Stuff kept from hot-reload attempt

Need to rewrite the ps1 lua metaprogram sometime soonish. Getting too bloated... need to consolidate code paths.

In this push codebase structure is starting to get a bit more realized. Decided todo now to match Pikuma's linking module files vods beginning to reorganize its codebase as well.
Atoms & atom components are not in their on *.atom.c files. (Not calling it tape.c as I don't really bake tapes like that outside of the unity c file so far...)

The lua metaprogram has had additional features added to it yet again to avoid hardcoding module handling and supporting multiple atom files per-module.
Either after the camera or cd-rom section I'll be most likely pausing to fully refactor the metaprogram. Possibly as a full re-write to get the loc minimal.

.gitignore

@@ -20,3 +20,4 @@ toolchain/lpeg
 
 scratch
 toolchain/libpsn00b
+scripts/pcsx_debug_helper.zip

.vscode/launch.json

@@ -142,6 +142,41 @@
 				"tbreak main",
 				"continue"
 			]
+		},
+		{
+			"name": "Debug: Hello Camera!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${workspaceRoot}",
+			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
+			"stopAtConnect": true,
+			"gdbpath": "gdb-multiarch",
+			"windows": {
+				"gdbpath": "gdb-multiarch.exe"
+			},
+			"osx": {
+				"gdbpath": "gdb"
+			},
+			"executable": "${workspaceRoot}/build/hello_camera.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_camera.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_camera.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"tbreak main",
+				"continue"
+			]
 		}
 	]
 }

code/duffle/dsl.atom.h

@@ -1,5 +1,5 @@
 /*
- * atom_dsl.h
+ * dsl.atom.h
  * ============================================================================
  *
  * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
@@ -57,7 +57,6 @@
 
 #ifdef INTELLISENSE_DIRECTIVES
 #pragma once
-// #include <stdint.h>
 #endif
 
 /* ============================================================================
@@ -148,12 +147,12 @@
  *   ... body ...
  *   atom_label(bounds_chk)                   ← another anchor
  *
- *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
+ *   atom_offset(culling, bounds_chk)         ← resolved by gen/offsets.h
  *
- * The metaprogram generates gen/atom_offsets.h with one #define with the offset value per atom_offset(F, T) call.
+ * The metaprogram generates gen/offsets.h with one #define with the offset value per atom_offset(F, T) call.
  * The preprocessor then expands the call to the right immediate value.
  *
- * If gen/atom_offsets.h is stale (or atom_label(name) is undefined), `atom_offset_F_T` becomes an undefined macro and the C build fails.
+ * If gen/offsets.h is stale (or atom_label(name) is undefined), `atom_offset_F_T` becomes an undefined macro and the C build fails.
  * ============================================================================*/
 #define atom_offset(F, T) atom_offset_ ## F ## _ ## T
 // atom_label is a pure annotation for the metaprogram's offset calculations.

code/duffle/dsl.h

@@ -43,7 +43,9 @@
 
 #define R_    restrict 
 #define V_    volatile 
-// Fictional, used for intiution.
+
+#pragma region Fictional //, used for intiution
+
 #define EUB_  restrict // Execute Unit Bound:    Data is siloed in the ALU Register File. The Load/Store Unit is bypassed. (Route to Execution Unit.  Keep in registers)
 #define ISO_  restrict // Isolated Provenance:   Alternative to Exu_. Guarantees electrical memory isolation, 
                        //                        unlocking the compiler’s ability to safely pack data across multiple parallel SIMD lanes (vectorization).
@@ -67,7 +69,8 @@
 #define latch_load_anchor(ptr)       //__atomic_load_n(ptr, ooo_anchor_)
 #define latch_store_drain(ptr, val)  //__atomic_store_n(ptr, val, ooo_drain_)
 #define pulse_xchg_weld(ptr, val)    //__atomic_exchange_n(ptr, val, ooo_weld_)
-//end of: Fictional.
+
+#pragma endreigon Fictional
 
 
 // R_ (restrict) establishes an "Eigen" or "Proprius" mapping.

code/duffle/gen/duffle.macs.h

@@ -1,134 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#pragma once
-#endif
-// Auto-generated by ps1_meta.lua — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
-// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
-
-#ifndef WORD_COUNT
-#define WORD_COUNT(name, count)  enum { words_##name = (count) };
-#endif
-
-/* 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)) \
-,	jump_reg( R_AtomJmp) \
-,	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)) \
-,	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)) \
-,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
-WORD_COUNT(mac_load_tri_indices, 3)
-
-/* 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) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY0) \
-,	gte_mv_to_data_r(R_V1, C2_VZ0) \
-,	shift_lleft(R_AT, R_T1, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY1) \
-,	gte_mv_to_data_r(R_V1, C2_VZ1) \
-,	shift_lleft(R_AT, R_T2, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY2) \
-,	gte_mv_to_data_r(R_V1, C2_VZ2)
-WORD_COUNT(mac_gte_load_tri_verts, 18)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-#define mac_insert_ot_tag_f3(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (5 - 1) << 24 = 4 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_f3, 11)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-#define mac_insert_ot_tag_g4(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (9 - 1) << 24 = 8 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_g4, 11)
-
-/* 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(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
-WORD_COUNT(mac_gte_store_f3, 3)
-
-#define mac_format_g4_color(r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
-,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
-,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
-,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
-WORD_COUNT(mac_format_g4_color, 12)
-
-/* 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). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
- * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
-#define mac_gte_store_g4_p012(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
-WORD_COUNT(mac_gte_store_g4_p012, 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 earlier RTPT.
- */
-#define mac_gte_store_g4_p3(...) \
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
-WORD_COUNT(mac_gte_store_g4_p3, 1)
-

code/duffle/gen/duffle.offsets.h

@@ -1,9 +0,0 @@
-// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
-#pragma once
-
-#pragma region lottes_tape
-
-
-#pragma endregion lottes_tape
-

code/duffle/gen/macs.h

@@ -0,0 +1,184 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\duffle/
+//   source: C:\projects\Pikuma\ps1\code\duffle\word_count.metadata.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\memory.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gcc_asm.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.atom.c
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#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).
+//   - mac_yield() is the safe default for atom-endings: 4 words, BD-slot of jr is mandatory nop.
+//   - mac_yield_load() + mac_yield_tail():
+//     - unconditional branch: mac_yield_load fills the branch's BD-slot (replaces a nop);
+//     - mac_yield_tail runs at the branch target (does NOT re-load R_AtomJmp).
+#define mac_yield(...) \
+	load_word(R_AtomJmp, R_TapePtr, 0) \
+,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
+,	jump_reg( R_AtomJmp) \
+,	nop
+WORD_COUNT(mac_yield, 4)
+
+/* atom_dbg_skip */
+#define mac_yield_load(...) \
+	load_word(R_AtomJmp, R_TapePtr, 0)
+WORD_COUNT(mac_yield_load, 1)
+
+/* atom_dbg_skip */
+#define mac_yield_tail(...) \
+	add_ui_self(R_TapePtr, S_(MipsCode)) \
+,	jump_reg(  R_AtomJmp) \
+,	nop
+WORD_COUNT(mac_yield_tail, 3)
+
+/* atom_dbg_skip */
+#define mac_load_v2s2(rs_x, rs_y, r_base, offset) \
+	load_half( rs_x, r_base, O_(V3_S2,x)) \
+,	load_half( rs_y, r_base, O_(V3_S2,y))
+WORD_COUNT(mac_load_v2s2, 2)
+
+/* atom_dbg_skip */
+#define mac_store_v2s2(rt_x, rt_y, base, offset) \
+	store_half(rt_x, base, offset + O_(V2_S2,x)) \
+,	store_half(rt_y, base, offset + O_(V2_S2,y))
+WORD_COUNT(mac_store_v2s2, 2)
+
+/* atom_dbg_skip */
+#define mac_store_rects2(rt_x, rt_y, rt_width, rt_height, base, offset) \
+	store_half(rt_x,      base, offset + O_(Rect_S2,x)) \
+,	store_half(rt_y,      base, offset + O_(Rect_S2,y)) \
+,	store_half(rt_width,  base, offset + O_(Rect_S2,width)) \
+,	store_half(rt_height, base, offset + O_(Rect_S2,height))
+WORD_COUNT(mac_store_rects2, 4)
+
+/* atom_dbg_skip */
+#define mac_load_tri_indices(r_face_cusor, r_i0, r_i1, r_i2) \
+	load_half_u(r_i0, r_face_cusor, 0 * S_(S2)) \
+,	load_half_u(r_i1, r_face_cusor, 1 * S_(S2)) \
+,	load_half_u(r_i2, r_face_cusor, 2 * S_(S2))
+WORD_COUNT(mac_load_tri_indices, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_store_f3(r_primitive_cursor) \
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_F3,p0)) \
+,	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_F3,p1)) \
+,	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_F3,p2))
+WORD_COUNT(mac_gte_store_f3, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_load_tri_verts(r_vert_base, r_v0, r_v1, r_v2) \
+	shift_lleft(R_AT, r_v0, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	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_v1, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	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_v2, v3s2_byteoff) \
+,	add_u_self(R_AT, r_vert_base) \
+,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
+,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
+,	gte_mv_to_data_r(R_V0, C2_VXY2) \
+,	gte_mv_to_data_r(R_V1, C2_VZ2)
+WORD_COUNT(mac_gte_load_tri_verts, 18)
+
+/* atom_dbg_skip */
+#define mac_gte_store_g4_p012(r_primitive_cursor) \
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_G4,p0)) \
+,	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_G4,p1)) \
+,	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p2))
+WORD_COUNT(mac_gte_store_g4_p012, 3)
+
+/* atom_dbg_skip */
+#define mac_gte_store_g4_p3(r_primitive_cursor) \
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3))
+WORD_COUNT(mac_gte_store_g4_p3, 1)
+
+#define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
+	load_upper_i(reg_transfer, cmd >> 16) \
+,	or_i_self(   reg_transfer, cmd & 0xFFFF) \
+,	store_word(  reg_transfer, reg_base, port)
+WORD_COUNT(mac_gcmd_push, 3)
+
+/* atom_dbg_skip */
+#define mac_store_rgb8(rr, rg, rb, base, offset) \
+	store_byte(rr, base, offset + O_(RGB8,r)) \
+,	store_byte(rg, base, offset + O_(RGB8,g)) \
+,	store_byte(rb, base, offset + O_(RGB8,b))
+WORD_COUNT(mac_store_rgb8, 3)
+
+/* atom_dbg_skip */
+#define mac_pack_color_word(r_base, 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_base, (off))
+WORD_COUNT(mac_pack_color_word, 3)
+
+/* atom_dbg_skip */
+#define mac_format_f3_color(r_base, r, g, b) \
+	mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
+WORD_COUNT(mac_format_f3_color, 3)
+
+#define mac_format_g4_color(r_prim_cursor, r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c1), 0,               r1,g1,b1) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c2), 0,               r2,g2,b2) \
+,	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c3), 0,               r3,g3,b3)
+WORD_COUNT(mac_format_g4_color, 12)
+
+#define mac_insert_ot_tag_f3(r_ot_base, r_prim_cursor) \
+	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
+,	add_u_self(  R_T1, r_ot_base)                             /* T1 = & OrderingTable[OTZ] */ \
+,	load_word(   R_AT, R_T1,          O_(PolyTag,code))       /* AT = old_ot_head */ \
+,	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (5 - 1) << 24 = 4 << 24 */ \
+,	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits))   /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
+,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
+,	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code))       /* prim->tag = packed(prim_length, old_addr) */ \
+,	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits))   /* AT = (prim_length << 24) | old_addr */ \
+,	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)) \
+,	store_word(  R_AT, R_T1,          O_(PolyTag,code))       /* OrderingTable[OTZ] = PrimCursor */
+WORD_COUNT(mac_insert_ot_tag_f3, 11)
+
+#define mac_insert_ot_tag_g4(r_ot_base, r_prim_cursor) \
+	shift_lleft( R_T1, R_T1, S_(U4)/2)                         /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
+,	add_u_self(  R_T1, r_ot_base)                              /* T1 = & OrderingTable[OTZ] */ \
+,	load_word(   R_AT, R_T1,          O_(PolyTag,code))        /* AT = old_ot_head */ \
+,	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (9 - 1) << 24 = 8 << 24 */ \
+,	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
+,	or_u(        R_AT, R_AT, R_V0)                             /* Merge length */ \
+,	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
+,	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
+,	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)) \
+,	store_word(  R_AT, R_T1,          O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
+WORD_COUNT(mac_insert_ot_tag_g4, 11)
+

code/duffle/gen/offsets.h

@@ -0,0 +1,53 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\duffle\
+//   source: C:\projects\Pikuma\ps1\code\duffle\word_count.metadata.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\memory.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gcc_asm.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\dsl.atom.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\lottes_tape.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\math.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\mips.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gte.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\gp.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.atom.c
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.atom.c
+#pragma once
+
+#pragma region duffle
+
+
+// --- atom: pad_bios_snapshot (78 words) ---
+
+#define _atom_offset_snap_root_skip_disconnected      8
+#define _atom_offset_disconnected_snap_end            61
+#define _atom_offset_case_2_id_dispatch               8
+#define _atom_offset_pending_snap_end                 51
+#define _atom_offset_id_dispatch_try_analog_stick     11
+#define _atom_offset_id_dispatch_snap_end             38
+#define _atom_offset_try_analog_stick_try_analog_pad  12
+#define _atom_offset_analog_stick_snap_end            24
+#define _atom_offset_try_analog_pad_try_unsupported   11
+#define _atom_offset_analog_pad_snap_end              10
+
+enum {
+    atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
+    atom_offset_disconnected_snap_end = _atom_offset_disconnected_snap_end,
+    atom_offset_case_2_id_dispatch = _atom_offset_case_2_id_dispatch,
+    atom_offset_pending_snap_end = _atom_offset_pending_snap_end,
+    atom_offset_id_dispatch_try_analog_stick = _atom_offset_id_dispatch_try_analog_stick,
+    atom_offset_id_dispatch_snap_end = _atom_offset_id_dispatch_snap_end,
+    atom_offset_try_analog_stick_try_analog_pad = _atom_offset_try_analog_stick_try_analog_pad,
+    atom_offset_analog_stick_snap_end = _atom_offset_analog_stick_snap_end,
+    atom_offset_try_analog_pad_try_unsupported = _atom_offset_try_analog_pad_try_unsupported,
+    atom_offset_analog_pad_snap_end = _atom_offset_analog_pad_snap_end,
+};
+
+#pragma endregion duffle
+

code/duffle/gp.atom.c

@@ -0,0 +1,82 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "dsl.h"
+#	include "gp.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(gp_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_gcmd_push(U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_gcmd_push, {
+	load_upper_i(reg_transfer, cmd >> 16),
+	or_i_self(   reg_transfer, cmd & 0xFFFF),
+	store_word(  reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_store_rgb8(U1 rr, U1 rg, U1 rb, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rgb8, {
+	store_byte(rr, base, offset + O_(RGB8,r)),
+	store_byte(rg, base, offset + O_(RGB8,g)),
+	store_byte(rb, base, offset + O_(RGB8,b)),	
+})
+
+/* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
+ * byte offset. Internal helper used by the *_format_*_color macros. */
+FI_ Slice_MipsCode ac_pack_color_word(U4 r_base, U4 off, U4 cmd, U1 r, U1 g, U1 b)
+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_base, (off)),
+})
+
+/* Words: 3; Emits the F3 command+color word (cmd byte | BLUE | GREEN | RED)
+ * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields). */
+FI_ Slice_MipsCode ac_format_f3_color(U4 r_base, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+
+/* Words: 12; Emits the four (code|color) words of a Poly_G4.
+ * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
+FI_ Slice_MipsCode ac_format_g4_color(U4 r_prim_cursor,
+	U1 r0, U1 g0, U1 b0,
+  U1 r1, U1 g1, U1 b1,
+  U1 r2, U1 g2, U1 b2,
+  U1 r3, U1 g3, U1 b3)
+MipsAtomComp_Proc_(ac_format_g4_color, {
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c1), 0,               r1,g1,b1),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c2), 0,               r2,g2,b2),
+	mac_pack_color_word(r_prim_cursor, O_(Poly_G4,c3), 0,               r3,g3,b3),
+})
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
+I_ Slice_MipsCode ac_insert_ot_tag_f3(U4 r_ot_base, U4 r_prim_cursor) MipsAtomComp_Proc_(ac_insert_ot_tag_f3, {
+	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
+	add_u_self(  R_T1, r_ot_base),                             // T1 = & OrderingTable[OTZ]
+	load_word(   R_AT, R_T1,          O_(PolyTag,code)),       // AT = old_ot_head
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (5 - 1) << 24 = 4 << 24
+	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits)),   // Strip upper 8 bits (length from prev cell) → keep only low 24
+	or_u(        R_AT, R_AT, R_V0),                            // Merge length
+	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code)),       // prim->tag = packed(prim_length, old_addr)
+	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits)),   // AT = (prim_length << 24) | old_addr
+	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)),
+	store_word(  R_AT, R_T1,          O_(PolyTag,code)),       // OrderingTable[OTZ] = PrimCursor
+})
+
+/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
+ * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
+I_ Slice_MipsCode ac_insert_ot_tag_g4(U4 r_ot_base, U4 r_prim_cursor) MipsAtomComp_Proc_(ac_insert_ot_tag_g4, {
+	shift_lleft( R_T1, R_T1, S_(U4)/2),                         // T1 = otz * S_(U4) (otz arg is implicit R_T1)
+	add_u_self(  R_T1, r_ot_base),                              // T1 = & OrderingTable[OTZ]
+	load_word(   R_AT, R_T1,          O_(PolyTag,code)),        // AT = old_ot_head
+	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (9 - 1) << 24 = 8 << 24
+	mask_upper(  R_AT, R_AT,          S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
+	or_u(        R_AT, R_AT, R_V0),                             // Merge length
+	store_word(  R_AT, r_prim_cursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
+	shift_lleft( R_AT, r_prim_cursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
+	shift_lright(R_AT, R_AT,          S_(PolyTag_len_bits)),
+	store_word(  R_AT, R_T1,          O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
+})
+
+#pragma endregion MACs (Mips Atom Components)

code/duffle/gte.atom.c

@@ -0,0 +1,76 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "gte.h"
+#	include "gp.h"
+#	include "lottes_tape.h"	
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(gte_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+/* Words: 3; Loads 3 S2 indices from the face array */
+FI_ Slice_MipsCode ac_load_tri_indices(U4 r_face_cusor, U4 r_i0, U4 r_i1, U4 r_i2) atom_dbg_skip MipsAtomComp_Proc_(ac_load_tri_indices, {
+	load_half_u(r_i0, r_face_cusor, 0 * S_(S2)),
+	load_half_u(r_i1, r_face_cusor, 1 * S_(S2)),
+	load_half_u(r_i2, r_face_cusor, 2 * S_(S2)),
+})
+
+/* 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). */
+FI_ Slice_MipsCode ac_gte_store_f3(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_f3, {
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_F3,p0)),
+	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_F3,p1)),
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_F3,p2)),
+})
+
+/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
+I_ Slice_MipsCode ac_gte_load_tri_verts(U4 r_vert_base, U4 r_v0, U4 r_v1, U4 r_v2) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_load_tri_verts, {
+	shift_lleft(R_AT, r_v0, v3s2_byteoff), add_u_self(R_AT, r_vert_base), 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_v1, v3s2_byteoff), add_u_self(R_AT, r_vert_base), 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_v2, v3s2_byteoff), add_u_self(R_AT, r_vert_base), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+})
+
+/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
+ * G4 triangle portion to p0/p1/p2.
+ * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
+ * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
+ * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
+FI_ Slice_MipsCode ac_gte_store_g4_p012(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_g4_p012, {
+	gte_sw(C2_SXY0, r_primitive_cursor, O_(Poly_G4,p0)),
+	gte_sw(C2_SXY1, r_primitive_cursor, O_(Poly_G4,p1)),
+	gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p2)),
+})
+
+/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
+ * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its single-vertex result to SXY2;
+ * SXY0 still holds v0.screen from the earlier RTPT.
+ */
+FI_ Slice_MipsCode ac_gte_store_g4_p3(U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ac_gte_store_g4_p3, { gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3)) })
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Bsked Atoms
+
+typedef Struct_(Binds_SetGteWorld) {
+	M3_S2* transform;
+};
+internal MipsAtom_(set_gte_world) atom_info(
+		atom_bind(Binds_SetGteWorld)
+	,	atom_reads(R_TapePtr)
+){
+	/* Pop matrix address from tape into R_T3 ($11) */
+	load_word(R_T3, R_TapePtr, O_(Binds_SetGteWorld,transform)),
+	add_ui_self(    R_TapePtr, S_(Binds_SetGteWorld)),
+	/* Load 3x3 Rotation + 3x1 Translation from R_T3 into GTE CONTROL Regs (ctc2) */
+	load_word(R_T0, R_T3, 0),  load_word(R_T1, R_T3,  4),
+	gte_mv_to_ctrl_r(R_T0, gte_cr_RT11), gte_mv_to_ctrl_r(R_T1, gte_cr_RT12),
+	load_word(R_T0, R_T3, 8),  load_word(R_T1, R_T3, 12), load_word(R_T2, R_T3, 16),
+	gte_mv_to_ctrl_r(R_T0, gte_cr_RT13), gte_mv_to_ctrl_r(R_T1, gte_cr_RT21), gte_mv_to_ctrl_r(R_T2, gte_cr_RT22),
+	load_word(R_T0, R_T3, 20), load_word(R_T1, R_T3, 24), load_word(R_T2, R_T3, 28),
+	gte_mv_to_ctrl_r(R_T0, gte_cr_TRX),  gte_mv_to_ctrl_r(R_T1, gte_cr_TRY),  gte_mv_to_ctrl_r(R_T2, gte_cr_TRZ),
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/duffle/lottes_tape.h

@@ -1,36 +1,83 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
 #	include "dsl.h"
 #	include "gcc_asm.h"
 # include "mips.h"
 # include "gte.h"
 # include "memory.h"
-#	include "atom_dsl.h"
+#	include "dsl.atom.h"
-#	include "gen/duffle.macs.h"
-#	include "gen/duffle.offsets.h"
 #endif
 
-typedef U4 const MipsCode; // Underlying type to mips asm words.
+#pragma region Tape Drive
-typedef Slice_(MipsCode);
+/* -----------------------------------------------------------------------------
-
+ *  TAPE DRIVE ABI
-typedef U4 const MipsAtom; // Underlying type to an array of mips asm words that must terminate with an ac_yield.
+ * -----------------------------------------------------------------------------
-#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+ * Note(Ed): One of the main purposes of this codebase is to help me
-
+ * learn this, as such the information below may be entirely realized 
-// Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
+ * or finalized conceptually.
-//   MipsAtomComp_(ac_X) { body }
+ * -----------------------------------------------------------------------------
-// expands to:
+ * This ABI and its associated legos were directly inspired by researching
-//   MipsCode ac_X[] align_(4) = { body };
+ * the work of Timothy Lottes and Onat Türkçüoğlu; along with many others.
-#define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+ * It's the simplest bootstrap of a a directly executed chain of assemby
-
+ * arrays (Atoms) that terminate with a yield sequence to the next atom.
-// Used for components with value-args (e.g., ac_format_f3_color).
+ * These eventually lead to a terminal atom for the tape which is defined
-//   FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+ * below as "tape_exit".
-// expands to:
+ * 
-//   FI_ Slice_MipsCode ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+ * This behaves as one of the simplest runtime harnesses ontop of a 
-#define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
+ * host-enviornment's execution engine to author and compose programs with.
-
+ * From here various conventions can be further applied.
-// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
+ * To make things easier to understand it may be better to focus on what this
-
+ * ABI does not have. It does not have have any branching within the tape but
-/* Register aliases */
+ * relative branches within atoms or between atoms.
+ * Branching nearly is always downstream. Stack usage is non-existent.
+ * Push/Pop, FIFO, or Arena/Bump data structures are used by atoms explicitly.
+ * In it's current form withe C11 macro dsl, the user also has to do manual register 
+ * allocation per atom.
+ * 
+ * One of the remarkable things about utilizing this abi is its essentially
+ * interopable with CPUs, GPUs, FPGA, or, basically anything
+ * from the 5th generation consoles and onward.
+ * The ABI directly reflects how all computational hardware must be architected
+ * in order to execute digital logic effectively on current era tech.
+ * On the PS1 we don't have access to a few features like multi-threading,
+ * speculative execution, or L3 cache; but, we can set the foundation for legoing
+ * whats required for eventually expanding this ABI's paradigm and core atoms
+ * to take those newer hardware features into account. For example, you can easily
+ * expand this to support wave-based execution model on a PS2 or PS3.
+ * Not having a stack or automatic register allocation means the user can't ignore
+ * excessive argument shuffle across workload or waves and thier phases.
+ * Crossing ABI boundaries to other runtimes that do has obviouss penalties.
+ * 
+ * Learning data-oreinted code becomes a natural progression. Your not fighting
+ * a stack-based procedural paradigm that wants to argument shuffle on by lack of 
+ * constraints on how the user may "call" a procedure. The user doesn't
+ * have to hammer down "rules" or patterns to know how to massage the compiler
+ * to dissolve those call frames to get the asesmbly into its desired form.
+ * The form is obvious, and once the user gets to author these compoonents
+ * it becomes a game of tetris.
+ * 
+ * Another feature is this ABI is very compatible with bootstrapping and developing
+ * simple toolchains built off of bit-packed annotated command streams the user can
+ * directly author, maintatain, and immediately execute. That being like a color forth.
+ * This can make the tetris less of a chore with some helpful policy generation for
+ * allocation of registers, helping to choose resuable components, designing DSL on
+ * the fly, etc.
+ * -----------------------------------------------------------------------------
+ * TODO(Ed): We need pretty ascii diagrams and proper guides, articles, etc.
+ * -----------------------------------------------------------------------------
+ * For now this ideation is just started functioning. I'm abusing C11 & a lua metaprogram
+ * to help establish a hybrid toolchain to ideate on a traditional text-based
+ * authoring UX for this paradigm.
+ * If pcsx-redux provides viable hot-reload and persistent data storage beyond
+ * save-states (just copying ram to filesystem), I can author a color forth to
+ * mess around with. With either an editor in-emulator or on the actual machine itself.
+ * Assembly is tedius, but I think this codebase most likely has a pretty ergonomic 
+ * flavor worst case...
+ * */
+/* Register Allocation Info */
 enum {
 	R_AtomJmp  = R_T8 atom_reg,  /* debug-visible; tape yield handshake scratch */
 	R_TapePtr  = R_T9 atom_reg,  /* The Instruction Stream Pointer */
@@ -59,28 +106,53 @@ enum {
 	R_TScratch6  = R_T6,
 	R_TScratch7  = R_T7,
 	R_TScratch8 =  R_T8,
-	R_TScratch10 = R_V0,
+	R_TScratch10 = R_V0, // Tend to be used with gte DMAs
-	R_TScratch11 = R_V1,
+	R_TScratch11 = R_V1, // Tend to be used with gte DMAs
 	// Note(Ed): We can technically clobber these, but don't unless we hit a bottleneck.
-	// R_TScratch12 = R_A0,
+	// A 0-2
-	// R_TScratch13 = R_A1,
+	// S 0-7
-	// R_TScratch14 = R_A3,
-	// TODO(Ed): Review S0-S7, they are technically avaialble, we just have to snapshot them at the ABI boundary.
-	// TODO(Ed): This is technically a waste of cycles for most work? so maybe only do this for expensive atoms on-demand or atom phases.
-// TODO(Ed): Sort out the other available registers... (Not sure how much is left avail)
 };
 
