCC pop_options")

code/duffle/gen/duffle.macs.h

@@ -9,6 +9,12 @@
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
 #endif
 
+/* atom_dbg_skip */
+/* ---------------------------------------------------------------------------
+ *  MACRO ATOM Components (Reusable Assembly Components)
+ *  These do NOT yield. They are expanded inline inside Tape Atoms.
+ * ---------------------------------------------------------------------------*/
+// The 'Yield' sequence for Tape Atoms (mac_yield).
 #define mac_yield(...) \
 	load_word(R_AtomJmp, R_TapePtr, 0) \
 ,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
@@ -16,6 +22,7 @@
 ,	nop
 WORD_COUNT(mac_yield, 4)
 
+/* atom_dbg_skip */
 /* Words: 3; Loads 3 S2 indices from the face array */
 #define mac_load_tri_indices(...) \
 	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)) \
@@ -23,6 +30,8 @@ WORD_COUNT(mac_yield, 4)
 ,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
 WORD_COUNT(mac_load_tri_indices, 3)
 
+/* atom_dbg_skip */
+/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
 #define mac_gte_load_tri_verts(...) \
 	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
 ,	add_u_self(R_AT, R_VertBase) \
@@ -74,16 +83,19 @@ WORD_COUNT(mac_insert_ot_tag_f3, 11)
 ,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
 WORD_COUNT(mac_insert_ot_tag_g4, 11)
 
+/* atom_dbg_skip */
 #define mac_pack_color_word(off, cmd, r, g, b) \
 	load_upper_i(R_AT, (cmd) << 8  | (b)) \
 ,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
 ,	store_word(  R_AT, R_PrimCursor, (off))
 WORD_COUNT(mac_pack_color_word, 3)
 
+/* atom_dbg_skip */
 #define mac_format_f3_color(r, g, b) \
 	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
 WORD_COUNT(mac_format_f3_color, 3)
 
+/* atom_dbg_skip */
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
  * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
 #define mac_gte_store_f3_post_rtpt(...) \
@@ -99,6 +111,7 @@ WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
 ,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
 WORD_COUNT(mac_format_g4_color, 12)
 
+/* atom_dbg_skip */
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
  * G4 triangle portion to p0/p1/p2.
  * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
@@ -113,6 +126,7 @@ WORD_COUNT(mac_format_g4_color, 12)
 ,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
 WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
 
+/* atom_dbg_skip */
 /* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
  * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
  * single-vertex result to SXY2; SXY0 still holds v0.screen from the

code/duffle/gp_vendor_sym.h

@@ -2,7 +2,7 @@
  *  duffle DSL — GPU Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the PSYQ-style CamelCase aliases for the canonical duffle GPU primitive setters and OT operations.
+ *  Provides the PSYQ-style CamelCase aliases for the duffle GPU primitive setters and OT operations.
  *  The duffle snake_case names are primary; this header is for users who prefer the PSYQ SDK function names from the legacy C API.
  *
  *  USAGE:   #include "duffle/gp_vendor_sym.h"   // after gp.h
@@ -24,7 +24,7 @@
  *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and descriptive; no vendor alias is provided for them.
  *
  *  The vendor mnemonics are NOT registered with the duffle word-count metadata (word_counts.metadata.h).
- *  They expand to the duffle canonical macros which DO have word-count entries
+ *  They expand to the duffle macros which DO have word-count entries
  *  (the ones emitted by mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13 (objdump byte-identical) holds.
  * ============================================================================ */
 

code/duffle/gte.h

@@ -353,41 +353,34 @@ enum { _C2_TX_SUBS_ = 0
 
 /* GTE command words for the common cases.
  *
- * These are pure compile-time integer constants — the C compiler
+ * These are pure compile-time integer constants — the C compiler constant-folds them into `.word` directives in .rodata.
- * constant-folds them into `.word` directives in .rodata. Use them
+ * Use them inside `asm_inline(...)` blocks (see `gte_rtpt` below for the idiom).
- * inside `asm_inline(...)` blocks (see `gte_rtpt` below for the
- * canonical idiom).
  *
  * Decomposition (per the `enc_gte_<field>` definitions above):
  *   gte_cmdw_<name> = gte_cmd_base | enc_gte_cmd(<cmd>)
  
- * The SF/MX/V/CV/LM fields are all zero in the common cases (standard
+ * The SF/MX/V/CV/LM fields are all zero in the common cases 
- * rotation-matrix, no scaling factor, V0 vector, translation vector,
+ * (standard rotation-matrix, no scaling factor, V0 vector, translation vector, no clamp),
- * no clamp), so the only varying bits are the `cmd` field.
+ * so the only varying bits are the `cmd` field.
  *
- * Naming follows the file's convention: `gte_cmd_*` is the raw
+ * Naming follows the file's convention: `gte_cmd_*` is the raw 6-bit `cmd` field id, `gte_cmdw_*`
- * 6-bit `cmd` field id, `gte_cmdw_*` is the fully-encoded 32-bit
+ *  is the fully-encoded 32-bit instruction word ready to drop into a `.word` directive.
- * instruction word ready to drop into a `.word` directive.
  *
  * --------------------------------------------------------------------------
  *  PsyQ-compatibility note (RTPS/RTPT):
- *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and
+ *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and RTPS as `cop2 0x0180001`.
- *  RTPS as `cop2 0x0180001`. Both have `0x20` set in the upper-reserved
+ *  Both have `0x20` set in the upper-reserved region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
- *  region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
+ *  Per psx-spec these bits are reserved/must-be-zero,
- *  Per psx-spec these bits are reserved/must-be-zero, but the real GTE
+ *  but the real GTE hardware and PCSX-Redux's GTE model both IGNORE them on these two commands
- *  hardware and PCSX-Redux's GTE model both IGNORE them on these two
+ *  (the perspective divide happens regardless of `sf`).
- *  commands (the perspective divide happens regardless of `sf`).
  *
- *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits
+ *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits clear),
- *  clear), PCSX-Redux's GTE checks those bits more strictly than the
+ *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops — 
- *  silicon does and RTPT silently no-ops — the floor's screen
+ *  the floor's screen coordinates come out as raw projection-of-rotation (Z never divided),
- *  coordinates come out as raw projection-of-rotation (Z never
+ *  `nclip` ends up wrong, and the triangle is culled.
- *  divided), `nclip` ends up wrong, and the triangle is culled.
  *
- *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to
+ *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to match the working bit pattern everyone has shipped for 25 years.
- *  match the working bit pattern everyone has shipped for 25 years.
+ *  NCLIP/OP/MVMVA stay spec-clean — their reserved bits really are zero in the original PsyQ source.
- *  NCLIP/OP/MVMVA stay spec-clean — their reserved bits really are
- *  zero in the original PsyQ source.
  * --------------------------------------------------------------------------
  */
 #define gte_cmdw_psyq_compat  (1u << 21 | enc_gte_sf(gte_sf_integer))
@@ -425,8 +418,8 @@ enum { _C2_TX_SUBS_ = 0
  * (XY at offset 0) and C2_VZ0 (Z at offset 4) using `lwc2`.
  *
  * Uses string-style GCC inline asm with `%0` substitution because the 
- * base register `r0` is a runtime GPR chosen by the compiler — it cannot
+ * base register `r0` is a runtime GPR chosen by the compiler.
- * be encoded into a static `.word` constant.
+ * It cannot be encoded into a static `.word` constant.
  *
  * Usage:
  *   asm_gte_load_v0(svector_ptr);
@@ -458,26 +451,21 @@ enum {
 
 /* gte_load_vN(r_ptr, base) — placeholder-punned lwc2 loaders
  *
- * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen
+ * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen GTE vector register, where `<base>` is the GPR number you pass in
- * GTE vector register, where `<base>` is the GPR number you pass in
  * (typically one of R_T4..R_T9 for the standard "3-pointer" pattern).
  *
  * The caller MUST bind `r_ptr` to that same GPR via a register variable:
  *     register V3_S2* p_in_12 __asm__("$12") = my_ptr;
  *     gte_load_v0(p_in_12, R_T4);   // R_T4 = 12, base is $12
  *
- * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register
+ * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register `p_in_12` can live in),
- * `p_in_12` can live in), which is exactly the register the .word
+ * which is exactly the register the .word constants expect. A `"$12"` clobber would conflict with the register-variable binding
- * constants expect. A `"$12"` clobber would conflict with the
+ * ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
- * register-variable binding ("asm specifier for variable conflicts
+ * The other ABI-clobbers ($2/$8/$9/$31) stay because the GTE instructions don't touch caller-saved GPRs but the kernel does treat them as volatile.
- * with asm clobber list"), so we omit it. The other ABI-clobbers
- * ($2/$8/$9/$31) stay because the GTE instructions don't touch
- * caller-saved GPRs but the kernel does treat them as volatile.
  *
  * WHICH REGISTER TO PICK
  * ----------------------
- * Any caller-saved GPR is safe. Recommended default for an RTPT-style
+ * Any caller-saved GPR is safe. Recommended default for an RTPT-style 3-pointer pipeline:
- * 3-pointer pipeline:
  *   gte_load_v0(p0, R_T4);  // $12
  *   gte_load_v1(p1, R_T5);  // $13
  *   gte_load_v2(p2, R_T6);  // $14
@@ -490,8 +478,7 @@ enum {
  *   clobbers section   : "$2", "$8", ..., "memory"      (from asm_clobber)
  *   3 colons total, GCC-legal. No string-syntax mnemonics in the .word body.
  *
- * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that
+ * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that starts the clobbers section. */
- * starts the clobbers section. */
 #define gte_load_v0(r_ptr, base) asm volatile(       \
 	asm_words( gte_lw_v0_xy(base), gte_lw_v0_z(base) ) \
 	asm_rpins,   r_use(r_ptr)                          \
@@ -510,11 +497,11 @@ enum {
 	asm_clobber: rlit(R_V0), rlit(R_T0), rlit(R_T1), rlit(R_RA), clb_mem_drain \
 )
 
-/* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — the canonical prelude to gte_cmd_rtpt.
+/* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — prelude to gte_cmd_rtpt.
  *
  * Loads all three GTE input vectors (6 words) from three separate pointers,
- * one per GTE vector register, each loaded from its own base GPR. Caller
+ * one per GTE vector register, each loaded from its own base GPR.
- * must bind each `pN` to `bN` via a register variable.
+ * Caller must bind each `pN` to `bN` via a register variable.
  *
  *   register V3_S2* p0 rgcc(R_T4) = verts[0].ptr;  // → __asm__("$12")
  *   register V3_S2* p1 rgcc(R_T5) = verts[1].ptr;  // → __asm__("$13")

code/duffle/gte_vendor_sym.h

@@ -2,7 +2,7 @@
  *  duffle DSL — GTE Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2 instructions as thin aliases to the canonical duffle macros in gte.h.
+ *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2 instructions as thin aliases to the duffle macros in gte.h.
  *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/gte_vendor_sym.h"   // after gte.h

code/duffle/lottes_tape.h

@@ -57,10 +57,10 @@ enum {
  * ---------------------------------------------------------------------------*/
 
 /* The 'Exit' Atom */
-MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
 /* Generalized Tape Engine Runner */
-FI_ void tape_run(Slice_U4 tape) { register U4* tp rgcc(R_TapePtr) = tape.ptr; asm volatile(
+NI_ void tape_run(Slice_MipsCode tape) { register U4* tp rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
 			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
 		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
@@ -91,8 +91,8 @@ FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }
 FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
 
-FI_ Slice_U4 tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_U4){ C_(U4*,tb->ptr), tb->used }; }
+FI_ Slice_MipsCode tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_MipsCode){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Slice_U4){ C_(U4*,tb.ptr),  tb.used }; }
+FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Slice_MipsCode){ C_(U4*,tb.ptr),  tb.used }; }
 #define tb_scope(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,tape_exit))
 
 #pragma endregion Tape Drive
@@ -104,22 +104,21 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
  * ---------------------------------------------------------------------------*/
 
 // The 'Yield' sequence for Tape Atoms (mac_yield). 
-MipsAtomComp_(ac_yield) {
+atom_dbg_skip MipsAtomComp_(ac_yield) {
 	load_word(R_AtomJmp, R_TapePtr, 0),
 	add_ui_self(         R_TapePtr, S_(MipsCode)),
 	jump_reg( R_AtomJmp), nop,
 };
 
 /* Words: 3; Loads 3 S2 indices from the face array */
-MipsAtomComp_(ac_load_tri_indices) {
+atom_dbg_skip MipsAtomComp_(ac_load_tri_indices) {
 	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
 	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
 	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
 };
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-atom_dbg_skip_over()
+atom_dbg_skip MipsAtomComp_(ac_gte_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),
@@ -158,7 +157,7 @@ MipsAtomComp_(ac_insert_ot_tag_g4) {
 /* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
  * byte offset. Internal helper used by the *_format_*_color macros. */
 FI_ MipsAtom ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
-MipsAtomComp_Proc_(ac_pack_color_word, {
+atom_dbg_skip MipsAtomComp_Proc_(ac_pack_color_word, {
 	load_upper_i(R_AT, (cmd) << 8  | (b)),
 	or_i_self(   R_AT, ((g)  << 8) | (r)),
 	store_word(  R_AT, R_PrimCursor, (off)),
@@ -167,11 +166,11 @@ MipsAtomComp_Proc_(ac_pack_color_word, {
 /* Words: 3; Emits the F3 command+color word (cmd byte | BLUE | GREEN | RED)
  * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields). */
 FI_ MipsAtom ac_format_f3_color(U1 r, U1 g, U1 b)
-MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
 
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
  * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
+atom_dbg_skip MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
 	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
 	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
 	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
@@ -199,7 +198,7 @@ MipsAtomComp_Proc_(ac_format_g4_color, {
  * three registers aligned with v0/v1/v2 you must store before RTPS).
  * The macro name declares the pipeline position; check #6 (GTE state-
  * machine validation) verifies the call site matches the declaration. */
-MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
+atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
 	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
 	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
 	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
@@ -211,7 +210,7 @@ MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
  * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
  * prevents).
  */
-MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
+atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
 
 #pragma endregion Macro Atom Components
 
@@ -295,7 +294,6 @@ internal MipsAtom_(set_gte_world) atom_info(
 /* DIAGNOSTIC 1: Pure tape loop test */
 internal MipsAtom_(diag_yield) { mac_yield() };
 
-// TODO(Ed): Reduce magic numbers/offsets
 /* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
 internal MipsAtom_(diag_color) {
 	store_word(  R_0, R_T7, 0), 
@@ -324,7 +322,6 @@ internal MipsAtom_(diag_color) {
 	mac_yield()
 };
 
-// TODO(Ed): Reduce magic numbers/offsets
 /* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
 internal MipsAtom_(diag_gte) {
 	/* Load 3 indices */

code/duffle/memory.h

@@ -103,7 +103,7 @@ FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
 	arena->used     = 0;
 }
 FI_ FArena farena_make(Slice mem) { FArena a; farena_init(& a, mem); return a; }
-I_ Slice   farena_push(FArena_R arena, U4 amount, Opt_farena o) {
+I_  Slice  farena_push(FArena_R arena, U4 amount, Opt_farena o) {
 	if (amount == 0) { return (Slice){}; }
 	U4 desired   = amount * (o.type_width == 0 ? 1 : o.type_width);
 	U4 to_commit = align_pow2(desired, o.alignment ?  o.alignment : MEM_ALIGNMENT_DEFAULT);

code/duffle/mips.h

@@ -436,7 +436,7 @@ enum { _BitOffsets = 0
 /* --- Shift-amount alias (matches the gas convention `\p3 = shamt`) --- */
 #define shift_amount(rd, rt, n)     shift_lleft(rd, rt, n)
 
-/* nop — canonical sll $0, $0, 0 */
+/* nop — sll $0, $0, 0 */
 #define nop  shift_lleft(rdiscard, rdiscard, 0)
 #define nop2 nop, nop
 

code/duffle/mips_vendor_sym.h

@@ -2,7 +2,7 @@
  *  duffle DSL — MIPS Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics as thin aliases to the canonical duffle macros in mips.h.
+ *  Provides the textbook MIPS assembly mnemonics as thin aliases to the duffle macros in mips.h.
  *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/mips_vendor_sym.h"   // after mips.h

code/gte_hello/gen/gte_hello.offsets.h

@@ -15,9 +15,9 @@ enum {
     atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
 };
 
-// --- atom: floor_f3_face (66 words) ---
+// --- atom: floor_f3_face (58 words) ---
 
-#define _atom_offset_culling_floor_f3_face_exit       29
+#define _atom_offset_culling_floor_f3_face_exit       25
 #define _atom_offset_bounds_chk_floor_f3_face_exit    13
 
 enum {

code/gte_hello/hello_gte.c

@@ -1,6 +1,6 @@
-#include "stdio.h"
+#include <stdio.h>
 #include <stdlib.h>
-#include "assert.h"
+#include <assert.h>
 // #include "libgpu.h"
 // #include "libetc.h"
 // #include "libgte.h"
@@ -99,8 +99,13 @@ typedef Struct_(Ent_Floor) {
 	A2_V3_S2 faces;
 };
 
-enum { scratchpad_size = 1024, };
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
 typedef Struct_(SMemory) {
+	U4 MemTape[MemTape_Len];
+
 	DoubleBuffer            screen_buf;
 	A2_OrderingTable_Buffer ordering_tbl;
 	PrimitiveArena          primitives;
@@ -182,6 +187,7 @@ void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_
 void render(void) {
 }
 
+GCC_OPTIMIZATION_DISABLE
 void update(PrimitiveArena* pa, U4* ordering_buf) 
 {
 	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
@@ -207,6 +213,8 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 	A2_S2 p;    //???
 	S4 flag; //????
 
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
 	// Draw Cube
 	if (0)
 	{
@@ -259,8 +267,7 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
 
-		LP_ U4 mem_temp_tape[512];
+		tb.used = 0; tb_scope(& tb) {
-		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
 			tb_emit(& tb, rbind_cube_g4_face);
 				tb_data(& tb, prim_cursor);
 				tb_data(& tb, u4_(smem.cube.faces));
@@ -344,12 +351,11 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
 
-		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris.
+		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris?
 		// The tape atoms in-flight should not need to care.
 
 		// Prepare the tape. (Push protocol to tape)
-		LP_ U4 mem_temp_tape[512];
+		tb.used = 0; tb_scope(& tb) {
-		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
 			tb_emit(& tb, set_gte_world);
 				tb_data(& tb, u4_(& smem.tform_world));
 
@@ -367,25 +373,18 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}
-
 		tape_run(tb_slice(tb));// Fire off the tape.
 