-#pragma region Tape Drive
+typedef U4 const MipsCode; // Underlying type to mips asm words.
-/* ---------------------------------------------------------------------------
+typedef Slice_(MipsCode);
- *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
+
- * ---------------------------------------------------------------------------*/
+typedef U4 const MipsAtom; // Underlying type to an array of mips asm words that must terminate with an ac_yield.
+#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+
+// Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
+//   MipsAtomComp_(ac_X) { body }
+// expands to:
+//   MipsCode ac_X[] align_(4) = { body };
+#define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+
+// Used for components with value-args (e.g., ac_format_f3_color).
+//   FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+// expands to:
+//   FI_ Slice_MipsCode ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+#define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
+
+/* Line-table anchor: gcc only adds a file to the .debug_line file table when the
+   file contains line-numbered content. Files containing only:
+     - `MipsAtomComp_` static-array declarations, or
+     - `MipsAtomComp_Proc_` (force-inline) function bodies whose line info gets
+       attributed to the call site at the include point are otherwise omitted from the file table,
+			 which breaks the DWARF injection when it tries to resolve atom-component provenance paths.
+
+   Place `ATOM_FILE_LINE_MARKER();` once at file scope in any `.atom.c` that defines atoms.
+	 The macro expands to a file-scope `internal U4 const` declaration keeps the file in the line table.
+	 The constant is in `.rodata` and unreferenced; the linker may eliminate it.
+	 The two-level concat + `__LINE__` suffix makes the identifier unique per call site
+	 (the identifier embeds the source line, so duplicates across `#include`d files don't collide). */
+#define ATOM_FILE_DEBUGGER_LINE_MARKER(file_name) internal U4 const tmpl(atom_file_debugger_line_marker,file_name) = 0
+
  typedef Slice_(MipsAtom); typedef Slice_MipsAtom Tape;
 
 /* The 'Exit' Atom */
 atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
-//TODO(Ed): Do we backup R_S0-7 here? Have it in a heavier tape run as a opt-in? Same with V0-1 and A0-3?
+// TODO(Ed): When we have a substantial workload/throughput, profile each of these to see impact at ABI boundaries.
-/* Generalized Tape Engine Runner */
+
+/* Tape Runner (Default) */
 FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
 		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
@@ -91,12 +163,32 @@ FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape
 	asm_rpins, r_use(tape_ptr)
 	asm_clobber: 
 		rlit(R_AT),
-		rlit(R_V0), rlit(R_V1),
+		rlit(R_V0), rlit(R_V1), // We clobber these for GTE ACs (that don't expose register selection, might expose them in the future...)
 		rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
 		rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8),
 		clb_mem_drain 
 ); }
 
+/* Tape Runner (Static and Arg Clobbers) */
+FI_ void tape_run_a02_s07(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
+	asm_words(
+		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
+		, add_ui_self(R_TapePtr, S_(MipsAtom)) /* Advance tape */
+		, call_reg(   R_AtomJmp)               /* jalr $t9 */
+		, nop                                  /* Branch delay slot */
+	)
+	asm_rpins, r_use(tape_ptr)
+	asm_clobber: 
+		rlit(R_AT),
+		rlit(R_V0), rlit(R_V1), rlit(R_A0), rlit(R_A1), rlit(R_A2),
+		rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+		rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8),
+		rlit(R_S0), rlit(R_S1), rlit(R_S2), rlit(R_S3), rlit(R_S4),
+		rlit(R_S5), rlit(R_S6), rlit(R_S7),
+		clb_mem_drain 
+); }
+
+// Procedural authoring of tapes:
 typedef Relative_(FArena) Struct_(TapeBuilder) { U4 ptr; U4 capacity; U4 used; };
 FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start; tb->used = 0; }
 FI_ TapeBuilder tb_make_old(             FArena* arena) { return (TapeBuilder){ arena->start, 0 }; }
@@ -113,7 +205,6 @@ FI_ Tape tb_slice(TapeBuilder  tb) {                        return (Tape){ C_(U4
 
 FI_ void tb_scope_run_end(TapeBuilder* tb) { tb_emit(tb,tape_exit); tape_run(tb_slice(tb[0])); }
 #define tb_scope_run(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_scope_run_end(tb))
-
 #pragma endregion Tape Drive
 
 #pragma region Macro Mips Atom Components
@@ -123,124 +214,30 @@ FI_ void tb_scope_run_end(TapeBuilder* tb) { tb_emit(tb,tape_exit); tape_run(tb_
  * ---------------------------------------------------------------------------*/
 
 // The 'Yield' sequence for Tape Atoms (mac_yield).
+//   - mac_yield() is the safe default for atom-endings: 4 words, BD-slot of jr is mandatory nop.
+//   - mac_yield_load() + mac_yield_tail():
+//     - unconditional branch: mac_yield_load fills the branch's BD-slot (replaces a nop);
+//     - mac_yield_tail runs at the branch target (does NOT re-load R_AtomJmp).
+
 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,
 };
 
-enum {
+atom_dbg_skip MipsAtomComp_(ac_yield_load) {
-	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	load_word(R_AtomJmp, R_TapePtr, 0),
-	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
-	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
-	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
-#define R_PrimCursor_Code  R_T7_Code
-#define R_FaceCursor_Code  R_T4_Code
-#define R_VertBase_Code    R_T5_Code
-#define R_OtBase_Code      R_T6_Code
 };
 
-/* Words: 3; Loads 3 S2 indices from the face array */
+atom_dbg_skip MipsAtomComp_(ac_yield_tail) {
-atom_dbg_skip MipsAtomComp_(ac_load_tri_indices) {
+	add_ui_self(R_TapePtr, S_(MipsCode)),
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	jump_reg(  R_AtomJmp), nop,
-	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 MipsAtomComp_(ac_gte_load_tri_verts) {
-	shift_lleft(R_AT, R_T0, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-	shift_lleft(R_AT, R_T1, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-	shift_lleft(R_AT, R_T2, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
-};
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-MipsAtomComp_(ac_insert_ot_tag_f3) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (5 - 1) << 24 = 4 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-MipsAtomComp_(ac_insert_ot_tag_g4) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (9 - 1) << 24 = 8 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
- * byte offset. Internal helper used by the *_format_*_color macros. */
-FI_ Slice_MipsCode ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
-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)),
-})
-
-/* 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_ Slice_MipsCode ac_format_f3_color(U1 r, U1 g, U1 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). */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_f3) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
-};
-
-/* Words: 12; Emits the four (code|color) words of a Poly_G4.
- * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-FI_ Slice_MipsCode ac_format_g4_color(
-	U1 r0, U1 g0, U1 b0,
-  U1 r1, U1 g1, U1 b1,
-  U1 r2, U1 g2, U1 b2,
-  U1 r3, U1 g3, U1 b3)
-MipsAtomComp_Proc_(ac_format_g4_color, {
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
-	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1),
-	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2),
-	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3),
-})
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
- * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p012) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
-};
-
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its single-vertex result to SXY2;
- * SXY0 still holds v0.screen from the earlier RTPT.
- */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
-
 #pragma endregion Macro Atom Components
 
 #pragma region Mips Atom Builder
-// This allows for runtime procedural authoring of mips atoms.
+// This helps with runtime procedural authoring of mips atoms.
 
 typedef Struct_(FMipsAtom512) { U4 data[512]; U4 used; };
 
@@ -263,57 +260,9 @@ FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
 }
 
 #define mipsatom_from_builder(ab) (Slice_MipsCode){ab.start, ab.used}
-
 #pragma endregion Mips Atom Builder
 
 #pragma region Baked Mips Atoms
 // These atoms are resolved at compile time and are (usually) statically linked readonly data.
 
-enum {
-	bios_flushcache = 0x44,
-	bios_table_addr = 0xA0,
-};
-
-/* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
- * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
- *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
- *   2. $a0 = bios_flushcache (arg0)
- *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
- *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
- *      nop                             ; branch delay slot
- *   5. lw $ra, 4($sp);  jr $ra         ; restore & return
- *   6. sp += 8
- */
-internal MipsAtom_(mips_flush_icache) {
-	add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment), // sp -= 8
-	store_word(rret_addr, rstack_ptr, S_(U4)),           // sw  $ra,   4($sp) 
-	add_ui(rret_0, rdiscard, bios_flushcache),           // addiu $a0, $0, 0x44 
-	add_ui(rtmp_0, rdiscard, bios_table_addr),           // addiu $t0, $0, 0xA0 
-	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
-	load_word(rret_addr, rstack_ptr, S_(U4)),            // lw   $ra, 4($sp) 
-	jump_reg(rret_addr),                                 // jr   $ra 
-	add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment),  // sp += 8 (BD) 
-	mac_yield(),
-};
-
-typedef Struct_(Binds_SetGteWorld) {
-	M3_S2* transform;
-};
-internal MipsAtom_(set_gte_world) atom_info(
-		atom_bind(Binds_SetGteWorld)
-	,	atom_reads(R_TapePtr)
-){
-	/* Pop matrix address from tape into R_T3 ($11) */
-	load_word(R_T3, R_TapePtr, O_(Binds_SetGteWorld,transform)),
-	add_ui_self(    R_TapePtr, S_(Binds_SetGteWorld)),
-	/* Load 3x3 Rotation + 3x1 Translation from R_T3 into GTE CONTROL Regs (ctc2) */
-	load_word(R_T0, R_T3, 0),  load_word(R_T1, R_T3,  4),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT11), gte_mv_to_ctrl_r(R_T1, gte_cr_RT12),
-	load_word(R_T0, R_T3, 8),  load_word(R_T1, R_T3, 12), load_word(R_T2, R_T3, 16),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT13), gte_mv_to_ctrl_r(R_T1, gte_cr_RT21), gte_mv_to_ctrl_r(R_T2, gte_cr_RT22),
-	load_word(R_T0, R_T3, 20), load_word(R_T1, R_T3, 24), load_word(R_T2, R_T3, 28),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_TRX),  gte_mv_to_ctrl_r(R_T1, gte_cr_TRY),  gte_mv_to_ctrl_r(R_T2, gte_cr_TRZ),
-	mac_yield()
-};
-
 #pragma endregion Baked Mips Atoms

code/duffle/math.atom.c

@@ -0,0 +1,29 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "math.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(math_atom_c);
+
+#pragma region MACs (Mips Atom Component)
+
+FI_ Slice_MipsCode ac_load_v2s2(U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_load_v2s2, {
+	load_half( rs_x, r_base, O_(V3_S2,x)),
+	load_half( rs_y, r_base, O_(V3_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_v2s2(U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_v2s2, {
+	store_half(rt_x, base, offset + O_(V2_S2,x)),
+	store_half(rt_y, base, offset + O_(V2_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_rects2(U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rects2, {
+	store_half(rt_x,      base, offset + O_(Rect_S2,x)),
+	store_half(rt_y,      base, offset + O_(Rect_S2,y)),
+	store_half(rt_width,  base, offset + O_(Rect_S2,width)),
+	store_half(rt_height, base, offset + O_(Rect_S2,height)),
+})
+
+#pragma endregion MACs (Mips Atom Component)

code/duffle/math.h

@@ -61,10 +61,5 @@ FI_ void add_a3s4_fp(A3_S4_R out_a, A3_S4 b) {
 	(out_a[0])[2] += b[2] >> 1;
 }
 
-FI_ void add_v3s4(V3_S4_R out_a, V3_S4 b) { 
+FI_ void add_v3s4   (V3_S4_R out_a, V3_S4 b) {  add_a3s4   (pcast(A3_S4_R, out_a), pcast(A3_S4, b));  }
-	add_a3s4(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) {  add_a3s4_fp(pcast(A3_S4_R, out_a), pcast(A3_S4, b));  }
-}
-
-FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) { 
-	add_a3s4_fp(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
-}

code/duffle/mips.atom.c

@@ -0,0 +1,38 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(mips_atom_c);
+
+#pragma region Baked Atoms
+
+enum {
+	bios_flushcache = 0x44,
+	bios_table_addr = 0xA0,
+};
+
+/* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
+ * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
+ *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
+ *   2. $a0 = bios_flushcache (arg0)
+ *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
+ *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
+ *      nop                             ; branch delay slot
+ *   5. lw $ra, 4($sp);  jr $ra         ; restore & return
+ *   6. sp += 8
+ */
+internal MipsAtom_(mips_flush_icache) {
+	add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment), // sp -= 8
+	store_word(rret_addr, rstack_ptr, S_(U4)),           // sw  $ra,   4($sp) 
+	add_ui(rret_0, rdiscard, bios_flushcache),           // addiu $a0, $0, 0x44 
+	add_ui(rtmp_0, rdiscard, bios_table_addr),           // addiu $t0, $0, 0xA0 
+	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
+	load_word(rret_addr, rstack_ptr, S_(U4)),            // lw   $ra, 4($sp) 
+	jump_reg(rret_addr),                                 // jr   $ra 
+	add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment),  // sp += 8 (BD) 
+	mac_yield(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/mips.h

@@ -362,10 +362,28 @@ enum { _BitOffsets = 0
 /* call_reg rs — jump-and-link to register-held address; link in $ra. */
 #define call_reg(rs)               jump_link((rs), R_RA)
 
-/* j target — absolute jump within the current 256MB region. */
+/* j target — absolute jump within the current 256MB region.
+ * WARNING: `jump(off)` CANNOT BE USED for within-atom jumps in the current pipeline.
+ * The MIPS j opcode encodes `(target_addr >> 2)` in its 26-bit immediate field; an  ABSOLUTE byte address, not a relative word offset.
+ * The metaprogram computes `off` as a relative word offset (`target_word_idx - branch_word_idx - 1`), which the assembler/linker does NOT resolve.
+ *
+ * `jump(off)` is only safe when the BUILD PIPELINE owns the absolute position of the emitted code — i.e. when: s
+ *   - the build emits a symbol-relative `.word` expression that the linker resolvess via `R_MIPS_26`, OR
+ *   - the code is hand-assembled with explicit absolute targets, OR a custom post-build patcher resolves the 26-bit field.
+ */
 #define jump(off)                  enc_i(op_j,     R_0, R_0, (off))
 
-/* call_addr off — jump-and-link to immediate address. */
+/* jump_rel off — unconditional relative jump (the within-atom-safe `jump`).
+ * MIPS I R3000A has no "branch always" opcode. The idiom for an unconditional relative jump is `beq $0, $0, off`.
+ */
+#define jump_rel(off)              branch_equal(R_0, R_0, (off))
+
+/* call_addr off — jump-and-link to immediate address.
+ *
+ * Same WARNING as `jump(off)` above: the jal opcode also encodes an absolute 26-bit target.
+ * For within-atom calls, the current pipeline has no equivalent always-taken call-and-link idiom.
+ * Workaround: `branch_link` (always-taken branch + explicit `la $ra, next_word_addr; jr $ra`), or just use `call_reg($tmp)` after loading the target into a register.
+ */
 #define call_addr(off)             enc_i(op_jal,   R_0, R_0, (off))
 
 /* --- Store family (mirrors the load family) --- */

code/duffle/pad.atom.c

@@ -0,0 +1,184 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#		include "gen/macs.h"
+#		include "gen/offsets.h"
+#	include "mips.h"
+#	include "dsl.atom.h"
+#	include "lottes_tape.h"
+#	include "pad.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(pad_atom_c);
+
+#pragma region Baked Atoms
+
+/* ----- pad_bios_snapshot -----
+ * Per-frame snapshot of one BIOS pad buffer into PadState.
+ * Decoder (branch ladder on raw[0] status + raw[1] id):
+ *   1. raw[0] == 0xFF         -> Disconnected (buttons=0, axes=0x80)
+ *   2. raw[0]==0 && raw[1]==0 -> Pending      (buttons=0, axes=0x80)
+ *   3. raw[1] == 0x41         -> Digital      (buttons normalized; axes=0x80)
+ *   4. raw[1] == 0x53         -> AnalogStick  (buttons normalized; axes from raw[4..7])
+ *   5. raw[1] in 0x7x         -> AnalogPad    (buttons normalized; axes from raw[4..7])
+ *   6. else                   -> Unsupported  (buttons=0, axes=0x80)
+ *
+ * Buttons normalization: byte_swap16((~raw_buttons) & 0xFFFF).
+ *   raw_buttons    = load_half_u(raw, 2) = raw[2] | (raw[3] << 8).
+ *   byte_swap16(x) = (x >> 8) | (x << 8); nor(x, R_0) = ~x. store_half truncates to 16 bits so the upper-16 mask is implicit in the store.
+ *
+ * Register use (atom-local; no wave-context touched):
+ *   R_T0 = raw base      (kept throughout; axes loads read raw[4..7] from R_T0)
+ *   R_T1 = state base    (kept throughout; all stores go through R_T1)
+ *   R_T2 = raw[0] status (alive across the disc/pending/id dispatch, then dead)
+ *   R_T3 = raw[1] id     (alive across the id dispatch, then dead)
+ *   R_T4 = scratch       (shifts, compares, immediate loads, store values)
+ *   R_T5 = scratch       (parallel lui+ori for the 0x80808080 axes constant + byte-swap target)
+ */
+enum {
+	R_PadRaw    = R_T0 atom_reg atom_type(U1),
+	R_PadState  = R_T1 atom_reg,
+	R_RawStatus = R_T2 atom_reg,
+	R_RawId     = R_T3 atom_reg,
+};
+typedef Struct_(Binds_PadBiosSnapshot) {
+    PadBiosRaw* raw;
+    PadState*   state;
+};
+internal MipsAtom_(pad_bios_snapshot) atom_info(atom_bind(Binds_PadBiosSnapshot)
+, atom_reads( R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+, atom_writes(R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+) {
+	/* === Bind consumption: T0 = raw, T1 = state, advance R_TapePtr by 8. */
+	load_word(R_PadRaw,   R_TapePtr, O_(Binds_PadBiosSnapshot,raw)),
+	load_word(R_PadState, R_TapePtr, O_(Binds_PadBiosSnapshot,state)),
+	add_ui_self(          R_TapePtr, S_(Binds_PadBiosSnapshot)),
+
+	/* === Read raw[0] (status) + raw[1] (id) */
+	load_byte_u(R_RawStatus, R_PadRaw, 0),
+	load_byte_u(R_RawId,     R_PadRaw, 1),
+
+atom_label(snap_root) /* === Case 1: Disconnected (status == 0xFF). */
+	add_ui(R_T4, R_0, 0xFF), branch_ne(R_RawStatus, R_T4, atom_offset(snap_root, skip_disconnected)),
+	/* BD-slot: pre-compute PadStatus_Disconnected. Branch reads R_T4=0xFF in EX before this WB completes.
+	 * If branch NOT taken (fall through to pending/id_dispatch), R_T4 is overwritten by the next case body's add_ui — harmless. */
+
+atom_label(disconnected) /* === Disconnected body. */
+	/* R_T4 = PadStatus_Disconnected from snap_root BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(disconnected, snap_end)),
+		/* BD-slot: load next atom's entry point (replaces the nop).
+		 * The unconditional branch always jumps to snap_end, where mac_yield_tail()
+		 * transfers control to R_AtomJmp without re-loading it. */
+		mac_yield_load(),
+atom_label(skip_disconnected)
+
+	/* === Case 2: Pending (status == 0 && id == 0)
+	 * Combined check: if (status | id) != 0 then skip to id_dispatch.
+	 * Falls through to the Pending case only when both are zero. */
+	or_u_self(R_RawStatus, R_RawId), branch_ne(R_RawStatus, R_0, atom_offset(case_2, id_dispatch)),
+	/* BD-slot: pre-compute PadStatus_Pending. Branch reads R_RawStatus in EX before this WB completes.
+	 * If branch NOT taken (fall through to id_dispatch), R_T4 is overwritten by the digital/analog body add_ui — harmless. */
+
+atom_label(pending) /* === Pending body */
+	/* R_T4 = PadStatus_Pending from case_2 BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(pending, snap_end)),
+		mac_yield_load(),
+
+atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
+	add_ui(R_T4, R_0, 0x41), branch_ne(R_RawId, R_T4, atom_offset(id_dispatch, try_analog_stick)),
+	/* BD-slot: pre-compute PadStatus_Digital. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_stick), R_T4 is overwritten by the analog body add_ui. */
+
+	/* === Digital body (status, buttons normalize, axes=0x80, id, branch. */
+	/* R_T4 = PadStatus_Digital from id_dispatch BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),
+	/* Fill R_T4's load-delay slot with the 0x80808080 axes constant into R_T5
+	 * (R_T5 is dead on this path; it's only consumed at the analog_pad range check). */
+	load_upper_i(R_T5, 0x8080), or_i_self(R_T5, 0x8080),
+	nor_u(      R_T4, R_T4, R_0), /* raw_buttons is already in host bit order; no swap needed */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	store_word(  R_T5, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0x41),
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+
+	jump_rel(atom_offset(id_dispatch, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_stick) /* === Case 4: AnalogStick (id == 0x53)*/
+	add_ui(R_T4, R_0, 0x53), branch_ne(R_RawId, R_T4, atom_offset(try_analog_stick, try_analog_pad)),
+	/* BD-slot: pre-compute PadStatus_AnalogStick. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_pad), R_T4 is overwritten by the analog_pad body add_ui. */
+
+atom_label(analog_stick) /* === AnalogStick body
+	* Axes are loaded as two halfwords: raw[6..7] → left_xy (sh at offset 8), raw[4..5] → right_xy (sh at offset 10).
+	* R_T5 holds left_xy / id-value in turn (it's dead on this path — only consumed at the analog_pad range check). */
+	/* R_T4 = PadStatus_AnalogStick from try_analog_stick BD-slot. */
+	store_word(  R_T4, R_PadState, O_(PadState,status)),
+	load_half_u( R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u( R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot (doesn't read R_T4) */
+	nor_u(       R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half(  R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u( R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half(  R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half(  R_T4, R_PadState, O_(PadState,right_x)),
+	add_ui(      R_T5, R_0, 0x53),                        /* R_T5 = id value (clobbers left_xy, already stored) */
+	store_byte(  R_T5, R_PadState, O_(PadState,id)),
+	jump_rel(atom_offset(analog_stick, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_analog_pad) /* === Case 5-6: AnalogPad (id & 0xF0 == 0x70) */
+	and_i(    R_T4, R_RawId, 0xF0),
+	add_ui(   R_T5, R_0,     0x70),
+	branch_ne(R_T4, R_T5, atom_offset(try_analog_pad, try_unsupported)),
+	/* BD-slot: pre-compute PadStatus_AnalogPad. Branch reads R_T4 in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_unsupported), R_T4 is overwritten by the unsupported body add_ui. */
+
+atom_label(analog_pad) /* === AnalogPad body
+	* Same shape as AnalogStick with AnalogPad status. R_T5 holds left_xy (it's dead on this path). */
+	/* R_T4 = PadStatus_AnalogPad from try_analog_pad BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u(R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot */
+	nor_u(      R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u(R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half( R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half( R_T4, R_PadState, O_(PadState,right_x)),
+	store_byte( R_RawId, R_PadState, O_(PadState,id)),
+
+	jump_rel(atom_offset(analog_pad, snap_end)),
+		mac_yield_load(),
+
+atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the AnalogPad range-check miss. */
+	add_ui(    R_T4, R_0, PadStatus_Unsupported),
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0xFF),                      /* 0xFF sentinel: "unknown id" */
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+	/* Fall through to snap_end. */
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
+atom_label(snap_end)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the case-exit branch. */
+	mac_yield_tail(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/psyq.atom.c

@@ -0,0 +1,7 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+#	include "psyq.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(pysq_atom_c);

code/duffle/psyq.h

@@ -1,8 +1,8 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
-#	include "duffle/dsl.h"
+#	include "dsl.h"
-#	include "duffle/math.h"
+#	include "math.h"
-#	include "duffle/gp.h"
+#	include "gp.h"
 #endif
 
 typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };

code/duffle/word_count.metadata.h

@@ -6,7 +6,7 @@
 // One line per macro that appears in your atom sources.
 //
 // This file is encoding-macros-only. 
-// The auto-generated component macros (mac_X) live in duffle/gen/<dir>.macs.h (included separately by the unity build).
+// The auto-generated component macros (mac_X) live in the source directory's own gen/macs.h (per-directory aggregation; included separately by the unity build).
 // The unity build should include THIS file and the .macs.h file in the same TU, with both wrapped 
 // (or the include guard order handled) to avoid WORD_COUNT redeclaration.
 //

code/hello_camera/gen/macs.h

@@ -2,53 +2,23 @@
 #pragma once
 #endif
 // Auto-generated by ps1_meta.lua — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.tape.c
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera/
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.c
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.h
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.atom.c
 // Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
 
 #ifndef WORD_COUNT
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
 #endif
 
-/* atom_dbg_skip */
-#define mac_load_v2s2(rs_x, rs_y, r_base, offset) \
-	load_half( rs_x, r_base, O_(V3_S2,x)) \
-,	load_half( rs_y, r_base, O_(V3_S2,y))
-WORD_COUNT(mac_load_v2s2, 2)
-
-/* atom_dbg_skip */
-#define mac_store_v2s2(rt_x, rt_y, base, offset) \
-	store_half(rt_x, base, offset + O_(V2_S2,x)) \
-,	store_half(rt_y, base, offset + O_(V2_S2,y))
-WORD_COUNT(mac_store_v2s2, 2)
-
-/* atom_dbg_skip */
-#define mac_store_rects2(rt_x, rt_y, rt_width, rt_height, base, offset) \
-	store_half(rt_x,      base, offset + O_(Rect_S2,x)) \
-,	store_half(rt_y,      base, offset + O_(Rect_S2,y)) \
-,	store_half(rt_width,  base, offset + O_(Rect_S2,width)) \
-,	store_half(rt_height, base, offset + O_(Rect_S2,height))
-WORD_COUNT(mac_store_rects2, 4)
-
-/* atom_dbg_skip */
-#define mac_store_rgb8(rr, rg, rb, base, offset) \
-	store_byte(rr, base, offset + O_(DrawEnv,initial_bg_color.r)) \
-,	store_byte(rg, base, offset + O_(DrawEnv,initial_bg_color.g)) \
-,	store_byte(rb, base, offset + O_(DrawEnv,initial_bg_color.b))
-WORD_COUNT(mac_store_rgb8, 3)
-
-#define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
-	load_upper_i(reg_transfer, cmd >> 16) \
-,	or_i_self(   reg_transfer, cmd & 0xFFFF) \
-,	store_word(  reg_transfer, reg_base, port)
-WORD_COUNT(mac_gcmd_push, 3)
-
 #define mac_put_disp_env(reg_transfer, reg_base, port) \
 	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
 ,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port) \
 ,	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port) \
 ,	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
 ,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port)
-WORD_COUNT(mac_put_disp_env, 15)
+WORD_COUNT(mac_put_disp_env, 5)
 
 #define mac_put_draw_env(reg_transfer, reg_base, port) \
 	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port) /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */ \
@@ -67,5 +37,5 @@ WORD_COUNT(mac_put_disp_env, 15)
 ,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port)	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */ \
 ,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
 ,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)
-WORD_COUNT(mac_put_draw_env, 48)
+WORD_COUNT(mac_put_draw_env, 16)
 

code/hello_camera/gen/offsets.h

@@ -0,0 +1,50 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera\
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.c
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.h
+//   source: C:\projects\Pikuma\ps1\code\hello_camera\hello_camera.atom.c
+#pragma once
+
+#pragma region hello_camera
+
+
+// --- atom: pad_apply_input (60 words) ---
+
+#define _atom_offset_dpad_left_exit_dpad_left         6
+#define _atom_offset_dpad_right_exit_dpad_right       6
+#define _atom_offset_dead_zone_low_check_dead_low_active  8
+#define _atom_offset_dead_zone_high_check_dead_high_active  15
+#define _atom_offset_dead_zone_skip_exit_stick        24
+#define _atom_offset_end_low_exit_stick               12
+
+enum {
+    atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
+    atom_offset_dpad_right_exit_dpad_right = _atom_offset_dpad_right_exit_dpad_right,
+    atom_offset_dead_zone_low_check_dead_low_active = _atom_offset_dead_zone_low_check_dead_low_active,
+    atom_offset_dead_zone_high_check_dead_high_active = _atom_offset_dead_zone_high_check_dead_high_active,
+    atom_offset_dead_zone_skip_exit_stick = _atom_offset_dead_zone_skip_exit_stick,
+    atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
+};
+
+// --- atom: cube_g4_face (76 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           41
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+#pragma endregion hello_camera
+

code/hello_camera/hello_camera.atom.c

@@ -0,0 +1,468 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#		include "duffle/gen/macs.h"
+#		include "duffle/gen/offsets.h"
+#	include "duffle/dsl.atom.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/mips.h"
+#	include "duffle/gte.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#	include "duffle/word_count.metadata.h"
+#	include "duffle/psyq.h"
+#	include "duffle/math.atom.c"
+#	include "duffle/mips.atom.c"
+#	include "duffle/gte.atom.c"
+#	include "duffle/gp.atom.c"
+#	include "duffle/psyq.atom.c"
+#		include "gen/offsets.h"
+#		include "gen/macs.h"
+#	include "hello_camera.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(hello_joypad_atom_c);
+
+#pragma region MACs (Mips Atom components)
+
+FI_ Slice_MipsCode ac_put_disp_env(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_disp_env, {
+	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
+	// Sequence per libpsyx PutDispEnv: DrawArea TL → DrawArea BR → Mask → DrawArea TL → DrawArea BR
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_put_draw_env(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_draw_env, {
+	/*
+	* ORIGIN: each code word corresponds to the EXACT value libpsyx's PutDrawEnv function would compute for the same DrawEnv settings.
+	* References:
+	*   - libpsyx source: `toolchain/psyq-4_7/lib/libgpu.a` (binary, function `PutDrawEnv`)
+	*   - PSX-SPX doc:     https://problemkaputt.de/psx-spx.htm#gputdrawingcommands
+	*   - PSYQ SDK:        `setdrawenv` / `makelongdr_env` source
+	*   - NOCASH PSX spec: §"GP0(E1h) Draw Mode setting" through §"DR_ENV"
+	*
+	* The 16-word format is documented in the PSYQ SDK manual and on NOCASH's PSX-spec.txt. The libpsyx reference is at:
+	*   ./toolchain/psyq-4_7/lib/libgpu.a
+	*   (binary; the PutDrawEnv implementation builds the 16-word DR_ENV from the user's DRAWENV struct and emits it via GP0 GPU commands.)
+	*
+	* Word indices (libpsyx PutDrawEnv / SetDrawEnv order):
+	*   tag      = (length << 24) | addr                — 16-word packet (1 tag + 15 code)
+	*   code[0]  = DrawMode (dfe=1, dtd=0, tpage=0)     — must come first per libpsyx
+	*   code[1]  = TextureWindow (tw=(0,0))             — bare-cmd word; GPU uses current state
+	*   code[2]  = DrawArea top-left (clip.x=0, clip.y=240)
+	*   code[3]  = DrawArea bottom-right (clip.x+w=320, clip.y+h=480)
+	*   code[4]  = DrawOffset (ofs=(0,0))               — bare-cmd word
+	*   code[5]  = Mask (dtd=0, dfe=1, isbg=1)          — 0xE6 cmd + isbg bit
+	*   code[6]  = Initial-bg-color (isbg=1, r=7, g=7, b=7)
+	*   code[7]  = DrawMode (isbg=1, tpage=0)           — re-asserts DrawMode with isbg
+	*   code[8..10]  = padding (NOP)                    — 3 words to fill the packet
+	*   code[11..12] = TextureWindow bottom-right       — defaults to (0,0,0,0)
+	*   code[13..14] = padding (NOP)                    — completes the 16-word packet
+	*/
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port), /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */
+	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port), /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */
+	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port), /* code[1]  TextureWindow (tw=(0,0)) */
+	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port), /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port), /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */
+
+	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port), /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */
+	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port), /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */
+	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port), /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */
+	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port), /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */
+
+	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+
+	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+
+	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+})
+
+#pragma endregion MACs
+
+#pragma region Baked Atoms
+
+enum {
+	R_ScreenX   = R_T5 atom_reg atom_type(U2),
+	R_ScreenY   = R_T6 atom_reg atom_type(U2),
+	R_ScreenBuf = R_T7 atom_reg, /* Caller-pinned: & smem.screen_buf */
+#define R_ScreenBuf_Code R_T7_Code
+};
+//screen_env_init. Mirrors the libpsyx's SetDefDispEnv + SetDefDrawEnv + the manual enable_auto_clear / initial_bg_color writes.
+internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
+,	atom_reads(R_T0, R_ScreenX, R_ScreenY, R_ScreenBuf)
+, atom_writes(R_T0, R_ScreenX, R_ScreenY)
+) {
+	/* display[0] = (0, 0, 320, 240); rest of struct zeroed. */
+	add_ui(R_ScreenX, R_0, ScreenRes_X), add_ui(R_ScreenY, R_0, ScreenRes_Y),
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area.width) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)       + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace)   + OA_(DoubleBuffer,display,0)),
+
+	/* display[1] = (0, 240, 320, 240); rest of struct zeroed. */
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)     + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace) + OA_(DoubleBuffer,display,1)),
+
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area)         + OA_(DoubleBuffer,draw,0)), /* draw[0].clip_area = (0, 240, 320, 240). C11's SetDefDrawEnv writes clip.y = y_arg. */
+	mac_store_v2s2(  R_0, R_ScreenY,                       R_ScreenBuf, O_(DrawEnv,drawing_offset[0]) + OA_(DoubleBuffer,draw,0)), /* draw[0].drawing_offset[0] = (0, 240); C11 passes y_arg as ofs. */
+
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_window = (0, 0, 0, 0); two word-zeroes cover the full 8-byte tw field. */
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,0)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,0)),
+
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,drawing_offset[0].x)  + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_page = 10 (gp0_tpage_default). C11 SetDefDrawEnv at C11_only.elf:0x8001273C writes the same 0x0A. . */
+	add_ui(R_T0, R_0, gp0_tpage_default), 
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,0)),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0] control bytes: flag_dither=1, flag_draw_on_display=1 (the dfe bit per psx-spx; libpsyx sets it via `SetDefDrawEnv`'s conditional at C11_only.elf:0x80012728), enable_auto_clear=1. Each byte is named;
+	 * the previous `store_word(R_0, ..., +20)` overwrote all four with zero. */
+	add_ui(R_T0, R_0, 1),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].initial_bg_color = (r=7, g=7, b=7). */
+	add_ui(R_T0, R_0, 7), 
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,0)),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,1)),
+
+	mac_yield(),
+};
+
+enum {
+	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+#define R_IO_BaseAddr_Code R_T4_Code
+};
+internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads(R_IO_BaseAddr)) {
+	store_word(R_0, R_IO_BaseAddr, GPIO_PORT1_OFFSET),                                   /* GP1(00h) Reset */
+	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(01h) ClearFIFO */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(02h) AckIRQ */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(03h) Display ON */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(04h) DMADirection=2 (CPU→GPU). libpsyx's per-frame PutDrawEnv/DrawOTag use DMA2; without this the DMA queue never drains. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(05h) StartDisplayArea (X=0, Y=0) */
+
+	/* GP1: DisplayMode + Display Ranges */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+
+	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
+	load_upper_i(R_T5, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_T5, gte_cr_OFX_Code),
+	load_upper_i(R_T5, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_T5, gte_cr_OFY_Code),
+
+	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
+	add_ui(R_T5, R_0, ScreenZ), gte_mv_to_ctrl_r(R_T5, gte_cr_H_Code),
+
+	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
+	mac_gcmd_push(gp1_word_DisplayOn(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_yield(),
+};
+
+/* ----- pad_apply_input -----
+ * Reads pad[0].buttons + pad[0].left_x; 
+ * Applies the input-semantics deltas to cube_rot.y + floor_rot.y:
+ *   - D-pad Left:  cube_rot.y += 30, floor_rot.y += 5
+ *   - D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5
+ *   - Analog stick X (dead zone 0x70..0x90):
+ *       cube delta  = (0x80 - left_x) >> 2  (range approx -32..+32)
+ *       floor delta = (0x80 - left_x) >> 5  (range approx -4..+4)
+ *   - D-pad + analog deltas add when used together.
+ *
+ * Convention:
+ *   pad_state = 0 means no buttons active. 
+ *   The fail-safe zero-button value flows through unchanged, so a disconnected/fresh pad produces no rotation.
+ *   The branch_le_zero pattern below matches the existing pad_input_demo convention (atom body lines 248/257).
+ *
+ * Signed-delta trick:
+ * load_byte_u zero-extends left_x to 32 bits; sub_u from 0x80 wraps to a SIGNED two's-complement value in the negative range;
+ * shift_aright (sra) then correctly sign-extends the shift for both positive (left_x < 0x80) and negative (left_x > 0x80) cases.
+ * Digital pads publish left_x = 0x80 → delta = 0 → no rotation, so the analog step is naturally a no-op for digital controllers.
+ */
+typedef Struct_(Binds_PadApplyInput) {
+    PadState* state;
+    V3_S2*    cube_rot;
+    V3_S2*    floor_rot;
+};
+enum {
+	R_PadStateT5 = R_T5 atom_reg,
+	R_CubeRot    = R_T1 atom_reg,
+	R_FloorRot   = R_T2 atom_reg,
+};
+internal MipsAtom_(pad_apply_input) atom_info(atom_bind(Binds_PadApplyInput)
+, atom_reads(R_T0, R_CubeRot, R_FloorRot, R_T3, R_T4, R_PadStateT5, R_TapePtr)
+, atom_writes(     R_CubeRot, R_FloorRot)
+) {
+	/* Pop Binds from tape (state, cube_rot, floor_rot) */
+	load_word(R_PadStateT5, R_TapePtr, O_(Binds_PadApplyInput,state)),
+	load_word(R_CubeRot,    R_TapePtr, O_(Binds_PadApplyInput,cube_rot)),
+	load_word(R_FloorRot,   R_TapePtr, O_(Binds_PadApplyInput,floor_rot)),
+	add_ui_self(            R_TapePtr, S_(Binds_PadApplyInput)),
+
+	/* Load pad[0].buttons into R_T0. */
+	load_word(R_T0, R_PadStateT5, O_(PadState,buttons)), nop,
+	// Note(Ed): Potential op with delay slot?
+
+	/* D-pad Left: cube_rot.y += 30, floor_rot.y += 5.  */
+	and_i(R_T3, R_T0, pad0_(Pad_Left)), branch_le_zero(R_T3, atom_offset(dpad_left, exit_dpad_left)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, 30),
+		add_si(    R_T3, R_T3, 5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_left)
+
+	/* D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5. */
+	and_i(R_T3, R_T0, pad0_(Pad_Right)), branch_le_zero(R_T3, atom_offset(dpad_right, exit_dpad_right)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, -30),
+		add_si(    R_T3, R_T3, -5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_right)
+
+	/* Analog left-stick X: dead zone 0x70..0x90.
+	 * Cube delta = (0x80 - left_x) >> 2; floor delta = (0x80 - left_x) >> 5. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)),
+
+	/* Dead-zone check: skip analog if left_x in [0x70, 0x90] inclusive. Outside dead zone on LOW side: left_x < 0x70 (strictly).
+	 * set_lt_u(R_T4, R_T3, R_T4=0x70) → R_T4 = (left_x < 0x70) ? 1 : 0. */
+	add_ui(R_T4, R_0, 0x70), set_lt_u(R_T4, R_T3, R_T4), branch_ne(R_T4, R_0, atom_offset(dead_zone_low_check, dead_low_active)),
+	add_ui(R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_low_active */
+
+atom_label(dead_check_upper)
+	/* left_x >= 0x70 → check upper bound. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)), /* reload */
+	add_ui(     R_T4, R_0, 0x90),
+
+	/* R_T4 = (0x90 < left_x) ? 1 : 0 → (left_x > 0x90) ? 1 : 0 */
+	set_lt_u(R_T4, R_T4, R_T3), branch_ne(R_T4, R_0, atom_offset(dead_zone_high_check, dead_high_active)),
+	add_ui(  R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_high_active */
+	jump_rel(atom_offset(dead_zone_skip, exit_stick)),
+		mac_yield_load(),
+
+atom_label(dead_low_active)
+	/* R_T3 = left_x (from line 632 lbu; not clobbered between dead_zone_low_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_low_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3), /* R_T3 = 0x80 - left_x */
+	/* delta = 0x80 - left_x (positive). */
+
+	/* R_T4 = cube_delta */
+	shift_aright(R_T4, R_T3, 2),
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)),
+	nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+	/* R_T4 = floor_delta — moved into the load-delay slot of the floor load below (fills the 1-instruction gap;
+	 * doesn't read R_T0; R_T4 settles by the subsequent add_u). */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)),
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+	jump_rel(atom_offset(end_low, exit_stick)),
+		mac_yield_load(),
+
+atom_label(dead_high_active)
+	/* R_T3 = left_x (from line 641 lbu in dead_check_upper; not clobbered between dead_zone_high_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * The earlier `load_byte_u(R_T3, ...)` reload was redundant and introduced a load-use hazard on the next `sub_u`.
+	 * R_T4 = 0x80 from the BD-slot of `dead_zone_high_check`'s branch_ne. */
+	sub_u( R_T3, R_T4, R_T3),
+	/* delta = 0x80 - left_x (signed negative). */
+
+	shift_aright(R_T4, R_T3, 2),                      /* R_T4 = cube_delta (signed) */
+	load_half(   R_T0, R_CubeRot, O_(V3_S2,y)),
+	nop,
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_CubeRot, O_(V3_S2,y)),
+
+	/* R_T4 = floor_delta (signed) — moved into the load-delay slot of the floor load below. */
+	load_half(   R_T0, R_FloorRot, O_(V3_S2,y)),
+	shift_aright(R_T4, R_T3, 5),
+	add_u(       R_T0, R_T0, R_T4),
+	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
+
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
+atom_label(exit_stick)
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the dead-zone/exit branch. */
+	mac_yield_tail(),
+};
+
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
+#define R_PrimCursor_Code  R_T7_Code
+#define R_FaceCursor_Code  R_T4_Code
+#define R_VertBase_Code    R_T5_Code
+#define R_OtBase_Code      R_T6_Code
+};
+
+typedef Struct_(Binds_CubeTri) {
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
+	mac_yield()
+};
+
+ // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
+
+	mac_gte_load_tri_verts(R_VertBase, R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
+		/* BD-slot: write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
+		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded — 
+		 * harmless because the OT entry that points to this prim is created later, only on the body path. */
+		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(R_PrimCursor),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(R_PrimCursor),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+			mac_insert_ot_tag_g4(R_OtBase, R_PrimCursor),
+			mac_format_g4_color(R_PrimCursor,
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, R_TapePtr)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
+	mac_yield()
+};
+
+// atom_dbg_skip
+internal
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(  R_FaceCursor, R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase,   R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
+		/* Format Primitive */
+		mac_gte_store_f3(R_PrimCursor),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
+			mac_format_f3_color(R_PrimCursor, 0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(R_OtBase, R_PrimCursor),         /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+	// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/hello_camera/hello_camera.c

@@ -0,0 +1,336 @@
+#pragma region Vendors
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+// #include "libgpu.h"
+// #include "libetc.h"
+// #include "libgte.h"
+#pragma endregion Vendors
+
+#pragma region Duffle Headers
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+
+#include "duffle/word_count.metadata.h"
+
+#include "duffle/dsl.h"
+#include "duffle/memory.h"
+#include "duffle/math.h"
+
+#include "duffle/gcc_asm.h"
+#include "duffle/mips.h"
+#include "duffle/gp.h"
+#include "duffle/gte.h"
+#include "duffle/pad.h"
+
+#include "duffle/dsl.atom.h"
+#include "duffle/lottes_tape.h"
+
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/math.atom.c"
+#include "duffle/mips.atom.c"
+#include "duffle/gte.atom.c"
+#include "duffle/gp.atom.c"
+#include "duffle/pad.atom.c"
+#include "duffle/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Hello Camera Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
+#include "hello_camera.h"
+#pragma endregion Hello Camera Headers
+
+#pragma region Hello Joypad TUs
+#include "hello_camera.atom.c"
+#pragma endregion Hello Joypad TUs
+
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
+typedef Struct_(SMemory) {
+	PrimitiveArena          primitives;
+	A2_OrderingTable_Buffer ordering_tbl;
+	DoubleBuffer            screen_buf;
+	S4                      active_buf_id;
+
+	U4 MemTape[MemTape_Len];
+
+	M3_S2 tform_world;
+
+	Ent_Cube  cube;
+	Ent_Floor floor;
+
+	PadBiosRaw pad_raw[2];
+	PadState   pad[2];
+
+	U4_V scratchpad; // d-cache
+};
+global SMemory smem;
+extern SMemory smem;
+
+I_ B1* prim__alloc(U4 type_width, Str8 type_name) {
+	gknown PrimitiveArena* pa  = & smem.primitives;
+	gknown B1*             buf = (B1*) r_(smem.primitives.buf)[smem.active_buf_id];
+	assert(pa->used + type_width < PrimitiveBuff_Len);
+	B1* next  = buf + pa->used;
+	pa->used += type_width;
+	return next;
+}
+#define prim_alloc(type) (type*)prim__alloc(S_(type), slit( stringify(type)))
+
+/* Uses ONE 8-byte frame allocated via the compiler's standard prologue.
+ * The 4 wasted-arg words for B(12h) InitPAD2 live at [SP+0..15] but are not explicitly allocated.
+ * The compiler handles the MIPS O32 "wasted stack" convention for us by treating the B-call as a 4-arg call.
+ *
+ * The buffer pointers are passed as arguments so the compiler keeps them in callee-saved registers;
+ * The B(12h) asm volatile block does NOT clobber those registers (it clobbers only the volatile GPRs + the B-table arg registers explicitly).
+ * The C-level writes after the call re-load the pointers from their callee-saved homes.
+ *
+ * The clobber list for both B-calls names the full BIOS destroy set documented in kernelbios.md:167-174 (R1..R15, R24..R25, R31, HI/LO).
+ * The kernel-ABI "volatile GPRs" subset is clb_system; the rest of the destroy set is enumerated explicitly here. */
+NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
+{
+	/* Pin raw0 + raw1 to $a0 + $a1 via rgcc; the B(12h) call uses these directly.
+	 * The `(void)` casts mark them as unread after the call so the compiler doesn't need to move them back. */
+	register PadBiosRaw* p0 rgcc(R_A0) = raw0;
+	register PadBiosRaw* p1 rgcc(R_A1) = raw1;
+	(void)p0; (void)p1;
+
+	// TODO(Ed): Properly annotate the raw values in the inline asm instructions.
+	// Use enums.
+
+	/* B(12h) InitPAD2(raw0, 0x22, raw1, 0x22)
+	 *   $a0 = raw0 (rgcc-bound; survives the sequence below)
+	 *   $a1 = raw1 (preserved into $a2 before $a1 is overwritten)
+	 *   $a2 = raw1 (moved from $a1; survives $a1's overwrite)
+	 *   $a3 = 0x22 (immediate)
+	 *   $t1 = 0x12 (function number)
+	 *   $t2 = 0xB0 (BIOS B-table address) */
+	asm volatile(
+		asm_words(
+			or_u(    rarg_2, rarg_1, rdiscard), /* $a2 = $a1 = raw1 */
+			add_ui(  rarg_1, rdiscard, 0x22),   /* $a1 = 0x22 */
+			add_ui(  rarg_3, rdiscard, 0x22),   /* $a3 = 0x22 */
+			add_ui(  rtmp_1, rdiscard, 0x12),   /* $t1 = 0x12 */
+			add_ui(  rtmp_2, rdiscard, 0xB0),   /* $t2 = 0xB0 */
+			call_reg(rtmp_2),                   /* jalr $t2, $ra */
+			nop                                 /* BD slot */
+		)
+		asm_rpins, r_use(p0), r_use(p1)
+		asm_clobber: 
+			rlit(R_AT), 
+			rlit(R_V0), rlit(R_V1), 
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+
+	/* The C-level writes re-load the pointers via the parameter names and write 0xFF to each 
+	 * buffer's status byte to mark the initial-state hazard documented in kernelbios.md:1621-1624. */
+	u1_v(raw0)[0] = 0xFF;
+	u1_v(raw1)[0] = 0xFF;
+
+	/* B(13h) StartPAD2() — no args. The BIOS preserves $sp. */
+	asm volatile(
+		asm_words(
+			add_ui(  rtmp_1, rdiscard, 0x13), /* $t1 = 0x13 */
+			add_ui(  rtmp_2, rdiscard, 0xB0), /* $t2 = 0xB0 (re-load) */
+			call_reg(rtmp_2),                 /* jalr $t2, $ra */
+			nop                               /* BD slot */
+		)
+		asm_clobber:
+			rlit(R_AT),
+			rlit(R_V0), rlit(R_V1),
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+}
+
+GCC_OPTIMIZATION_DISABLE
+void update(PrimitiveArena* pa, U4* ordering_buf) 
+{
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
+	if (1) // Pad Input
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			/* BIOS-owned polling: per-frame snapshot of both ports. */
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw, & smem.pad_raw[0]);
+				tb_data_(state, & smem.pad[0]);
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw,   & smem.pad_raw[1]);
+				tb_data_(state, & smem.pad[1]);
+			/* Per-frame rotation apply: consume pad[0].buttons + pad[0].left_x */
+			tb_emit_(pad_apply_input);
+				tb_data_(state,     & smem.pad[0]);
+				tb_data_(cube_rot,  & smem.cube.rot);
+				tb_data_(floor_rot, & smem.floor.rot);
+		}
+	}
+
+	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
+
+	// Update the position based on acceleration and velocity
+	gknown V3_S4_R pos = & smem.cube.pos;
+	gknown V3_S4_R vel = & smem.cube.vel;
+	gknown V3_S4_R acc = & smem.cube.accel;
+	add_v3s4(vel, acc[0]);
+	add_v3s4_fp(pos, vel[0]);
+	// vel->x += acc->x;
+	// vel->y += acc->y;
+	// vel->z += acc->z;
+	// pos->x += vel->x;
+	// pos->y += vel->y;
+	// pos->z += vel->z;
+
+	if (pos->y + 150 > smem.floor.pos.y) vel->y *= -1;
+
+	// Prep
+	S4 nclip = 0;
+	S4 orderingtbl_z = 0;
+	A2_S2 p;    //???
+	S4 flag; //????
+
+
+	// Draw cube
+	if (1)
+	{
+		m3s2_rotation   (& smem.cube.rot,    & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.cube.pos);
+		m3s2_scale      (& smem.tform_world, & smem.cube.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.cube.faces));
+				tb_data(& tb, u4_(smem.cube.verts));
+				tb_data(& tb, u4_(ordering_buf));
+
+			for (U4 i = 0; i < Cube_num_faces; i++) {
+				// Two triangles per quad face: (x,y,z) and (x,z,w)
+				tb_emit(& tb, cube_g4_face);
+			}
+
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));
+
+		// smem.cube.rot.y += 30;
+	}
+	// Draw floor
+	if (1)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+
+		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?
+		// The tape atoms in-flight should not need to care.
+
+		// Prepare the tape. (Push protocol to tape)
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, set_gte_world);
+				tb_data(& tb, u4_(& smem.tform_world));
+
+			tb_emit(& tb, rbind_floor_f3_face);
+			// TODO(Ed): Just use a single context struct ref
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.floor.faces));
+				tb_data(& tb, u4_(smem.floor.verts));
+				tb_data(& tb, u4_(ordering_buf));
+			for (U4 i = 0; i < Floor_num_faces; i++) {
+				tb_emit(& tb, floor_f3_face);
+			}
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		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;
+	}
+}
+GCC_OPTIMIZATION_ENABLE
+
+void render(void) {
+}
+
+void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
+	draw_sync(0);
+	vsync(0);
+	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
+	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
+	{
+		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
+		pa->used = 0;
+	}
+	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
+}
+
+GCC_OPTIMIZATION_DISABLE
+int main(void)
+{
+	smem = (SMemory){0};
+	smem.scratchpad = C_(U4_V, 0x1F800000);
+	// smem.primitives.used = 0;
+	// smem.active_buf_id   = 0;
+	/*Persistent Entity Setup*/{
+		ent_cube128_init(& smem.cube.verts, & smem.cube.faces); {
+			Ent_Cube* cube = & smem.cube;
+			cube->rot    = v3s2(0, 0, 0);
+			cube->scale  = v3s4_fp_one();
+			cube->accel  = v3s4(0, 1, 0);
+			cube->pos    = v3s4(0, -400, 1800);
+		}
+		ent_floor_init(& smem.floor.verts, & smem.floor.faces); {
+			Ent_Floor* floor = & smem.floor;
+			floor->rot   = v3s2(0, 0, 0);
+			floor->pos   = v3s4(0, 450, 1800);
+			floor->scale = v3s4_fp_one();
+		}
+	}
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape)); {
+		reset_graph(0);
+		/* Direct BIOS: poll both ports during VBlank. */
+		pad_bios_init_start(& smem.pad_raw[0], & smem.pad_raw[1]);
+		/* Pinned registers for the GPU init atom. */
+		register U4*           io_base_addr rgcc(R_IO_BaseAddr) = u4_r(IO_BASE_ADDR);
+		register DoubleBuffer* screen_buf   rgcc(R_ScreenBuf)   = & smem.screen_buf;
+		tb.used = 0; tb_scope_run(& tb) {
+			tb_emit(& tb, screen_env_init);
+			tb_emit(& tb, gp_screen_init);
+		}
+	}
+	while (1) {
+		gknown S4* active_buf_id  = & smem.active_buf_id;
+		gknown U4* ordering_buf   = r_(smem.ordering_tbl)[active_buf_id[0]];
+		gknown PrimitiveArena* pa = & smem.primitives;
+		update(pa, ordering_buf);
+		render();
+		gp_display_frame(& smem.screen_buf, active_buf_id, ordering_buf, pa);
+	};
+	return 0;
+}
+GCC_OPTIMIZATION_ENABLE