 		// C-side state (pa->used) has already been updated by the tape!
 		smem.floor.rot.y += 5;
 	}
 	// --- TAPE DIAGNOSTICS ---
-	if (1)
+	if (0)
 	{
 		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
 		TapeBuilder tb = tb_make_old(& tape_arena); tb_scope(& tb) {
 			// Skip set_gte_world atom for diagnostics to isolate the triangle loop
 			for (U4 i = 0; i < Floor_num_faces; i++) {
-				// =======================================================
-				// SWAP EMIT TO TEST DIFFERENT PARTS OF THE PIPELINE:
-				// =======================================================
-				// 1. code_diag_yield -> Tests Tape Engine jump logic
-				// 2. code_diag_color -> Tests OT and Prim Arena memory
-				// 3. code_diag_gte   -> Tests Vertex arrays and GTE Math
 				// tb_emit(& tb, code_diag_yield);
 				// tb_emit(& tb, code_diag_color);
 				// tb_emit(& tb, code_diag_gte); 
@@ -394,9 +393,9 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		B1* prim_cursor = (B1*)r_(pa->buf)[smem.active_buf_id] + pa->used;
 		tape_run(tb_slice(tb));
 		pa->used = (U4)prim_cursor - (U4)r_(pa->buf)[smem.active_buf_id];
-		smem.floor.rot.y += 5;
 	}
 }
+GCC_OPTIMIZATION_ENABLE
 
 int main(void)
 {

code/gte_hello/hello_gte_tape.c

@@ -103,28 +103,27 @@ MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(f
 	mac_yield()
 };
 
+atom_dbg_skip
 internal
-atom_dbg_skip_over()
 MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-	, atom_writes(R_PrimCursor, R_FaceCursor)
+	, atom_writes(R_PrimCursor, R_FaceCursr)
 ) {
 	mac_load_tri_indices(  R_T0, R_T1, R_T2),
 	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
-	nop2, gte_cmdw_nclip,
+	gte_cmdw_nclip,
 
 	/* Culling (Branch forward if Backface) */
-	nop2, gte_mv_from_data_r(R_T0, C2_MAC0),
+	gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
-	branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop,
 	/* Format Primitive */
 	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
 	mac_gte_store_f3_post_rtpt(),
 
 	/* Calculate Depth */
-	nop2, gte_avg_sort_z3,
+	gte_avg_sort_z3,
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
+	gte_mv_from_data_r(R_T1, C2_OTZ),
 	/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
 	add_ui(      R_AT, R_0,  OrderingTbl_Len),
 	set_lt_u(    R_AT, R_T1, R_AT),

scripts/build_psyq.ps1

@@ -81,19 +81,6 @@ $path_psyq          = join-path $path_toolchain  'psyq-4_7'
 $path_psyq_iwyu     = join-path $path_toolchain  'psyq_iwyu'
 $path_psyq_imyu_inc = join-path $path_psyq_iwyu  'include'
 
-function Get-SourceFiles { param([Parameter(Mandatory=$true)] [string[]]$paths, [Parameter(Mandatory=$true)] [string[]]$extensions)
-	$files = @()
-	foreach ($p in $paths) {
-		if (-not (test-path $p)) { continue }
-		foreach ($ext in $extensions) {
-			Get-ChildItem -Path $p -File -Recurse -Filter "*$ext" -ErrorAction SilentlyContinue | ForEach-Object {
-				$files += $_.FullName
-			}
-		}
-	}
-	return ($files | Sort-Object -Unique)
-}
-
 function assemble-unit { param(
 	[string]  $unit,
 	[string]  $link_module,
@@ -153,7 +140,7 @@ function compile-unit { param(
 		$f_arch_no_shared,
 		$f_arch_no_stack_prot
 	)
-	# $compile_args    += $f_std_c23
+	$compile_args    += $f_std_c11
 	$compile_args    += ($f_include + $path_psyq_imyu_inc)
 	$compile_args    += ($f_include + $path_nugget)
 
@@ -243,6 +230,52 @@ function make-binary { param([string]$elf, [string]$exe)
 	if ($LASTEXITCODE -ne 0) { Write-Error "Objcopy failed. Aborting."; exit 1 }
 }
 
+function ps1-meta { param(
+		[string]$unity_root,
+		[string[]]$sources,
+		[Parameter(Mandatory=$true)][string]$metadata,
+		[string]$out_root = (join-path $path_build 'gen'),
+		[string[]]$passes = @('--pre-link'),
+		[string[]]$extra_args = @()
+	)
+	# `--unity-root` and `--source` are
+	# mutually exclusive. Exactly one of `$unity_root` / `$sources` must
+	# be supplied; the other must be absent.
+	if ($null -ne $unity_root -and $unity_root -ne '') 
+	{
+		if ($null -ne $sources -and $sources.Count -gt 0) {
+			write-error 'ps1-meta: -unity_root and -sources are mutually exclusive'
+			exit 2
+		}
+	} 
+	elseif ($null -eq $sources -or $sources.Count -eq 0) {
+		write-error 'ps1-meta: either -unity_root <file> or -sources <file...> is required'
+		exit 2
+	}
+
+	$script = join-path $path_scripts 'ps1_meta.lua'
+	$input_summary = if ($null -ne $unity_root -and $unity_root -ne '') {
+		"unity=$unity_root"
+	}
+	else {
+		"$($sources.Count) source(s)"
+	}
+	write-host "ps1-meta  $input_summary, passes=$($passes -join ',')" ` -ForegroundColor Magenta
+
+	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root) + @($extra_args)
+	if ($null -ne $unity_root -and $unity_root -ne '') {
+		$arg_list += @('--unity-root', $unity_root)
+	}
+	else {
+		foreach ($s in $sources) { $arg_list += @('--source', $s) }
+	}
+	& luajit $script @arg_list
+	if ($LASTEXITCODE -ne 0) {
+		write-error "ps1-meta failed (exit $LASTEXITCODE). Aborting."
+		exit $LASTEXITCODE
+	}
+}
+
 function build-hello_psyqo {
 	$includes += @()
 
@@ -317,34 +350,16 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function ps1-meta { param(
-		[Parameter(Mandatory=$true)][string[]]$sources,
-		[Parameter(Mandatory=$true)][string]$metadata,
-		[string]$out_root = (join-path $path_build 'gen'),
-		[string[]]$passes = @('--pre-link'),
-		[string[]]$extra_args = @()
-	)
-	$script = join-path $path_scripts 'ps1_meta.lua'
-	write-host "ps1-meta  $($sources.Count) source(s), passes=$($passes -join ',')" ` -ForegroundColor Magenta
-	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root) + @($extra_args)
-	foreach ($s in $sources) { $arg_list += @('--source', $s) }
-	& luajit $script @arg_list
-	if ($LASTEXITCODE -ne 0) {
-		write-error "ps1-meta failed (exit $LASTEXITCODE). Aborting."
-		exit $LASTEXITCODE
-	}
-}
-
 function build-gte_hello {
 	$includes += @()
 
 	$path_module        = join-path $path_code   'gte_hello'
 	$path_duffle        = join-path $path_code   'duffle'
 	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build   'gen'
 
-	$source_dirs  = @($path_duffle, $path_module)
+	$src_c = join-path $path_module 'hello_gte.c'
-	$atom_sources = Get-SourceFiles -paths $source_dirs -extensions @('.h', '.c')
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
-	ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -360,7 +375,6 @@ function build-gte_hello {
 
 	# assemble-unit $src_asm $module_asm $includes $assemble_args
 
-	$src_c    = join-path $path_module 'hello_gte.c'
 	$module_c = join-path $path_build  'hello_gte_c.o'
 
 	$compile_args = @()
@@ -386,27 +400,17 @@ function build-gte_hello {
 	make-binary $elf $exe
 
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
-	ps1-meta -sources $atom_sources -metadata $path_atom_metadata `
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
-		-out_root (join-path $path_build 'gen') `
-		-passes @('--post-link') `
-		-extra_args @('--elf', $elf)
 
-	# F' + G' splice: collapse 9 objcopy subprocess invocations into 3.
+	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
-	# - 1 call: 3x --update-section for F' (line / aranges / rnglists)
+	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
-	# - 1 call: 3x --update-section for G' (info / abbrev / str)
+	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
-	# - 1 call: 2x --add-section for G' (loc / loclists — these don't exist in the source ELF)
-	# - 1 call: 1x --set-section-flags (.rodata / .data enable code flag)
-	# = 4 objcopy calls (was 9; saved 5 spawns).
-	$dwarfLineBin     = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_line.bin'
-	$dwarfArangesBin  = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_rnglists.bin'
 	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
 	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
 	{
 		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
 		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-
+		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		# Single objcopy call: 3x --update-section for F' (line, aranges, rnglists).
 		$f_args = @(
 			"--update-section=.debug_line=$dwarfLineBin",
 			"--update-section=.debug_aranges=$dwarfArangesBin",
@@ -419,12 +423,11 @@ function build-gte_hello {
 			return;
 		}
 
-		# G' 5-section splice: 3 update-section (info / abbrev / str) + 2 add-section (loc / loclists).
+		$dwarfInfoBin     = join-path $path_build_gen 'hello_gte.dwarf_info.bin'
-		$dwarfInfoBin     = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_info.bin'
+		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_gte.dwarf_abbrev.bin'
-		$dwarfAbbrevBin   = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_abbrev.bin'
+		$dwarfStrBin      = join-path $path_build_gen 'hello_gte.dwarf_str.bin'
-		$dwarfStrBin      = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_str.bin'
+		$dwarfLocBin      = join-path $path_build_gen 'hello_gte.dwarf_loc.bin'
-		$dwarfLocBin      = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_loc.bin'
+		$dwarfLoclistsBin = join-path $path_build_gen 'hello_gte.dwarf_loclists.bin'
-		$dwarfLoclistsBin = join-path (join-path $path_build 'gen') 'hello_gte.dwarf_loclists.bin'
 		$g_args = @(
 			"--update-section=.debug_info=$dwarfInfoBin",
 			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
@@ -441,7 +444,7 @@ function build-gte_hello {
 
 		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
 		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
-		# The shipping ELF and PS-EXE remain byte/flag unchanged.
+		# The original ELF and PS-EXE remain byte/flag unchanged.
 		& $Objcopy `
 			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
 			--set-section-flags ".data=alloc,load,data,code,contents" `
@@ -449,7 +452,8 @@ function build-gte_hello {
 		if ($LASTEXITCODE -ne 0) {
 			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
 			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-		} else {
+		} 
+		else {
 			Write-Host "[build] DWARF-injected ELF: $injectElf"
 		}
 	}

scripts/duffle_paths.lua

@@ -11,11 +11,10 @@
 ---   ```
 ---
 --- That small bootstrap: (a) locates this helper via `arg[0]` / `debug.getinfo`,
---- (b) loads it (which sets `package.path` + `package.cpath` via cached `git rev-parse`),
+--- (b) loads it (which sets `package.path` + `package.cpath`),
 --- (c) at the bottom calls `require("duffle")` (now resolvable since `package.path` was just set) and returns the duffle M.
 --- Net effect: the caller gets the duffle module in one statement; no separate `dofile(...)` + `require("duffle")` dance.
 ---
---- Replaces the prior 2-line (entry) or 4-line (pass) pattern that had the call site do its own path resolution + duplicated setup.
 
 local M = {}
 
@@ -27,9 +26,6 @@ local CACHE_KEY = "__duffle_repo_root__"
 --- parent of the directory containing this script. We derive it directly from `debug.getinfo(1, "S").source`
 --- (returns `@<path>` for the currently-running chunk).
 ---
---- Replaces the prior `io.popen("git rev-parse --show-toplevel")` approach, which cost ~100-180ms per
---- LuaJIT process on Windows due to git's CLI startup. The path-derive approach costs <1ms.
----
 --- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source),
 --- return nil and let `M.setup()` fail loud.
 --- @return string|nil
@@ -61,7 +57,12 @@ end
 function M.setup()
 	local repo_root = find_repo_root()
 	if not repo_root then
-		io.stderr:write("[duffle_paths] git rev-parse failed -- not in a git repo?\n")
+		-- Unreachable in practice: find_repo_root() derives the repo root from this script's
+		-- own source path via debug.getinfo(1, "S").source (no subprocess, no git CLI, <1ms).
+		-- A nil return means the source path did not match the expected
+		-- <repo>/scripts/duffle_paths.lua layout — a packaging bug, not a "missing git repo"
+		-- condition. os.exit(2) is retained so a real failure surfaces loud rather than
+		-- silently producing an unconfigured module table.
 		os.exit(2)
 	end
 

scripts/elf_dwarf.lua

@@ -240,7 +240,7 @@ end
 
 -- Pure-Lua 5.3 LEB128 readers (no `bit` library). `2^shift` arithmetic matches the existing parser.
 -- Offsets are 0-based; returns (value, next_pos).
--- Track A Task 10: promoted from `local function` to M.* exports so passes/dwarf_injection.lua
+-- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua
 -- can import them as file-scope locals per the 2nd-caller lift precedent
 -- (the uleb128 + sleb128 encoders were promoted the same way).
 function M.read_uleb128_at(buf, pos)
@@ -404,10 +404,8 @@ local function read_form_value(buf, str_buf, pos, form)
 end
 
 --- Read a `DW_FORM_ref_sig8` value at 0-based offset `pos` from `buf`.
---- Returns the low 4 bytes (LE) as `low`, the high 4 bytes (LE) as `high`, and
+--- Returns the low 4 bytes (LE) as `low`, the high 4 bytes (LE) as `high`, and the cursor position after the 8-byte value as `next_pos`.
---- the cursor position after the 8-byte value as `next_pos`.
+--- Callers that need the full type-unit + type-offset pair (e.g. to resolve a type identifier embedded as a signature)
---- Callers that need the full type-unit + type-offset pair
---- (e.g. to resolve a type identifier embedded as a signature)
 --- should use this directly rather than going through `read_form_value`,
 --- which only exposes the low 4 bytes to preserve its existing (value, next_pos) return shape.
 --- @param buf string
@@ -427,8 +425,7 @@ end
 --
 -- Unit header layout (from pos 0):
 --   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
---   -- followed by type_unit_specific fields:
+--   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
---   type_signature(8) + type_offset(4)
 -- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
 -- @param info string  -- the .debug_info section bytes
 -- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
@@ -619,7 +616,7 @@ end
 --- - ELF32 symtab entry = 16 bytes (`st_name:4 + st_value:4 + st_size:4 + st_info:1 + st_other:1 + st_shndx:2`); offsets within each entry are zero-based wire offsets.
 --- - Direct Lua `string.byte`/`string.sub`/`string.find` boundaries receive `+ 1`.
 --- - We filter on STB_GLOBAL (high nibble of st_info = 1) to match `nm`'s default (external symbols only). STB_WEAK excluded.
---- - We strip the `code_` prefix to match the previous `read_nm` output. 
+--- - The `code_` prefix is stripped (MipsAtom_ macros emit bare atom names, no `code_` prefix).
 --- - `st_size > 0` filter excludes undefined/imported symbols.
 --- @param elf_path Path
 --- @return table<string, {integer, integer}>
@@ -657,7 +654,7 @@ function M.read_nm(elf_path)
 				-- Extract the name from .strtab (null-terminated C string).
 				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
 				local name     = strtab:sub(st_name_off + 1, name_end - 1)
-				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` since the `code_` prefix was removed from the MipsAtom_ macro).
+				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — MipsAtom_ macros strip the `code_` prefix).
 				-- The atoms_source_map pass already filters out non-atom symbols via the source-map.txt cross-ref.
 				if name and #name > 0 then
 					local st_value = M.read_u32_le(symtab, entry_off + SYM_ST_VALUE)
@@ -794,132 +791,4 @@ end
 -- I/O helpers: atoms source-map + native directory glob
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Parse a FORMAT_VERSION <expected_version> atoms-meta file (sourcemap or provenance).
---- Shared by M.parse_source_map_file + M.parse_provenance_file.
---- The two callers differ only in how they parse WORD lines; that's `extract_word(line)`.
---- Returns the standard `{name -> {total, words}}` shape.
---- Returns `{}` on format-version mismatch (and logs to stderr).
---- @param path             string
---- @param expected_version integer
---- @param extract_word     fun(line: string): table|nil  -- caller-supplied per-line parser
---- @return table<string, table>
-function M.parse_atom_records(path, expected_version, extract_word)
-	local out = {}
-	local cur_name, cur_words = nil, {}
-	for raw in io.lines(path) do
-		local line = raw
-		if line:match("^#") then
-			local ver = line:match("^# FORMAT_VERSION%s+(%d+)")
-			if ver and tonumber(ver) ~= expected_version then
-				io.stderr:write(string.format(
-					"[elf_dwarf.parse_atom_records] version mismatch (got %s, expected %d) in %s\n",
-					ver, expected_version, path))
-				return {}
-			end
-			-- skip other comments
-		elseif line:sub(1, 4) == "ATOM" then
-			-- ATOM <name> "<abs-source-path>" <total>
-			local _, _, name = line:find("ATOM%s+(%S+)%s+\"[^\"]*\"%s+(%d+)")
-			if name then
-				cur_name  = name
-				cur_words = {}
-				out[name] = { total = 0, words = cur_words }
-			end
-		elseif line == "ENDATOM" then
-			-- Update the recorded total from the entries count
-			-- (matches the `lines[1] = lines[1]:gsub(" 0$", " " .. total)` patch in atoms_source_map.lua:170).
-			if cur_name and out[cur_name] then
-				out[cur_name].total = #cur_words
-			end
-			cur_name, cur_words = nil, {}
-		elseif line:sub(1, 4) == "WORD" and cur_name then
-			local field = extract_word(line)
-			if field then
-				cur_words[#cur_words + 1] = field
-			end
-		end
-	end
-	return out
-end
-
---- Parse a FORMAT_VERSION <expected_version> `*.atoms.sourcemap.txt` file.
---- Returns `{name -> {total = N, words = {{pos, line}, ...}}}`.
---- Returns `{}` on format-version mismatch (and logs to stderr).
----
---- **Wire format** (emitted by `passes/atoms_source_map.lua`):
---- ```
---- # FORMAT_VERSION <n>
---- ATOM <name> "<abs-source-path>" <total>
---- WORD <n> LINE <line> TEXT <text...>
---- ...
---- ENDATOM
---- ```
----
---- **Conventions:** the in-memory shape uses `{pos, line, text}`
---- (`atoms_source_map.lua:142`); the `.txt` file uses `WORD <n>` so the parser maps `n` → `pos` field name.
---- @param sm_path Path
---- @param expected_version integer  -- expected FORMAT_VERSION line
---- @return table<string, table>
-function M.parse_source_map_file(sm_path, expected_version)
-	return M.parse_atom_records(sm_path, expected_version, function(line)
-		local _, n, _, src_line = line:find("WORD%s+(%d+)%s+LINE%s+(%d+)")
-		if n and src_line then
-			return { pos = tonumber(n), line = tonumber(src_line) }
-		end
-	end)
-end
-
---- Parse a FORMAT_VERSION <expected_version> `*.atoms.provenance.txt` file.
---- Returns `{name -> {total = N, words = {{pos, call_file, call_line, comp_name, comp_file, comp_line}, ...}}}`.
---- Returns `{}` on format-version mismatch (and logs to stderr).
----
---- **Wire format** (emitted by `passes/atoms_source_map.lua`):
---- ```
---- # FORMAT_VERSION <n>
---- ATOM <name> "<abs-source-path>" <total>
---- WORD <n>  CALL <src-file>:<src-line>  RAW
---- WORD <n>  CALL <src-file>:<src-line>  MACRO <comp_name> "<comp-file>:<comp-line>"
---- ...
---- ENDATOM
---- ```
----
---- **Used by** `passes/dwarf_injection.lua` to:
----   - group consecutive MACRO rows into component invocations (one `DW_TAG_inlined_subroutine` each)
----   - emit abstract `DW_TAG_subprogram` per unique component name
----   - extend `.debug_line` so stepping into a `mac_X(...)` lands on the component's source line.
---- @param prov_path string  -- path to *.atoms.provenance.txt
---- @param expected_version integer  -- expected FORMAT_VERSION line
---- @return table<string, table>
-function M.parse_provenance_file(prov_path, expected_version)
-	return M.parse_atom_records(prov_path, expected_version, function(line)
-		-- Two accepted shapes:
-		--   WORD <n>  CALL <call-file>:<call-line>  RAW
-		--   WORD <n>  CALL <call-file>:<call-line>  MACRO <comp_name> "<comp-file>:<comp-line>"
-		local pos, call_file, call_line, comp_name, comp_file, comp_line =
-			line:match('WORD%s+(%d+)%s+CALL%s+(.-):(%d+)%s+MACRO%s+(%S+)%s+"([^"]*):(%d+)"')
-		if pos then
-			return {
-				pos       = tonumber(pos),
-				call_file = call_file,
-				call_line = tonumber(call_line),
-				comp_name = comp_name,
-				comp_file = comp_file,
-				comp_line = tonumber(comp_line),
-			}
-		end
-		-- RAW row.
-		local raw_pos, raw_file, raw_line = line:match('WORD%s+(%d+)%s+CALL%s+(.-):(%d+)%s+RAW')
-		if raw_pos then
-			return {
-				pos       = tonumber(raw_pos),
-				call_file = raw_file,
-				call_line = tonumber(raw_line),
-				comp_name = nil,
-				comp_file = nil,
-				comp_line = nil,
-			}
-		end
-	end)
-end
-
 return M

scripts/launch_pcsx_debug.ps1

@@ -28,11 +28,6 @@ param(
 
 $ErrorActionPreference = 'Stop'
 