code/hello_camera/hello_camera.h

@@ -0,0 +1,102 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "duffle/dsl.h"
+#	include "duffle/math.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#endif
+
+enum {
+	// PrimitiveBuff_Len = 4096,
+	// OrderingTbl_Len   = 2048,
+	PrimitiveBuff_Len = 131072,
+	OrderingTbl_Len   = 8192,
+};
+
+enum {
+	ScreenRes_X = 320,
+	ScreenRes_Y = 240,
+	ScreenZ     = 320,
+	ScreenRes_CenterX = (ScreenRes_X >> 1),
+	ScreenRes_CenterY = (ScreenRes_Y >> 1),
+};
+
+enum {
+	fp_one = (1 << 12),
+};
+
+#define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)
+
+typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
+typedef Array_(OrderingTable_Buffer, 2);
+
+typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
+typedef Array_(PrimitiveBuffer, 2);
+typedef Struct_(PrimitiveArena) {
+	A2_PrimitiveBuffer buf;
+	U4                 used;
+};
+
+#define Cube_num_verts 8
+typedef Array_(V3_S2, Cube_num_verts);
+#define Cube_num_faces 6
+typedef Array_(V4_S2, Cube_num_faces);
+I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
+	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
+		{ -128, -128, -128 },
+		{  128, -128, -128 },
+		{  128, -128,  128 },
+		{ -128, -128,  128 },
+		{ -128,  128, -128 },
+		{  128,  128, -128 },
+		{  128,  128,  128 },
+		{ -128,  128,  128 }
+	};
+	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
+		{ 3, 2, 0, 1 },
+		{ 0, 1, 4, 5 },
+		{ 4, 5, 7, 6 },
+		{ 1, 2, 5, 6 },
+		{ 2, 3, 6, 7 },
+		{ 3, 0, 7, 4 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
+	return;
+}
+typedef Struct_(Ent_Cube) {
+	V3_S4 accel;
+	V3_S4 vel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A8_V3_S2 verts;
+	A6_V4_S2 faces;
+};
+
+#define Floor_num_verts 4
+typedef Array_(V3_S2, Floor_num_verts);
+#define Floor_num_faces 2
+typedef Array_(V3_S2, Floor_num_faces);
+I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
+	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
+		{ -900, 0, -900 },
+		{ -900, 0,  900 },
+		{  900, 0, -900 },
+		{  900, 0,  900 },
+	};
+	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
+		{ 0, 1, 2 },
+		{ 1, 3, 2 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
+};
+typedef Struct_(Ent_Floor) {
+	V3_S4 accel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};

code/hello_gte/hello_gte.c

@@ -14,13 +14,13 @@
 #include "duffle/gp.h"
 #include "duffle/gte.h"
 
-#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/macs.h"
-#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/gen/offsets.h"
 #include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
 #include "duffle/word_count.metadata.h"
 
-#	include "gen/hello_gte.offsets.h"
+#	include "gen/offsets.h"
 #include "hello_gte.h"
 
 #include "hello_gte.tape.c"

code/hello_gte/hello_gte.tape.c

@@ -1,10 +1,10 @@
 #ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/macs.h"
-#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/gen/offsets.h"
 #	include "duffle/atom_dsl.h"
 #	include "duffle/lottes_tape.h"
 #	include "duffle/word_count.metadata.h"
-#	include "gen/hello_gte.offsets.h"
+#	include "gen/offsets.h"
 #	include "hello_gte.h"
 #endif
 
@@ -125,9 +125,9 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
 
 		gte_cmdw_avg_sort_z4,
 		gte_mv_from_data_r(R_T1, C2_OTZ),
+
 		add_ui(      R_AT, R_0, OrderingTbl_Len),
 		set_lt_u(    R_AT, R_T1, R_AT), 
-
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
 			mac_insert_ot_tag_g4(),
 			mac_format_g4_color(
@@ -184,7 +184,7 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 		/* Calculate Depth */
 		gte_avg_sort_z3,
 		gte_mv_from_data_r(R_T1, C2_OTZ),
-		/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
+		/* Bounds Check OTZ < OrderingTbl_Len (Branch forward to skip insertion) */
 		add_ui(      R_AT, R_0,  OrderingTbl_Len),
 		set_lt_u(    R_AT, R_T1, R_AT),
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,

code/hello_joypad/gen/macs.h

@@ -0,0 +1,41 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_joypad/
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.c
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.h
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.atom.c
+// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)
+
+#ifndef WORD_COUNT
+#define WORD_COUNT(name, count)  enum { words_##name = (count) };
+#endif
+
+#define mac_put_disp_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_disp_env, 5)
+
+#define mac_put_draw_env(reg_transfer, reg_base, port) \
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port) /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */ \
+,	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port) /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port) /* code[1]  TextureWindow (tw=(0,0)) */ \
+,	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port) /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */ \
+,	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port) /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */ \
+,	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port) /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */ \
+,	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port) /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */ \
+,	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port) /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */ \
+,	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port) /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */ \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port)	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */ \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port) \
+,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)
+WORD_COUNT(mac_put_draw_env, 16)
+

code/hello_joypad/gen/offsets.h

@@ -1,13 +1,16 @@
 // Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.tape.c
+// Directory: C:\projects\Pikuma\ps1\code\hello_joypad\
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.c
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.h
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.atom.c
 #pragma once
 
-#pragma region hello_joypad.tape
+#pragma region hello_joypad
 
 
-// --- atom: cube_g4_face (77 words) ---
+// --- atom: cube_g4_face (76 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_cull_cube_g4_face_exit           41
 #define _atom_offset_bounds_chk_cube_g4_face_exit     24
 
 enum {
@@ -28,15 +31,15 @@ enum {
 // --- atom: pad_bios_snapshot (78 words) ---
 
 #define _atom_offset_snap_root_skip_disconnected      8
-#define _atom_offset_disconnected_snap_end            60
+#define _atom_offset_disconnected_snap_end            61
 #define _atom_offset_case_2_id_dispatch               8
-#define _atom_offset_pending_snap_end                 50
+#define _atom_offset_pending_snap_end                 51
 #define _atom_offset_id_dispatch_try_analog_stick     11
-#define _atom_offset_id_dispatch_snap_end             37
+#define _atom_offset_id_dispatch_snap_end             38
 #define _atom_offset_try_analog_stick_try_analog_pad  12
-#define _atom_offset_analog_stick_snap_end            23
+#define _atom_offset_analog_stick_snap_end            24
 #define _atom_offset_try_analog_pad_try_unsupported   11
-#define _atom_offset_analog_pad_snap_end              9
+#define _atom_offset_analog_pad_snap_end              10
 
 enum {
     atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
@@ -57,8 +60,8 @@ enum {
 #define _atom_offset_dpad_right_exit_dpad_right       6
 #define _atom_offset_dead_zone_low_check_dead_low_active  8
 #define _atom_offset_dead_zone_high_check_dead_high_active  15
-#define _atom_offset_dead_zone_skip_exit_stick        23
+#define _atom_offset_dead_zone_skip_exit_stick        24
-#define _atom_offset_end_low_exit_stick               11
+#define _atom_offset_end_low_exit_stick               12
 
 enum {
     atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
@@ -69,5 +72,5 @@ enum {
     atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
 };
 
-#pragma endregion hello_joypad.tape
+#pragma endregion hello_joypad
 

code/hello_joypad/hello_joypad.atom.c

@@ -1,51 +1,29 @@
 #ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
+#	pragma once
-#	include "duffle/gen/duffle.offsets.h"
+#		include "duffle/gen/macs.h"
-#	include "duffle/atom_dsl.h"
+#		include "duffle/gen/offsets.h"
+#	include "duffle/dsl.atom.h"
 #	include "duffle/lottes_tape.h"
 #	include "duffle/mips.h"
 #	include "duffle/gte.h"
 #	include "duffle/gp.h"
 #	include "duffle/pad.h"
 #	include "duffle/word_count.metadata.h"
-#	include "psyq.h"
+#	include "duffle/psyq.h"
-#	include "gen/hello_joypad.offsets.h"
+#	include "duffle/math.atom.c"
-#	include "gen/hello_joypad.macs.h"
+#	include "duffle/mips.atom.c"
+#	include "duffle/gte.atom.c"
+#	include "duffle/gp.atom.c"
+#	include "duffle/psyq.atom.c"
+#		include "gen/offsets.h"
+#		include "gen/macs.h"
 #	include "hello_joypad.h"
 #endif
 
+ATOM_FILE_DEBUGGER_LINE_MARKER(hello_joypad_atom_c);
+
 #pragma region MACs (Mips Atom components)
 
-FI_ Slice_MipsCode ac_load_v2s2(U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_load_v2s2, {
-	load_half( rs_x, r_base, O_(V3_S2,x)),
-	load_half( rs_y, r_base, O_(V3_S2,y)),
-})
-
-FI_ Slice_MipsCode ac_store_v2s2(U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_v2s2, {
-	store_half(rt_x, base, offset + O_(V2_S2,x)),
-	store_half(rt_y, base, offset + O_(V2_S2,y)),
-})
-
-FI_ Slice_MipsCode ac_store_rects2(U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rects2, {
-	store_half(rt_x,      base, offset + O_(Rect_S2,x)),
-	store_half(rt_y,      base, offset + O_(Rect_S2,y)),
-	store_half(rt_width,  base, offset + O_(Rect_S2,width)),
-	store_half(rt_height, base, offset + O_(Rect_S2,height)),
-})
-
-FI_ Slice_MipsCode ac_store_rgb8(U1 rr, U1 rg, U1 rb, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rgb8, {
-	store_byte(rr, base, offset + O_(DrawEnv,initial_bg_color.r)),
-	store_byte(rg, base, offset + O_(DrawEnv,initial_bg_color.g)),
-	store_byte(rb, base, offset + O_(DrawEnv,initial_bg_color.b)),	
-})
-
-FI_ Slice_MipsCode ac_gcmd_push(U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
-MipsAtomComp_Proc_(ac_gcmd_push, {
-	load_upper_i(reg_transfer, cmd >> 16),
-	or_i_self(   reg_transfer, cmd & 0xFFFF),
-	store_word(  reg_transfer, reg_base, port),
-})
-
 FI_ Slice_MipsCode ac_put_disp_env(U4 reg_transfer, U4 reg_base, U2 port)
 MipsAtomComp_Proc_(ac_put_disp_env, {
 	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
@@ -114,68 +92,6 @@ MipsAtomComp_Proc_(ac_put_draw_env, {
 
 #pragma region Baked Atoms
 
-/* DIAGNOSTIC 1: Pure tape loop test */
-internal MipsAtom_(diag_yield) { mac_yield() };
-
-/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
-internal MipsAtom_(diag_color) {
-	store_word(  R_0, R_T7, 0), 
-	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
-	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
-	store_word(  R_AT, R_T7, 4),
-	
-	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
-	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
-
-	add_ui(          R_T1, R_0,  10),
-	shift_lleft_self(R_T1,       S_(U4)/2), 
-	add_u_self(      R_T1,       R_T6),         
-	
-	load_word(   R_AT, R_T1, 0),        
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
-	store_word(  R_AT, R_T7, 0),       
-	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
-	or_u_self(  R_AT, R_V0),          
-	store_word( R_AT, R_T1, 0),       
-
-	add_ui(R_T7, R_T7, 20),          
-
-	mac_yield()
-};
-
-/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
-internal MipsAtom_(diag_gte) {
-	/* Load 3 indices */
-	load_half_u(R_T0, R_T4, 0),
-	load_half_u(R_T1, R_T4, 2),
-	load_half_u(R_T2, R_T4, 4),
-
-	/* Load Vertices into GTE */
-	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-
-	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-
-	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
-
-	/* Run Math */
-	nop2, gte_cmdw_rtpt,
-	nop2, gte_cmdw_nclip,
-	nop2,
-
-	/* Advance Face Cursor and Yield */
-	add_ui(R_T4, R_T4, 8),
-
-	mac_yield()
-};
-
 enum {
 	R_ScreenX   = R_T5 atom_reg atom_type(U2),
 	R_ScreenY   = R_T6 atom_reg atom_type(U2),
@@ -264,6 +180,17 @@ internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads
 	mac_yield(),
 };
 
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
+#define R_PrimCursor_Code  R_T7_Code
+#define R_FaceCursor_Code  R_T4_Code
+#define R_VertBase_Code    R_T5_Code
+#define R_OtBase_Code      R_T6_Code
+};
+
 typedef Struct_(Binds_CubeTri) {
 	U4     PrimCursor;
 	V4_S2* FaceCursor;
@@ -294,20 +221,23 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
 	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
 	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
 
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase, R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
 	gte_cmdw_nclip,
 
 	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
+		/* BD-slot: write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
+		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded — 
+		 * harmless because the OT entry that points to this prim is created later, only on the body path. */
 		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
 		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
 		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
 		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
 
-		mac_gte_store_g4_p012(),
+		mac_gte_store_g4_p012(R_PrimCursor),
 		gte_cmdw_rotate_translate_perspective_single,
-		mac_gte_store_g4_p3(),
+		mac_gte_store_g4_p3(R_PrimCursor),
 
 		gte_cmdw_avg_sort_z4,
 		gte_mv_from_data_r(R_T1, C2_OTZ),
@@ -315,8 +245,8 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
 		set_lt_u(    R_AT, R_T1, R_AT),
 
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
-			mac_insert_ot_tag_g4(),
+			mac_insert_ot_tag_g4(R_OtBase, R_PrimCursor),
-			mac_format_g4_color(
+			mac_format_g4_color(R_PrimCursor,
 				/* c0 magenta */ 0xFF, 0x00, 0xFF,
 				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
 				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
@@ -356,8 +286,8 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
 	, atom_writes(R_PrimCursor, R_FaceCursor)
 ) {
-	mac_load_tri_indices(  R_T0, R_T1, R_T2),
+	mac_load_tri_indices(  R_FaceCursor, R_T0, R_T1, R_T2),
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase,   R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
 	gte_cmdw_nclip,
 
@@ -365,7 +295,7 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 	gte_mv_from_data_r(R_T0, C2_MAC0),
 	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
 		/* Format Primitive */
-		mac_gte_store_f3(),
+		mac_gte_store_f3(R_PrimCursor),
 
 		/* Calculate Depth */
 		gte_avg_sort_z3,
@@ -374,8 +304,8 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 		add_ui(      R_AT, R_0,  OrderingTbl_Len),
 		set_lt_u(    R_AT, R_T1, R_AT),
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
-			mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_format_f3_color(R_PrimCursor, 0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-			mac_insert_ot_tag_f3(),                 /* Insert into Ordering Table Linked List */
+			mac_insert_ot_tag_f3(R_OtBase, R_PrimCursor),         /* Insert into Ordering Table Linked List */
 			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
 				// Note(Ed): No bounds checking, should be checked before atom runs.
 		// end: branch(bounds_chk)
@@ -459,9 +389,11 @@ atom_label(disconnected) /* === Disconnected body. */
 	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
 	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
 	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
-	branch_equal(R_0, R_0, atom_offset(disconnected, snap_end)), nop,
+	jump_rel(atom_offset(disconnected, snap_end)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		/* BD-slot: load next atom's entry point (replaces the nop).
-	// jump(atom_offset(disconnected, snap_end)), nop,
+		 * The unconditional branch always jumps to snap_end, where mac_yield_tail()
+		 * transfers control to R_AtomJmp without re-loading it. */
+		mac_yield_load(),
 atom_label(skip_disconnected)
 
 	/* === Case 2: Pending (status == 0 && id == 0)
@@ -479,9 +411,8 @@ atom_label(pending) /* === Pending body */
 	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
 	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
 	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
-	branch_equal(R_0, R_0, atom_offset(pending, snap_end)), nop,
+	jump_rel(atom_offset(pending, snap_end)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		mac_yield_load(),
-	// jump(atom_offset(pending, snap_end)), nop,
 
 atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
 	add_ui(R_T4, R_0, 0x41), branch_ne(R_RawId, R_T4, atom_offset(id_dispatch, try_analog_stick)),
@@ -503,9 +434,8 @@ atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
 	add_ui(      R_T4, R_0, 0x41),
 	store_byte(  R_T4, R_PadState, O_(PadState,id)),
 
-	branch_equal(R_0, R_0, atom_offset(id_dispatch, snap_end)), nop,
+	jump_rel(atom_offset(id_dispatch, snap_end)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		mac_yield_load(),
-	// jump(atom_offset(id_dispatch, snap_end)), nop,
 
 atom_label(try_analog_stick) /* === Case 4: AnalogStick (id == 0x53)*/
 	add_ui(R_T4, R_0, 0x53), branch_ne(R_RawId, R_T4, atom_offset(try_analog_stick, try_analog_pad)),
@@ -526,9 +456,8 @@ atom_label(analog_stick) /* === AnalogStick body
 	store_half(  R_T4, R_PadState, O_(PadState,right_x)),
 	add_ui(      R_T5, R_0, 0x53),                        /* R_T5 = id value (clobbers left_xy, already stored) */
 	store_byte(  R_T5, R_PadState, O_(PadState,id)),
-	branch_equal(R_0, R_0, atom_offset(analog_stick, snap_end)), nop,
+	jump_rel(atom_offset(analog_stick, snap_end)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		mac_yield_load(),
-	// jump(atom_offset(analog_stick, snap_end)), nop,
 
 atom_label(try_analog_pad) /* === Case 5-6: AnalogPad (id & 0xF0 == 0x70) */
 	and_i(    R_T4, R_RawId, 0xF0),
@@ -550,9 +479,8 @@ atom_label(analog_pad) /* === AnalogPad body
 	store_half( R_T4, R_PadState, O_(PadState,right_x)),
 	store_byte( R_RawId, R_PadState, O_(PadState,id)),
 
-	branch_equal(R_0, R_0, atom_offset(analog_pad, snap_end)), nop,
+	jump_rel(atom_offset(analog_pad, snap_end)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		mac_yield_load(),
-	// jump(atom_offset(analog_pad, snap_end)), nop,
 
 atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the AnalogPad range-check miss. */
 	add_ui(    R_T4, R_0, PadStatus_Unsupported),
@@ -565,8 +493,12 @@ atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the
 	store_byte(  R_T4, R_PadState, O_(PadState,id)),
 	/* Fall through to snap_end. */
 
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
 atom_label(snap_end)
-	mac_yield(),
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the case-exit branch. */
+	mac_yield_tail(),
 };
 
 /* ----- pad_apply_input -----
@@ -650,10 +582,8 @@ atom_label(dead_check_upper)
 	/* R_T4 = (0x90 < left_x) ? 1 : 0 → (left_x > 0x90) ? 1 : 0 */
 	set_lt_u(R_T4, R_T4, R_T3), branch_ne(R_T4, R_0, atom_offset(dead_zone_high_check, dead_high_active)),
 	add_ui(  R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_high_active */
-	branch_equal(R_0, R_0, atom_offset(dead_zone_skip, exit_stick)), nop,
+	jump_rel(atom_offset(dead_zone_skip, exit_stick)),
-	/* Fall-through = left_x in [0x70, 0x90] (dead zone); skip analog entirely. */
+		mac_yield_load(),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
-	// jump(atom_offset(dead_zone_skip, exit_stick)), nop,
 
 atom_label(dead_low_active)
 	/* R_T3 = left_x (from line 632 lbu; not clobbered between dead_zone_low_check branch + its BD-slot `add_ui R_T4, 0x80`).
@@ -675,9 +605,8 @@ atom_label(dead_low_active)
 	add_u(       R_T0, R_T0, R_T4),
 	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
 
-	branch_equal(R_0, R_0, atom_offset(end_low, exit_stick)), nop,
+	jump_rel(atom_offset(end_low, exit_stick)),
-	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+		mac_yield_load(),
-	// jump(atom_offset(end_low, exit_stick)), nop,
 
 atom_label(dead_high_active)
 	/* R_T3 = left_x (from line 641 lbu in dead_check_upper; not clobbered between dead_zone_high_check branch + its BD-slot `add_ui R_T4, 0x80`).
@@ -698,8 +627,12 @@ atom_label(dead_high_active)
 	add_u(       R_T0, R_T0, R_T4),
 	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
 
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
 atom_label(exit_stick)
-	mac_yield(),
+	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the dead-zone/exit branch. */
+	mac_yield_tail(),
 };
 
 #pragma endregion Baked Atoms

code/hello_joypad/hello_joypad.c

@@ -1,9 +1,17 @@
+#pragma region Vendors
 #include <stdio.h>
 #include <stdlib.h>
 #include <assert.h>
 // #include "libgpu.h"
 // #include "libetc.h"
 // #include "libgte.h"
+#pragma endregion Vendors
+
+#pragma region Duffle Headers
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+
+#include "duffle/word_count.metadata.h"
 