-$gdbInitPath = [System.IO.Path]::GetFullPath((Join-Path $PSScriptRoot '..\build\gen\hello_gte.gdbinit'))
-if (-not (Test-Path -LiteralPath $gdbInitPath -PathType Leaf)) {
-    Write-Warning "Generated GDB skip sidecar missing (non-fatal): $gdbInitPath. Run the GTE build to regenerate it; debugger launch will continue without generated skip-over commands."
-}
-
 # ── Pre-checks ──
 foreach ($p in @($PcsxPath, $ExePath, $HelperZip)) {
     if (-not (Test-Path $p)) {

scripts/passes/annotation.lua

@@ -1,27 +1,26 @@
 --- passes/annotation.lua — Atom-annotation DSL validator.
 ---
 --- Validates `MipsAtom_(name) atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...)) { ... }` declarations in source files.
---- Also reads: `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`)
+--- Also reads `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`).
 ---
---- Source scanning: done ONCE upstream by `duffle.scan_source()` (ps1_meta.lua pre-scans each source and stashes the result in `src.scan`).
+--- `duffle.scan_source()` scans each source once upstream; `ps1_meta.lua` stores that result in `src.scan`.
 ---
---- Writes:
+--- Ownership: the canonical `ctx.shared.corpus` supplies cross-source registries, while each `src.scan` supplies its source's declarations and bodies.
----   - `<ctx.out_root>/<dir_basename>.errors.h` — one per module, with `#error` directives on findings (the C compile will surface the error)
+--- A context without `ctx.shared.corpus` is rejected with an explicit canonical-corpus message.
----   - The annotations.txt report is rendered by `passes/report.lua` from the per-module results stashed in `ctx.flags._annot_results`
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible
+--- Writes `<ctx.out_root>/<dir_basename>.errors.h` once per module, with `#error` directives for findings that the C compile surfaces.
+--- `passes/report.lua` renders annotations.txt from `corpus.sources_by_dir`, re-validating each source through `M.validate()`.
+---
+--- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
 
--- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
+-- Bootstrap follows the entry scripts; `scripts/duffle_paths.lua` sets package.path and package.cpath. See `ps1_meta.lua` for the rationale.
--- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- `debug.getinfo(1, "S").source` locates this file for standalone and orchestrated runs, then `duffle_paths.lua` returns the loaded `duffle` module.
--- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
--- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 local write_file     = duffle.write_file
 local ensure_dir     = duffle.ensure_dir
 
--- The annotation pass now consults the source-derived registries built by scan_source:
+-- The annotation pass reads the source-derived registries from scan_source:
 --   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
 --   * pipe_ctx.type_name_registry      — for atom_dbg_reg_default(<T>, ...) and atom_reg_types(<T>, ...) type-identity checks
 
@@ -45,8 +44,6 @@ local ensure_dir     = duffle.ensure_dir
 --- @field project_root       string
 --- @field upstream           table<string, table>
 --- @field flags              table
---- @field flags._annot_results table[] -- stashed by annotation pass; consumed by report.lua
---- @field dry_run            boolean
 --- @field verbose            boolean
 
 --- @class PassResult
@@ -64,15 +61,15 @@ local ensure_dir     = duffle.ensure_dir
 --- @field writes  string[]      -- R_* names (write targets)
 --- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
 
---- @class SkipOverMarker  -- sub-shape of scan_source.lua's @class SkipOverMarker
+--- @class DebugSkipMarker  -- sub-shape of scan_source.lua's @class DebugSkipMarker
---- @field marker_kind    string                 -- exact marker ident (always "atom_dbg_skip_over")
+--- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
 --- @field marker_line    integer
 --- @field args           string|nil             -- trimmed text inside the parens (nil when has_parens is false)
 --- @field has_parens     boolean
+--- @field is_bare        boolean                -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
 --- @field pending        boolean                -- true while awaiting the following declaration
 --- @field superseded_by_marker_line integer|nil -- set on a marker that was bumped out of the pending slot
 --- @field target_kind    string|nil             -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
---- @field declaration_line integer|nil
 
 --- @class Finding
 --- @field line integer -- source line (or 0 for pass-level)
@@ -106,14 +103,11 @@ local ensure_dir     = duffle.ensure_dir
 -- Per-check functions (the CHECK_RULES table's payload)
 -- ════════════════════════════════════════════════════════════════════════════
 --
--- Each check has a uniform `append_to_findings` shape (errors[] / warnings[] / info[]).
+--- The dispatcher in `validate()` routes each result by convention: existence checks write errors[] and shape checks write warnings[].
--- The dispatcher in `validate()` decides which findings list each check writes to — by convention,
+--- `macro_word_drift` writes errors[] for missing or mismatched metadata and info[] for a match.
--- "existence" checks (declaration must exist, struct must exist) write errors[]; "shape" checks
--- (writes/reads must be wave-context) write warnings[].
--- The `macro_word_drift` check writes both errors[] (missing/mismatch) and info[] (match).
 
 --- Check: every annotated atom must have a matching MipsAtom_(name) declaration.
---- @param a AtomAnnotation
+--- @param a        AtomAnnotation
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_atom_decl_exists(a, pipe_ctx, findings)
@@ -141,10 +135,8 @@ local function check_unique_annotation(pipe_ctx, findings)
 end
 
 --- Check: BIND atoms must reference a real Binds_* struct.
---- Emitting a warning here keeps the annotation pass from being stop-on-error for the common test-fixture case,
+--- I keep this as a warning so the annotation pass can report the common test-fixture case; `check_abi_handoff` in static analysis supplies the build-stopping error.
---- while still surfacing the issue in the report.
+--- @param a        AtomAnnotation
---- The static-analysis report remains the source of truth for build-stopping errors.
---- @param a AtomAnnotation
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_binds_struct_exists(a, pipe_ctx, findings)
@@ -160,7 +152,7 @@ end
 
 --- Check: TAPE_WORDS(mac_X, N) ↔ WORD_COUNT(mac_X, N) drift.
 --- Three outcomes: missing (error), mismatch (error), match (info).
---- @param m MacroEntry
+--- @param m  MacroEntry
 --- @param wc table<string, integer> -- the shared word-count table (from ctx.shared.word_counts)
 --- @param findings Findings
 local function check_macro_word_drift(m, wc, findings)
@@ -185,10 +177,9 @@ local function check_macro_word_drift(m, wc, findings)
 	}
 end
 
---- Check: atom_dbg_reg_default(R_X, <type>) must target a register declared as a debug-visible alias in `pipe_ctx.register_alias_registry`,
+--- Check: atom_dbg_reg_default(R_X, <type>) targets an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
---- with a type name found in `pipe_ctx.type_name_registry`.
+--- Pointer depth remains bounded to 0 or 1, and duplicate defaults remain errors.
---- Pointer depth is still bounded to 0 or 1. Duplicate defaults are still detected.
+--- @param _src     SourceFile -- unused (kept for the per_source shape)
---- @param _src SourceFile        -- unused (kept for the per_source shape)
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_semantic_reg_defaults(_src, pipe_ctx, findings)
@@ -236,11 +227,9 @@ local function check_semantic_reg_defaults(_src, pipe_ctx, findings)
 	end
 end
 
---- Check: atom_reg_types(R_X, <type>) entries must point to a register declared in `pipe_ctx.register_alias_registry`, with a type name found in `pipe_ctx.type_name_registry`. 
+--- Check: atom_reg_types(R_X, <type>) entries target an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
---- The alias ident `R_<n>` now encodes the GPR identity only for entries that are explicitly opted in via the bare `atom_reg` marker.
+--- A bare `atom_reg` marker opts the `R_<n>` alias into GPR identity; references to R_T0..R_T3 require the same explicit marker.
---- R_T0..R_T3 are intentionally NOT auto-included (per the prototype principle: no auto-include of wave-context; explicit opt-in only). 
+--- @param _src     SourceFile
---- The check fires for any R_T0..R_T3 reference that hasn't been opted in via `#define atom_reg`.
---- @param _src SourceFile
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_atom_reg_types(_src, pipe_ctx, findings)
@@ -270,8 +259,8 @@ local function check_atom_reg_types(_src, pipe_ctx, findings)
 	end
 end
 
---- Check: atom_view(Binds_X) entries must reference a real Binds_* struct and that struct must declare at least one field.
+--- Check: atom_view(Binds_X) entries reference a Binds_* struct with at least one field.
---- @param _src SourceFile
+--- @param _src     SourceFile
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_atom_view_layout(_src, pipe_ctx, findings)
@@ -299,8 +288,7 @@ local function check_atom_view_layout(_src, pipe_ctx, findings)
 	end
 end
 
---- Check: Binds_* structs may not have duplicate field names
+--- Check: Binds_* structs require unique field names because atom_view uses those names for typed-field lookup in gdb.
---- (they would defeat the typed-field name lookup that atom_view exposes in gdb).
 --- @param _src SourceFile
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
@@ -323,26 +311,30 @@ local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
 	end
 end
 
--- Check: skip-over markers must satisfy shape + placement constraints.
+-- Check: debug-skip markers must satisfy shape + placement constraints.
---- Walks the priority list once; at most one error is appended per marker so that a single source-level defect does not cascade into multiple findings.
+--- Walks the priority list once; each marker produces at most one error, so one source defect yields one finding.
 --- Priority order (first defect wins):
----   1. has_parens == false         -> requires parentheses: marker()
+---   1. marker_kind ~= "atom_dbg_skip" -> legacy/renamed spelling (use `atom_dbg_skip`)
----   2. args  ~= ""                 -> takes no arguments
+---   2. marker_kind == "atom_dbg_skip" AND has_parens -> parenthesized form (the marker is bare-only)
----   3. superseded_by_marker_line   -> duplicate marker (cite superseding line)
+---   3. args  ~= ""                 -> takes no arguments
----   4. pending + no target_kind    -> dangling (no following declaration)
+---   4. superseded_by_marker_line   -> duplicate marker (cite superseding line)
----   5. unsupported target_kind     -> marker precedes an unrelated declaration
+---   5. pending + no target_kind    -> dangling (no following declaration)
---- Valid markers before whole-atom / bare-component / proc-component declarations emit no error and remain in src.scan.skip_over.atoms / .components.
+---   6. unsupported target_kind     -> marker precedes an unrelated declaration
---- @param marker SkipOverMarker
+--- Valid markers stamp `debug_skip` on whole-atom, bare-component, and proc-component declaration records in scan_source.lua.
+--- @param marker DebugSkipMarker
 --- @param _pipe_ctx PipeCtx     -- unused today; kept for plex-shape consistency with per_annot
 --- @param findings Findings
 local function check_skip_marker(marker, _pipe_ctx, findings)
 	local kind = marker.marker_kind
 	local line = marker.marker_line
 
-	if not marker.has_parens then
+	-- Left `scan.debug_skip_markers` with production records for `atom_dbg_skip` only; other identifiers take the walker's unrelated branch.
+
+	if marker.has_parens then
 		findings.errors[#findings.errors + 1] = {
 			line = line,
-			msg  = string.format("%s marker at line %d requires parentheses: marker()", kind, line),
+			msg  = string.format("%s marker at line %d must be bare; the parenthesized form is no longer accepted (use `atom_dbg_skip MipsAtom_(name) { ... }`)",
+				kind, line),
 		}
 		return
 	end
@@ -385,15 +377,11 @@ local function check_skip_marker(marker, _pipe_ctx, findings)
 	end
 end
 
---- Migration warning emitted alongside the new registry-membership check.
+--- Warn when a source references an unregistered alias.
 ---
---- R_TapePtr / R_AtomJmp / R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase
+--- R_TapePtr, R_AtomJmp, R_PrimCursor, R_FaceCursor, R_VertBase, and R_OtBase opt in through `#define atom_reg` in lottes_tape.h.
---- are the wave-context aliases opted in via `#define atom_reg` in lottes_tape.h (Task 21).
+--- When a source uses an unregistered R_X, this check emits one pass-level info entry for that source and directs C-ABI register names to explicit alias registration.
---- Any source referencing an R_X that's NOT in the registry will trip the new check; a single pass-level info entry
+--- @param _src     SourceFile
---- (emitted only when at least one such rejection lands in this source) tells users where to look.
----
---- This check is a stop-gap until users migrate off raw C-ABI register names.
---- @param _src SourceFile
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_wave_context_migration(_src, pipe_ctx, findings)
@@ -425,7 +413,7 @@ end
 --   per_annot(annot, pipe_ctx, findings)        -- runs once per AtomAnnotation
 --   post(pipe_ctx, findings)                    -- runs once after all per_annot calls complete (full-corpus aggregation)
 --   per_macro(macro, wc, findings)              -- runs once per TAPE_WORDS / _Pragma macro declaration
---   per_skip_marker(marker, pipe_ctx, findings) -- runs once per src.scan.skip_over.markers entry
+--   per_skip_marker(marker, pipe_ctx, findings) -- runs once per src.scan.debug_skip_markers entry
 --
 -- Adding a new check = 1 row here + 1 function above. The `validate()` dispatch loop never needs editing.
 
@@ -446,14 +434,56 @@ local CHECK_RULES = {
 -- Validation
 -- ════════════════════════════════════════════════════════════════════════════
 --
--- Pure check: read from src.scan, run validations, emit findings.
+-- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
--- No source walking; no parsing. The scan was done once upstream.
 
---- Validate one source against its pre-scanned SourceScan payload.
+--- Builds one pass-wide pipe_ctx from the merged `corpus.*` registries and source-ordered `corpus.atom_infos`; per-source declarations and bodies remain in `src.scan`.
+--- The module ownership contract above requires callers to construct `ctx.shared.corpus` through `build_ctx`; the error message below enforces that gate.
 --- @param ctx PassCtx
---- @param src SourceFile
+--- @return PipeCtx
+local function build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared and ctx.shared.corpus
+	if not corpus then
+		error("annotation requires ctx.shared.corpus "
+			.. "(the canonical corpus is the source of truth; "
+			.. "no per-source fallback is supported)", 0)
+	end
+
+	-- `corpus.atom_infos` preserves source order and duplicates; I precompute counts here for `check_unique_annotation` and the per-source checks.
+	local annot_counts = {}
+	for _, info in ipairs(corpus.atom_infos or {}) do
+		if info and info.atom_name then
+			annot_counts[info.atom_name] = (annot_counts[info.atom_name] or 0) + 1
+		end
+	end
+
+	-- Every consumer of these fields observes mutations via the canonical corpus without independently mutable registry construction.
+	return {
+		-- Cross-source lookup tables from corpus.
+		register_alias_registry  = corpus.register_alias_registry or {},
+		type_name_registry       = corpus.type_name_registry      or {},
+		atom_views               = corpus.atom_views              or {},
+		atom_ctxs                = corpus.atom_ctxs               or {},
+		atom_phases              = corpus.atom_phases             or {},
+		binds_by_name            = corpus.binds_by_name           or {},
+		atoms_by_name            = corpus.atoms_by_name           or {},
+		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
+		atom_infos_list          = corpus.atom_infos              or {},
+		-- Corpus-wide annotation count aggregation (post-rule consumes this).
+		annot_counts             = annot_counts,
+		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
+		collisions               = corpus.collisions              or {},
+		-- `check_macro_word_drift` reads `corpus.word_counts`, populated by word_count_eval.run.
+		word_counts              = corpus.word_counts or {},
+	}
+end
+
+--- Validate one source against its pre-scanned SourceScan payload + the corpus-wide pipe_ctx.
+--- @param ctx             PassCtx
+--- @param src             SourceFile
+--- @param corpus_pipe_ctx PipeCtx|nil  -- built once per pass from corpus registries; nil builds the same projection here.
 --- @return AnnotatedResult
-local function validate(ctx, src)
+local function validate(ctx, src, corpus_pipe_ctx)
+	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
 	local scan = src.scan
 
 	-- Project the pre-scanned atoms to the AtomEntry shape this pass needs.
@@ -479,9 +509,7 @@ local function validate(ctx, src)
 		}
 	end
 
-	-- Build pipe_ctx (Fleury: expose structure). Pre-compute everything the per-check functions need.
+	-- Build a per-source pipe_ctx: shared lookups come from `corpus_pipe_ctx`, while declarations, bodies, types, views, defaults, and occurrences come from `src.scan`.
-	-- Single source of truth for atom / binds / annotation-count lookups.
-	-- pipe_ctx.types / pipe_ctx.atom_views / pipe_ctx.seen_defaults are projected from the scan payload so per_source check rules can iterate.
 	local seen_defaults = {}
 	for reg, _ in pairs(scan.types or {}) do
 		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
@@ -494,33 +522,25 @@ local function validate(ctx, src)
 	local pipe_ctx = {
 		atom_index               = {},
 		binds_index              = {},
-		annot_counts             = {},
+		annot_counts             = corpus_pipe_ctx.annot_counts,
 		types                    = scan.types or {},
 		type_occurrences         = scan.type_occurrences or {},
 		atom_views               = scan.atom_views or {},
 		seen_defaults            = seen_defaults,
 		atom_infos_list          = atom_infos_list,
 		binds_list               = scan.binds or {},
-		-- Project the source-derived registries from the scan payload so per_source checks consult them instead of the deleted
+		-- See the module ownership contract; these shared lookup tables come from corpus_pipe_ctx.
-		-- SEMANTIC_DEFAULT_REGS / KNOWN_REG_DEFAULT_TYPES / etc.
+		register_alias_registry  = corpus_pipe_ctx.register_alias_registry,
-		register_alias_registry  = scan.register_alias_registry or {},
+		type_name_registry       = corpus_pipe_ctx.type_name_registry,
-		type_name_registry       = scan.type_name_registry or {},
 	}
 	for _, a in ipairs(atoms)      do pipe_ctx.atom_index [a.name] = a end
 	for _, b in ipairs(scan.binds) do pipe_ctx.binds_index[b.name] = b end
-	for _, a in ipairs(annots)     do
-		if a.name then
-			pipe_ctx.annot_counts[a.name] = (pipe_ctx.annot_counts[a.name] or 0) + 1
-		end
-	end
 
 	-- Findings live in a single struct with three lists (errors / warnings / info).
 	-- Each check writes to the list appropriate for its severity.
 	local findings = { errors = {}, warnings = {}, info = {} }
 
-	-- Propagate parse-time errors from scan_source's atom_info parsing.
+	-- Lift parse-time errors already recorded in scan_source's atom_info payload into this pass's findings list.
-	-- These are errors found in the atom_info(...) body itself (e.g., malformed args).
-	-- They are pre-existing in the scan payload — we just lift them into our findings list.
 	for _, a in ipairs(annots) do
 		if a.errors then
 			for _, msg in ipairs(a.errors) do
@@ -544,11 +564,9 @@ local function validate(ctx, src)
 		if rule.post then rule.post(pipe_ctx, findings) end
 	end
 
-	-- Per-skip-marker rules. 
+	-- scan_source records each marker in scan.debug_skip_markers; this loop validates each record independently and emits at most one error per marker.
-	-- Each raw marker recorded by scan_source (in scan.skip_over.markers) is validated independently; 
+	-- Valid markers stamp `debug_skip = true` on the following atom or component declaration, which downstream consumers read directly.
-	-- the check emits at most one error per marker.
+	local skip_markers = scan.debug_skip_markers or {}
-	-- Valid markers stay attached to scan.skip_over.atoms /.components for dwarf_injection.lua consumer.
-	local skip_markers = scan.skip_over and scan.skip_over.markers or {}
 	for _, marker in ipairs(skip_markers) do
 		for _, rule in ipairs(CHECK_RULES) do
 			if rule.per_skip_marker then rule.per_skip_marker(marker, pipe_ctx, findings) end
@@ -556,7 +574,7 @@ local function validate(ctx, src)
 	end
 
 	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
-	local wc = ctx.shared.word_counts
+	local wc = corpus_pipe_ctx.word_counts
 	for _, m in ipairs(scan.macros) do
 		for _, rule in ipairs(CHECK_RULES) do
 			if rule.per_macro then rule.per_macro(m, wc, findings) end
@@ -572,8 +590,8 @@ local function validate(ctx, src)
 	-- Information summary (always emitted).
 	findings.info[#findings.info + 1] = {
 		line = 0,
-		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)",
+		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)"
-			#atoms, #annots, #scan.macros, #scan.binds),
+			, #atoms, #annots, #scan.macros, #scan.binds),
 	}
 
 	return {
@@ -594,7 +612,6 @@ end
 --- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
 --- Empty directories (no errors, no atoms) produce no file.
 local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
-	if ctx.dry_run then return nil end
 	if atoms_count == 0 and #errors == 0 then
 		return nil
 	end
@@ -622,17 +639,6 @@ local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sour
 	return out_path
 end
 