 #include "duffle/dsl.h"
 #include "duffle/memory.h"
@@ -15,109 +23,43 @@
 #include "duffle/gte.h"
 #include "duffle/pad.h"
 
-#	include "duffle/gen/duffle.macs.h"
+#include "duffle/dsl.atom.h"
-#	include "duffle/gen/duffle.offsets.h"
-#include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
-#include "duffle/word_count.metadata.h"
 
-#include "psyq.h"
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/math.atom.c"
+#include "duffle/mips.atom.c"
+#include "duffle/gte.atom.c"
+#include "duffle/gp.atom.c"
+#include "duffle/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Joypade Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
 
-#	include "gen/hello_joypad.macs.h"
-#	include "gen/hello_joypad.offsets.h"
 #include "hello_joypad.h"
+#pragma region Joypad Headers
 
-#include "psyq.c"
+#pragma region Hello Joypad TUs
-#include "hello_joypad.tape.c"
+#include "hello_joypad.atom.c"
-
+#pragma endregion Hello Joypad TUs
-
-typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
-typedef Array_(OrderingTable_Buffer, 2);
-
-typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
-typedef Array_(PrimitiveBuffer, 2);
-typedef Struct_(PrimitiveArena) {
-	A2_PrimitiveBuffer buf;
-	U4                 used;
-};
-
-#define Cube_num_verts 8
-typedef Array_(V3_S2, Cube_num_verts);
-#define Cube_num_faces 6
-typedef Array_(V4_S2, Cube_num_faces);
-I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
-	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
-		{ -128, -128, -128 },
-		{  128, -128, -128 },
-		{  128, -128,  128 },
-		{ -128, -128,  128 },
-		{ -128,  128, -128 },
-		{  128,  128, -128 },
-		{  128,  128,  128 },
-		{ -128,  128,  128 }
-	};
-	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
-		{ 3, 2, 0, 1 },
-		{ 0, 1, 4, 5 },
-		{ 4, 5, 7, 6 },
-		{ 1, 2, 5, 6 },
-		{ 2, 3, 6, 7 },
-		{ 3, 0, 7, 4 },
-	};
-	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
-	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
-	return;
-}
-typedef Struct_(Ent_Cube) {
-	V3_S4 accel;
-	V3_S4 vel;
-	V3_S4 pos;
-	V3_S4 scale;
-	V3_S2 rot;
-	A8_V3_S2 verts;
-	A6_V4_S2 faces;
-};
-
-#define Floor_num_verts 4
-typedef Array_(V3_S2, Floor_num_verts);
-#define Floor_num_faces 2
-typedef Array_(V3_S2, Floor_num_faces);
-I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
-	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
-		{ -900, 0, -900 },
-		{ -900, 0,  900 },
-		{  900, 0, -900 },
-		{  900, 0,  900 },
-	};
-	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
-		{ 0, 1, 2 },
-		{ 1, 3, 2 },
-	};
-	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
-	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
-};
-typedef Struct_(Ent_Floor) {
-	V3_S4 accel;
-	V3_S4 pos;
-	V3_S4 scale;
-	V3_S2 rot;
-	A4_V3_S2 verts;
-	A2_V3_S2 faces;
-};
-
 
 enum { 
 	Scratchpad_Len = 1024, 
 	MemTape_Len    = 512,
 };
 typedef Struct_(SMemory) {
-	U4 MemTape[MemTape_Len];
-
-	DoubleBuffer            screen_buf;
-	A2_OrderingTable_Buffer ordering_tbl;
 	PrimitiveArena          primitives;
+	A2_OrderingTable_Buffer ordering_tbl;
+	DoubleBuffer            screen_buf;
 	S4                      active_buf_id;
 
+	U4 MemTape[MemTape_Len];
+
 	M3_S2 tform_world;
 
 	Ent_Cube  cube;
@@ -212,50 +154,6 @@ NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
 	);
 }
 
-void gp_screen_init_c11(DoubleBuffer* screen_buf, S4* active_buf_id)
-{
-	reset_graph(0);
-
-	// Set the current initial buffer
-	active_buf_id[0] = 0;
-
-	// Just setting env data, not interacting with console hw.
-	// First buffer area
-	displayenv_init(& r_(screen_buf->display)[0], 0, 0,           ScreenRes_X, ScreenRes_Y);
-	drawenv_init   (& r_(screen_buf->draw   )[0], 0, ScreenRes_Y, ScreenRes_X, ScreenRes_Y);
-	// Second buffer area
-	displayenv_init(& r_(screen_buf->display)[1], 0, ScreenRes_Y, ScreenRes_X, ScreenRes_Y);
-	drawenv_init   (& r_(screen_buf->draw   )[1], 0, 0,           ScreenRes_X, ScreenRes_Y);
-	// Set the back/drawing buffer
-	screen_buf->draw[0].enable_auto_clear = true;
-	screen_buf->draw[1].enable_auto_clear = true;
-	// Set the background clear color
-	screen_buf->draw[0].initial_bg_color = rgb8( .r = 7, .g = 7,  .b = 7 );
-	screen_buf->draw[1].initial_bg_color = rgb8( .r = 7, .g = 7,  .b = 7 );
-	// screen_buf->draw[1].initial_bg_color = rgb8( .r = 47, .g = 13, .b = 0 );
-	displayenv_put(& r_(screen_buf->display)[ active_buf_id[0] ]);
-	drawenv_put   (& r_(screen_buf->draw   )[ active_buf_id[0] ]);
-
-	// Initialize and setup the GTE geometry offsets
-	geom_init();
-	geom_set_offset(ScreenRes_CenterX, ScreenRes_CenterY);
-	geom_set_screen(ScreenZ);
-
-	set_display_enabled(1); // gp_DisplayEnabled
-}
-
-void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
-	draw_sync(0);
-	vsync(0);
-	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
-	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
-	{
-		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
-		pa->used = 0;
-	}
-	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
-}
-
 GCC_OPTIMIZATION_DISABLE
 void update(PrimitiveArena* pa, U4* ordering_buf) 
 {
@@ -478,28 +376,25 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		// C-side state (pa->used) has already been updated by the tape!
 		// smem.floor.rot.y += 5;
 	}
-	// --- TAPE DIAGNOSTICS ---
-	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++) {
-				// tb_emit(& tb, code_diag_yield);
-				// tb_emit(& tb, code_diag_color);
-				// tb_emit(& tb, code_diag_gte); 
-			}
-		}
-		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];
-	}
 }
 GCC_OPTIMIZATION_ENABLE
 
 void render(void) {
 }
 
+void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
+	draw_sync(0);
+	vsync(0);
+	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
+	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
+	{
+		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
+		pa->used = 0;
+	}
+	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
+}
+
+GCC_OPTIMIZATION_DISABLE
 int main(void)
 {
 	smem = (SMemory){0};
@@ -543,3 +438,4 @@ int main(void)
 	};
 	return 0;
 }
+GCC_OPTIMIZATION_ENABLE

code/hello_joypad/hello_joypad.h

@@ -7,8 +7,10 @@
 #endif
 
 enum {
-	PrimitiveBuff_Len = 4096,
+	// PrimitiveBuff_Len = 4096,
-	OrderingTbl_Len   = 2048
+	// OrderingTbl_Len   = 2048,
+	PrimitiveBuff_Len = 131072,
+	OrderingTbl_Len   = 8192,
 };
 
 enum {
@@ -24,3 +26,77 @@ enum {
 };
 
 #define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)
+
+typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
+typedef Array_(OrderingTable_Buffer, 2);
+
+typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
+typedef Array_(PrimitiveBuffer, 2);
+typedef Struct_(PrimitiveArena) {
+	A2_PrimitiveBuffer buf;
+	U4                 used;
+};
+
+#define Cube_num_verts 8
+typedef Array_(V3_S2, Cube_num_verts);
+#define Cube_num_faces 6
+typedef Array_(V4_S2, Cube_num_faces);
+I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
+	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
+		{ -128, -128, -128 },
+		{  128, -128, -128 },
+		{  128, -128,  128 },
+		{ -128, -128,  128 },
+		{ -128,  128, -128 },
+		{  128,  128, -128 },
+		{  128,  128,  128 },
+		{ -128,  128,  128 }
+	};
+	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
+		{ 3, 2, 0, 1 },
+		{ 0, 1, 4, 5 },
+		{ 4, 5, 7, 6 },
+		{ 1, 2, 5, 6 },
+		{ 2, 3, 6, 7 },
+		{ 3, 0, 7, 4 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
+	return;
+}
+typedef Struct_(Ent_Cube) {
+	V3_S4 accel;
+	V3_S4 vel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A8_V3_S2 verts;
+	A6_V4_S2 faces;
+};
+
+#define Floor_num_verts 4
+typedef Array_(V3_S2, Floor_num_verts);
+#define Floor_num_faces 2
+typedef Array_(V3_S2, Floor_num_faces);
+I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
+	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
+		{ -900, 0, -900 },
+		{ -900, 0,  900 },
+		{  900, 0, -900 },
+		{  900, 0,  900 },
+	};
+	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
+		{ 0, 1, 2 },
+		{ 1, 3, 2 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
+};
+typedef Struct_(Ent_Floor) {
+	V3_S4 accel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};

code/hello_joypad/psyq.c

@@ -1,3 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#	include "psyq.h"
-#endif

scripts/build_psyq.ps1

@@ -180,12 +180,9 @@ function link-modules { param([string[]]$link_modules, [string]  $elf, [string[]
 	$link_args += ($f_link_pass_through_prefix + $f_link_mapfile + $map)
 
 	$link_args += ($f_link_pass_through_prefix + $f_link_start_group)
-	# raw_sio_pad_poll_20260802 — Task 5.1c surgical library-list trim.
+	# 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math, mcrd, mcx, press, sio, snd, spu, tap)
-	# The 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math,
+	# had LOAD lines in the map but ZERO .o files pulled in — they were unused.
-	# mcrd, mcx, press, sio, snd, spu, tap) had LOAD lines in the map but
+	# 5 kept libraries (api, c, etc, gpu, gte) are required by the C-side calls in hello_joypad.c (reset_graph, draw_sync, vsync, etc.).
-	# ZERO .o files pulled in — they were unused. The 5 kept libraries
-	# (api, c, etc, gpu, gte) are required by the C-side calls in
-	# hello_joypad.c (reset_graph, draw_sync, vsync, etc.).
 	$libraries = @(
 		"api",
 		"c",
@@ -227,9 +224,7 @@ function ps1-meta { param(
 		[string[]]$passes = @('--pre-link'),
 		[string[]]$extra_args = @()
 	)
-	# `--unity-root` and `--source` are
+	# `--unity-root` and `--source` are mutually exclusive. Exactly one of `$unity_root` / `$sources` must be supplied; the other must be absent.
-	# mutually exclusive. Exactly one of `$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) {
@@ -265,6 +260,73 @@ function ps1-meta { param(
 	}
 }
 
+function inject-dwarf { param(
+	[string]$elf,
+	[string]$path_gen
+)
+	$base_name               = [System.IO.Path]::GetFileNameWithoutExtension($elf)
+	$path_dwarf_line_bin     = join-path $path_gen  "$base_name.dwarf_line.bin"
+	$path_dwarf_aranges_bin  = join-path $path_gen  "$base_name.dwarf_aranges.bin"
+	$path_dwarf_rnglists_bin = join-path $path_gen  "$base_name.dwarf_rnglists.bin"
+	$path_dwarf_info_bin     = join-path $path_gen  "$base_name.dwarf_info.bin"
+	$path_dwarf_abbrev_bin   = join-path $path_gen  "$base_name.dwarf_abbrev.bin"
+	$path_dwarf_str_bin      = join-path $path_gen  "$base_name.dwarf_str.bin"
+	$path_dwarf_loc_bin      = join-path $path_gen  "$base_name.dwarf_loc.bin"
+	$path_dwarf_loclists_bin = join-path $path_gen  "$base_name.dwarf_loclists.bin"
+	$path_inject_elf         = join-path $path_build "$base_name.dwarf-injected.elf"
+
+	if (-not (Test-Path $path_dwarf_line_bin))     { return }
+	if (-not (Test-Path $path_dwarf_aranges_bin))  { return }
+	if (-not (Test-Path $path_dwarf_rnglists_bin)) { return }
+
+	Write-Host "[build] DWARF-injecting $elf -> $path_inject_elf"
+	Copy-Item -LiteralPath $elf -Destination $path_inject_elf -Force
+
+	# Objcopy call 1: 3x --update-section for the PC-mapping tables (line, aranges, rnglists).
+	$objcopy_args_dwarf_pc = @(
+		"--update-section=.debug_line=$path_dwarf_line_bin",
+		"--update-section=.debug_aranges=$path_dwarf_aranges_bin",
+		"--update-section=.debug_rnglists=$path_dwarf_rnglists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_pc $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-pc splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Objcopy call 2: 3x --update-section + 2x --add-section for the debug-data tables (info, abbrev, str, loc, loclists).
+	$objcopy_args_dwarf_info = @(
+		"--update-section=.debug_info=$path_dwarf_info_bin",
+		"--update-section=.debug_abbrev=$path_dwarf_abbrev_bin",
+		"--update-section=.debug_str=$path_dwarf_str_bin",
+		"--add-section=.debug_loc=$path_dwarf_loc_bin",
+		"--add-section=.debug_loclists=$path_dwarf_loclists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_info $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-info splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
+	# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
+	# The original ELF and PS-EXE remain byte/flag unchanged.
+	& $Objcopy `
+		--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
+		--set-section-flags ".data=alloc,load,data,code,contents" `
+		$path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+	}
+	else {
+		Write-Host "[build] DWARF-injected ELF: $path_inject_elf"
+	}
+}
+# inject-dwarf
+
 function build-hello_psyqo {
 	$includes += @()
 
@@ -391,61 +453,7 @@ function build-hello_gte {
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
 	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
 
-	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
+	inject-dwarf $elf $path_build_gen
-	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
-	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
-	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
-	{
-		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
-		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		$f_args = @(
-			"--update-section=.debug_line=$dwarfLineBin",
-			"--update-section=.debug_aranges=$dwarfArangesBin",
-			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
-		)
-		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		$dwarfInfoBin     = join-path $path_build_gen 'hello_gte.dwarf_info.bin'
-		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_gte.dwarf_abbrev.bin'
-		$dwarfStrBin      = join-path $path_build_gen 'hello_gte.dwarf_str.bin'
-		$dwarfLocBin      = join-path $path_build_gen 'hello_gte.dwarf_loc.bin'
-		$dwarfLoclistsBin = join-path $path_build_gen 'hello_gte.dwarf_loclists.bin'
-		$g_args = @(
-			"--update-section=.debug_info=$dwarfInfoBin",
-			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
-			"--update-section=.debug_str=$dwarfStrBin",
-			"--add-section=.debug_loc=$dwarfLocBin",
-			"--add-section=.debug_loclists=$dwarfLoclistsBin"
-		)
-		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
-		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
-		# The original ELF and PS-EXE remain byte/flag unchanged.
-		& $Objcopy `
-			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
-			--set-section-flags ".data=alloc,load,data,code,contents" `
-			$injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-		} 
-		else {
-			Write-Host "[build] DWARF-injected ELF: $injectElf"
-		}
-	}
 }
 # build-hello_gte
 
@@ -496,63 +504,59 @@ function build-hello_joypad {
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
 	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
 
-	$dwarfLineBin     = join-path $path_build_gen 'hello_joypad.dwarf_line.bin'
+	inject-dwarf $elf $path_build_gen
-	$dwarfArangesBin  = join-path $path_build_gen 'hello_joypad.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path $path_build_gen 'hello_joypad.dwarf_rnglists.bin'
-	$injectElf        = join-path $path_build 'hello_joypad.dwarf-injected.elf'
-	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
-	{
-		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
-		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		$f_args = @(
-			"--update-section=.debug_line=$dwarfLineBin",
-			"--update-section=.debug_aranges=$dwarfArangesBin",
-			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
-		)
-		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
 }
+# build-hello_joypad
 
-		$dwarfInfoBin     = join-path $path_build_gen 'hello_joypad.dwarf_info.bin'
+function build-hello_camera {
-		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_joypad.dwarf_abbrev.bin'
+	$includes += @()
-		$dwarfStrBin      = join-path $path_build_gen 'hello_joypad.dwarf_str.bin'
-		$dwarfLocBin      = join-path $path_build_gen 'hello_joypad.dwarf_loc.bin'
-		$dwarfLoclistsBin = join-path $path_build_gen 'hello_joypad.dwarf_loclists.bin'
-		$g_args = @(
-			"--update-section=.debug_info=$dwarfInfoBin",
-			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
-			"--update-section=.debug_str=$dwarfStrBin",
-			"--add-section=.debug_loc=$dwarfLocBin",
-			"--add-section=.debug_loclists=$dwarfLoclistsBin"
-		)
-		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
 
-		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
+	$path_module        = join-path $path_code   'hello_camera'
-		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
+	$path_duffle        = join-path $path_code   'duffle'
-		# The original ELF and PS-EXE remain byte/flag unchanged.
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
-		& $Objcopy `
+	$path_build_gen     = join-path $path_build  'gen'
-			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
+
-			--set-section-flags ".data=alloc,load,data,code,contents" `
+	$src_c = join-path $path_module 'hello_camera.c'
-			$injectElf 2>&1 | Out-Null
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--pre-link')
-		if ($LASTEXITCODE -ne 0) {
+
-			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
+	$assemble_args = @()
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
+	$assemble_args += $f_debug
+	$assemble_args += $f_optimize_none
+	$assemble_args += ($f_include + $path_code)
+
+	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
+	$module_asm_crt = join-path $path_build         'crt0.o'
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+
+	$module_c = join-path $path_build  'hello_camera_c.o'
+
+	$compile_args = @()
+	$compile_args += $f_debug
+	$compile_args += $f_optimize_none
+	# $compile_args += $f_optimize_intrinsics
+	# $compile_args += $f_optimize_size
+	# $compile_args += $f_optimize_debug
+	$compile_args += ($f_include + $path_code)
+	compile-unit $src_c $module_c $includes $compile_args
+
+	$elf = join-path $path_build 'hello_camera.elf'
+	$exe = join-path $path_build 'hello_camera.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
+
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
+
+	inject-dwarf $elf $path_build_gen
 }
-		else {
+build-hello_camera
-			Write-Host "[build] DWARF-injected ELF: $injectElf"
-		}
-	}
-}
-build-hello_joypad
 
 # NO idea if this works yet...
 function Send-ToEmulator { param( [string]$exePath )

scripts/duffle.lua

@@ -1,18 +1,12 @@
 --- duffle.lua — shared primitives + domain tables for the tape-atom metaprograms.
----
+---   * Character classification: `is_space`, `is_alpha`, `is_alnum`, `is_digit`, plus the byte-fast `_byte` variants.
---- One ownership statement, then the rest is signal:
+---   * String / path primitives: `trim`, `dirname`, `basename_no_ext`, `normalize_path`, `canonical_path_key`, `find_byte`.
----   * **Character classification** (`is_space`, `is_alpha`, `is_alnum`, `is_digit`, plus the byte-fast `_byte` variants).
+---   * I/O primitives:           `read_file`, `write_file`, `ensure_dir`.
----   * **String / path primitives** (`trim`, `dirname`, `basename_no_ext`, `normalize_path`, `canonical_path_key`, `find_byte`).
+---   * Corpus resolution:        `parse_direct_quoted_includes`, `resolve_source_corpus`.
----   * **I/O primitives** (`read_file`, `write_file`, `ensure_dir`).
+---   * C-language scanner:       `skip_ws_and_cmt`, `skip_str_or_cmt`, `read_ident`, `read_parens`, `read_braces`, `read_brackets`, `read_balanced`, `scan_to_char`, `split_top_level_commas`.
----   * **Corpus resolution** (`parse_direct_quoted_includes`, `resolve_source_corpus`).
+---   * Word-count loader:        `load_word_counts` for `WORD_COUNT(...)` metadata files.
----   * **C-language scanner** (`skip_ws_and_cmt`, `skip_str_or_cmt`, `read_ident`, `read_parens`, `read_braces`, `read_brackets`,
+---   * Line lookup:              `LineIndex` returns an O(log N) `line_of(pos)` closure for source-mapping.
----     `read_balanced`, `scan_to_char`, `split_top_level_commas`).
+---   * Domain tables:            `TAPE_ATOM_MACROS`, `GTE_PIPELINE_LATENCY`, `GP0_CMD_SIZE`, `GP0_CMD_BY_SHAPE`, `INSTRUCTION_LATENCY`.
----   * **Word-count loader** (`load_word_counts` for `WORD_COUNT(...)` metadata files).
----   * **Line lookup** (`LineIndex` returns an O(log N) `line_of(pos)` closure for source-mapping).
----   * **Domain tables** (`TAPE_ATOM_MACROS`, `GTE_PIPELINE_LATENCY`, `GP0_CMD_SIZE`, `GP0_CMD_BY_SHAPE`,
----     `GP0_MACRO_CONTRIB`, `INSTRUCTION_LATENCY`).
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex.
 
 local M = {}
 
@@ -24,7 +18,7 @@ local lfs = require("lfs")
 -- Cross-file type aliases
 -- ════════════════════════════════════════════════════════════════════════════
 
---- @alias Path        string   -- absolute or CWD-relative file path
+--- @alias Path        string   -- Absolute or CWD-relative file path
 --- @alias LineNum     integer  -- 1-indexed source line number
 --- @alias ByteOff     integer  -- 0-indexed byte offset within a source string
 --- @alias MacroName   string   -- lower_snake_case macro identifier (e.g. "mac_yield")
@@ -32,10 +26,10 @@ local lfs = require("lfs")
 --- @alias Severity    string   -- "error" | "warning" | "info"
 
 --- @class SourceFile
---- @field path     Path      -- absolute path to the source file
+--- @field path     Path      -- Absolute path to the source file
---- @field text     string    -- the full source text
+--- @field text     string    -- Full source text
---- @field dir      string    -- the directory containing the source
+--- @field dir      string    -- Directory containing the source
---- @field basename string    -- filename without extension
+--- @field basename string    -- Filename without extension
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- ASCII byte constants
@@ -73,20 +67,12 @@ local BYTE_DIGIT_9 = 0x39       -- '9'
 -- ════════════════════════════════════════════════════════════════════════════
 -- Section -1: Bootstrap (path-setup at module load)
 -- ════════════════════════════════════════════════════════════════════════════
---
--- Path setup runs through `scripts/duffle_paths.lua`, which derives the repo root from `debug.getinfo(1, "S").source`
--- (no subprocess, ~0ms) and then calls `require("duffle")`.
--- Entry and pass scripts load `duffle_paths.lua` first; a `find_repo_root` / `setup_package_path` defined here was dead code in practice.
--- `git rev-parse` costs ~100-180ms per subprocess spawn on Windows; `debug.getinfo` is <1ms, so we keep only the fast path.
---
--- To load `duffle.lua` outside `duffle_paths.lua`, set `package.path` manually before `require`.
--- See `docs/guide_metaprogram_ssdl.md` §"I/O primitives" for the pattern.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Section 0: LPeg patterns (compiled once at module load)
 -- ════════════════════════════════════════════════════════════════════════════
 --
--- LPeg is a required dependency (PEG library, no regex). It's loaded via `package.cpath` — `duffle_paths.lua` wires the path to `toolchain/lpeg/lpeg.dll`.
+-- LPeg is a required dependency (PEG library). It's loaded via `package.cpath` — `duffle_paths.lua` wires the path to `toolchain/lpeg/lpeg.dll`.
 -- LPeg handles the high-level scanner; the byte-by-byte helpers in Section 1 handle classification primitives that LPeg's CPython-level cost would dominate.
 --
 -- If the require fails, fail loud with an actionable message. The build script (`update_deps.ps1`) builds lpeg.dll into `toolchain/lpeg/`; run it when the dll is missing.
@@ -108,24 +94,18 @@ local lpeg_alnum_pat  = alpha_pat + digit_pat
 local lpeg_alpha_pat = alpha_pat
 local lpeg_ident_pat = lpeg.C(alpha_pat * lpeg_alnum_pat^0)
 
--- String literal: "..." with backslash escapes.
+local lpeg_str_pat        = P('"') * (P(1) - S('"\\') + P('\\') * P(1))^0 * P('"') -- String literal: "..." with backslash escapes.
-local lpeg_str_pat = P('"') * (P(1) - S('"\\') + P('\\') * P(1))^0 * P('"')
+local lpeg_chr_pat        = P("'") * (P(1) - S("'\\") + P('\\') * P(1))^0 * P("'") -- Char literal: '...' with backslash escapes.
--- Char literal: '...' with backslash escapes.
+local lpeg_line_cmt_pat   = P("//") * (P(1) - S("\n"))^0                           -- Line comment: // ... to end-of-line.
-local lpeg_chr_pat = P("'") * (P(1) - S("'\\") + P('\\') * P(1))^0 * P("'")
+local lpeg_block_cmt_pat  = P("/*") * (P(1) - P("*/"))^0 * P("*/")                 -- Block comment: /* ... */ (no nesting per C standard).
--- Line comment: // ... to end-of-line.
+local lpeg_str_or_cmt_pat = lpeg_str_pat + lpeg_chr_pat + lpeg_line_cmt_pat + lpeg_block_cmt_pat -- String or comment (any of the four forms).
-local lpeg_line_cmt_pat = P("//") * (P(1) - S("\n"))^0
--- Block comment: /* ... */ (no nesting per C standard).
-local lpeg_block_cmt_pat = P("/*") * (P(1) - P("*/"))^0 * P("*/")
--- String or comment (any of the four forms).
-local lpeg_str_or_cmt_pat = lpeg_str_pat + lpeg_chr_pat + lpeg_line_cmt_pat + lpeg_block_cmt_pat
 
 -- Whitespace + comment skipper: zero+ (whitespace run | string | comment).
 local ws_pat              = S(" \t\n\r\v\f")
 local lpeg_ws_and_cmt_pat = (ws_pat + lpeg_str_or_cmt_pat)^0
 
 -- Generic "skip until target, but step over balanced groups" matcher.
--- Used by scan_to_char for non-ident / non-bracket chars.
+-- Used by scan_to_char for non-ident / non-bracket chars. We accept any single char except the target.
--- We accept any single char except the target.
 -- The balanced-group stepping is handled by the caller (via read_balanced).
 local lpeg_scan_to_target_pat = function(target) return (P(1) - P(target))^0  end
 
@@ -148,12 +128,10 @@ end
 
 -- Single digit.
 function M.is_digit_byte(b) return b and b >= BYTE_DIGIT_0 and b <= BYTE_DIGIT_9 end
-
 -- Letter OR digit OR underscore.
 function M.is_alnum_byte(b) return M.is_alpha_byte(b) or M.is_digit_byte(b) end
 
--- String-based wrappers (kept for callers that already have a single-char string;
+-- String-based wrappers (kept for callers that already have a single-char string; the byte versions are what the hot loops should call).
--- the byte versions are what the hot loops should call).
 function M.is_space(c)
 	if type(c) == "number" then return M.is_space_byte(c) end
 	return c == " "  or c == "\t" or c == "\n" or c == "\r" or c == "\v" or c == "\f"
@@ -309,7 +287,7 @@ local function absolute_normalized_path(path)
 end
 
 --- Return the normalized absolute, Windows-case-folded comparison key for a path.
---- -Ordinary relative paths resolve against the process cwd. Drive-relative paths are rejected because LuaFileSystem does not expose Windows per-drive current directories.
+--- Ordinary relative paths resolve against the process cwd. Drive-relative paths are rejected because LuaFileSystem does not expose Windows per-drive current directories.
 --- @param path Path
 --- @return string
 function M.canonical_path_key(path)
@@ -810,7 +788,6 @@ function M.resolve_source_corpus(options)
 end
 
 -- Split a brace-body into top-level comma-separated tokens. Honors nested parens/braces/brackets and skips strings/comments.
---
 -- Splits at top-level NEWLINES and SEMICOLONS too, AND emits a token break after a top-level comment/string.
 -- Pure-comment / pure-string chunks contribute 0 words.
 function M.split_top_level_commas(body)
@@ -846,7 +823,7 @@ function M.split_top_level_commas(body)
 				if has_real_content(chunk) then
 					tokens[#tokens + 1] = chunk
 				elseif #tokens > 0 then
-					-- Pure comment/string chunk at top level.
+					-- comment/string chunk at top level.
 					-- Append it to the LAST token so emit-context callers (components.lua build_component_lines) can convert
 					-- `// trailing comment` to `/* */` and emit it with the macro body.
 					-- count_token_words only inspects the leading ident, so a trailing comment does not affect the count.
@@ -920,8 +897,7 @@ function M.tokenize_body(body)
 		while scan <= len do
 			local c = body:byte(scan)
 			-- Terminator bytes (delimit a token at the top level): ',' = 0x2C, '\n' = 0x0A, ';' = 0x3B.
-			-- These also appear as separators between argument lists inside the parens/braces/brackets,
+			-- These also appear as separators between argument lists inside the parens/braces/brackets, so we stop the scan when we hit any of them.
-			-- so we stop the scan when we hit any of them.
 			if     c == BYTE_COMMA   then break end
 			if     c == BYTE_NEWLINE then break end
 			if     c == BYTE_SEMI    then break end
@@ -1088,8 +1064,8 @@ M.GTE_COMMAND_ALIASES = {
 --   * "Store delays are counted in numbers of clock cycles (not in numbers of opcodes).
 --    For 3 cycle delay, one must usually insert 3 cached opcodes (or one uncached opcode)."
 --
--- Per PSX-SPX `docs/psx-spx/docs/gtepipelinetimings.md` (the per-instruction input-latch measurement, which is the same
+-- Per PSX-SPX `docs/psx-spx/docs/gtepipelinetimings.md` (the per-instruction input-latch measurement, which is the same phenomenon modeled from the command side), 
--- phenomenon modeled from the command side), the values are:
+-- the values are:
 --   rtps:  every data register, every control register (RT / TR / OFX / OFY / H / DQA / DQB)
 --   rtpt:  same superset (rtpt reads V0..V2, the RT matrix, the TR vector, OFX / OFY, H, DQA, DQB)
 --   nclip: SXY0, SXY1, SXY2 (no RT / TR / OFX inputs)
@@ -1099,7 +1075,7 @@ M.GTE_COMMAND_ALIASES = {
 --   avsz3 / avsz4: SZ0..SZ3 + ZSF3/ZSF4
 --
 -- We model the data-register + control-register superset. Every relevant input is in this set per PSX-SPX `gtepipelinetimings.md`;
--- the per-input latching values there describe the same number's command-side view
+-- The per-input latching values there describe the same number's command-side view
 -- (a recent mtc2/ctc2 to that register must retire the same number of cycles before the command issues).
 -- Anything outside this set is safe to clobber immediately after a prior command.
 M.GTE_COMMAND_INPUTS = {
@@ -1163,7 +1139,6 @@ M.GTE_COMMAND_INPUTS = {
 }
 
 -- GTE command output-set + semantic role table.
---
 -- For each command, the set of C2 data registers the command writes as outputs, paired with the SEMANTIC ROLE of each output.
 -- The semantic role is the basis for the `_post_<cmd>` contract validation.
 -- The contract says "after <cmd>, the latest screen-XY is C2_SXY2" (C2_SXY0 is wrong; the FIFO side effects leave SXY0 as an older FIFO entry, never the newest).
@@ -1242,16 +1217,15 @@ M.GTE_COMMAND_OUTPUTS = {
 -- the latched value in the pipeline gets overwritten by the CPU before the pipeline consumes it.
 --
 -- This relation is the command -> register direction (the command is the producer; MTC2 / CTC2 is the consumer).
--- It is the inverse of the MTC2 -> command input propagation (register -> command direction), which is staged by the
+-- It is the inverse of the MTC2 -> command input propagation (register -> command direction), which is staged by the producer step of `analyze_hardware_relations`.
--- producer step of `analyze_hardware_relations`.
 --
--- The schema mirrors the producer-side relations (`direction`, `evidence`, `violation_kind`); `required` counts the
+-- The schema mirrors the producer-side relations (`direction`, `evidence`, `violation_kind`); 
--- emitted words strictly between the command's last output word and the overwrite.
+-- `required` counts the emitted words strictly between the command's last output word and the overwrite.
 -- `required = 0` permits the immediately following overwrite; `required = 4` requires four intervening words.
 --
 -- Per PSX-SPX `gtepipelinetimings.md` the per-command input latching measurements are the same numbers inverted.
--- They describe when a recent MTC2/CTC2 must retire before the command issues; this table describes when a recent
+-- They describe when a recent MTC2 / CTC2 must retire before the command issues. 
--- command's outputs latch into the pipeline before a later MTC2/CTC2 overwrites them.
+-- This table describes when a recent command's outputs latch into the pipeline before a later MTC2/CTC2 overwrites them.
 --
 -- Consumers:
 --   * passes/static_analysis.lua::analyze_hardware_relations (stages post-command latch relations in `pending` after a GTE command).
@@ -1266,9 +1240,7 @@ M.GTE_COMMAND_LATCH_WINDOWS = {
 		{ register = "C2_OTZ",  required = 4 },
 		{ register = "C2_IR0",  required = 4 },
 	},
-	-- RTPT: same latching as RTPS (the LAST projection in SXY2 is the
+	-- RTPT: same latching as RTPS (the LAST projection in SXY2 is the newest one; the earlier SXY0 / SXY1 entries are part of the batched triple).
-	-- newest one; the earlier SXY0 / SXY1 entries are part of the
-	-- batched triple).
 	["gte_cmdw_rtpt"] = {
 		{ register = "C2_SXY0", required = 4 },
 		{ register = "C2_SXY1", required = 4 },
@@ -1300,11 +1272,7 @@ M.GTE_COMMAND_LATCH_WINDOWS = {
 	},
 }
 
--- GTE component result contracts were removed: the `_post_<cmd>` naming convention was a soft convention
--- (the user did not want it formalized via static-analysis enforcement). A proper `atom_info` directive for ordering semantics is a future TODO.
-
 -- Operand-class table for the COP2->GPR load-delay check.
---
 -- Maps each emitting-token ident to the set of GPR operand positions it reads.
 -- Covers the current encoder vocabulary (`code/duffle/mips.h` + `code/duffle/gte.h`); add rows here as new encoders land.
 --
@@ -1430,34 +1398,7 @@ M.GP0_CMD_BY_SHAPE = {
 	["g4"]  = 0x38, ["gt4"] = 0x3C,
 }
 
--- TODO(Ed): REMOVE THIS HARDCODE, THIS SHOULD BE RESOLVED AUTOMATICALLY
+-- Per-instruction cycle cost (best-case, no stalls). Used by the static-analysis pass to emit per-atom cycle budgets.
--- Per-macro prim-buffer contribution: how many 32-bit words each macro writes to the primitive being built in main RAM.
--- (This counts RAM-side prim-buffer words, not .text instruction words.)
--- The sum across `mac_format_X_color` + `mac_gte_store_X_post_*` + `mac_insert_ot_tag_X` calls in an atom body must equal
--- `GP0_CMD_SIZE[GP0_CMD_BY_SHAPE[shape]]`.
-M.GP0_MACRO_CONTRIB = {
-	["mac_format_f3_color"]      = 1,
-	["mac_format_g3_color"]      = 3,
-	["mac_format_g4_color"]      = 4,
-	["mac_gte_store_f3"]         = 3,
-	["mac_gte_store_g3"]         = 3,
-	["mac_gte_store_g4_p012"]    = 3,
-	["mac_gte_store_g4_p3"]      = 1,
-	["mac_insert_ot_tag_f3"]     = 1,
-	["mac_insert_ot_tag_g4"]     = 1,
-}
-
--- Per-macro cycle cost (best-case, no stalls). Used by the static-analysis pass to emit per-atom cycle budgets.
--- The counts cover the expanded instruction sequence the macro emits (not just the surface token in source).
--- Worked example — `mac_pack_color_word(off, cmd, r, g, b)` expands to:
---     load_upper_i(R_AT, (cmd << 8) | b)        -- 1 cycle
---     or_i_self(R_AT, (g << 8) | r)             -- 1 cycle
---     store_word(R_AT, R_PrimCursor, off)       -- 1 cycle
---                                                = 3 cycles total
---
--- `mac_yield` emits a control-transfer sequence (load_word, add_ui_self, jump_reg, nop). The atom body's cycle budget excludes
--- the yield's cost (we model it as 0); the runtime cost lands in the next atom's prologue.
---
 -- GTE command values are the GTE instruction's intrinsic cycles — the latency after any pre-cmd `nop2` has retired.
 -- When the source emits `nop2, gte_cmdw_X`, the nops' cycles are added separately (1+1) plus the gte_cmdw_X value here:
 --   rtpt  = 23 + 2 nops = 25 total cycles  (PSX-SPX says 23 cycles for the cmd itself; the nops are pre-fill)
@@ -1471,8 +1412,15 @@ M.GP0_MACRO_CONTRIB = {
 -- PSX-SPX reports the GTE intrinsic cycles as the total execution time of the command itself (rtpt=23, rtps=15, nclip=8, etc.).
 -- The pre-fill nops are a codebase convention for retiring preceding C2 writes.
 -- See `docs/psx-spx/docs/geometrytransformationenginegte.md` for per-command cycle counts and
--- `docs/psx-spx/docs/gtepipelinetimings.md` for the hardware-verified input-latch boundaries (most inputs become
+-- `docs/psx-spx/docs/gtepipelinetimings.md` for the hardware-verified input-latch boundaries
--- safe to clobber after 0-4 cycles).
+-- (most inputs become safe to clobber after 0-4 cycles).
+--
+-- Per-macro cycle costs (`mac_yield`, `mac_pack_color_word`, ...) and per-macro prim-buffer contributions
+-- (`mac_format_*_color`, `mac_gte_store_*`, `mac_insert_ot_tag_*`) are NOT hardcoded here.
+-- `passes/components.lua::compute_components_metadata` derives both from each `MipsAtomComp_(ac_X)` body in
+-- `code/duffle/lottes_tape.h`, stores the values on `corpus.components[name].cycle_cost` and
+-- `corpus.components[name].gp0_contrib`, and `passes/static_analysis.lua` reads those fields directly.
+-- The `mac_yield` cost is 0 by convention (the runtime cost lands in the next atom's prologue).
 M.INSTRUCTION_LATENCY = {
 	-- CPU ALU (single-cycle R3000A ops)
 	["nop"]                 = 1,
@@ -1517,17 +1465,19 @@ M.INSTRUCTION_LATENCY = {
 	["store_word"]          = 1,
 	["store_half"]          = 1,
 	["store_byte"]          = 1,
-	-- Branches (branch + BD slot nop = 2 cycles; the BD slot's nop is
+	-- Branches (branch + BD slot nop = 2 cycles; the BD slot's nop is counted as part of the branch's cost)
-	-- counted as part of the branch's cost)
 	["branch_equal"]        = 2,  ["branch_ne"]          = 2,
 	["branch_le_zero"]      = 2,  ["branch_lt_zero"]     = 2,
 	["branch_ge_zero"]      = 2,  ["branch_gt_zero"]     = 2,
+	-- `jump_rel(off)` is the within-atom-safe unconditional-jump alias for `branch_equal(R_0, R_0, off)` (see `code/duffle/mips.h`).
+	-- Same cost as the underlying branch (1 instruction + 1 mandatory BD-slot nop = 2 cycles).
+	["jump_rel"]            = 2,
 	-- Jumps (jump + BD slot nop = 2 cycles)
 	["jump"]                = 2,  ["jump_reg"]           = 2,
 	["jump_link"]           = 2,  ["call_reg"]           = 2,
 	["call_addr"]           = 2,
-	-- COP2 transfers (mtc2/mfc2/ctc2/cfc2 = 1 cycle + COP2 latency; the
+	-- COP2 transfers (mtc2/mfc2/ctc2/cfc2 = 1 cycle + COP2 latency; 
-	-- COP2 latency is usually absorbed by subsequent nops or by the next
+	-- The COP2 latency is usually absorbed by subsequent nops or by the next
 	-- GTE command's pre-fill nops, so we count 1)
 	["gte_mv_to_data_r"]    = 1,
 	["gte_mv_from_data_r"]  = 1,
@@ -1568,22 +1518,6 @@ M.INSTRUCTION_LATENCY = {
 	["gte_load_v2"]         = 2,
 	["gte_load_v0v1v2"]     = 6,
 
-	-- TODO(Ed): REMOVE THIS HARDCODE, THIS SHOULD BE RESOLVED AUTOMATICALLY
-	-- mac_* helpers (cycle cost = sum of the expanded instructions)
-	-- mac_yield transfers control; cycle budget is 0 (the next atom absorbs the cost).
-	["mac_yield"]              = 0,
-	["mac_pack_color_word"]    = 3,  -- lui + ori + sw
-	["mac_format_f3_color"]    = 3,  -- = mac_pack_color_word
-	["mac_format_g4_color"]    = 12, -- 4 x mac_pack_color_word
-	["mac_load_tri_indices"]   = 3,  -- 3 x lhu
-	["mac_gte_load_tri_verts"] = 18, -- 3 x {sll, addu, lw, lw, mtc2, mtc2}
-	["mac_gte_store_f3"]       = 3,
-	["mac_gte_store_g3"]       = 3,
-	["mac_gte_store_g4_p012"]  = 3,
-	["mac_gte_store_g4_p3"]    = 1,
-	["mac_insert_ot_tag_f3"]   = 11,  -- 11 .word slots in the macro body
-	["mac_insert_ot_tag_g4"]   = 11,
-
 	-- Annotation markers (emit no code; pure metaprogram hints)
 	["atom_label"]          = 0,
 	["atom_offset"]         = 0,
@@ -1600,8 +1534,7 @@ M.UNKNOWN_INSTRUCTION_CYCLES = 1
 
 -- Hardware-relation policy table.
 --
--- The forward walker in `passes/static_analysis.lua::analyze_hardware_relations` reads every emitted word_event,
+-- The forward walker in `passes/static_analysis.lua::analyze_hardware_relations` reads every emitted word_event, matches its `encoder` against `row.token`, and:
--- matches its `encoder` against `row.token`, and:
 --   * stages the event as a producer in `atom.paths.forward_state`; or
 --   * matches it as a consumer against pending producers and records a hazard on `atom.paths.hazards` when the gap is below `visibility.required`.
 --
@@ -1621,8 +1554,8 @@ M.UNKNOWN_INSTRUCTION_CYCLES = 1
 --     and is reserved for future "self-retires" relations.
 --
 -- Evidence:
---   * `evidence.confidence` is one of `"exact"`, `"conservative"`, `"unknown"`. The severity comes from `violation_kind`; a hardware
+--   * `evidence.confidence` is one of `"exact"`, `"conservative"`, `"unknown"`. The severity comes from `violation_kind`;
---     measurement that the vendor caveats may still classify as `"conservative"` even when the underlying timing is numerically known.
+--     A hardware measurement that the vendor caveats may still classify as `"conservative"` even when the underlying timing is numerically known.
 --   * `evidence.source` is the upstream reference (file + line range) the row is sourced from. New rows must carry this citation.
 --
 -- Consumers:
@@ -1732,22 +1665,42 @@ M.HARDWARE_RELATIONS = {
 	},
 	-- Memory -> COP2 data register (LWC2).
 	-- The memory-side timing is not measured by the vendored GTE latch experiment, so this relation has no numeric retirement threshold.
-	-- The forward walker emits one info edge at the first command-input consumer and then clears the pending relation.
+	-- The LWC2 destination has TWO retirement regimes (per PSX-SPX):
+	--   * GTE-command consumer (`gte_cmdw_*`): the GTE pipeline LATCHES the LWC2 result, so a `gte_cmdw_*`
+	--     in the very next slot uses the latched value. Gap = 0 is allowed. (Per `docs/psx-spx/docs/gtepipelinetimings.md:271-274`.)
+	--   * Any other consumer: standard MIPS load delay applies. Gap = 1 required. (Per `docs/psx-spx/docs/cpuspecifications.md:407-419`.)
+	-- Two separate relations so the walker can dispatch by consumer type and emit different severities
+	-- (the GTE-command path is `info` because the latch is intentional; the non-GTE-consumer path is `error` because the missing nop is a real bug).
 	{
-		id             = "lwc2_unknown_visibility",
+		id             = "lwc2_to_gte_command",
-		semantic       = "LWC2",
+		semantic       = "LWC2_to_GTE",
 		token          = "gte_lw",
 		direction      = "memory_to_cop2_data",
 		reads          = { domain = "memory",    arg = 2 },
 		writes         = { domain = "cop2.data", arg = 1 },
-		visibility     = { kind = "unknown_consumer", required = nil },
+		required       = 0,  -- GTE-command consumer: gap = 0 OK (latched).
 		evidence       = {
-			confidence = "unknown",
+			confidence = "measured",
 			source     = "gtepipelinetimings.md:271-274",
 		},
 		violation_kind = "info",
 		clear_on_consumer = true,
 	},
+	{
+		id             = "lwc2_to_other_consumer",
+		semantic       = "LWC2_to_other",
+		token          = "gte_lw",
+		direction      = "memory_to_cop2_data",
+		reads          = { domain = "memory",    arg = 2 },
+		writes         = { domain = "cop2.data", arg = 1 },
+		required       = 1,  -- Non-GTE-consumer: standard MIPS load delay.
+		evidence       = {
+			confidence = "inferred",
+			source     = "cpuspecifications.md:407-419",
+		},
+		violation_kind = "error",
+		clear_on_consumer = true,
+	},
 	-- COP2 data register -> memory (SWC2). A read of C2 state, not a CPU-to-COP2 write.
 	-- The policy row stays in for direction/provenance; staging it as a later command-input producer is suppressed.
 	{
@@ -2079,14 +2032,14 @@ local E_MAC_PREFIX_LEN = 4
 --- Expand a body entry into the flat sequence of emitted machine-word events.
 ---
 --- Semantics (one event per emitted machine word):
----   * **Direct one-word encoders** (`load_word`, `add_ui`, `nop`, `gte_lw`, ...): one event with `ident` = leading ident, `args` = parsed top-level args.
+---   * Direct one-word encoders `load_word`, `add_ui`, `nop`, `gte_lw`, ...: One event with `ident` = leading ident, `args` = parsed top-level args.
----   * **`nop2`** (2-word pseudo-instruction): two events, both with `ident = "nop"` so the recognized "this slot is a no-op" semantic is visible to downstream analyses.
+---   * `nop2` (2-word pseudo-instruction): Two events, both with `ident = "nop"` so the recognized "this slot is a no-op" semantic is visible to downstream analyses.
----   * **Any other N-word token** in `word_counts`: N events sharing the same `ident` + `args` so useful CPU words retire slots in the cycle budget.
+---   * Any other N-word token in `word_counts`: N events sharing the same `ident` + `args` so useful CPU words retire slots in the cycle budget.
----   * **Known `mac_X(...)` calls**: recursively expand the indexed component body, including nested components. Every event from the expansion carries:
+---   * Known `mac_X(...)` calls: Recursively expand the indexed component body, including nested components. Every event from the expansion carries:
 ---     - `source` / `line` = the COMPONENT'S source path + the line of the token within the component body (i.e. "definition site").
 ---     - `call_source` / `call_line` = the ROOT atom's source path + call-site line, PRESERVED across recursion so nested events still point at the original root.
----   * **Unknown `mac_X`** (not in `component_index`): fall back to `word_counts[ident]` if present; otherwise emit one opaque event so the cycle budget accounts for the word.
+---   * Unknown `mac_X` (not in `component_index`): fall back to `word_counts[ident]` if present; otherwise emit one opaque event so the cycle budget accounts for the word.
----   * **Marker tokens** (`atom_label(...)` / `atom_offset(...)`): zero events (they are pure metaprogram hints, not emitted machine words).
+---   * Marker Tokens (`atom_label(...)` / `atom_offset(...)`): Zero events (they are pure metaprogram hints).
 ---
 --- Cycle protection: a per-expansion `visiting` set tracks components currently on the expansion stack; a re-entry produces a deterministic `{kind = "cycle", ...}` error and aborts that branch (does NOT hang, does NOT recurse).
 ---
@@ -2110,12 +2063,12 @@ local E_MAC_PREFIX_LEN = 4
 -- word_counts table is authored-metadata + current-component count table.
 