---- Stash aggregated per-module results for the report pass to consume.
-local function emit_module_annotations_stub(ctx, dir, dir_basename, atoms_count)
-	ctx.flags                = ctx.flags                or {}
-	ctx.flags._annot_results = ctx.flags._annot_results or {}
-	ctx.flags._annot_results[#ctx.flags._annot_results + 1] = {
-		dir          = dir,
-		dir_basename = dir_basename,
-		atoms_count  = atoms_count,
-	}
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- M.run — orchestrator entry
 -- ════════════════════════════════════════════════════════════════════════════
@@ -651,22 +657,22 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- Per-DIRECTORY (per-module) aggregation. Group sources by `src.dir`, validate every source in the dir, then emit ONE errors.h per dir.
+	-- Build the shared pipe_ctx once for this run; every validate() call sees the same cross-source registries.
-	-- `ctx.by_dir` is pre-computed in build_ctx (shared across all passes).
+	-- The corpus owns the canonical cross-source registries; per-source scans retain body / declaration ownership.
-	local by_dir = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	local corpus_pipe_ctx = build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared.corpus
+
+	-- Group `corpus.sources_by_dir` by module, validate every source in each bucket, and emit one errors.h per directory.
+	local by_dir = (corpus and corpus.sources_by_dir) or {}
 
 	for dir, dir_sources in pairs(by_dir) do
 		local dir_basename = dir:match("([^/\\]+)$") or dir
 		local dir_atoms    = 0
 		local dir_errors   = {}
 		local dir_warnings = {}
-		-- Per-source validate() results, cached for the report pass (it reads from this instead of re-validating each source).
-		ctx.flags                       = ctx.flags                       or {}
-		ctx.flags._annot_source_results = ctx.flags._annot_source_results or {}
 		for _, src in ipairs(dir_sources) do
-			local result  = validate(ctx, src)
+			local result  = validate(ctx, src, corpus_pipe_ctx)
 			result.source = src.path                           -- tag for downstream rendering
-			ctx.flags._annot_source_results[src.path] = result -- stash so report.lua reads from cache instead of re-running validate()
 			dir_atoms = dir_atoms + #result.atoms
 			for _, e in ipairs(result.errors) do
 				dir_errors[#dir_errors + 1] = { line = e.line, msg = e.msg, source = src.path }
@@ -682,8 +688,6 @@ function M.run(ctx)
 		if err_path then
 			table.insert(outputs, { errors_h = err_path })
 		end
-
-		emit_module_annotations_stub(ctx, dir, dir_basename, dir_atoms)
 	end
 
 	return { outputs = outputs, errors = errors, warnings = warnings }

scripts/passes/atoms_source_map.lua

@@ -1,26 +1,23 @@
 --- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
 ---
---- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
+--- Writer: this pass, given `atom.paths` (the per-atom mutable surface owned by `emission_model`). Readers:
---- for `MipsAtom_(name)` (kind="atom"), `MipsAtomComp_` / `MipsAtomComp_Proc_` (kind="comp_*"),
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and the gdb-runtime
---- and `MipsCode code_<name>` (kind="raw_atom") declarations. 
+--- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
---- Walks each atom's pre-tokenized body (`{{tok=string, rel=integer}, ...}` from `duffle.tokenize_body`), 
---- counts per-token word contributions via `ctx.shared.word_counts`, and emits one
---- `WORD N LINE L TEXT T` line per `.word` to `<out_root>/<basename>.atoms.sourcemap.txt`.
 ---
---- **Two output forms** (per the workspace's per-emission-form pattern from
+--- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: one
---- `guide_metaprogram_ssdl.md`):
+--- `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
----   1. **Canonical text form** — `<out_root>/<basename>.atoms.sourcemap.txt`.
+---
----      Format-version-tagged for forward-compat. 
+--- **Two output forms** (per the workspace's per-emission-form pattern from `guide_metaprogram_ssdl.md`):
----      Lives in `<out_root>/` (build/gen).
+---   1. **Sourcemap.txt form** — `<out_root>/<basename>.atoms.sourcemap.txt`. Format-version-tagged for forward-compat.
----      Matches the convention used by `annotation.lua` (`<out_root>/<basename>.errors.h`) + `static_analysis.lua` (`<out_root>/<basename>.static_analysis.txt`).
+---      Lives in `<out_root>/` (build/gen). Mirrors the convention used by `annotation.lua`
+---      (`<out_root>/<basename>.errors.h`) and `static_analysis.lua` (`<out_root>/<basename>.static_analysis.txt`).
 ---      Compile artifacts (`*.macs.h`, `*.offsets.h`) stay in `<source_dir>/gen/`.
----   2. **gdb-runtime form** — `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`
+---   2. **gdb-runtime form** — `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`. A pure gdb command script — addresses come
----      (pure gdb command script; addresses pre-computed via `nm`; the 9 user commands defined as `define ... end` blocks).
+---      from `nm`, the 9 user commands are static `define ... end` blocks. Emitted when `ctx.flags.gdb_runtime` is true
----      Emitted ONLY when `ctx.flags.gdb_runtime` is true AND `ctx.flags.elf_path` points to an existing ELF.
+---      AND `ctx.flags.elf_path` points to an existing ELF. Useful for `gdb-multiarch --without-python` users
----      The gdb runtime form lets `gdb-multiarch --without-python` users (the common case on Windows MinGW builds)
+---      (the common case on Windows MinGW builds) — `source <path>` loads it with no Python / Tcl / Guile required.
----      load the source-map data via `source <path>` — no Python/Tcl/Guile required.
 ---
---- **Output format** (canonical text form):
+--- **Output format** (sourcemap.txt form):
 ---   ```
 ---   # FORMAT_VERSION 1
 ---   # auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT
@@ -34,13 +31,9 @@
 ---   ENDATOM
 ---   ```
 ---
---- Marker calls (`atom_label(...)`, `atom_offset(...)`) emit 0 `.word`s. 
+--- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
---- They share the same walking convention as `passes/offsets.lua :: scan_atom_body`: 
---- Markers do NOT advance the word-offset counter, but if a marker is bundled on the same token with a trailing instruction 
---- (e.g. `atom_label(foo) load_half_u(...)`), the trailing instruction's word count is added. This matches `offsets.lua :: count_marker_rest`.
 ---
---- **Conventions:** tabs (1/level), EmmyLua annotations, no regex,
+--- **Conventions:** tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -50,10 +43,8 @@
 -- (works both standalone + when require'd). `duffle_paths.lua` sets package.path then returns `require("duffle")` 
 -- at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle           = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle    = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local elf_dwarf        = require("elf_dwarf")
+local elf_dwarf = require("elf_dwarf")
-local word_count_eval  = require("word_count_eval")
-local count_token_words = word_count_eval.count_token_words
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -63,199 +54,88 @@ local count_token_words = word_count_eval.count_token_words
 -- the gdb runtime loader rejects mismatches (E2).
 local FORMAT_VERSION = 1
 
--- Marker-call identifiers (mirrors offsets.lua:33-34).
-local LABEL_MARKER  = "atom_label"
-local OFFSET_MARKER = "atom_offset"
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class AtomSourceMapCtx
---- @field sources            table[] -- SourceScan payload per source (from `ctx.sources`)
 --- @field shared             table   -- `ctx.shared`
---- @field shared.word_counts table   -- macro name -> word count (populated by word-counts + components passes)
+--- @field shared.corpus      table   -- source-order registry; single writer is build_ctx
+--- @field shared.word_counts table
 --- @field out_root           string  -- output root (e.g. "build/gen")
---- @field dry_run            boolean -- if true, compute but don't write
 --- @field flags              table   -- `ctx.flags`; reads `flags.gdb_runtime` + `flags.elf_path`
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Helpers
+-- Atom-path renderers
 -- ════════════════════════════════════════════════════════════════════════════
 
--- ════════════════════════════════════════════════════════════════════════════
+--- Join word boundaries (from `items`) to per-word call text + source lines (from `word_events`).
--- Provenance emission
--- ════════════════════════════════════════════════════════════════════════════
-
--- Component-macro invocation prefix (mirrors components.lua's MAC_PREFIX).
-local MAC_PREFIX     = "mac_"
-local MAC_PREFIX_LEN = 4
-
---- Strip the `mac_` prefix from a token's leading identifier.
---- Returns nil if the identifier doesn't start with `mac_`
---- (so non-component tokens like `load_half_u`, `nop2`, `gte_cmdw_*` fall through cleanly).
---- @param tok string
---- @return string|nil
-local function strip_mac_prefix_from_token(tok)
-	local leading = duffle.read_ident(tok, 1)
-	if not leading then return nil end
-	if leading:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
-		return leading:sub(MAC_PREFIX_LEN + 1)
-	end
-	return nil
-end
-
---- Fetch the per-word body lines for a `mac_X(...)` invocation.
---- Walks the component's pre-tokenized body in lockstep with `count_token_words` and attributes each emitted `.word`
---- to a source line via `idx.line_of(...)`. 
---- Atom labels (`atom_label(...)`) emit 0 `.word`s and are skipped.
---- @param bare string|nil          -- the bare component name (e.g. `gte_load_tri_verts`)
---- @param comp_body_index table
---- @param wc table
---- @return table|nil                -- list of source lines, 1-based by word position
-local function fetch_body_lines(bare, comp_body_index, wc)
-	if not (bare and comp_body_index) then return nil end
-	local idx = comp_body_index[bare]
-	if not (idx and idx.body_tokens and idx.line_of) then return nil end
-	local lines = {}
-	for _, bt in ipairs(idx.body_tokens) do
-		local bt_tok = duffle.trim(bt.tok or "")
-		if bt_tok ~= "" then
-			local leading = duffle.read_ident(bt_tok, 1)
-			local bt_words
-			if leading == "atom_label" or leading == "atom_offset" then
-				bt_words = 0
-			else
-				bt_words = count_token_words(bt_tok, wc)
-			end
-			if bt_words > 0 then
-				local body_line = idx.line_of(idx.body_off + bt.rel)
-				for _ = 1, bt_words do lines[#lines + 1] = body_line end
-			end
-		end
-	end
-	return lines
-end
-
---- Unified per-word entry walker. `mode` is "sourcemap" (3 fields) or "provenance" (8 fields including component + body-line lookup).
---- Returns (entries, total_words). Markers contribute 0 entries.
 --- @param atom table
---- @param src  table
---- @param wc   table
---- @param mode string           -- "sourcemap" | "provenance"
---- @param comp table|nil        -- shared.components map (provenance only)
---- @param comp_body_index table|nil  -- per-source body index (provenance only)
 --- @return table[], integer
-local function compute_word_entries(atom, src, wc, mode, comp, comp_body_index)
+local function canonical_word_entries(atom)
-	local entries = {}
+	local paths = atom.paths or {}
-	local pos     = 0
+	local events = paths.word_events or {}
-	for _, t in ipairs(atom.body_tokens) do
+	local word_items = {}
-		local tok = t.tok
+	for _, item in ipairs(paths.items or {}) do
-		local rel = t.rel
+		if item.kind == "word" then word_items[#word_items + 1] = item end
-
-		local words
-		if duffle.is_marker_token(tok) then
-			words = duffle.count_marker_rest(tok, wc, count_token_words)
-		else
-			words = count_token_words(tok, wc)
-		end
-
-		-- Provenance-only: resolve component + body_lines (one fetch per token).
-		local comp_name, comp_line, comp_path, comp_kind
-		local body_lines
-		if mode == "provenance" then
-			local bare = strip_mac_prefix_from_token(tok)
-			if bare and comp and comp[bare] then
-				comp_name = bare
-				comp_line = comp[bare].line
-				comp_path = comp[bare].path
-				comp_kind = comp[bare].kind
-			end
-			if comp_name then body_lines = fetch_body_lines(bare, comp_body_index, wc) end
-		end
-
-		if words > 0 then
-			local line = src.scan.line_of(atom.body_off + rel)
-			local text = duffle.trim(tok):gsub("[\t\r\n]+", " ")
-			for i = 1, words do
-				local entry
-				if mode == "provenance" then
-					entry = {
-						pos       = pos,
-						line      = line,
-						text      = text,
-						comp_name = comp_name,
-						comp_line = comp_line,
-						comp_path = comp_path,
-						comp_kind = comp_kind,
-						body_line = body_lines and body_lines[i],
-					}
-				else  -- "sourcemap" (default)
-					entry = { pos = pos, line = line, text = text }
-				end
-				entries[#entries + 1] = entry
-				pos = pos + 1
-			end
-		end
 	end
-	return entries, pos
+
+	local entries = {}
+	for index, event in ipairs(events) do
+		local item = word_items[index] or {}
+		entries[#entries + 1] = {
+			pos        = event.i or (index - 1),
+			line       = event.call_line or item.line or 0,
+			text       = event.call_text or item.call_text or "",
+			body_line  = event.body_line or item.body_line or item.line or 0,
+			invocation = (event.outermost_invocation_id
+				and paths.invocations
+				and paths.invocations[event.outermost_invocation_id]) or nil,
+		}
+	end
+	return entries, #events
 end
 
---- Render one atom's provenance stanza. Format:
+--- Render one atom's provenance stanza. Format 1 line shapes:
----   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>"  [BODY <line>]`  (for component words)
+---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>" BODY <line>` (component invocation)
----   `WORD N  CALL <src-path>:<src-line>  RAW`                                                  (for direct instructions)
+---   `WORD N  CALL <src-path>:<src-line>  RAW` (raw `.word` outside any mac_* component)
---- `BODY <line>` is the source line of THIS specific word within the macro body
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was
---- (lottes_tape.h:N where N is the per-word body line).
+--- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
---- Absent for RAW rows and for component rows whose component declaration could not be indexed (older pass combinations / external macros).
+--- @param src table
---- Downstream consumers (dwarf_injection, tests) fall back to DefLine / comp_line when BODY is absent.
+--- @param atom table
---- Returns (lines, total_words).
+--- @param wc table -- identity alias of corpus.word_counts
---- @param src    table
---- @param atom   table
---- @param wc     table
---- @param comp   table  -- shared.components map
---- @param comp_body_index table  -- per-source component body index: bare_name -> {body_off, body_tokens, line_of}
 --- @return string[], integer
-local function emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
+local function emit_provenance_stanza(src, atom, wc)
 	local lines    = {}
-	local rel_path = src.path:gsub("\\", "/")
+	local rel_path = src.path:gsub("\\\\", "/")
-	local entries, total = compute_word_entries(atom, src, wc, "provenance", comp, comp_body_index)
+	local entries, total = canonical_word_entries(atom)
 
-	-- ATOM header line with placeholder total (patched after we know it).
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
 
-	for _, pe in ipairs(entries) do
+	for _, entry in ipairs(entries) do
-		if pe.comp_name then
+		local inv = entry.invocation
-			local body_suffix = ""
+		local macro_count = inv and wc["mac_" .. inv.component_name]
-			if pe.body_line then
+		if inv and macro_count ~= nil then
-				body_suffix = "  BODY " .. tostring(pe.body_line)
+			lines[#lines + 1] = string.format(
-			end
+				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
-			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d"%s',
+				entry.pos, rel_path, entry.line, inv.component_name,
-				pe.pos, rel_path, pe.line, pe.comp_name, pe.comp_path, pe.comp_line, body_suffix)
+				inv.def_path or "", inv.def_line or 0, entry.body_line)
 		else
-			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", pe.pos, rel_path, pe.line)
+			lines[#lines + 1] = string.format(
+				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
 		end
 	end
 
-	-- Patch the placeholder total in the ATOM header line.
 	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
 	lines[#lines + 1] = "ENDATOM"
 	return lines, total
 end
 
---- Build a per-source component body index keyed by the bare component name (e.g. `gte_load_tri_verts`).
+--- Render the full provenance file content for one source.
---- Each entry holds the data we need to map each emitted `.word` to its actual source line within the macro body:
----   body_off    -- byte offset of the `{` (start of body) in the component's source file.
----   body_tokens -- list of {tok, rel} pairs; `rel` is the byte offset within the body.
----   line_of     -- closure resolving byte offsets in the component's source file to lines.
---- Only `comp_bare` + `comp_proc` declarations contribute (a macro invocation can only resolve to one of those).
---- First declaration wins (subsequent redeclarations would collide; today's sources declare each component exactly once).
---- Render the full provenance file content for one source (one `.atoms.provenance.txt` per source).
 --- @param src table
 --- @param wc table
---- @param comp table  -- shared.components map
---- @param comp_body_index table  -- cross-source component body index (built once in M.run; may be empty)
 --- @return string
-local function render_provenance(src, wc, comp, comp_body_index)
+local function render_provenance(src, wc)
 	local lines = {}
 	lines[#lines + 1] = "# FORMAT_VERSION 1"
 	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
@@ -265,62 +145,60 @@ local function render_provenance(src, wc, comp, comp_body_index)
 	lines[#lines + 1] = "# dwarf_injection to synthesize DW_TAG_inlined_subroutine instances + per-word"
 	lines[#lines + 1] = "# line program rows for native source-level step into component bodies."
 
-	-- The cross-source component body index is passed in from M.run (one global lookup shared across every source's provenance file).
+	local function append(atom)
-	-- A per-source lookup would miss every component whose declaration is in another source (e.g. `gte_load_tri_verts` is declared in `lottes_tape.h` but invoked from `hello_gte_tape.c`).
+		local stanza = emit_provenance_stanza(src, atom, wc)
-
-	for _, atom in ipairs(src.scan.atoms or {}) do
-		local stanza = emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
 		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
 	end
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		if atom.paths then append(atom) end
+	end
 	for _, atom in ipairs(src.scan.raw_atoms or {}) do
-		local stanza = emit_provenance_stanza(src, atom, wc, comp, comp_body_index)
+		if atom.paths then append(atom) end
-		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
 	end
 
 	return table.concat(lines, "\n") .. "\n"
 end
 
---- Render one atom's stanza for the canonical text form (ATOM header line, N WORD lines, ENDATOM marker).
+--- Render one atom's stanza for the sourcemap.txt form (ATOM header line, N WORD lines, ENDATOM marker).
 --- Returns (lines, total_words).
 --- @param src    table
 --- @param atom   table
 --- @param wc     table
 --- @return string[], integer
-local function emit_atom_stanza(src, atom, wc)
+local function emit_atom_stanza(src, atom)
-	local lines              = {}
+	local lines    = {}
-	local rel_path           = src.path:gsub("\\", "/")
+	local rel_path = src.path:gsub("\\\\", "/")
-	local entries, total     = compute_word_entries(atom, src, wc)
+	local entries, total = canonical_word_entries(atom)
 
-	-- ATOM header line with placeholder total (patched after we know it).
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
-	for _, we in ipairs(entries) do
+	for _, entry in ipairs(entries) do
 		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
-			we.pos, we.line, we.text)
+			entry.pos, entry.line, entry.text)
 	end
-
-	-- Patch the placeholder total in the ATOM header line.
 	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
 	lines[#lines + 1] = "ENDATOM"
 	return lines, total
 end
 
---- Render the full source map file content for one source (one .atoms.sourcemap.txt per source).
+--- Render the full source map file content for one source (one .atoms.sourcemap.txt per source). Mirrors offsets.lua's
---- Mirrors offsets.lua's `project_atoms` shape: scan.atoms + scan.raw_atoms, no kind filter.
+--- `project_atoms` shape: scan.atoms + scan.raw_atoms, no kind filter.
 --- @param src table
 --- @param wc table
 --- @return string
-local function render_source_map(src, wc)
+local function render_source_map(src)
 	local lines = {}
 	lines[#lines + 1] = "# FORMAT_VERSION " .. FORMAT_VERSION
 	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
 
-	for   _, atom in ipairs(src.scan.atoms or {}) do
+	local function append(atom)
-		local stanza = emit_atom_stanza(src, atom, wc)
+		local stanza = emit_atom_stanza(src, atom)
 		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
 	end
-	for   _, atom in ipairs(src.scan.raw_atoms or {}) do
+	for _, atom in ipairs(src.scan.atoms or {}) do
-		local stanza = emit_atom_stanza(src, atom, wc)
+		if atom.paths then append(atom) end
-		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		if atom.paths then append(atom) end
 	end
 
 	return table.concat(lines, "\n") .. "\n"
@@ -338,69 +216,48 @@ local function gdb_escape(s)
 	return (s:gsub("\\", "\\\\"):gsub('"', '\\"'))
 end
 
---- Build the list of atoms with addresses + word entries.
+--- Build the list of atoms with addresses + word entries. Shared helper for the gdb-runtime file emission.
---- Shared helper for the gdb-runtime file emission.
 --- @param ctx PassCtx
 --- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
 local function build_atom_table(ctx)
-	local wc    = (ctx.shared and ctx.shared.word_counts) or {}
 	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
-
+	local corpus = ctx.shared and ctx.shared.corpus
 	local matched = {}
-	for _, src in ipairs(ctx.sources) do
+
-		if src.scan then
+	for _, src in ipairs(corpus.source_order or {}) do
-			local file_base = src.path:match("([^/\\]+)$") or src.path
+		local file_base = src.path:match("([^/\\\\]+)$") or src.path
-			for _, atom in ipairs(src.scan.atoms or {}) do
+		local function append(atom)
-				if atom.kind == nil or atom.kind == "atom" then
+			if not atom.paths then return end
-					local name    = atom.raw_name or atom.name
+			local name = atom.raw_name or atom.name
-					local info    = addrs[name]
+			local info = addrs[name]
-					if info then
+			if not info then return end
-						local entries, total = compute_word_entries(atom, src, wc)
+			local entries, total = canonical_word_entries(atom)
-						matched[#matched + 1] = {
+			matched[#matched + 1] = {
-							name       = name,
+				name       = name,
-							src_path   = src.path,
+				src_path   = src.path,
-							file_base  = file_base,
+				file_base  = file_base,
-							addr       = info[1],
+				addr       = info[1],
-							size_bytes = info[2],
+				size_bytes = info[2],
-							words      = total,
+				words      = total,
-							entries    = entries,
+				entries    = entries,
-						}
+			}
-					end
-				end
-			end
-			for _, atom in ipairs(src.scan.raw_atoms or {}) do
-				local name = atom.name
-				local info = addrs[name]
-				if info then
-					local entries, total = compute_word_entries(atom, src, wc)
-					matched[#matched + 1] = {
-						name       = name,
-						src_path   = src.path,
-						file_base  = file_base,
-						addr       = info[1],
-						size_bytes = info[2],
-						words      = total,
-						entries    = entries,
-					}
-				end
-			end
 		end
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do append(atom) end
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do append(atom) end
 	end
 
 	-- Deterministic order: sort by address (matches `nm` output ordering).
 	table.sort(matched, function(a, b) return a.addr < b.addr end)
-	for i, a in ipairs(matched) do
+	for i, a in ipairs(matched) do a.idx = i - 1 end
-		a.idx = i - 1
-	end
 	return matched
 end
 
---- Append the 9 gdb command definitions to `lines`. Pure gdb scripting no Python, no Tcl, no Guile required. 
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`, the convenience
---- **Fully hardcoded per-atom** because gdb doesn't do nested `$` substitution in var names
+--- vars set in `emit_gdb_runtime` provide printf args, and each command is a static sequence of `printf` / `tbreak` /
---- `$__atom_name_$__i` inside a `while` loop is treated as one literal identifier, not a concat.
+--- `if ... end` blocks. The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
 ---
---- Each command is a static sequence of `printf` / `tbreak` / `if ... end` blocks.
+--- Why hardcoded per-atom: gdb's `$` substitution doesn't concat inside var names — `$__atom_name_$__i` in a `while`
---- The Lua pass emits N atoms' worth of lines — no runtime iteration. 
+--- loop resolves to one literal identifier, not `name_i`. Compile-time emission is the only path.
 --- @param lines   table  -- output line buffer (mutated in place)
 --- @param matched table  -- list of atom records from `build_atom_table`
 local function append_gdb_commands(lines, matched)
@@ -529,31 +386,6 @@ local function append_gdb_commands(lines, matched)
 	lines[#lines + 1] = "end"
 	lines[#lines + 1] = ""
 
-	-- ── show_c2 ──
-	-- GTE data regs (COP2). pcsx-redux's gdb stub doesn't expose COP2 (only 72 regs: 32 GPR + COP0 + FPR).
-	--   curl http://localhost:8080/api/v1/lua/gte
-	-- We keep the command definition as a stub that points the user at the plugin.
-	lines[#lines + 1] = "define show_c2"
-	lines[#lines + 1] = '  echo "[gdb_tape_atoms] show_c2: gdb stub does not expose COP2 in this build."'
-	lines[#lines + 1] = '  echo "[gdb_tape_atoms] Use scripts/pcsx_debug_helper.zip + curl http://localhost:8080/api/v1/lua/gte"'
-	lines[#lines + 1] = '  echo "[gdb_tape_atoms] (or pcsx-redux Debug > Registers window for a native view)"'
-	lines[#lines + 1] = "end"
-	lines[#lines + 1] = "document show_c2"
-	lines[#lines + 1] = "  Stub. The gdb stub in this pcsx-redux build does not expose COP2 regs."
-	lines[#lines + 1] = "  For GTE data + control state, use the pcsx_debug_helper Lua plugin or the"
-	lines[#lines + 1] = "  pcsx-redux Debug > Registers window."
-	lines[#lines + 1] = "end"
-	lines[#lines + 1] = ""
-
-	-- ── show_c2ctl ──
-	lines[#lines + 1] = "define show_c2ctl"
-	lines[#lines + 1] = '  echo "[gdb_tape_atoms] show_c2ctl: see show_c2 for the same workaround."'
-	lines[#lines + 1] = "end"
-	lines[#lines + 1] = "document show_c2ctl"
-	lines[#lines + 1] = "  Stub. Same workaround as show_c2."
-	lines[#lines + 1] = "end"
-	lines[#lines + 1] = ""
-
 	-- ── wave_ctx ──
 	lines[#lines + 1] = "define wave_ctx"
 	lines[#lines + 1] = '  printf "$t4 = R_FaceCursor     0x%08x\\n", $t4'
@@ -566,9 +398,9 @@ local function append_gdb_commands(lines, matched)
 	lines[#lines + 1] = "end"
 end
 
---- Emit the gdb-runtime file (post-link). Pure gdb scripting — no Python.
+--- Emit the gdb-runtime file (post-link). Pure gdb scripting — addresses come from `mipsel-none-elf-nm -S`, get embedded
---- Reads ELF addresses via `mipsel-none-elf-nm -S`, embeds them in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb
---- so gdb loads the data via `set $var = ...` + `define ... end` blocks at source-time.
+--- source-time.
 --- @param ctx PassCtx
 local function emit_gdb_runtime(ctx)
 	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
@@ -627,12 +459,9 @@ local function emit_gdb_runtime(ctx)
 	lines[#lines + 1] = 'printf "[gdb_tape_atoms] runtime loaded %d atoms from %s\\n", $__atom_count, $__elf_path'
 
 	local out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
-	if not ctx.dry_run then
+	duffle.ensure_dir(duffle.dirname(out_path))
-		duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
-		duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
+	-- io.stderr:write(string.format("[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
-	end
-	io.stderr:write(string.format(
-		"[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -641,46 +470,10 @@ end
 
 local M = {}
 
---- Build the cross-source component body index used by `render_provenance` to attribute each emitted `.word` to its actual line within the macro body.
+--- Pass entry. For each source that declares at least one `MipsAtom_(name)` / `MipsCode code_<name>`, emit two files
----
+--- in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
---- Components are declared in one source (the header that contains `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`) 
+--- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation). When `ctx.flags.gdb_runtime`
---- but invoked from many source files (every atom body that calls `mac_X(...)`).
+--- is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
---- The body_offset + body_tokens + line_of live with the declaration source, so a per-source index would miss invocations from other sources.
----
---- The cross-source index is keyed by the bare component name (`gte_load_tri_verts`, NOT `ac_gte_load_tri_verts`)
---- `strip_mac_prefix_from_token` strips the `mac_` prefix from call-site identifiers and yields that exact bare name;
---- matching it here keeps the lookup aligned with the `ctx.shared.components` map's keying convention.
---- First declaration wins (subsequent redeclarations would collide; today's sources declare each component exactly once).
---- @param ctx PassCtx
---- @return table<string, table>  -- {[comp_name] = {body_off, body_tokens, line_of}}
-local function build_cross_source_component_body_index(ctx)
-	local index = {}
-	for _, src in ipairs(ctx.sources or {}) do
-		if src.scan and src.scan.atoms then
-			local line_of = src.scan.line_of
-			for _, atom in ipairs(src.scan.atoms) do
-				if atom.kind == "comp_bare" or atom.kind == "comp_proc" then
-					-- Prefer `atom.name` (stripped of `ac_` prefix); fall back to `raw_name`
-					-- only if the stripped name is absent (defensive — current scan-source always sets both).
-					local name = atom.name or atom.raw_name
-					if name and not index[name] then
-						index[name] = {
-							body_off    = atom.body_off,
-							body_tokens = atom.body_tokens,
-							line_of     = line_of,
-						}
-					end
-				end
-			end
-		end
-	end
-	return index
-end
-
---- Pass entry: emit one `<out_root>/<basename>.atoms.sourcemap.txt` per source file that contains at least one `MipsAtom_(name)` / `MipsCode code_<name>` declaration.
---- Also emits `<out_root>/<basename>.atoms.provenance.txt`:
---- per-.word provenance with `mac_X(...)` component resolution back to the component's definition file:line + the per-word body line.
---- Optionally also emit `<ctx.out_root>/gdb_tape_atoms_runtime.gdb` when `ctx.flags.gdb_runtime` is true.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -688,54 +481,47 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- word-counts + components passes must have populated shared.word_counts.
+	local corpus = ctx.shared and ctx.shared.corpus
-	-- If absent, the orchestrator wired the deps wrong — fail loud.
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
-	local wc = (ctx.shared and ctx.shared.word_counts) or {}
+		error("atoms_source_map.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
-	if not wc or not next(wc) then
+	end
+
+	-- Word counts come from `corpus.word_counts` (populated by word_count_eval + components passes).
+	local wc = corpus.word_counts or {}
+	if not next(wc) then
 		warnings[#warnings + 1] = {
 			line = 0,
-			msg  = "atoms_source_map: ctx.shared.word_counts is empty; the word-counts + components passes may not have populated it. Check the PASSES dep edges.",
+			msg  = "atoms_source_map: corpus.word_counts is empty; the word-counts + components passes may not have populated it. Check the PASSES dep edges.",
 		}
 	end
 
-	-- shared.components map is populated by `passes/components.lua`.
+	-- Always emit the text form (per-source).
-	-- Used to attribute each emitted `.word` to either a component macro or the enclosing atom body.
+	for _, src in ipairs(corpus.source_order) do
-	-- If absent, all words fall through as RAW (correct behavior — provenance is additive).
+		local has_projection = false
-	local comp = (ctx.shared and ctx.shared.components) or {}
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do
-
+			if (atom.kind == "atom" or atom.kind == "raw_atom") and atom.paths then
-	-- Cross-source component body index.
+				has_projection = true; break
-	-- Built ONCE so every source's provenance writer can resolve `mac_X(...)` invocations back to the macro's body tokens (regardless of which source declared the component).
-	-- Per-source copies were insufficient — the atom file (`hello_gte_tape.c`) does not contain the `MipsAtomComp_(...)` declarations, 
-	-- so the body data would be missing for every component invocation the atom file emitted.
-	local comp_body_index = build_cross_source_component_body_index(ctx)
-
-	-- Always emit the canonical text form (per-source).
-	for _, src in ipairs(ctx.sources) do
-		if src.scan then
-			local n_atoms     = src.scan.atoms and #src.scan.atoms or 0
-			local n_raw_atoms = src.scan.raw_atoms and #src.scan.raw_atoms or 0
-			if n_atoms + n_raw_atoms > 0 then
-				local basename = duffle.basename_no_ext(src.path)
-
-				-- (1) atoms.sourcemap.txt — per-.word line map (unchanged contract).
-				local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
-				local sourcemap_body = render_source_map(src, wc)
-
-				-- (2) atoms.provenance.txt — per-.word provenance with `mac_X(...)` component resolution back to the component's definition file:line.
-				--     Consumed by `passes/dwarf_injection.lua` to synthesize `DW_TAG_inlined_subroutine` instances for source-level Step Into on component invocations.
-				local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
-				local prov_body = render_provenance(src, wc, comp, comp_body_index)
-
-				if not ctx.dry_run then
-					duffle.ensure_dir(duffle.dirname(sourcemap_path))
-					duffle.write_file_lf(sourcemap_path, sourcemap_body)
-					duffle.write_file_lf(prov_path,    prov_body)
-				end
-
-				outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
-				outputs[#outputs + 1] = { kind = "report", path = prov_path }
 			end
 		end
+		if not has_projection then
+			for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
+				if atom.paths then has_projection = true; break end
+			end
+		end
+		if has_projection then
+			local basename = duffle.basename_no_ext(src.path)
+			-- (1) atoms.sourcemap.txt — format-1 per-word call-site map.
+			local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
+			local sourcemap_body = render_source_map(src)
+			-- (2) atoms.provenance.txt — format-1 per-word definition/body map.
+			local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
+			local prov_body = render_provenance(src, wc)
+			duffle.ensure_dir(duffle.dirname(sourcemap_path))
+			duffle.write_file_lf(sourcemap_path, sourcemap_body)
+			duffle.write_file_lf(prov_path,    prov_body)
+			outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
+			outputs[#outputs + 1] = { kind = "report", path = prov_path }
+		end
 	end
 
 	-- Optionally emit the gdb-runtime form (post-link, one file per build).

scripts/passes/components.lua

@@ -1,25 +1,16 @@
 --- passes/components.lua — Component-macro header generator.
 ---
---- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
+--- Ownership: `corpus.word_counts`, `corpus.components`, and `corpus.component_body_index`.
---- for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations, then does per-source backward lookups
+--- Scanner owns `declaration_comment` and `debug_skip` on each declaration record; this pass projects both forward.
---- for the function-args string (from the preceding `FI_ MipsAtom ac_X(...)` function declaration)
---- and the preceding comment block (for LSP/IntelliSense signature docs).
 ---
---- Emits a per-directory `<dir_basename>.macs.h` containing one `#define mac_X(sig) \` macro per component + `WORD_COUNT(mac_X, N)`
+--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations,
---- entries for downstream offset computation.
+--- then resolves the function-args string from the preceding `FI_ MipsAtom ac_X(...)` declaration via a backward walk.
+---
+--- Emits one `<dir_basename>.macs.h` per source with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
 
---- @class Component
---- @field name    string
---- @field body    string
---- @field args    string|nil
---- @field line    integer
---- @field comment string|nil
-
---- @class M
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
@@ -31,7 +22,6 @@
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir  = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle          = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -69,12 +59,10 @@ local GEN_SUBDIR      = "gen"
 --- @field sources            SourceFile[]           -- all source files in the build
 --- @field metadata_path      string                 -- path to word_count.metadata.h
 --- @field shared             table                  -- cross-pass shared state
---- @field shared.word_counts table<string, integer> -- populated by word-counts + components
 --- @field out_root           string                 -- output root (e.g. "build/gen")
 --- @field project_root       string                 -- project root (e.g. "code/")
 --- @field upstream           table<string, table>   -- per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @field dry_run            boolean                -- if true, compute but don't write
 --- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
@@ -83,11 +71,13 @@ local GEN_SUBDIR      = "gen"
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class Component
---- @field name    string     -- atom name (without `ac_` prefix)
+--- @field name       string     -- atom name (without `ac_` prefix)
---- @field body    string     -- brace-delimited body (without the braces)
+--- @field body       string     -- brace-delimited body (without the braces)
---- @field args    string|nil -- function-args string (function form only)
+--- @field args       string|nil -- function-args string (function form only)
---- @field line    integer    -- source line of the declaration
+--- @field line       integer    -- source line of the declaration
---- @field comment string|nil -- preceding `/* */` or `//` comment block (signature doc)
+--- @field comment    string|nil -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @field kind       string     -- "comp_bare" | "comp_proc"
+--- @field debug_skip boolean    -- mirror of `a.debug_skip` (scanner-owned); true iff a bare `atom_dbg_skip` marker immediately preceded the declaration
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Local helpers (file I/O + path normalization)
@@ -96,7 +86,11 @@ local GEN_SUBDIR      = "gen"
 local M = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Back-walk helpers (composed into the 2 entry points below: find_function_args_for + preceding_comment_block)
+-- Back-walk helpers (composed into the entry point below: find_function_args_for)
+--
+-- Only the function-args lookup for proc components occurs here.
+-- The preceding-comment walk occur in `scan_source.lua` — `a.declaration_comment` carries the resolved comment,
+-- so this file reads it forward rather than re-walking the source.
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation of the given name.
@@ -108,8 +102,8 @@ local M = {}
 --- We then verify the preceding context ends with `MipsAtom`
 --- (the function-decl keyword with possible qualifiers between).
 ---
---- @param source string
+--- @param source     string
---- @param name string
+--- @param name       string
 --- @param before_pos integer
 --- @return string|nil
 local function find_function_args_for(source, name, before_pos)
@@ -143,79 +137,6 @@ local function find_function_args_for(source, name, before_pos)
 	return inner
 end
 