 --- @class EmissionProjection
---- @field items       table[] -- ordered stream of word|label|offset|invoke_begin|invoke_end
+--- @field items       table[] -- Ordered stream of word|label|offset|invoke_begin|invoke_end
---- @field word_events table[] -- dense view of items where kind == "word"
+--- @field word_events table[] -- Dense view of items where kind == "word"
---- @field markers     table[] -- dense view of items where kind == "label"|"offset"
+--- @field markers     table[] -- Dense view of items where kind == "label"|"offset"
 --- @field invocations InvocationRecord[] -- dense view of items where kind == "invoke_begin"|"invoke_end"
---- @field errors      table[] -- token-resolution failures surfaced without fail-loud
+--- @field errors      table[] -- Token-resolution failures surfaced without fail-loud
---- @field warnings    table[] -- opaque warnings (e.g. unknown uncounted macro)
+--- @field warnings    table[] -- Opaque warnings (e.g. unknown uncounted macro)
 
 --- @class InvocationRecord
 --- Lives at `atom.paths.invocations[*]`. Constructed once at the single invocation-construction site
@@ -2123,20 +2076,20 @@ local E_MAC_PREFIX_LEN = 4
 --- @field id              integer  -- 1-based, monotonic per-atom invocation id (0 is reserved for "no open invocation")
 --- @field parent_id       integer  -- 0 for the outermost (root) call; otherwise the id of the immediately enclosing invocation
 --- @field kind            string   -- "comp_bare" | "comp_proc" (component form that triggered the expansion)
---- @field component_name  string   -- the bare component name without the `mac_` prefix
+--- @field component_name  string   -- Bare component name without the `mac_` prefix
---- @field call_text       string   -- the immediate `mac_X(...)` token text (or root call text for the outermost entry)
+--- @field call_text       string   -- Immediate `mac_X(...)` token text (or root call text for the outermost entry)
---- @field root_call_text  string   -- the IMMUTABLE outermost `mac_X(...)` token text for every word emitted in this call's expansion
+--- @field root_call_text  string   -- IMMUTABLE outermost `mac_X(...)` token text for every word emitted in this call's expansion
---- @field call_path       string   -- source path of the call site (root atom source for direct calls, component source for nested expansions)
+--- @field call_path       string   -- Source path of the call site (root atom source for direct calls, component source for nested expansions)
---- @field call_line       integer  -- source line of the call site
+--- @field call_line       integer  -- Source line of the call site
---- @field def_path        string   -- source path of the component definition
+--- @field def_path        string   -- Source path of the component definition
---- @field def_line        integer  -- source line of the component declaration
+--- @field def_line        integer  -- Source line of the component declaration
 --- @field start_pos       integer  -- 0-based emitted-word position of the FIRST word inside this invocation (the value of `word_idx` AT `emit_invoke_begin` time, BEFORE the first word is emitted). Words emitted inside this invocation occupy `start_pos..start_pos+#body_lines-1` (inclusive, 0-based). Downstream DWARF/provenance consumers MUST read this; do NOT reconstruct it from `start_word` (which is the 1-based items index including `invoke_begin`/`invoke_end` markers).
 --- @field end_pos         integer  -- 0-based position of the LAST word inside this invocation (set by `emit_invoke_end` to `word_idx - 1` AFTER all body words are emitted).
 --- @field start_word      integer  -- 1-based items index of the `invoke_begin` item
 --- @field end_word        integer  -- 1-based items index of the `invoke_end` item (set by `emit_invoke_end`)
---- @field word_count      integer  -- number of `word` items emitted between `start_word` and `end_word` (inclusive)
+--- @field word_count      integer  -- Number of `word` items emitted between `start_word` and `end_word` (inclusive)
 --- @field debug_skip      boolean  -- `debug_skip` stamp; true iff `corpus.components[name].debug_skip` is true at construction. Always boolean (never `nil`).
---- @field errors          table[]  -- per-invocation construction errors (cycle / count_mismatch); does not include pass-level errors
+--- @field errors          table[]  -- Per-invocation construction errors (cycle / count_mismatch); does not include pass-level errors
 
 -- Internal recursive walker. The items stream holds every emitted event in order; `word_events`, `markers`,
 -- `invocations`, `errors`, `warnings` are dense views / side outputs appended alongside.
@@ -2212,10 +2165,15 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 	end
 
 	local function emit_marker(kind, name, target, line,
-		immediate_call_text, root_call_text_w)
+		immediate_call_text, root_call_text_w,
+		consuming_encoder, consuming_arg_pos)
 		local inv_ids   = open_invocation_ids_snapshot()
 		local outermost = inv_ids[1] or 0
 		-- Markers carry the open invocation stack snapshot. `call_text` / `root_call_text` belong to words, not markers — markers are zero-width and skip per-word call-site attribution.
+		-- `consuming_encoder` + `consuming_arg_pos` carry the surrounding control-transfer instruction context
+		-- (e.g. `branch_le_zero` consuming its 3rd argument, or `jump` / `call_addr` consuming their only argument).
+		-- `passes/offsets.lua` reads these to dispatch per-consuming-instruction offset encoding.
+		-- nil for top-level markers (where the marker is the entire token — no surrounding consuming instruction).
 		local it = {
 			kind                     = kind,
 			name                     = name,
@@ -2225,6 +2183,8 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 			outermost_invocation_id  = outermost,
 		}
 		if target ~= nil then it.target = target end
+		if consuming_encoder then it.consuming_encoder = consuming_encoder end
+		if consuming_arg_pos  then it.consuming_arg_pos  = consuming_arg_pos  end
 		items[#items + 1] = it
 		markers[#markers + 1] = {
 			kind              = kind,
@@ -2232,39 +2192,102 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 			line              = line,
 			word_index        = word_idx,
 			target            = target,
+			consuming_encoder = consuming_encoder,
+			consuming_arg_pos = consuming_arg_pos,
 		}
 	end
 
-	local function emit_embedded_markers(tok, tok_line)
+	-- Count top-level commas in `tok` between position `from_pos` (inclusive) and `to_pos` (exclusive).
+	-- Tracks paren depth so commas inside nested () don't count. Skips string literals + comments.
+	-- Used by `emit_embedded_markers` to compute `consuming_arg_pos` for each embedded marker.
+	local function count_top_level_commas(tok, from_pos, to_pos)
+		local depth  = 0
+		local count  = 0
+		local i      = from_pos
+		while i < to_pos do
+			local c = tok:sub(i, i)
+			if c == "'" or c == '"' then
+				local next_pos = M.skip_str_or_cmt(tok, i)
+				i = (next_pos > i) and next_pos or (i + 1)
+			elseif c == "/" and tok:sub(i + 1, i + 1) == "/" then
+				-- line comment: skip to end of line
+				local nl = tok:find("\n", i, true)
+				i = (nl and nl + 1) or (#tok + 1)
+			elseif c == "/" and tok:sub(i + 1, i + 1) == "*" then
+				-- block comment: skip to matching */
+				local close = tok:find("*/", i + 2, true)
+				i = (close and close + 2) or (#tok + 1)
+			elseif c == "(" then
+				depth = depth + 1
+				i = i + 1
+			elseif c == ")" then
+				depth = depth - 1
+				i = i + 1
+			elseif c == "," and depth == 0 then
+				count = count + 1
+				i = i + 1
+			else
+				i = i + 1
+			end
+		end
+		return count
+	end
+
+	-- Find the position of the consuming instruction's open paren (the `(` that starts the consuming instruction's argument list).
+	-- Returns nil if the token's leading text isn't an ident followed by `(` (e.g. the ident is at the start of a non-instruction token).
+	local function find_consuming_paren(tok)
+		local i = 1
+		while i <= #tok do
+			local c = tok:sub(i, i)
+			if    c == "(" then return i end
+			if not c:match("[%w_]") and c ~= " " then return nil end
+			i = i + 1
+		end
+		return nil
+	end
+
+	local function emit_embedded_markers(tok, tok_line, consuming_encoder)
+		-- When called with a non-nil `consuming_encoder`, the marker is nested inside that instruction's argument list.
+		-- We compute each marker's arg position by counting top-level commas between the consuming instruction's `(` and the marker's start.
+		local consuming_paren = nil
+		if consuming_encoder then consuming_paren = find_consuming_paren(tok) end
 		local pos = 1
 		while pos <= #tok do
-			-- trim leading whitespace and comments before each scan.
+			-- Trim leading whitespace and comments before each scan.
 			pos = M.skip_ws_and_cmt(tok, pos)
 			if pos > #tok then break end
 			local  ident, after = M.read_ident(tok, pos)
 			if not ident then
-				-- not an ident: token is a string or comment; skip or one-step.
+				-- Not an ident: token is a string or comment; skip or one-step.
 				local next_pos = M.skip_str_or_cmt(tok, pos)
 				pos = (next_pos > pos) and next_pos or (pos + 1)
 				goto continue_loop
 			end
 			if ident ~= "atom_label" and ident ~= "atom_offset" then
-				-- ordinary ident; nothing to emit, step past the ident only.
+				-- Ordinary ident; nothing to emit, step past the ident only.
 				pos = after
 				goto continue_loop
 			end
-			-- marker ident: parse the (...) arguments.
+			-- Marker ident: parse the (...) arguments.
 			local open               = M.skip_ws_and_cmt(tok, after)
 			local inner, after_paren = M.read_parens(tok, open)
 			if not inner then
-				-- (...) unreadable: fall back to non-marker behavior.
+				-- (...) Unreadable: fall back to non-marker behavior.
 				pos = after
 				goto continue_loop
 			end
-			-- commit: label takes 1 arg, offset takes 2.
+			-- Commit: label takes 1 arg, offset takes 2.
+			-- For embedded markers, propagate the consuming_encoder + the marker's arg position
+			-- (1-based) so `passes/offsets.lua` can dispatch per-consuming-instruction offset encoding.
+			-- Top-level markers (no consuming_encoder) get nil for both — the offsets pass treats
+			-- them as branch-equivalent for backward compatibility.
+			local arg_pos = nil
+			if consuming_encoder and consuming_paren then
+				arg_pos = count_top_level_commas(tok, consuming_paren + 1, pos) + 1
+			end
 			local args = split_top_level_args(inner)
-			if ident == "atom_label" then emit_marker("label",  args[1] or "", nil,           tok_line) 
+			if ident == "atom_label" then emit_marker("label",  args[1] or "", nil,           tok_line, nil, nil, consuming_encoder, arg_pos)
-			else                          emit_marker("offset", args[1] or "", args[2] or "", tok_line)
+			else                          emit_marker("offset", args[1] or "", args[2] or "", tok_line, nil, nil, consuming_encoder, arg_pos)
 			end
 			pos = after_paren
 			::continue_loop::
@@ -2276,7 +2299,6 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 		next_inv_id = next_inv_id + 1
 		-- Invocation-level debug_skip stamp: Emission pass owns `atom.paths.invocations[*].debug_skip`.
 		-- The stamp is resolved from the `corpus.components[name]` registry (passed in via `ctx_table.components` by `emission_model.run`), 
-		-- NOT from a parallel skip map, source-text re-parse, or second pass over `invocations`.
 		-- Unmarked components stamp `false` (not `nil`) so consumers can dispatch on the boolean without nil checks.
 		--
 		-- The walker has already found the component body in `ctx_table.component_index[component_name]`, so the matching entry MUST exist in `ctx_table.components[component_name]`
@@ -2347,7 +2369,6 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 	end
 
 	-- Resolve the per-token word count. If unresolved, surface ONE warning
-	-- (NOT an error; the build does not fail-loud on an uncounted opaque word)
 	-- and fall back to 1 opaque word so the cycle budget still accounts for the slot.
 	local function resolve_count(ident, tok_line)
 		local wc = ctx_table.word_counts
@@ -2371,10 +2392,10 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 	end
 
 	-- Recursive walker: walk one body entry, possibly descending into components.
-	-- `walk_parent_inv_id` is the invocation ID of the enclosing call (0 for the root call).
+	-- walk_parent_inv_id:       Invocation ID of the enclosing call (0 for the root call).
-	-- `walk_root_call_text` is the outermost `mac_X(...)` token text (preserved across recursion).
+	-- walk_root_call_text:      Outermost `mac_X(...)` token text (preserved across recursion).
-	-- `walk_immediate_call_text` is the IMMEDIATE outer `mac_X(...)` token text for words emitted in this body — nil for the root atom body.
+	-- walk_immediate_call_text: IMMEDIATE outer `mac_X(...)` token text for words emitted in this body — nil for the root atom body.
-	-- The two trackers are propagated as separate parameters so words deep inside nested expansions correctly identify both their immediate call site and the outermost call site.
+	-- Two trackers are propagated as separate parameters so words deep inside nested expansions correctly identify both their immediate call site and the outermost call site.
 	local function walk_body_entry(body_entry, walk_parent_inv_id,
 		walk_root_call_text, walk_immediate_call_text)
 		local tokens     = body_entry.body_tokens or {}
@@ -2391,8 +2412,18 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 			local _, args  = token_ident_and_args(tok)
 			local tok_line = line_of(body_off + bt.rel) or 0
 			-- embedded markers live only in non-marker tokens.
-			if ident ~= "atom_label" and ident ~= "atom_offset" then emit_embedded_markers(tok, tok_line) end
+			-- Pass `ident` as the consuming instruction so `emit_embedded_markers` can compute each marker's arg position + record the consuming_encoder for the offsets pass.
+			-- Canonicalize `jump_rel` to `branch_equal` (its preprocessor-expanded form) so the `consuming_encoder` metadata in marker records is canonical. 
+			-- `jump_rel`: unconditional jump alias from `code/duffle/mips.h`.
+			local consuming_encoder_for_markers = (ident == "jump_rel") and "branch_equal" or ident
+			if ident ~= "atom_label" and ident ~= "atom_offset" then
+				emit_embedded_markers(tok, tok_line, consuming_encoder_for_markers)
+			end
 			-- atom_label / atom_offset: terminal markers, no further descent.
+			-- Top-level markers (the marker IS the entire token) have no consuming instruction;
+			-- nil for both `consuming_encoder` and `consuming_arg_pos`.
+			-- The offsets pass treats these as branch-equivalent for backward compatibility.
+			-- TODO(Ed): Review this don't want legacy cruft here..
 			if     ident == "atom_label"  then emit_marker("label",  args[1] or "", nil,           tok_line); return
 			elseif ident == "atom_offset" then emit_marker("offset", args[1] or "", args[2] or "", tok_line); return
 			end
@@ -2402,7 +2433,7 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 				if comp then
 					local invocation_root_call_text = walk_root_call_text or tok
 					if ctx_table.visiting[bare] then
-						-- cycle: still allocate inv_id, emit zero-width begin/end, record the cycle error; do NOT recurse.
+						-- Cycle: still allocate inv_id, emit zero-width begin/end, record the cycle error; do NOT recurse.
 						local inv = emit_invoke_begin(comp.kind or "comp_bare", bare, tok, invocation_root_call_text, def_source, tok_line)
 						inv.parent_id = walk_parent_inv_id
 						inv.call_text = tok
@@ -2417,14 +2448,14 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 						emit_invoke_end(inv)
 						return
 					end
-					-- first visit: descend + count + count_mismatch-check below.
+					-- First visit: descend + count + count_mismatch-check below.
 					ctx_table.visiting[bare] = true
 					local inv = emit_invoke_begin(comp.kind or "comp_bare", bare, tok, invocation_root_call_text, def_source, tok_line)
 					inv.parent_id = walk_parent_inv_id
 					inv.call_text = tok
 					inv.def_path  = comp.source
 					inv.def_line  = comp.declaration
-					-- propagate trackers into the recursive walk:
+					-- Propagate trackers into the recursive walk:
 					--   immediate_call_text = this call's tok (the IMMEDIATE outer call for words emitted in this body)
 					--   root_call_text      = the OUTERMOST call (immutable across the recursion)
 					walk_body_entry({
@@ -2439,7 +2470,7 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 						tok)
 					ctx_table.visiting[bare] = nil
 					emit_invoke_end(inv)
-					-- count `word` items inside [start_word, end_word].
+					-- Count `word` items inside [start_word, end_word].
 					local wc_inside = 0
 					for i = inv.start_word, inv.end_word do
 						local it = items[i]
@@ -2465,8 +2496,8 @@ local function _project_emission_inner(root_body_entry, ctx_table)
 				end
 				-- mac_X NOT in component_index: fall through to opaque emit.
 			end
-			-- direct encoder, or mac_X-without-component: resolve count + emit n words.
+			-- Direct encoder, or mac_X-without-component: resolve count + emit n words.
-			-- resolve_count may emit a warning if the count is unresolved.
+			-- Resolve_count may emit a warning if the count is unresolved.
 			local n         = resolve_count(ident, tok_line)
 			local out_ident = (ident == "nop2") and "nop" or ident
 			for _ = 1, n do

scripts/elf32.lua

@@ -0,0 +1,418 @@
+-- elf32.lua — Pure-Lua ELF32 format helpers with no lfs / no lpeg dependency.
+-- The reload helper's `parse_manifest` (scripts/pcsx_debug_helper/reload.lua) 
+-- and the metaprogram's `read_elf_sections` + `read_nm` (scripts/elf_dwarf.lua)
+-- both parsed ELF32 headers from wire bytes.
+--
+-- This module contains the format constants and the byte-level walker.
+--- The metaprogram side keeps `read_u32_le` / `read_u16_le` as local forwarders; the helper side calls `E.*` directly.
+--
+-- **Adapter contract (explicit pass style):**
+-- The helper VM's `Support.File` exposes byte-read methods that require `self` (fileffi.lua:225-227),
+-- so callers wrap once in a 1-line adapter that strips `self`.
+-- The parsers here operate on the unwrapped form.
+-- Reads are flat function calls — `E.read_u8(adapter, off)`, `E.read_u32(adapter, off)`, `E.size(adapter)`.
+--   read_u8(adapter, off)   -> integer | nil
+--   read_u16(adapter, off)  -> integer | nil
+--   read_u32(adapter, off)  -> integer | nil
+--   size(adapter)           -> integer
+--
+-- **Convention:** every offset in the constants tables is a zero-based wire offset.
+-- The `+ 1` conversion happens only at the `string.byte` boundary inside the readers.
+--
+-- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
+-- spec: System V ABI gABI v1.2 §"Symbol Table" (Elf32_Sym layout)
+
+local M = {}
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Little-endian readers (bit-weighted accumulator, math.floor only)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read a 4-byte little-endian unsigned integer from `adapter` at zero-based wire offset `off`.
+---
+--- Bit weights are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
+--- byte 0 contributes its value directly;
+--- byte 1 is shifted left by 8; byte 2 by 16; byte 3 by 24.
+---
+--- math.floor (not LuaJIT's `>>`) keeps the body portable across LuaJIT 2.0/2.1 and plain Lua 5.x. `string.byte` receives `+ 1` at the boundary.
+---
+--- **Call form:** explicit-pass. The reader receives `adapter` as the first positional argument and the offset as the second; no `self` is passed.
+--- Test fixtures declare `function(offset) ... end` and the parsers call them via dot syntax `adapter.read_u8_at(off)`.
+--- The colon form `adapter:read_u8_at(off)` would prepend the adapter table as `offset` and break the contract.
+--- @param adapter table
+--- @param off     integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u32(adapter, off)
+	return adapter.read_u8_at(off)
+		+ adapter.read_u8_at(off + 0x01) * 0x00000100
+		+ adapter.read_u8_at(off + 0x02) * 0x00010000
+		+ adapter.read_u8_at(off + 0x03) * 0x01000000
+end
+
+--- Read a 2-byte little-endian unsigned integer from `adapter` at zero-based wire offset `off`.
+--- @param adapter table
+--- @param off integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u16(adapter, off)
+	return adapter.read_u8_at(off)
+		+ adapter.read_u8_at(off + 0x01) * 0x00000100
+end
+
+--- Read a 1-byte unsigned integer from `adapter` at zero-based wire offset `off`.
+--- @param adapter table
+--- @param off integer  -- zero-based wire offset
+--- @return integer|nil
+function M.read_u8(adapter, off)
+	return adapter.read_u8_at(off)
+end
+
+--- Total adapter byte length.
+--- @param adapter table
+--- @return integer
+function M.size(adapter)
+	return adapter.read_size()
+end
+
+--- Forwarders kept for backward compat with scripts/elf_dwarf.lua.
+--- The metaprogram side keeps `read_u32_le` / `read_u16_le`;
+--- both layers now use the same byte-level helpers under the hood.
+function M.read_u32_le(buf, off)
+	local byte_off = off + 1
+	return buf:byte(byte_off)
+		+ buf:byte(byte_off + 0x01) * 0x00000100
+		+ buf:byte(byte_off + 0x02) * 0x00010000
+		+ buf:byte(byte_off + 0x03) * 0x01000000
+end
+
+--- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
+--- @param buf string
+--- @param off integer  -- zero-based wire offset
+--- @return integer
+function M.read_u16_le(buf, off)
+	local byte_off = off + 1
+	return buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Format constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- ELF format constants (System V ABI gABI v1.2).
+M.ELFCLASS32  = 1    -- spec: gABI v1.2 §"ELF Header" — EI_CLASS byte
+M.ELFDATA2LSB = 1    -- spec: gABI v1.2 §"ELF Header" — EI_DATA byte
+M.EM_MIPS     = 8    -- spec: gABI v1.2 §"Machine Information" — MIPS architecture
+
+-- Section type constants (System V ABI gABI v1.2 §"Section Header Table").
+M.SHT_SYMTAB = 2     -- spec: gABI v1.2 §"Section Types" — symbol table
+M.SHT_STRTAB = 3     -- spec: gABI v1.2 §"Section Types" — string table
+M.SHT_NOBITS = 8     -- spec: gABI v1.2 §"Section Types" — no space in file
+
+-- Section flag constants (System V ABI gABI v1.2 §"Section Header Table").
+M.SHF_WRITE     = 0x1 -- spec: gABI v1.2 §"Section Attributes" — writable
+M.SHF_ALLOC     = 0x2 -- spec: gABI v1.2 §"Section Attributes" — occupies memory
+M.SHF_EXECINSTR = 0x4 -- spec: gABI v1.2 §"Section Attributes" — executable
+
+-- ---------------------------------------------------------------------------
+-- ELF32 header layout (System V ABI gABI v1.2 §"ELF Header" Table 1)
+-- ---------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets. The header is 52 bytes total (header_bytes = 0x34 = 52).
+M.ELF32_HEADER = {
+	magic_offset         = 0x00,  -- 4 bytes; expected "\127ELF"
+	magic                = "\127ELF",
+	class_offset         = 0x04,  -- 1 byte;  1 = ELF32, 2 = ELF64
+	endian_offset        = 0x05,  -- 1 byte;  1 = little-endian, 2 = big-endian
+	header_bytes         = 0x34,  -- ELF32 header is 52 bytes total
+	e_entry_offset       = 0x18,  -- 4-byte LE; entry-point virtual address
+	e_shoff_offset       = 0x20,  -- 4-byte LE; section-header table file offset
+	e_shentsize_offset   = 0x2E,  -- 2-byte LE; section-header entry size in bytes
+	e_shnum_offset       = 0x30,  -- 2-byte LE; number of section headers
+	e_shstrndx_offset    = 0x32,  -- 2-byte LE; index of section-name string table
+}
+
+-- ---------------------------------------------------------------------------
+-- ELF32 section-header layout (System V ABI gABI v1.2 §"Section Header Table")
+-- ---------------------------------------------------------------------------
+-- Each entry is 40 bytes (sh_entsize_bytes = 0x28 = 40);
+-- zero-based, field offsets relative to the start of the entry.
+M.ELF32_SECTION = {
+	sh_name_offset    = 0x00,  -- 4-byte LE; offset into .shstrtab
+	sh_type_offset    = 0x04,  -- 4-byte LE; section type (SHT_*)
+	sh_flags_offset   = 0x08,  -- 4-byte LE; section flags (SHF_*)
+	sh_addr_offset    = 0x0C,  -- 4-byte LE; virtual address at execution
+	sh_offset_offset  = 0x10,  -- 4-byte LE; section's file offset
+	sh_size_offset    = 0x14,  -- 4-byte LE; section's size in bytes
+	sh_link_offset    = 0x18,  -- 4-byte LE; link to a related section
+	sh_entsize_bytes  = 0x28,  -- spec: gABI v1.2 §"Section Header Table" — 40 bytes per entry
+}
+
+-- ---------------------------------------------------------------------------
+-- ELF32 symbol-table entry layout (System V ABI gABI v1.2 §"Symbol Table")
+-- ---------------------------------------------------------------------------
+-- Each entry is 16 bytes (sym_entry_bytes = 0x10 = 16);
+-- zero-based, field offsets relative to the start of the entry.
+M.ELF32_SYM = {
+	st_name          = 0x00,  -- 4-byte LE; offset into the linked string table
+	st_value         = 0x04,  -- 4-byte LE; symbol value (address / absolute)
+	st_size          = 0x08,  -- 4-byte LE; symbol size in bytes
+	st_info          = 0x0C,  -- 1 byte;  binding (high nibble) + type (low nibble)
+	sym_entry_bytes  = 0x10,  -- spec: gABI v1.2 §"Symbol Table" — 16 bytes per entry
+}
+
+-- DWARF32 initial-length terminator (DWARF4 §7.4) — kept here so the metaprogram's elf_dwarf.lua can drop its own copy of the same constant.
+M.dw_dwarf32_terminator = 0xFFFFFFFF
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Adapter validation
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Validate that `adapter` exposes the byte-read surface.
+--- Returns true on success, false + a stable error code on failure.
+--- The helper side calls this before parse_manifest to reject callers before any byte is read.
+--- @param adapter any
+--- @return boolean, string|nil
+function M.validate_adapter(adapter)
+	if type(adapter) ~= "table" then return false, "bad_file_adapter" end
+	if type(adapter.read_u8_at)  ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_u16_at) ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_u32_at) ~= "function" then return false, "bad_file_adapter" end
+	if type(adapter.read_size)   ~= "function" then return false, "bad_file_adapter" end
+	return true, nil
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- String-table reader
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Extract a NUL-terminated C string from `strtab` at zero-based offset `off`.
+--- Returns nil if `off` is out of range or the string is not NUL-terminated.
+--- @param strtab string
+--- @param off integer
+--- @return string|nil
+function M.get_str(strtab, off)
+	if off < 0 or off >= #strtab then return nil end
+	local  end_pos = strtab:find("\0", off + 1, true)
+	if not end_pos then return nil end
+	return strtab:sub(off + 1, end_pos - 1)
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Header / section / symbol walkers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Read the ELF32 header through `adapter` and validate the magic, class, and data encoding.
+--- Returns a table on success:
+---   { e_entry, e_shoff, e_shentsize, e_shnum, e_shstrndx, error = nil }
+--- On failure returns nil + a stable error code:
+---   bad_magic, unsupported_elf_class, unsupported_elf_data, truncated_header
+--- The header's machine field is NOT validated here — callers (e.g. the helper's prime path) decide whether to require EM_MIPS before symbol reads.
+--- @param adapter table
+--- @return table|nil, string|nil
+function M.parse_elf32_headers(adapter)
+	local  ok, err = M.validate_adapter(adapter)
+	if not ok then return nil, err end
+
+	-- 4-byte magic: 0x7F 'E' 'L' 'F'.
+	-- The byte readers take the adapter explicitly.
+	-- The production `Support.File` adapter is wrapped by the caller to drop its implicit `self` so the parser shape is flat pass-style.
+	local b1 = M.read_u8(adapter, 0)
+	local b2 = M.read_u8(adapter, 1)
+	local b3 = M.read_u8(adapter, 2)
+	local b4 = M.read_u8(adapter, 3)
+	if not (b1 and b2 and b3 and b4)
+		or not (b1 == 0x7f and b2 == 0x45 and b3 == 0x4c and b4 == 0x46) then
+		return nil, "bad_magic"
+	end
+
+	local class = M.read_u8(adapter, M.ELF32_HEADER.class_offset)
+	if    class ~= M.ELFCLASS32 then
+		return nil, "unsupported_elf_class"
+	end
+
+	local data = M.read_u8(adapter, M.ELF32_HEADER.endian_offset)
+	if    data ~= M.ELFDATA2LSB then
+		return nil, "unsupported_elf_data"
+	end
+
+	local e_entry     = M.read_u32(adapter, M.ELF32_HEADER.e_entry_offset)
+	local e_shoff     = M.read_u32(adapter, M.ELF32_HEADER.e_shoff_offset)
+	local e_shentsize = M.read_u16(adapter, M.ELF32_HEADER.e_shentsize_offset)
+	local e_shnum     = M.read_u16(adapter, M.ELF32_HEADER.e_shnum_offset)
+	local e_shstrndx  = M.read_u16(adapter, M.ELF32_HEADER.e_shstrndx_offset)
+	if not (e_entry and e_shoff and e_shentsize and e_shnum and e_shstrndx) then
+		return nil, "truncated_header"
+	end
+
+	return {
+		e_entry     = e_entry,
+		e_shoff     = e_shoff,
+		e_shentsize = e_shentsize,
+		e_shnum     = e_shnum,
+		e_shstrndx  = e_shstrndx,
+		error       = nil,
+	}
+end
+
+--- Read one section-header entry from `adapter` at `sh_off`.
+--- Returns a table with the wire fields plus a (yet-unresolved) `name` field.
+--- @param adapter table
+--- @param sh_off  integer
+--- @return table|nil, string|nil  -- entry, error
+local function read_section_entry(adapter, sh_off)
+	local entry = {
+		sh_name   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_name_offset),
+		sh_type   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_type_offset),
+		sh_flags  = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_flags_offset),
+		sh_addr   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_addr_offset),
+		sh_offset = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_offset_offset),
+		sh_size   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_size_offset),
+		sh_link   = M.read_u32(adapter, sh_off + M.ELF32_SECTION.sh_link_offset),
+		name      = "",
+	}
+	if not (entry.sh_name and entry.sh_type and entry.sh_flags and entry.sh_addr
+		and entry.sh_offset and entry.sh_size and entry.sh_link) then
+		return nil, "truncated_section_headers"
+	end
+	return entry, nil
+end
+
+--- Walk every section header in `hdr` and return a 1-based array of entries
+--- (the section at logical index 0 is at array position 1, etc.).
+--- Each entry has the wire fields plus a resolved `name` derived from `.shstrtab`.
+--- Returns nil + a stable error code on failure: truncated_section_headers, missing_shstrtab, truncated_strtab
+--- @param adapter table
+--- @param hdr     table  -- the table returned by parse_elf32_headers
+--- @return table|nil, string|nil
+function M.walk_sections(adapter, hdr)
+	if not hdr or hdr.error then return nil, hdr and hdr.error or "truncated_section_headers" end
+
+	local file_size = M.size(adapter)
+	if hdr.e_shoff + hdr.e_shnum * hdr.e_shentsize > file_size then
+		return nil, "truncated_section_headers"
+	end
+
+	-- Read every section header first; we need .shstrtab to resolve names.
+	local sections = {}
+	for i = 0, hdr.e_shnum - 1 do
+		local sh_off = hdr.e_shoff + i * hdr.e_shentsize
+		local entry, err = read_section_entry(adapter, sh_off)
+		if not entry then return nil, err end
+		sections[i + 1] = entry
+	end
+
+	if hdr.e_shstrndx >= hdr.e_shnum then
+		return nil, "missing_shstrtab"
+	end
+
+	local  shstrtab = sections[hdr.e_shstrndx + 1]
+	if not shstrtab or shstrtab.sh_type ~= M.SHT_STRTAB then
+		return nil, "missing_shstrtab"
+	end
+	if shstrtab.sh_offset + shstrtab.sh_size > file_size then
+		return nil, "truncated_section_headers"
+	end
+	local  shstrtab_bytes = M.read_section_bytes(adapter, shstrtab)
+	if not shstrtab_bytes then return nil, "truncated_section_headers" end
+
+	for _, s in ipairs(sections) do
+		s.name = M.get_str(shstrtab_bytes, s.sh_name) or ""
+	end
+
+	return sections, nil
+end
+
+--- Read the bytes of one section. Returns a string, or nil if the adapter returns nil for any byte (out-of-bounds).
+--- The caller is responsible fors sizing the buffer (the section's sh_offset + sh_size must fit in adapter.size).
+--- @param adapter table
+--- @param section table  -- one entry from walk_sections
+--- @return string|nil
+function M.read_section_bytes(adapter, section)
+	local size = section.sh_size
+	if    size == 0 then return "" end
+	local out = {}
+	for i = 0, size - 1 do
+		local b = M.read_u8(adapter, section.sh_offset + i)
+		if    b == nil then return nil end
+		out[#out + 1] = string.char(b)
+	end
+	return table.concat(out)
+end
+
+--- Convenience: walk sections, then look up the named section, then read its bytes.
+--- Returns nil + a stable error code if the section is absent or out-of-bounds.
+--- @param adapter  table
+--- @param sections table  -- 1-based array from walk_sections
+--- @param name     string
+--- @return string|nil, string|nil
+function M.read_named_section(adapter, sections, name)
+	if not sections then return nil, "missing_section" end
+	for _, s in ipairs(sections) do
+		if   s.name == name then
+			local bytes = M.read_section_bytes(adapter, s)
+			if not bytes then return nil, "truncated_section_data" end
+			return bytes, nil
+		end
+	end
+	return nil, "missing_section"
+end
+
+--- Walk every SHT_SYMTAB section in `sections` and accumulate symbols by name.
+--- Each stored entry is `{ value = st_value, size = st_size, info = st_info, shndx = st_shndx }`.
+--- Both STB_LOCAL and STB_GLOBAL symbols are included; the live ELF stores `smem` as a local symbol.
+--- Returns nil + a stable error code on failure: missing_symtab_strtab, truncated_section_headers
+--- @param adapter  table
+--- @param sections table
+--- @return table|nil, string|nil
+function M.collect_symbols(adapter, sections)
+	if not sections then return nil, "missing_sections" end
+	local symbols = {}
+	local file_size = M.size(adapter)
+	for _, s in ipairs(sections) do
+		if s.sh_type == M.SHT_SYMTAB then
+			local  strtab = sections[s.sh_link + 1]
+			if not strtab or strtab.sh_type ~= M.SHT_STRTAB then
+				return nil, "missing_symtab_strtab"
+			end
+			if strtab.sh_offset + strtab.sh_size > file_size then
+				return nil, "truncated_section_headers"
+			end
+			local  strtab_bytes = M.read_section_bytes(adapter, strtab)
+			if not strtab_bytes then return nil, "truncated_section_headers" end
+			if s.sh_offset + s.sh_size > file_size then
+				return nil, "truncated_section_headers"
+			end
+			local  symtab_bytes = M.read_section_bytes(adapter, s)
+			if not symtab_bytes then return nil, "truncated_section_headers" end
+			local n = #symtab_bytes / M.ELF32_SYM.sym_entry_bytes
+			for j = 0, n - 1 do
+				local e = s.sh_offset + j * M.ELF32_SYM.sym_entry_bytes
+				local st_name  = M.read_u32(adapter, e + M.ELF32_SYM.st_name)
+				if    st_name then
+					local st_value = M.read_u32(adapter, e + M.ELF32_SYM.st_value)
+					local st_size  = M.read_u32(adapter, e + M.ELF32_SYM.st_size)
+					local st_info  = M.read_u8(adapter, e + M.ELF32_SYM.st_info)
+					-- st_shndx is at offset 14 (2 bytes) — derived from the layout
+					-- the metaprogram reads too. Inline the read to keep the
+					-- adapter as the only I/O surface.
+					local b1 = M.read_u8(adapter, e + 14)
+					local b2 = M.read_u8(adapter, e + 15)
+					if not (b1 and b2) then
+						return nil, "truncated_section_headers"
+					end
+					local st_shndx = b1 + b2 * 0x100
+					local name = M.get_str(strtab_bytes, st_name) or ""
+					if    name ~= "" then
+						symbols[name] = {
+							value = st_value,
+							size  = st_size,
+							info  = st_info,
+							shndx = st_shndx,
+						}
+					end
+				end
+			end
+		end
+	end
+	return symbols, nil
+end
+
+return M