---- Find the contiguous comment block immediately preceding `pos` in `source`.
---- Returns the comment text (with the `/* */` or `//` markers preserved) or an empty string if no comment is adjacent.
----
---- Used to copy signature comments from the source declaration (`MipsAtomComp_` / `MipsAtomComp_Proc_` / function decl)
---- over to the generated `mac_X` macro, so LSP/IntelliSense displays the args doc.
---- @param source string
---- @param pos integer
---- @return string
-local function preceding_comment_block(source, pos)
-	local scan_pos = pos
-	local pieces   = {}
-	while true do
-		-- Skip whitespace (space/tab/newline/CR) backward from `scan_pos`,
-		-- returning the position of the first non-whitespace char.
-		local non_ws = scan_pos - 1
-		while non_ws > 0 do
-			local ch = source:sub(non_ws, non_ws)
-			if ch == " " or ch == "\t" or ch == "\n" or ch == "\r" then
-				non_ws = non_ws - 1
-			else
-				break
-			end
-		end
-		if non_ws == 0 then break end
-
-		local is_block_close = non_ws >= 2 and source:sub(non_ws - 1, non_ws) == "*/"
-		local is_line_end    = source:sub(non_ws, non_ws) == "\n" or source:sub(non_ws, non_ws) == "\r"
-
-		if is_block_close then
-			-- Find the opening `/*` for a block comment whose `*/` ends at `non_ws`.
-			-- Walk back from `non_ws` over `/*` candidates.
-			local prefix  = source:sub(1, non_ws - 1)
-			local open_at = nil
-			for scan = #prefix - 1, 1, -1 do
-				if prefix:sub(scan, scan + 1) == "/*" then
-					open_at = scan
-					break
-				end
-			end
-			if not open_at then break end
-			-- Walk back from `open_at` over leading spaces + tabs to include the indentation before the `/*`.
-			local block_start = open_at
-			while block_start > 1 do
-				local ch = source:sub(block_start - 1, block_start - 1)
-				if ch == " " or ch == "\t" then
-					block_start = block_start - 1
-				else
-					break
-				end
-			end
-			table.insert(pieces, 1, source:sub(block_start, non_ws))
-			scan_pos = block_start
-		elseif is_line_end then
-			-- Walk back from `non_ws` to the start of the source line (the most recent `\n` or position 1).
-			local line_start = non_ws
-			while line_start > 1 and source:sub(line_start - 1, line_start - 1) ~= "\n" do
-				line_start = line_start - 1
-			end
-			local line = source:sub(line_start, non_ws)
-			if line:sub(1, 2) == "//" then
-				table.insert(pieces, 1, line)
-				scan_pos = line_start - 1
-			else
-				break
-			end
-		else
-			break
-		end
-	end
-	if #pieces == 0 then return "" end
-	return table.concat(pieces, "\n")
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Argument-name extraction
 -- ════════════════════════════════════════════════════════════════════════════
@@ -235,7 +156,6 @@ local function extract_arg_names(args_str)
 		if trimmed ~= "" then
 			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
 			-- then walk back over the identifier chars (alnum + `_`).
-			-- Plex: inlined the 2 single-caller helpers (no 2-caller rule met).
 			local ident_end = #trimmed
 			while ident_end > 0 do
 				local ch = trimmed:sub(ident_end, ident_end)
@@ -248,7 +168,7 @@ local function extract_arg_names(args_str)
 			local ident_start = ident_end
 			while ident_start > 0 do
 				local ch = trimmed:sub(ident_start, ident_start)
-				if duffle.is_alnum(ch) or ch == "_" then
+				if duffle.is_alnum_byte(string.byte(ch)) or ch == "_" then
 					ident_start = ident_start - 1
 				else
 					break
@@ -267,26 +187,34 @@ end
 -- Component projection (read from pre-scanned SourceScan)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
+--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
--- Does per-source backward lookups for args (preceding function decl) and comment (preceding comment block).
+--- Reads the scanner-owned `declaration_comment` (resolved by scan_source.lua, skipping backward across an associated bare `atom_dbg_skip` marker when present).
--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table instead of calling duffle.tokenize_body again.
+--- Per-source backward lookups remain in place only for the function `args` of proc components.
--- @param source string  -- the full source text (needed for backward lookups)
+--- That lookup is unique to components.lua and stays separate from the declaration-comment walk.
--- @param scan  table   -- SourceScan from duffle.scan_source
+--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table instead of calling duffle.tokenize_body again.
--- @return Component[]
+--- Carries the scanner-owned `debug_skip` flag forward so the generated projection can emit `/* atom_dbg_skip */`
+--- before the authored comment and so `update_canonical_components` can mirror the same field onto `corpus.components[name]`.
+--- @param source string  -- the full source text (needed for backward lookups)
+--- @param scan   table   -- SourceScan from duffle.scan_source
+--- @return Component[]
 local function project_components(source, scan)
 	local out = {}
 	for _, a in ipairs(scan.atoms) do
 		if a.kind == "comp_bare" or a.kind == "comp_proc" then
-			local args    = find_function_args_for(source, a.raw_name, a.ident_pos)
+			local args = find_function_args_for(source, a.raw_name, a.ident_pos)
-			local comment = preceding_comment_block(source, a.ident_pos)
+			-- Comment ownership: scan_source.lua stamps `declaration_comment` on the record by walking backward past any associated bare marker.
+			-- The pass reads `declaration_comment` directly.
+			local comment = a.declaration_comment or ""
 			out[#out + 1] = {
 				line        = a.line,
 				name        = a.name,
 				body        = a.body,
+				body_off    = a.body_off,
 				body_tokens = a.body_tokens,
 				args        = args,
 				comment     = comment,
 				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
+				debug_skip  = a.debug_skip == true,
 			}
 		end
 	end
@@ -340,11 +268,11 @@ end
 -- Word-count computation (memoized recursive lookup)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
+--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
--- Returns the ident unchanged if it doesn't start with the prefix 
+--- Returns the ident unchanged if it doesn't start with the prefix 
--- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
+--- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
--- @param ident string|nil
+--- @param ident string|nil
--- @return string|nil
+--- @return string|nil
 local function strip_mac_prefix(ident)
 	if not ident then return nil end
 	if ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
@@ -353,13 +281,13 @@ local function strip_mac_prefix(ident)
 	return ident
 end
 
--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
+--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
--- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
+--- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
--- @param name string -- the component name (without `mac_`)
+--- @param name         string -- the component name (without `mac_`)
--- @param comp_by_name table<string, Component>
+--- @param comp_by_name table<string, Component>
--- @param wc table<string, integer>
+--- @param wc           table<string, integer>
--- @param cache table<string, integer>
+--- @param cache        table<string, integer>
--- @return integer
+--- @return integer
 local function word_count_rec(name, comp_by_name, wc, cache)
 	if cache[name] ~= nil then return cache[name] end
 	cache[name] = -1  -- mark in-progress (cycle detection)
@@ -398,7 +326,7 @@ end
 --- references hit memoized values instead of re-walking the body.
 --- Cycle detection (A -> B -> A) is preserved via the in-progress `-1` sentinel in `cache`.
 --- @param components Component[]
---- @param wc table<string, integer>
+--- @param wc         table<string, integer>
 --- @return table<string, integer>  -- map of component name (without `mac_`) -> word count
 local function count_all_components(components, wc)
 	local comp_by_name = {}
@@ -471,13 +399,23 @@ end
 
 --- Build the list of lines for one component
 --- (signature comment, `#define mac_X(...)` line with backslash-continued tokens, then `WORD_COUNT(mac_X, N)` entry).
---- @param c Component
+--- For skipped components, a `/* atom_dbg_skip */` marker comment is emitted immediately before the authored comment block.
+--- The marker is a single line, the comment comes next, and the `#define` line follows. The `debug_skip` stamp is scanner-owned
+--- (`a.debug_skip == true` on the declaration record); the components pass projects it directly.
+--- @param c          Component
 --- @param components Component[]
---- @param wc table<string, integer>
+--- @param wc         table<string, integer>
 --- @return string[]  -- list of lines for this component
 local function build_component_lines(c, counts)
 	local lines = {}
 
+	-- Marker comment: emitted once for every skipped component.
+	-- The marker is scanner-owned (declared by `atom_dbg_skip` immediately before the declaration in the source);
+	-- the components pass projects `c.debug_skip` and emits the marker as a generated comment.
+	if c.debug_skip then
+		lines[#lines + 1] = "/* atom_dbg_skip */"
+	end
+
 	if c.comment and c.comment ~= "" then
 		for _, line in ipairs(split_comment_lines(c.comment)) do
 			lines[#lines + 1] = line
@@ -505,10 +443,10 @@ end
 -- Per-source emit logic
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
+--- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
--- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
+--- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
--- @param src SourceFile
+--- @param src SourceFile
--- @return string[]
+--- @return string[]
 local function header_boilerplate(src)
 	return {
 		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
@@ -530,13 +468,13 @@ local function header_boilerplate(src)
 	}
 end
 
--- Compute the output path for one source's `.macs.h` file.
+--- Compute the output path for one source's `.macs.h` file.
--- The pre-rework convention uses the *directory* basename
+--- The pre-rework convention uses the *directory* basename (not the source file basename)
--- (not the source file basename) e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+---  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
--- This matches what the C codebase #includes.
+--- This matches what the C codebase #includes.
--- @param src SourceFile
+--- @param src SourceFile
--- @return string  -- the output directory
+--- @return string  -- the output directory
--- @return string  -- the full output path
+--- @return string  -- the full output path
 local function compute_macs_h_path(src)
 	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
 	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
@@ -545,11 +483,10 @@ end
 
 --- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
 --- Writes in BINARY mode so LF line endings are preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
---- Honors `ctx.dry_run`: prints the intended path but does not write the file.
+--- @param ctx        PassCtx
---- @param ctx PassCtx
+--- @param src        SourceFile
---- @param src SourceFile
 --- @param components Component[]
---- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
 --- @return string|nil  -- path to the written file (nil if no components)
 local function emit_component_macros_h(ctx, src, components, counts)
 	if #components == 0 then return nil end
@@ -563,11 +500,6 @@ local function emit_component_macros_h(ctx, src, components, counts)
 	end
 
 	local content = table.concat(lines, "\n") .. "\n"
-	if ctx.dry_run then
-		print(string.format("  -> %s (dry-run)", out_path))
-		return out_path
-	end
-
 	duffle.ensure_dir(out_dir)
 	duffle.write_file_lf(out_path, content)
 	print(string.format("  -> %s", out_path))
@@ -578,41 +510,90 @@ end
 -- Pass entry
 -- ════════════════════════════════════════════════════════════════════════════
 
--- (internal) Extend `ctx.shared.word_counts` with this source's component macros so offsets sees them without re-reading the file.
+--- (internal) Extend `corpus.word_counts` with this source's component macros so offsets sees them without re-reading the file.
--- @param ctx PassCtx
+--- First declaration wins: a later caller's count is dropped (the existing entry from the first source is preserved).
--- @param components Component[]
+--- @param corpus     table  -- the corpus
--- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param components Component[]
-local function update_shared_word_counts(ctx, components, counts)
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
-	local wc = ctx.shared.word_counts
+local function update_canonical_word_counts(corpus, components, counts)
+	local wc = corpus.word_counts
 	for _, c in ipairs(components) do
-		wc["mac_" .. c.name] = counts[c.name]
+		local key = "mac_" .. c.name
+		if wc[key] == nil then
+			wc[key] = counts[c.name]
+		end
 	end
 end
 
 --- @class ComponentDef
---- @field name    string  -- bare name (without ac_/mac_ prefix)
+--- @field name       string  -- bare name (without ac_/mac_ prefix)
---- @field line    integer -- definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
+--- @field line       integer -- definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
---- @field path    string  -- absolute source path of the definition
+--- @field path       string  -- absolute source path of the definition
---- @field kind    string  -- "comp_bare" | "comp_proc"
+--- @field kind       string  -- "comp_bare" | "comp_proc"
+--- @field debug_skip boolean -- mirror of the scanner-owned `a.debug_skip`; consumers read this directly
 
---- (internal) Extend `ctx.shared.components` with this source's components-by-name map so downstream passes
+--- (internal) Populate `corpus.components` with this source's components-by-name map.
---- (atoms_source_map, dwarf_injection) can resolve `mac_X(...)` invocations back to their component definition file:line.
+--- First declaration wins; later declarations of the same bare name are dropped and recorded as a collision via `corpus.collisions` (kind = "component").
---- provenance emission uses this to attribute each emitted `.word` to either a component macro or the enclosing atom body.
+--- The pass does NOT write to `ctx.shared.components` (ownership follows the canonical contract).
--- @param ctx PassCtx
+--- The `debug_skip` field mirrors the scanner-owned declaration record (`c.debug_skip`).
--- @param src SourceFile
+--- No parallel skip map is built here; consumers that need the per-component skip state read `corpus.components[name].debug_skip` directly.
--- @param components Component[]
+--- @param corpus     table  -- the corpus
-local function update_shared_components(ctx, src, components)
+--- @param src        SourceFile
-	ctx.shared.components         = ctx.shared.components or {}
+--- @param components Component[]
+local function update_canonical_components(corpus, src, components)
 	local rel_path = src.path:gsub("\\", "/")
 	for _, c in ipairs(components) do
 		-- Keyed by bare name (e.g. `yield`, `load_tri_indices`).
-		-- The atoms_source_map pass strips the `mac_` prefix from the call site identifier before lookup.
+		-- The atoms_source_map pass looks up components by bare name from the corpus;
-		ctx.shared.components[c.name] = {
+		-- `mac_` prefix lives at the call-site identifier and is stripped before lookup.
-			name = c.name,
+		if corpus.components[c.name] == nil then
-			line = c.line,
+			corpus.components[c.name] = {
-			path = rel_path,
+				name       = c.name,
-			kind = c.kind or "comp_bare",
+				line       = c.line,
-		}
+				path       = rel_path,
+				kind       = c.kind or "comp_bare",
+				debug_skip = c.debug_skip == true,
+			}
+		else
+			-- A second declaration of the same bare name: record a typed collision so static-analysis + the report can surface it.
+			-- Identical-shape declarations (same path + line) reuse the first-wins entry without a collision record.
+			local existing = corpus.components[c.name]
+			if existing.path ~= rel_path or existing.line ~= c.line then
+				local kind     = c.kind or "comp_bare"
+				local first_kind = existing.kind or "comp_bare"
+				corpus.collisions[#corpus.collisions + 1] = {
+					kind              = "component",
+					name              = c.name,
+					first_site        = { path = existing.path, line = existing.line },
+					conflicting_site  = { path = rel_path,       line = c.line },
+					first_shape       = "kind=" .. first_kind,
+					conflicting_shape = "kind=" .. kind,
+				}
+			end
+		end
+	end
+end
+
+--- (internal) Populate `corpus.component_body_index` with this source's body index entries.
+--- First declaration wins; later declarations are dropped (no separate collision record: the components collision is already surfaced by `update_canonical_components`).
+--- The pass writes to `corpus.component_body_index` only (the corpus owns this projection).
+--- @param corpus     table  -- the corpus
+--- @param src        SourceFile
+--- @param components Component[]
+--- @param scan       table  -- the SourceScan payload (for line_of)
+local function update_canonical_component_body_index(corpus, src, components, scan)
+	local line_of = scan and scan.line_of
+	for _, c in ipairs(components) do
+		if corpus.component_body_index[c.name] == nil then
+			corpus.component_body_index[c.name] = {
+				body_tokens = c.body_tokens,
+				body_off    = c.body_off,
+				line_of     = line_of,
+				source      = src.path,
+				declaration = c.line,
+				kind        = c.kind,
+			}
+		end
 	end
 end
 
@@ -623,24 +604,41 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- Initialize shared component map.
+	-- Corpus ownership gate.
-	-- The atoms_source_map and dwarf_injection passes consume `ctx.shared.components` to resolve `mac_X(...)` 
+	local corpus = ctx.shared and ctx.shared.corpus
-	-- invocations back to the component's definition file:line.
+	if type(corpus) ~= "table" then
-	ctx.shared.components = ctx.shared.components or {}
+		error("components.run requires ctx.shared.corpus.", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("components.run requires ctx.shared.corpus.source_order.", 0)
+	end
+	if type(corpus.word_counts) ~= "table" then
+		error("components.run requires ctx.shared.corpus.word_counts; "
+			.. "word_count_eval.run must run before components.run "
+			.. "(see PASSES deps).", 0)
+	end
 
-	for _, src in ipairs(ctx.sources) do
+	-- Projection ownership:
+	--   * `corpus.word_counts["mac_"..name]` — current component count
+	--   * `corpus.components[name]`         — bare-name component definition
+	--   * `corpus.component_body_index[name]` — body / line_of / source index
+	-- The pass writes to the corpus only; consumers read from the corpus directly.
+
+	for _, src in ipairs(corpus.source_order) do
 		-- project_components reads from src.scan + does backward lookups on src.text
 		local components = project_components(src.text, src.scan)
 		if #components > 0 then
-			-- Compute word counts for ALL components once (was: rebuilt per call inside the helpers).
+			-- Compute all component word counts once per source.
-			local counts = count_all_components(components, ctx.shared.word_counts)
+			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
+			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
+			local counts = count_all_components(components, corpus.word_counts)
 			local macs_path = emit_component_macros_h(ctx, src, components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				update_shared_word_counts(ctx, components, counts)
+				-- Populate the projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
-				-- share component definitions with downstream passes.
+				update_canonical_word_counts(corpus, components, counts)
-				-- `mac_X(...)` invocations in atom bodies resolve back to (path, line) via this map.
+				update_canonical_components(corpus, src, components)
-				update_shared_components(ctx, src, components)
+				update_canonical_component_body_index(corpus, src, components, src.scan)
 			end
 		end
 	end

scripts/passes/emission_model.lua

@@ -0,0 +1,239 @@
+--- passes/emission_model.lua: Per-atom emission projection.
+---
+--- The `emission-model` pass owns `atom.paths`, the canonical per-atom mutable surface for atoms and raw atoms with bodies in `ctx.shared.corpus.source_order`.
+--- For each atom, the pass invokes `duffle.project_emission(body_text, component_index, word_counts, components)`.
+--- It stores the ordered `items` stream plus the dense `word_events` / `markers` / `invocations` views on `atom.paths`.
+---
+--- Public boundary:
+---   * `M.run(ctx)` is the only entry point.
+---   * The pass returns `{outputs = {}, errors = ..., warnings = ...}`.
+---     Pass kind = `validation` → `PASS_KIND_STOP_ON_ERROR.validation` preserves the existing build-stopping policy.
+---
+--- Source-order discipline:
+---   * `corpus.source_order` sets the source-record order.
+---   * Within each source, the pass visits `src.scan.atoms` and `src.scan.raw_atoms` in declaration order.
+---
+--- Per-atom projection fields on `atom.paths`:
+---   `tokens`, `line_in_body`, `items`, `word_events`, `markers`, `invocations`, `errors`, `warnings`.
+---   The construction walk appends `items` and derives each dense view from that ordered stream.
+---
+--- Component expansion and construction validation:
+---   * known `mac_X(...)` calls recursively expand component bodies;
+---   * invocation records retain monotonic IDs, parent IDs, immediate call text, and the immutable outermost root call text;
+---   * invocation construction stamps `debug_skip` from `corpus.components[name].debug_skip` at the construction site (no second pass, no source parse, no parallel lookup);
+---   * component cycles close balanced invocation boundaries and emit a `cycle` construction error at the recursive edge;
+---   * declared-vs-measured component word counts emit `count_mismatch` construction errors; opaque uncounted macros emit warnings.
+---
+--- `passes.scan_source` strips its private `_code_macros` / `_code_macro_bodies` tables before this pass runs.
+
+local M = {}
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Bootstrap: load `duffle_paths.lua` via debug.getinfo so the module works standalone (run as `luajit passes/emission_model.lua`) and when require'd from the orchestrator.
+-- ─────────────────────────────────────────────────────────────────────────
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Helpers
+-- ─────────────────────────────────────────────────────────────────────────
+
+-- Convert the recursive walk's body-relative line numbers into physical source lines once.
+-- The walker builds `line_of` from `body_text` and stamps body-relative line numbers (1..N) into `item.line` and `invocation.call_line`.
+-- This function converts those values to physical source lines at the close site with the forwarded source `line_of` closure.
+--
+-- `call_line` discipline:
+--   * ROOT invocations (`inv.parent_id == 0`) receive body-relative `call_line` values directly from `M.LineIndex(body_text)` in the walker.
+--     The source `line_of` closure supplies physical lines at the close site, so this function converts each root value exactly once.
+--   * INNER invocations (`inv.parent_id ~= 0`) receive physical `call_line` values directly from the COMPONENT's `line_of` in the walker.
+--     Recursive descent forwards that closure through `corpus.component_body_index[name].line_of`; those values arrive physical and remain unchanged.
+--
+-- After this function, every `inv.call_line` is physical. DWARF and provenance output read it directly.
+-- The word-event loop forwards the already-physical `outer_inv.call_line` into `we.call_line` for words inside an invocation.
+local function stamp_root_provenance(projection, atom_record, src, corpus)
+	local root_line_of   = src.scan and src.scan.line_of
+	assert(type(root_line_of) == "function"
+		, "emission_model: src.scan.line_of is required (canonical LineIndex closure over the source text) to stamp physical provenance")
+	assert(type(atom_record.body_off) == "number"
+		, "emission_model: atom_record.body_off (byte offset of the body's first byte in source) is required to derive `root_body_line`. The scanner must populate body_off for every atom record.")
+	-- `root_body_line` is the physical source line of the ATOM HEADER byte containing the opening `{`; that byte is one byte BEFORE `atom_record.body_off`.
+	-- The walker assigns line 2 to the body's first content line because line 1 is the trailing `\n` after `{`. Body-text line k therefore maps to `root_body_line + (k - 1)`.
+	-- `body_off - 1` points at the opening `{`, whose line index identifies the header line. `body_off` points after `{` and would shift every word row forward by one line.
+	local root_body_line = root_line_of(atom_record.body_off - 1)
+		or atom_record.line or 0
+	local component_index = corpus.component_body_index or {}
+	local word_items = {}
+
+	for _, item in ipairs(projection.items) do
+		if item.kind == "word" then word_items[#word_items + 1] = item end
+	end
+
+	-- Resolve one word's physical body line, where the byte containing that word appears in source.
+	--   * Component expansions carry `invocation_ids`; the component's full-file `line_of` leaves `item.line` physical.
+	--   * Raw tokens in the root atom body carry an empty `invocation_ids` list and a body-relative `item.line`; convert them here.
+	local function body_line_for(event, item)
+		local ids = event.invocation_ids or {}
+		-- The innermost open invocation identifies which line index the walker used.
+		-- A component `line_of` makes `item.line` physical; the atom's `body_text` line index makes it body-relative.
+		if ids and #ids > 0 then
+			local inner_id  = ids[#ids]
+			local inner_inv = inner_id and projection.invocations[inner_id]
+			if inner_inv then
+				local component = component_index[inner_inv.component_name]
+				if component and component.line_of then
+					-- Walker used `comp.line_of`, which is the source's physical LineIndex. item.line is already physical.
+					return item.line or 0
+				end
+			end
+		end
+		-- RAW root-body word: item.line is body-text's 1-based line number (the first content line is line 2 because line 1 is the trailing `\n` after `{`).
+		-- Convert body-text-relative → physical using `root_body_line + (item.line - 1)`.
+		return (root_body_line or 0) + (item.line or 1) - 1
+	end
+
+	-- Stamp the root source path onto invocation records whose `call_path` the walker left empty.
+	-- The walker passes `body_entry.source` to `emit_invoke_begin`; `M.project_emission` creates the root `body_entry` with source `""`, leaving its `call_path` empty.
+	-- This stamp gives every invocation a physical `call_path` matching `passes/atoms_source_map.lua`'s in-memory provenance projection.
+	local root_path = src.path or ""
+	for _, inv in ipairs(projection.invocations) do
+		if inv.call_path == nil or inv.call_path == "" then
+			inv.call_path = root_path
+		end
+	end
+
+	-- Normalize `inv.call_line` to a physical source line.
+	--   * ROOT invocations (`parent_id == 0`) carry body-relative `call_line` values from `M.LineIndex(body_text)`; convert them once with `root_body_line`.
+	--   * INNER invocations (`parent_id ~= 0`) carry physical `call_line` values from the component's `line_of`; retain them unchanged.
+	for _, inv in ipairs(projection.invocations) do
+		if inv.parent_id == 0 then
+			inv.call_line = (root_body_line or 0) + (inv.call_line or 1) - 1
+		end
+	end
+
+	-- Build `body_lines` for each invocation.
+	-- `atoms_source_map` and `dwarf_injection` read `inv.body_lines[k]` directly from the invocation record created here.
+	-- Component words already carry physical `item.line` values from the walker's COMPONENT line index, so `body_line_for` returns them unchanged.
+	for _, inv in ipairs(projection.invocations) do
+		local sw = inv.start_word
+		local ew = inv.end_word
+		local bls = {}
+		for i = sw, ew do
+			local it = projection.items and projection.items[i]
+			if it and it.kind == "word" then
+				local fake_event = { invocation_ids = { inv.id } }
+				bls[#bls + 1] = body_line_for(fake_event, it) or 0
+			end
+		end
+		inv.body_lines = bls
+	end
+
+	-- Resolve each `word_event`'s physical `body_line` and `call_line`.
+	-- For words inside an invocation, `we.call_line` identifies the OUTER atom source line containing the `mac_X(...)` token that triggered expansion.
+	-- The root-invocation conversion above makes every `inv.call_line` physical; forward it directly and use each raw word's `body_line` as the fallback.
+	for index, we in ipairs(projection.word_events) do
+		local item      = word_items[index] or {}
+		local body_line = body_line_for(we, item)
+		item.line    = body_line
+		we.body_line = body_line
+
+		local call_line = body_line
+		local outer_id  = we.outermost_invocation_id or 0
+		local outer_inv = projection.invocations[outer_id]
+		if outer_inv then
+			-- `outer_inv.call_line` is physical after the conversion loop above, so use it directly.
+			call_line = outer_inv.call_line
+		end
+		we.call_line = call_line
+
+		if we.def_path  == nil or we.def_path  == "" then we.def_path  = src.path         or "" end
+		if we.def_line  == nil or we.def_line  == 0  then we.def_line  = atom_record.line or 0  end
+		if we.call_path == nil or we.call_path == "" then we.call_path = src.path         or "" end
+	end
+end
+
+-- Project one atom record into `atom.paths`.
+-- Mutates the atom record in-place and returns the projection (for pass-level error/warning accumulation).
+local function project_atom(atom_record, src, corpus)
+	local body  = atom_record.body or ""
+	local wc    = corpus.word_counts or {}
+	local cbi   = corpus.component_body_index or {}
+	-- That construction site stamps `invocation.debug_skip` while appending each record to `proj.invocations`.
+	local proj  = duffle.project_emission(body, cbi, wc, corpus.components)
+	local paths = {
+		tokens       = atom_record.body_tokens or {},
+		line_in_body = duffle.build_body_line_index(body),
+		items        = proj.items,
+		word_events  = proj.word_events,
+		markers      = proj.markers,
+		invocations  = proj.invocations,
+		errors       = proj.errors,
+		warnings     = proj.warnings,
+	}
+	stamp_root_provenance(proj, atom_record, src, corpus)
+	atom_record.paths = paths
+	return proj
+end
+
+-- ─────────────────────────────────────────────────────────────────────────
+-- Run the emission-model pass.
+-- ─────────────────────────────────────────────────────────────────────────
+
+--- @param ctx PassCtx  -- { shared = { corpus = ... }, out_root, ... }
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx and ctx.shared and ctx.shared.corpus
+	if type(corpus)              ~= "table" then error("emission_model: ctx.shared.corpus is required (canonical projection)", 0) end
+	if type(corpus.source_order) ~= "table" then error("emission_model: ctx.shared.corpus.source_order is required", 0) end
+
+	-- Project once, collect errors + warnings for one atom.
+	-- Kind must be one of: atom | raw_atom | comp_bare | comp_proc.
+	local function process_atom(atom, src)
+		if not (atom and atom.body) then return end
+		local kind = atom.kind
+		if kind ~= "atom" and kind ~= "raw_atom" and kind ~= "comp_bare" and kind ~= "comp_proc" then
+			return
+		end
+		local proj = project_atom(atom, src, corpus)
+		for _, e in ipairs(proj.errors) do
+			-- Preserve `kind` (cycle / count_mismatch / unbalanced) so readers dispatch on the diagnostic class and leave the message string as display text.
+			errors[#errors + 1] = {
+				kind   = e.kind,
+				line   = e.line,
+				msg    = e.msg,
+				source = e.source or src.path,
+			}
+		end
+		for _, w in ipairs(proj.warnings) do
+			warnings[#warnings + 1] = {
+				kind = w.kind,
+				line = w.line,
+				msg  = w.msg,
+			}
+		end
+	end
+
+	-- Walk `corpus.source_order`; within each source, visit atoms followed by raw_atoms.
+	-- Recognized kinds (atom | raw_atom | comp_bare | comp_proc) each receive the atom.paths projection via duffle.project_emission.
+	-- Components are macros inlined into atom bodies; focused tests and isolated component analyses consume atom.paths directly.
+	for _, src in ipairs(corpus.source_order) do
+		local scan = src.scan or {}
+		for _, atom in ipairs(scan.atoms or {}) do
+			process_atom(atom, src)
+		end
+		for _, atom in ipairs(scan.raw_atoms or {}) do
+			process_atom(atom, src)
+		end
+	end
+
+	return {
+		outputs  = outputs,
+		errors   = errors,
+		warnings = warnings,
+	}
+end
+
+return M

scripts/passes/offsets.lua

@@ -21,18 +21,12 @@
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
-local count_token_words = word_count_eval.count_token_words
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Marker-call identifiers inside atom bodies.
-local LABEL_MARKER    = "atom_label"
-local OFFSET_MARKER   = "atom_offset"
-
 -- Offset macro/enum naming prefixes (the emitted header uses these).
 local OFFSET_MACRO_PREFIX = "_atom_offset_"
 local OFFSET_ENUM_PREFIX  = "atom_offset_"
@@ -52,16 +46,10 @@ local OFFSET_MACRO_COL = 44
 --- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
 --- @field shared             table                  -- cross-pass shared state
---- @field shared.word_counts table                  -- macro name -> word count
+--- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared.word_counts table
 --- @field out_root           string                 -- output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
---- @field flags              table                  -- CLI flags
---- @field dry_run            boolean                -- if true, compute but don't write
---- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -69,10 +57,10 @@ local OFFSET_MACRO_COL = 44
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class BranchOffset
---- @field tag    string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field tag         string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
---- @field target string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field target      string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field pos    integer  -- the branch's word position within the atom body
+--- @field branch_word integer  -- branch word position within the atom body
---- @field offset integer  -- computed `target_word - branch_word - 1`
+--- @field offset      integer  -- computed `target_word - branch_word - 1`
 
 --- @class AtomData
 --- @field name        string         -- atom name
@@ -80,169 +68,85 @@ local OFFSET_MACRO_COL = 44
 --- @field offsets     BranchOffset[] -- per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Per-token marker-call helpers (atom_label / atom_offset inside bodies)
+-- Canonical marker projection
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Extract comma-separated identifier args from a parenthesized group after a function-like macro call.
+-- MARKER_PROJECTORS is the marker-kind data table.
--- Returns (args, after_paren) where `after_paren` is the position just past the closing `)`, or nil if `token` did not start with `(`.
+-- The emission-model pass already records marker word positions;
--- @param token string
+-- this pass only projects those records into the label/branch lookup shape needed by offset computation.
--- @param after_ident integer
+local MARKER_PROJECTORS = {
--- @return string[], integer|nil
+	label = function(state, marker)
-local function extract_ident_args(token, after_ident)
+		state.labels[marker.name] = marker.word_index
-	local arg_start = duffle.skip_ws_and_cmt(token, after_ident)
+	end,
-	if token:sub(arg_start, arg_start) ~= "(" then return {}, nil end
+	offset = function(state, marker)
-	local inner, after_paren = duffle.read_parens(token, arg_start)
+		state.branches[#state.branches + 1] = {
-	-- scan: <marker>(<args>)
+			tag         = marker.name,
-
+			target      = marker.target,
-	local args      = {}
+			branch_word = marker.word_index,
-	local pos       = 1
+		}
-	local inner_len = #inner
+	end,
-	while pos <= inner_len do
-		pos = duffle.skip_ws_and_cmt(inner, pos)
-		if pos > inner_len then break end
-		local ident, after = duffle.read_ident(inner, pos)
-		if ident and ident ~= "" then
-			table.insert(args, ident)
-			pos = after
-		else
-			pos = pos + 1
-		end
-		pos = duffle.skip_ws_and_cmt(inner, pos)
-		if pos <= inner_len and inner:sub(pos, pos) == "," then pos = pos + 1 end
-	end
-
-	return args, after_paren
-end
-
--- (internal) Record a `atom_label(name)` marker — `at_pos` is the branch-free word position within the atom body.
--- @param labels table<string, integer>
--- @param args string[]
--- @param at_pos integer
-local function record_label_marker(labels, args, at_pos)
-	if #args >= 1 then labels[args[1]] = at_pos end
-end
-
--- (internal) Record a `atom_offset(tag, target)` marker.
--- @param branches table[]  -- list of {pos=, target=, tag=}
--- @param args string[]
--- @param at_pos integer
-local function record_offset_marker(branches, args, at_pos)
-	if #args >= 2 then
-		table.insert(branches, { pos = at_pos, target = args[2], tag = args[1] })
-	end
-end
-
--- MARKER_TO_HANDLER — data-driven marker dispatch (the plex pattern).
--- Maps the marker ident to its recorder function. Each handler takes (out_table, args, at_pos).
--- Adding a new marker type = 1 row + 1 recorder function.
-local MARKER_TO_HANDLER = {
-	[LABEL_MARKER]  = record_label_marker,
-	[OFFSET_MARKER] = record_offset_marker,
 }
 
---- Scan a single token for atom_label/atom_offset markers, walking through balanced groups transparently (so nested calls are found).
+--- Project canonical marker records into the two lookup tables used by the offset renderer.
---- @param token string
+--- No source text, body text, or body token is inspected.
---- @param at_pos integer    -- the branch-free word position of this token in the body
+--- @param markers table[] -- atom.paths.markers
---- @param labels table<string, integer>
+--- @return table<string, integer>, table[]
---- @param branches table[]
+local function project_markers(markers)
-local function scan_for_atom_markers(token, at_pos, labels, branches)
+	local state = { labels = {}, branches = {} }
-	local pos     = 1
+	for _, marker in ipairs(markers or {}) do
-	local tok_len = #token
+		local project = MARKER_PROJECTORS[marker.kind]
-	while pos <= tok_len do
+		if project then project(state, marker) end
-		pos = duffle.skip_ws_and_cmt(token, pos)
-		if pos > tok_len then break end
-		local ch = token:sub(pos, pos)
-		if duffle.is_alpha(ch) then
-			local ident, after = duffle.read_ident(token, pos)
-			local handler = MARKER_TO_HANDLER[ident]
-			if handler then
-				local args, after_paren = extract_ident_args(token, after)
-				-- Marker found — dispatch to its recorder. markers share labels and branches as
-				-- out-tables; the recorder picks which one(s) to write to based on its semantics.
-				-- (record_label_marker writes to labels; record_offset_marker writes to branches.)
-				handler(ident == LABEL_MARKER and labels or branches, args, at_pos)
-				pos = after_paren or after
-			else
-				pos = after
-			end
-		else
-			local nx = duffle.skip_str_or_cmt(token, pos)
-			pos = (nx > pos) and nx or (pos + 1)
-		end
 	end
-end
+	return state.labels, state.branches
-
---- Scan an atom body for labels + branches, count total words.
---- Returns (labels, branches, total_words).
---- @param body string
---- @param word_counts table
---- @return table<string, integer>, table[], integer
--- scan_atom_body: walk pre-tokenized body for atom_label/atom_offset markers + word counts.
--- Uses `atom.body_tokens` from the SourceScan payload (pre-tokenized by scan-source pass).
--- @param body_tokens table[]  -- {{tok=string, rel=integer}, ...} from duffle.tokenize_body
--- @param word_counts table
--- @return table, table, integer  -- labels, branches, total_words
-local function scan_atom_body(body_tokens, word_counts)
-	local pos      = 0
-	local labels   = {}
-	local branches = {}
-	for _, t in ipairs(body_tokens) do
-		local tok = t.tok
-		if duffle.is_marker_token(tok) then
-			-- Marker call: record at the current pos, do NOT advance pos.
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + duffle.count_marker_rest(tok, word_counts, count_token_words)
-		else
-			local words = count_token_words(tok, word_counts)
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + words
-		end
-	end
-	return labels, branches, pos
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
+--- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
--- @param labels table<string, integer>
+--- @param labels table<string, integer>
--- @param branches table[]
+--- @param branches table[]
--- @return BranchOffset[]
+--- @return BranchOffset[]
 local function compute_offsets(labels, branches)
 	local results = {}
 	for _, br in ipairs(branches) do
 		local target = labels[br.target]
 		if not target then
-			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.pos .. ")")
+			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")")
 		end
-		results[#results + 1] = { target = br.target, tag = br.tag, offset = target - br.pos - 1 }
+		results[#results + 1] = {
+			target      = br.target,
+			tag         = br.tag,
+			branch_word = br.branch_word,
+			offset      = target - br.branch_word - 1,
+		}
 	end
 	return results
 end
 
--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
+--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
--- @param s string
+--- @param s string
--- @param w integer
+--- @param w integer
--- @return string
+--- @return string
 local function pad_right(s, w)
 	return s .. string.rep(" ", math.max(0, w - #s))
 end
 
--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
+--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
--- @param r BranchOffset
+--- @param bo BranchOffset
--- @return table
+--- @return table
-local function make_offset_const(r)
+local function make_offset_const(bo)
 	return {
-		macro_name = OFFSET_MACRO_PREFIX .. r.tag .. "_" .. r.target,
+		macro_name = OFFSET_MACRO_PREFIX .. bo.tag .. "_" .. bo.target,
-		enum_name  = OFFSET_ENUM_PREFIX  .. r.tag .. "_" .. r.target,
+		enum_name  = OFFSET_ENUM_PREFIX  .. bo.tag .. "_" .. bo.target,
-		value      = r.offset,
+		value      = bo.offset,
 	}
 end
 
--- (internal) Emit one atom's offset constants + enum into the lines buffer.
+--- (internal) Emit one atom's offset constants + enum into the lines buffer.
--- @param add fun(s: string)
+--- @param add  fun(s: string)
--- @param atom AtomData
+--- @param atom AtomData
 local function emit_atom_offsets(add, atom)
 	if #atom.offsets == 0 then return end
 	add("// --- atom: " .. atom.name .. " (" .. atom.total_words .. " words) ---")
@@ -263,10 +167,10 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
--- Generate the per-source .offsets.h header.
+--- Generate the per-source .offsets.h header.
--- @param source_path string
+--- @param source_path string
--- @param atoms_data AtomData[]
+--- @param atoms_data  AtomData[]
--- @return string
+--- @return string
 local function generate_header(source_path, atoms_data)
 	local basename = duffle.basename_no_ext(source_path)
 
@@ -288,59 +192,41 @@ local function generate_header(source_path, atoms_data)
 	return table.concat(lines, "\n") .. "\n"
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- M — module exports
--- ════════════════════════════════════════════════════════════════════════════
-
 local M = {}
 
--- Project the pre-scanned SourceScan entries into the {name, body, body_tokens} shape this pass needs.
+--- (internal) Process one source: render offsets from canonical atom paths.
--- MipsAtom_ entries have kind="atom"; MipsCode code_<name> entries have kind="raw_atom".
+--- Returns the offsets_h path if a header was written, or nil.
--- `body_tokens` is set by scan-source on every `scan.atoms[i]` / `scan.raw_atoms[i]`; we carry it forward
+--- @param ctx PassCtx
--- so `scan_atom_body` reads from the precomputed table directly (no per-atom tokenize_body fallback).
+--- @param src SourceFile
--- @param scan table  -- SourceScan from duffle.scan_source
+--- @return string|nil  -- the offsets_h path
--- @return table[]  -- list of {name=, body=, body_tokens=}
-local function project_atoms(scan)
-	local out = {}
-	for _, a in ipairs(scan.atoms) do
-		out[#out + 1] = { name = a.raw_name, body = a.body, body_tokens = a.body_tokens }
-	end
-	for _, a in ipairs(scan.raw_atoms) do
-		out[#out + 1] = { name = a.name, body = a.body, body_tokens = a.body_tokens }
-	end
-	return out
-end
-
--- (internal) Process one source: project atoms from scan, scan bodies, write header.
--- Returns the offsets_h path if a header was written, or nil.
--- @param ctx PassCtx
--- @param src SourceFile
--- @return string|nil  -- the offsets_h path
 local function process_source(ctx, src)
-	local atoms = project_atoms(src.scan)
-	if #atoms == 0 then return nil end
-
 	local atoms_data = {}
-	for _, atom in ipairs(atoms) do
+	local scan = src.scan or {}
-		local labels, branches, total = scan_atom_body(atom.body_tokens, ctx.shared.word_counts)
+
+	local function append_atom(atom)
+		local paths = atom and atom.paths
+		if not paths then return end
+		local labels, branches = project_markers(paths.markers)
 		atoms_data[#atoms_data + 1] = {
-			name        = atom.name,
+			name        = atom.raw_name or atom.name,
-			total_words = total,
+			total_words = #(paths.word_events or {}),
 			offsets     = compute_offsets(labels, branches),
 		}
 	end
 
+	for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+	for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	if #atoms_data == 0 then return nil end
+
 	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
-	if not ctx.dry_run then
+	duffle.ensure_dir(duffle.dirname(out_path))
-		duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file(out_path, generate_header(src.path:gsub("/", "\\"), atoms_data))
-		duffle.write_file(out_path, generate_header(src.path, atoms_data))
-	end
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each source, emits a per-module `<dir_basename>.offsets.h` containing `#define _atom_offset_F_T = N` constants 
+--- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
---- for every `atom_offset(F, T)` reference in the source's atoms.
+--- containing constants for every marker recorded in atom.paths.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -348,7 +234,12 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	for _, src in ipairs(ctx.sources) do
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	end
+
+	for _, src in ipairs(corpus.source_order) do
 		local out_path = process_source(ctx, src)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }

scripts/passes/report.lua

@@ -5,8 +5,8 @@
 ---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
 ---   - `build/gen/annotation_validation.txt` — the project summary.
 ---
---- The annotation pass stashes per-MODULE summary entries in `ctx.flags._annot_results` (set by `passes/annotation.lua`). 
+--- The annotation pass emits `errors.h` files per module and the canonical `corpus.sources_by_dir` projection groups sources by directory.
---- This pass re-validates each source via `annotation.validate()` to get the detailed per-source results needed for the report. 
+--- This pass iterates the canonical dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
@@ -25,6 +25,11 @@
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
+-- Load the annotation pass so we can re-validate each source against the canonical corpus projection.
+-- The annotation pass exposes `M.validate`, which returns the per-source AnnotationResult (atoms / annots / macros / binds / errors / warnings)
+-- that the report pass renders into the per-module `<dir_basename>.annotations.txt` output.
+local annotation    = dofile(_bootstrap_dir .. "annotation.lua")
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
@@ -67,8 +72,6 @@ local PASS_NAME = "report"
 --- @field project_root       string               -- project root (e.g. "code/")
 --- @field upstream           table<string, table> -- per-pass upstream outputs
 --- @field flags              table                -- CLI flags + per-pass stash
---- @field flags._annot_results ModuleEntry[]      -- stashed by annotation pass
---- @field dry_run            boolean              -- if true, compute but don't write
 --- @field verbose            boolean              -- if true, log diagnostic info
 
 --- @class PassResult
@@ -138,31 +141,31 @@ local PASS_NAME = "report"
 -- Per-MODULE annotation report (aggregated across all sources in a dir)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
+--- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
--- @param path string
+--- @param path string
--- @return string
+--- @return string
 local function source_basename(path)
 	return path:match(BASENAME_PATTERN) or path
 end
 
--- (internal) Format a single annotation entry as one rendered line.
+--- (internal) Format a single annotation entry as one rendered line.
--- @param a AnnotEntry
+--- @param a AnnotEntry
--- @param src_name string
+--- @param src_name string
--- @return string
+--- @return string
 local function format_annot_line(a, src_name)
 	if a.error then
 		return string.format("  ✗ line %d  %s  [ERROR: %s]  [%s]", a.line, a.macro or "?", a.error, src_name)
 	end
 	local line = string.format("  ●  line %d  %s  [%s]", a.line, a.name, src_name)
-	if a.binds               then line = line .. "  binds=" .. a.binds end
+	if a.binds       then line = line .. "  binds="   .. a.binds end
-	if #a.reads > 0          then line = line .. "  reads={" .. table.concat(a.reads, ",") .. "}" end
+	if #a.reads  > 0 then line = line .. "  reads={"  .. table.concat(a.reads,  ",") .. "}" end
-	if #a.writes > 0         then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
+	if #a.writes > 0 then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
 	return line
 end
 
--- (internal) Tally totals across all results in a module.
+--- (internal) Tally totals across all results in a module.
--- @param results AnnotationResult[]
+--- @param results AnnotationResult[]
--- @return integer, integer, integer, integer, integer, integer
+--- @return integer, integer, integer, integer, integer, integer
 local function tally_module_totals(results)
 	local total_atoms, total_annots, total_binds, total_macros = 0, 0, 0, 0
 	local total_errors, total_warnings = 0, 0
@@ -315,9 +318,7 @@ local function render_module_report(dir, sources, results)
 	-- Each renderer writes its header + content via the `add` closure (pre-bound above).
 	-- Adding a new section = 1 row here + 1 render_<thing>_section function.
 	for _, section in ipairs(SECTION_RENDERERS) do
-		add(section.header)
 		section.render(add, results, totals)
-		add("")
 	end
 
 	return table.concat(lines, "\n") .. "\n"
@@ -377,21 +378,20 @@ end
 -- Orchestration helpers
 -- ════════════════════════════════════════════════════════════════════════════
 
--- (internal) Pull per-source validate() results from the annotation pass's stash.
+--- (internal) Re-validate every source in a directory against the canonical corpus projection.
--- The annotation pass runs first in the dep chain and caches results in `ctx.flags._annot_source_results`;
+--- Calls `annotation.validate()` per source to produce the per-source AnnotationResult (atoms / annots / macros / binds / errors / warnings)
--- we read from there instead of re-validating each source.
+--- that the report renderer consumes. Eeach report pass run is reproducible from the corpus.
--- Returns the list of module results + the flat list of all results (for the project-wide summary).
+--- Returns the list of module results + the flat list of all results (for the project-wide summary).
--- @param ctx PassCtx
+--- @param ctx PassCtx
--- @param dir_sources SourceFile[]
+--- @param dir_sources SourceFile[]
--- @return AnnotationResult[], AnnotationResult[]
+--- @return AnnotationResult[], AnnotationResult[]
 local function lookup_module_results(ctx, dir_sources)
-	local src_cache      = (ctx.flags and ctx.flags._annot_source_results) or {}
 	local module_results = {}
 	local all_results    = {}
 	for _, src in ipairs(dir_sources) do
-		local result = src_cache[src.path]
+		if src.scan then
-		if result then
+			local result = annotation.validate(ctx, src, nil)
-			result.source = src.path                                    -- defensive (annotation tags it too; this guards against cache misses from earlier iterations)
+			result.source = src.path                                    -- tag for downstream rendering
 			module_results[#module_results + 1] = result
 			all_results[#all_results + 1] = result
 		end
@@ -399,9 +399,9 @@ local function lookup_module_results(ctx, dir_sources)
 	return module_results, all_results
 end
 
--- (internal) Does this module's results contain anything worth emitting?
+--- (internal) Does this module's results contain anything worth emitting?
--- @param module_results AnnotationResult[]
+--- @param module_results AnnotationResult[]
--- @return boolean
+--- @return boolean
 local function module_has_content(module_results)
 	for _, r in ipairs(module_results) do
 		if #r.atoms  > 0 or #r.annots > 0 or #r.binds    > 0
@@ -412,8 +412,8 @@ local function module_has_content(module_results)
 	return false
 end
 
--- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
+--- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
--- @param fmt string
+--- @param fmt string
 local function debug_log(fmt, ...)
 	if _G[DEBUG_FLAG] then
 		io.stderr:write(string.format("[%s] " .. fmt, PASS_NAME, ...))
@@ -435,27 +435,27 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	local module_entries = (ctx.flags and ctx.flags._annot_results) or {}
+	-- Module grouping comes from `corpus.sources_by_dir` (the canonical projection).
-	local by_dir         = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	-- Iterate it directly; no private cache, no per-pass stash.
+	local corpus = ctx.shared and ctx.shared.corpus
+	local by_dir = (corpus and corpus.sources_by_dir) or {}
 
-	if not ctx.dry_run then duffle.ensure_dir(ctx.out_root) end
+	duffle.ensure_dir(ctx.out_root)
 
 	local all_results_for_summary = {}
-	for _, entry in ipairs(module_entries) do
+	for dir, dir_sources in pairs(by_dir) do
-		debug_log("entry: dir=%s basename=%s atoms_count=%d dir_sources=%d\n", entry.dir, entry.dir_basename, entry.atoms_count, #(by_dir[entry.dir] or {}))
+		local dir_basename = dir:match("([^/\\]+)$") or dir
+		debug_log("dir=%s basename=%s sources=%d\n", dir, dir_basename, #dir_sources)
 
-		if entry.atoms_count > 0 or #(by_dir[entry.dir] or {}) > 0 then
+		if #dir_sources > 0 then
-			local dir_sources                 = by_dir[entry.dir] or {}
 			local module_results, all_results = lookup_module_results(ctx, dir_sources)
 			for _, r in ipairs(all_results) do
 				all_results_for_summary[#all_results_for_summary + 1] = r
 			end
 
 			if module_has_content(module_results) then
-				local out_path = ctx.out_root .. "/" .. entry.dir_basename .. ".annotations.txt"
+				local out_path = ctx.out_root .. "/" .. dir_basename .. ".annotations.txt"
-				if not ctx.dry_run then
+				duffle.write_file(out_path, render_module_report(dir, dir_sources, module_results))
-					duffle.write_file(out_path, render_module_report(entry.dir, dir_sources, module_results))
-				end
 				outputs[#outputs + 1] = { annotations_txt = out_path }
 			else
 				debug_log("  -> no content; skipping\n")
@@ -463,7 +463,7 @@ function M.run(ctx)
 		end
 	end
 
-	if not ctx.dry_run and #all_results_for_summary > 0 then
+	if #all_results_for_summary > 0 then
 		local summary_path = ctx.out_root .. "/annotation_validation.txt"
 		duffle.write_file(summary_path, render_project_report(all_results_for_summary))
 		outputs[#outputs + 1] = { summary_txt = summary_path }

scripts/passes/word_count_eval.lua

@@ -1,11 +1,17 @@
 --- word_count_eval.lua — Word-counting logic for the tape-atom metaprogram pipeline.
 ---
---- Three responsibilities:
+--- Two responsibilities:
----   1. **Public utilities** (used by `passes/components.lua`, `passes/offsets.lua`, `passes/annotation.lua`):
+---   1. **Public utility** `M.count_token_words(token, wc)`: Used by `passes/offsets.lua`, `passes/annotation.lua`, and other passes.
----        - `M.count_token_words(token, wc)` — words emitted by one token
+---   2. **Pass entry** `M.run(ctx)`: Loads the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts` for downstream passes.
----        - `M.scan_dir(dir, suffix)` — glob walk for *.macs.h
+---      The generated `.macs.h` files are OUTPUT artifacts and are NOT inputs to this pass;
----   2. **Pass entry** `M.run(ctx)` — loads metadata.h + *.macs.h into `ctx.shared.word_counts` for downstream passes.
+---      Current component counts are owned by `passes/components.lua` (which populates `corpus.word_counts` and `corpus.component_body_index`
----   3. **Internal helpers** for the body scanner.
+---      AFTER computing each current count from the just-built body + `corpus.word_counts`).
+---
+--- **Canonical contract**:
+---   * `ctx.shared.corpus.word_counts` is the count table.
+---   * `corpus.word_counts` is the sole count table. Consumers read `corpus.word_counts` directly.
+---   * `ctx.shared.components` and `ctx.shared.component_body_index` are NOT created by this pass (projections only).
+---   * No `.macs.h` recursive discovery (no `scan_dir`, no scan cache, no `_invalidate_scan_cache`).
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
@@ -14,23 +20,11 @@
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
--- Bootstrap: see `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
 -- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-
--- ════════════════════════════════════════════════════════════════════════════
--- Constants
--- ════════════════════════════════════════════════════════════════════════════
-
--- Required native extension: lfs (LuaFileSystem). Built by `update_deps.ps1` to
--- `toolchain/lfs/lfs.dll` and wired into package.cpath by `scripts/duffle_paths.lua`.
--- If lfs is missing, `require` throws — fail loud per the build-tool convention.
-local lfs = require("lfs")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
@@ -49,12 +43,12 @@ local lfs = require("lfs")
 --- @field sources            SourceFile[]  -- all source files in the build
 --- @field metadata_path      string        -- path to word_count.metadata.h
 --- @field shared             table         -- cross-pass shared state
---- @field shared.word_counts WordCounts    -- populated by this pass
+--- @field shared.corpus      table         -- canonical corpus (required)
+--- @field shared.corpus.word_counts WordCounts -- canonical count table (populated by this pass)
 --- @field out_root           string        -- output root (e.g. "build/gen")
 --- @field project_root       string        -- project root (e.g. "code/")
 --- @field upstream           table<string, table>  -- per-pass upstream outputs
 --- @field flags              table         -- CLI flags
---- @field dry_run            boolean       -- if true, compute but don't write
 --- @field verbose            boolean       -- if true, log diagnostic info
 
 --- @class PassResult
@@ -76,8 +70,8 @@ local M = {}
 --- For most tokens (regular MIPS instructions) this returns 1.
 --- For `mac_X(...)` calls, this returns the resolved word count from `wc` (recursively if needed). For `nop2` etc., returns wc[name].
 --- For unknown macros, returns 1 and (optionally) warns.
---- @param token string        -- a single token from split_top_level_commas
+--- @param token string     -- a single token from split_top_level_commas
---- @param wc WordCounts       -- the shared word-count table
+--- @param wc    WordCounts -- the shared word-count table
 --- @return integer
 function M.count_token_words(token, wc)
 	local s = duffle.trim(token)
@@ -92,83 +86,42 @@ function M.count_token_words(token, wc)
 	return 1
 end
 
--- ┌────────────────────────────────────────────────────────────────────┐
--- │ Shared utility: scan_dir                                           │
--- └────────────────────────────────────────────────────────────────────┘
-
--- Cache the scan_dir result per (dir, suffix) in package.loaded.
--- The cache persists for the lifetime of the Lua process (cleared when ps1_meta.lua exits).
--- If a build removes/creates .macs.h files mid-process, the caller can invalidate by calling `M._invalidate_scan_cache()`.
-local SCAN_CACHE_KEY = "__word_count_eval_scan_cache__"
-
---- Scan `code/` for files matching `suffix` (e.g. `*.macs.h`).
---- Native directory enumeration via lfs (~2ms). Zero subprocess spawns.
---- @param dir string        -- project root directory
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
-function M.scan_dir(dir, suffix)
-	local key = dir .. "\0" .. suffix
-
-	local cache = package.loaded[SCAN_CACHE_KEY]
-	if    cache and cache[key] then return cache[key] end
-
-	local results = {}
-	local code_dir = dir .. "/code"
-	if lfs.attributes(code_dir, "mode") == "directory" then
-		for mod_name in lfs.dir(code_dir) do
-			if mod_name ~= "." and mod_name ~= ".." then
-				local gen_path = code_dir .. "/" .. mod_name .. "/gen"
-				if lfs.attributes(gen_path, "mode") == "directory" then
-					for fname in lfs.dir(gen_path) do
-						if fname:match("%.macs%.h$") then
-							results[#results + 1] = gen_path .. "/" .. fname
-						end
-					end
-				end
-			end
-		end
-	end
-
-	-- Cache the result (including empty results).
-	cache      = cache or {}
-	cache[key] = results
-	package.loaded[SCAN_CACHE_KEY] = cache
-
-	return results
-end
-
---- Invalidate the scan cache (call after creating new .macs.h files in the same Lua process — usually not needed).
-function M._invalidate_scan_cache() package.loaded[SCAN_CACHE_KEY] = nil end
-
 -- ┌────────────────────────────────────────────────────────────────────┐
 -- │ Pass entry: M.run(ctx) — "word-counts" pass                        │
 -- └────────────────────────────────────────────────────────────────────┘
 
---- Load metadata.h + scan for existing *.macs.h files into ctx.shared.word_counts.
+--- Load the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts`.
---- Loading the .macs.h files is idempotent: entries from later (current-build) .macs.h files override metadata.h entries of the same name.
+--- Generated `.macs.h` files are OUTPUT artifacts and are NOT scanned as inputs.
+--- Current component counts are computed and inserted by `passes/components.lua`
+--- after the components pass iterates `corpus.source_order` and writes each source's `<dir_basename>.macs.h` file.
+---
+--- Contract:
+---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).
+---   * `ctx.metadata_path` MUST be a readable file path to the authored `word_count.metadata.h`.
+---   * The pass assigns exactly one table to `corpus.word_counts`.
+---     Consumers read the corpus-owned table directly.
+---     Consumers must read `corpus.word_counts` directly.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
-	local wc = {}
+	-- 1. Canonical-corpus ownership gate.
-
+	local corpus = ctx.shared and ctx.shared.corpus
-	-- 1. Load metadata.h (the encoding-macro source of truth).
+	if type(corpus) ~= "table" then
-	local meta_counts = duffle.load_word_counts(ctx.metadata_path)
+		error("word_count_eval.run requires ctx.shared.corpus (canonical corpus). The fixture must install the corpus before running this pass.", 0)
-	for name, count in pairs(meta_counts) do wc[name] = count end
-
-	-- 2. Scan project_root recursively for *.macs.h files (component-macro source).
-	local macs_files = M.scan_dir(ctx.project_root, "*.macs.h")
-	for _, macs_path in ipairs(macs_files) do
-		local ok, mc = pcall(duffle.load_word_counts, macs_path)
-		if not ok then
-			io.stderr:write(string.format("[word_count_eval] parse error in '%s': %s\n", macs_path, tostring(mc)))
-		elseif type(mc) ~= "table" then
-			io.stderr:write(string.format("[word_count_eval] '%s' did not return a table (got %s)\n", macs_path, type(mc)))
-		else
-			for name, count in pairs(mc) do wc[name] = count end
-		end
 	end
 
-	ctx.shared.word_counts = wc
+	-- 2. metadata_path gate.
+	if type(ctx.metadata_path) ~= "string" or ctx.metadata_path == "" then
+		error("word_count_eval.run requires ctx.metadata_path (path to the authored word_count.metadata.h).", 0)
+	end
+
+	-- 3. Load authored metadata. Generated .macs.h files are NOT scanned
+	--    (the pass computes their counts from the just-built bodies after disk emission; see passes/components.lua).
+	local wc = duffle.load_word_counts(ctx.metadata_path)
+
+	-- 4. Assign the count table. ONE assignment, no copy. The assignment creates no secondary alias.
+	corpus.word_counts = wc
+
 	return { outputs = {}, errors = {}, warnings = {} }
 end
 

scripts/update_deps.ps1

@@ -39,13 +39,7 @@ pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
 # PCSX-Redux — built via MSBuild (VS2022)
-#
 # Requires: Visual Studio 2022 with the C++ desktop workload.
-# The .vcxproj files target platform toolset v145, but VS2022 ships v143;
-# we pass /p:PlatformToolset=v143 to retarget at build time (no file edits).
-# NuGet packages (glfw, luajit.native, libFFmpeg-lite, x64sentry) are
-# restored automatically by MSBuild on first build.
-#
 # Output: toolchain\pcsx-redux\vsprojects\x64\Debug\pcsx-redux.exe
 # ════════════════════════════════════════════════════════════════════════════
 
@@ -65,8 +59,7 @@ $path_pcsx_sln = join-path $path_pcsx_redux 'vsprojects\pcsx-redux.sln'
 &	$msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
 
 # Locate luajit via scoop. `luajit.exe` is on PATH via scoop's shim;
-# we use `scoop prefix` to find the install root for the include dir
+# we use `scoop prefix` to find the install root for the include dir (needed to compile lpeg against luajit's headers).
-# (needed to compile lpeg against luajit's headers).
 # If scoop or luajit is missing, fail fast with an actionable message.
 $luajit_prefix = & scoop prefix luajit 2>$null
 if (-not $luajit_prefix -or -not (Test-Path (Join-Path $luajit_prefix 'bin/luajit.exe'))) {
@@ -87,7 +80,6 @@ if (-not $lua_inc_dir) {
 # Generate lpeg.dll by compiling the 6 source files directly.
 # `gcc` is on PATH (scoop's shim puts it there).
 # The source files: lpcap.c lpcode.c lpcset.c lpprint.c lptree.c lpvm.c
-# (per the lpeg makefile — no `make.lua` template generator in this version).
 # Link against luajit's import library (`libluajit-5.1.a`) for the Lua C API symbols (lua_*, luaL_*).
 $luajit_lib_dir = Join-Path $luajit_prefix 'lib'
 $lpeg_sources = @('lpcap.c', 'lpcode.c', 'lpcset.c', 'lpprint.c', 'lptree.c', 'lpvm.c')
@@ -103,8 +95,6 @@ pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
 # lfs (LuaFileSystem) — compiled from pcsx-redux's vendored luafilesystem source.
-# Used by word_count_eval.lua :: scan_dir for native directory enumeration (~2ms)
-# instead of spawning `dir /b /s` as a subprocess (~56ms).
 # Source: toolchain/pcsx-redux/third_party/luafilesystem/src/lfs.c
 # Output: toolchain/lfs/lfs.dll
 # ════════════════════════════════════════════════════════════════════════════
@@ -118,11 +108,6 @@ $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 
 # ════════════════════════════════════════════════════════════════════════════
 # OpenBIOS — built from the PCSX-Redux source tree via make + mipsel-none-elf
-#
-# OpenBIOS is an open-source PS1 BIOS implementation (no retail BIOS dump needed). 
-# It builds with the MIPS cross-toolchain (`mipsel-none-elf-gcc`, on PATH via the `mips` toolchain installer) 
-# + `make` (on PATH via scoop).
-#
 # Output: toolchain\pcsx-redux\src\mips\openbios\openbios.bin
 # ════════════════════════════════════════════════════════════════════════════