scripts/elf_dwarf.lua

@@ -1,13 +1,8 @@
 --- elf_dwarf.lua — ELF32 + DWARF + atoms source-map utilities.
---- All ELF32 + DWARF-specific code lives here.
----
 --- **What this module contains:**
 ---   - **Format-constant tables** (the byte-offset / opcode / size encyclopedias for ELF32, DWARF4 aranges, DWARF5 rnglists, DWARF line-program, MIPS).
 ---     Every constant carries a spec:` comment naming the spec section that defines it.
 ---   - **I/O helpers**: little-endian byte read/write, ELF32 section walker, nm symbol reader, source-map parser, native directory glob.
----
---- **Conventions:** tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Native dependencies
@@ -16,6 +11,11 @@
 -- lfs is wired into package.cpath by `duffle_paths.lua` (vendored under `toolchain/lfs/lfs.dll`). 
 local lfs = require("lfs")
 
+-- scripts/elf32.lua contains format-constant tables + the byte-level walker.
+-- The this file re-exports `read_u32_le` / `read_u16_le` (and the DWARF32 terminator).
+-- TODO(Ed): Remove re-export.
+local E = require("elf32")
+
 local M = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -104,34 +104,16 @@ M.MIPS_BYTES_PER_WORD = 0x04
 -- ----------------------------------------------------------------------------
 -- ELF32 (System V ABI gABI v1.2)
 -- ----------------------------------------------------------------------------
---- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors,
+--- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors, and section-relative values are zero-based wire offsets. 
---- and section-relative values are zero-based wire offsets. Only Lua string APIs receive
+--- Only Lua string APIs receive a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
---- a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- ELF/DWARF field offsets are expressed in hex so they map directly to the zero-based byte positions in the binary file.
 ---
---- ELF/DWARF field offsets are expressed in hex so they map directly to the
+--- The ELF32 header / section / sym layout tables are within scripts/elf32.lua.
---- zero-based byte positions in the binary file.
+--- The metaprogram re-exports the DWARF32 initial-length terminator.
 
-
+--- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
---- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
+M.dw_dwarf32_terminator = E.dw_dwarf32_terminator
-M.ELF32 = {
+-- TODO(Ed): Remove re-export.
-	magic_offset          = 0x00,         -- 4-byte magic "\127ELF" at file offset 0x00
-	magic                 = "\127ELF",
-	class_offset          = 0x04,         -- 1-byte; 1 = ELF32, 2 = ELF64
-	class_elf32           = 1,
-	endian_offset         = 0x05,         -- 1-byte; 1 = little-endian, 2 = big-endian
-	endian_little         = 1,
-	header_bytes          = 0x34,         -- spec: gABI v1.2 §"ELF Header" — ELF32 header is 52 bytes total
-	e_shoff_offset        = 0x20,         -- 4-byte LE; section-header table file offset
-	e_shentsize_offset    = 0x2E,         -- 2-byte LE; section-header entry size in bytes
-	e_shnum_offset        = 0x30,         -- 2-byte LE; number of section headers
-	e_shstrndx_offset     = 0x32,         -- 2-byte LE; index of section-name string table
-	sh_size_bytes         = 0x28,         -- spec: gABI v1.2 §"Section Header Table" — each entry is 40 bytes
-	sh_name_offset        = 0x00,         -- 4-byte LE; offset into .shstrtab
-	sh_type_offset        = 0x04,         -- 4-byte LE; section type (SHT_*)
-	sh_offset_offset      = 0x10,         -- 4-byte LE; section's file offset
-	sh_size_offset        = 0x14,         -- 4-byte LE; section's size in bytes
-	dw_dwarf32_terminator = 0xFFFFFFFF,   -- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
-}
 
 -- ----------------------------------------------------------------------------
 -- DWARF4 .debug_aranges (per DWARF5 spec §7.4 — Address Range Table)
@@ -192,8 +174,7 @@ M.DWARF_LINE_OPS = {
 	DW_LNE_end_sequence = 1,    -- spec: §6.2.5.3
 	DW_LNE_set_address  = 2,    -- spec: §6.2.5.3
 	-- Standard opcode header (§6.2.5.1)
-	-- opcode_base + line_range are 1-byte header fields; hex so they map
+	-- opcode_base + line_range are 1-byte header fields; hex so they map directly to the line-program header byte sequence.
-	-- directly to the line-program header byte sequence.
 	-- line_base stays signed decimal (=-5) since 0xFB obscures the spec semantics.
 	opcode_base              = 0x0D,
 	line_base                = -5,
@@ -249,37 +230,31 @@ M.DWARF5_DEBUG_LINE = {
 --- Read a 4-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- Equivalent to `string.unpack("<I4", buf, off + 1)` but avoids the table-return shape + works under LuaJIT 2.1
 --- (which has partial `string.unpack` coverage).
----
 --- **Convention:** `off` is a zero-based wire offset; `+ 1` is applied only at the `string.byte` boundary.
 ---
---- **Byte weights** are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
+--- Thin forwarder: the canonical implementation lives in scripts/elf32.lua.
---- byte 0 contributes its value directly; byte 1 is shifted left by 8
+--- The "second caller lifts" pattern keeps the metaprogram side fluent
---- (= 0x100); byte 2 by 16 (= 0x10000); byte 3 by 24 (= 0x1000000).
+--- (`M.read_u32_le(buf, off)`) while the body is deduped.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u32_le(buf, off)
-	local byte_off = off + 1
+	return E.read_u32_le(buf, off)
-	return buf:byte(byte_off)
-		+ buf:byte(byte_off + 0x01) * 0x00000100
-		+ buf:byte(byte_off + 0x02) * 0x00010000
-		+ buf:byte(byte_off + 0x03) * 0x01000000
 end
 
 --- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- (`off` is zero-based; `+ 1` is applied only at the `string.byte` boundary.)
+--- Thin forwarder — see `M.read_u32_le` for the rationale.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u16_le(buf, off)
-	local byte_off = off + 1
+	return E.read_u16_le(buf, off)
-	return buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
 end
 
 -- Pure-Lua 5.3 LEB128 readers (no `bit` library). `2^shift` arithmetic matches the existing parser.
 -- Offsets are 0-based; returns (value, next_pos).
--- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua
+-- 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
--- can import them as file-scope locals per the 2nd-caller lift precedent
 -- (the uleb128 + sleb128 encoders were promoted the same way).
 function M.read_uleb128_at(buf, pos)
 	local value, shift = 0, 0
@@ -427,9 +402,9 @@ local function read_form_value(buf, str_buf, pos, form)
 		-- The constant is declared in the abbrev; no value bytes in the DIE.
 		return nil, pos
 	elseif form == M.DW_FORM.ref_sig8 then
-		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): an 8-byte value identifying a type
+		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): An 8-byte value identifying a type by signature.
-		-- by signature. The low 4 bytes (LE) are the type signature (content hash);
+		-- The low  4 bytes (LE) are the type signature (content hash);
-		-- the high 4 bytes (LE) are a CU-relative offset into the matching type unit.
+		-- The high 4 bytes (LE) are a CU-relative offset into the matching type unit.
 		-- Consumers use the low  4 to look up the type unit (see M.find_type_unit_by_signature)
 		--          then the high 4 to resolve the specific type within it.
 		-- Return the low  4 as the primary value to preserve the (value, next_pos) shape;
@@ -455,20 +430,20 @@ function M.read_ref_sig8(buf, pos)
 	return M.read_u32_le(buf, pos), M.read_u32_le(buf, pos + 4), pos + 8
 end
 
--- DWARF5 §7.5.6 (Type Entries).
+--- DWARF5 §7.5.6 (Type Entries).
--- Walk all units in `info` and return the 0-based offset of the first unit
+--- Walk all units in `info` and return the 0-based offset of the first unit whose `DW_AT_type_signature`
--- whose `DW_AT_type_signature` (8-byte value at the end of the unit header) equals `target_sig`.
+--- (8-byte value at the end of the unit header) equals `target_sig`.
--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
+--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
+--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
---
+---
--- Unit header layout (from pos 0):
+--- Unit header layout (from pos 0):
---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
+---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
+---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
+--- 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 info          string   -- the .debug_info section bytes
--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
+--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
+--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
+--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
 function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
 	local pos         = 0
 	local section_len = #info
@@ -577,69 +552,58 @@ function M.read_elf_sections(elf_path, section_names)
 		return result
 	end
 
-	-- Read the ELF32 header.
+	local file_size
-	local header = f:read(M.ELF32.header_bytes)
+	do
-	if not header or #header < M.ELF32.header_bytes then
+		f:seek("end", 0)
-		io.stderr:write("[elf_dwarf.read_elf_sections] ELF too small for ELF32 header\n")
+		file_size = f:seek("cur", 0)
+	end
+	local adapter = {
+		read_u8_at = function(offset)
+			f:seek("set", offset)
+			local b = f:read(1)
+			if not b then return nil end
+			return b:byte()
+		end,
+		read_u16_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			if not b1 or not b2 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+		end,
+		read_u32_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			local b3 = f:read(1)
+			local b4 = f:read(1)
+			if not b1 or not b2 or not b3 or not b4 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+				+ b3:byte() * 0x10000 + b4:byte() * 0x1000000
+		end,
+		read_size = function() return file_size end,
+	}
+
+	-- Delegate the header parse + section walk to E.*.
+	local hdr, hdr_err = E.parse_elf32_headers(adapter)
+	if not hdr then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] header parse failed: %s\n", tostring(hdr_err)))
 		f:close()
 		return result
 	end
 
-	-- Sanity-check magic + class + endianness.
+	local sections, walk_err = E.walk_sections(adapter, hdr)
-	if header:sub(M.ELF32.magic_offset + 1, M.ELF32.magic_offset + 0x04) ~= M.ELF32.magic then
+	if not sections then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not an ELF file\n")
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] section walk failed: %s\n", tostring(walk_err)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.class_offset + 1) ~= M.ELF32.class_elf32 then
-		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] not ELF32 (class=%d)\n", header:byte(M.ELF32.class_offset + 1)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.endian_offset + 1) ~= M.ELF32.endian_little then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not little-endian; unsupported\n")
 		f:close()
 		return result
 	end
 
-	-- Parse section-header table location + dimensions from the header.
+	-- Resolve the requested sections.
-	local e_shoff     = M.read_u32_le(header, M.ELF32.e_shoff_offset)
+	for _, s in ipairs(sections) do
-	local e_shentsize = M.read_u16_le(header, M.ELF32.e_shentsize_offset)
+		if wanted[s.name] then
-	local e_shnum     = M.read_u16_le(header, M.ELF32.e_shnum_offset)
+			local bytes = E.read_section_bytes(adapter, s)
-	local e_shstrndx  = M.read_u16_le(header, M.ELF32.e_shstrndx_offset)
+			if bytes then result[s.name] = bytes end
-
-	-- Read the section-header string table (.shstrtab) so we can resolve section names from their `sh_name` offsets.
-	f:seek("set", e_shoff + e_shstrndx * e_shentsize)
-	local strtab_hdr = f:read(e_shentsize)
-	if not strtab_hdr or #strtab_hdr < e_shentsize then
-		io.stderr:write("[elf_dwarf.read_elf_sections] could not read .shstrtab header\n")
-		f:close()
-		return result
-	end
-	local strtab_offset = M.read_u32_le(strtab_hdr, M.ELF32.sh_offset_offset)
-	local strtab_size   = M.read_u32_le(strtab_hdr, M.ELF32.sh_size_offset)
-	f:seek("set", strtab_offset)
-	local strtab = f:read(strtab_size) or ""
-
-	-- Walk all section headers; collect (offset, size) for the wanted names.
-	local function read_section_bytes(sh_offset, sh_size)
-		f:seek("set", sh_offset)
-		return f:read(sh_size) or ""
-	end
-
-	for sh_idx = 0, e_shnum - 1 do
-		f:seek("set", e_shoff + sh_idx * e_shentsize)
-		local sh = f:read(e_shentsize)
-		if not sh or #sh < e_shentsize then break end
-		local sh_name   = M.read_u32_le(sh, M.ELF32.sh_name_offset)
-		local sh_offset = M.read_u32_le(sh, M.ELF32.sh_offset_offset)
-		local sh_size   = M.read_u32_le(sh, M.ELF32.sh_size_offset)
-
-		-- Extract the name (null-terminated C string in strtab).
-		local name_end = strtab:find("\0", sh_name + 1, true) or (sh_name + 1)
-		local name     = strtab:sub(sh_name + 1, name_end - 1)
-		if wanted[name] then
-			result[name] = read_section_bytes(sh_offset, sh_size)
 		end
 	end
 
@@ -656,49 +620,87 @@ end
 --- - We filter on STB_GLOBAL (high nibble of st_info = 1) to match `nm`'s default (external symbols only). STB_WEAK excluded.
 --- - The `code_` prefix is stripped (MipsAtom_ macros emit bare atom names, no `code_` prefix).
 --- - `st_size > 0` filter excludes undefined/imported symbols.
+---
 --- @param elf_path Path
 --- @return table<string, {integer, integer}>
 function M.read_nm(elf_path)
 	local addrs = {}
 
-	-- Read .symtab + .strtab via the existing ELF walker (no subprocess).
+	-- Existence check first; an empty or missing ELF returns an empty map.
-	local sections = M.read_elf_sections(elf_path, {".symtab", ".strtab"})
+	if lfs.attributes(elf_path, "mode") ~= "file" then
-	local symtab   = sections[".symtab"]
-	local strtab   = sections[".strtab"]
-	if not symtab or not strtab or #symtab == 0 or #strtab == 0 then
-		-- No symbol table (e.g. stripped ELF). Return empty.
 		return addrs
 	end
 
-	-- Iterate the 16-byte ELF32 symtab entries.
+	local f = io.open(elf_path, "rb")
-	-- Each entry (zero-based): st_name at 0, st_value at 4, st_size at 8, st_info at 12, st_other at 13, st_shndx at 14.
+	if not f then
-	local SYM_ENTRY_BYTES = 0x10
+		return addrs
-	local SYM_ST_NAME     = 0x00
-	local SYM_ST_VALUE    = 0x04
-	local SYM_ST_SIZE     = 0x08
-	local SYM_ST_INFO     = 0x0C
-	local n_syms = #symtab / SYM_ENTRY_BYTES
-	for i = 0, n_syms - 1 do
-		local entry_off  = i * SYM_ENTRY_BYTES
-		local st_info    = symtab:byte(entry_off + SYM_ST_INFO + 1)
-		-- High nibble = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
-		-- Use math.floor(/16) instead of bit.rshift for LuaJIT 2.1 compat
-		-- (LuaJIT's `>>` is 5.3+, but math.floor(x/16) works on all versions).
-		local binding    = math.floor(st_info / 16)
-		if binding == 0 or binding == 1 then            -- STB_LOCAL or STB_GLOBAL
-			local st_size   = M.read_u32_le(symtab, entry_off + SYM_ST_SIZE)
-			if st_size > 0 then
-				local st_name_off = M.read_u32_le(symtab, entry_off + SYM_ST_NAME)
-				-- Extract the name from .strtab (null-terminated C string).
-				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
-				local name     = strtab:sub(st_name_off + 1, name_end - 1)
-				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — MipsAtom_ macros strip the `code_` prefix).
-				-- The atoms_source_map pass already filters out non-atom symbols via the source-map.txt cross-ref.
-				if name and #name > 0 then
-					local st_value = M.read_u32_le(symtab, entry_off + SYM_ST_VALUE)
-					addrs[name] = { st_value, st_size }
 	end
+
+	-- Build the file adapter for E.*.
+	local file_size
+	do
+		f:seek("end", 0)
+		file_size = f:seek("cur", 0)
 	end
+	local adapter = {
+		read_u8_at = function(offset)
+			f:seek("set", offset)
+			local b = f:read(1)
+			if not b then return nil end
+			return b:byte()
+		end,
+		read_u16_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			if not b1 or not b2 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+		end,
+		read_u32_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			local b3 = f:read(1)
+			local b4 = f:read(1)
+			if not b1 or not b2 or not b3 or not b4 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+				+ b3:byte() * 0x10000 + b4:byte() * 0x1000000
+		end,
+		read_size = function() return file_size end,
+	}
+
+	-- Delegate the header + section walk to E.*.
+	local hdr, hdr_err = E.parse_elf32_headers(adapter)
+	if not hdr then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] header parse failed: %s\n", tostring(hdr_err)))
+		f:close()
+		return addrs
+	end
+
+	local sections, walk_err = E.walk_sections(adapter, hdr)
+	if not sections then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] section walk failed: %s\n", tostring(walk_err)))
+		f:close()
+		return addrs
+	end
+
+	-- E.collect_symbols returns every defined symbol (no binding filter).
+	-- The metaprogram then applies its STB_LOCAL / STB_GLOBAL + size>0 filter, matching `nm`'s default (external symbols only).
+	local symbols, sym_err = E.collect_symbols(adapter, sections)
+	if not symbols then
+		io.stderr:write(string.format("[elf_dwarf.read_nm] symbol collection failed: %s\n", tostring(sym_err)))
+		f:close()
+		return addrs
+	end
+
+	f:close()
+
+	for name, entry in pairs(symbols) do
+		-- High nibble of st_info = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
+		-- math.floor(/16) is portable across LuaJIT 2.0/2.1 and plain Lua 5.x.
+		local binding = math.floor(entry.info / 16)
+		if (binding == 0 or binding == 1) and entry.size > 0 then
+			addrs[name] = { entry.value, entry.size }
 		end
 	end
 
@@ -836,12 +838,11 @@ end
 ---   * The `.debug_line` section may contain MULTIPLE line-program units
 ---     File indices are 1-based, **per unit**; we concatenate all units and the index ranges from 1..N₁ in unit 1, N₁+1..N₁+N₂ in unit 2, etc.
 ---     Per-unit indices (the way gcc emits them, and the way `DW_LNS_set_file` references them in the line program)
----     are returned via the `basename_to_index` map only when the unit boundary happens to align with the metaprogram's per-atom
+---     are returned via the `basename_to_index` map only when the unit boundary happens to align with the metaprogram's per-atom `inv.call_file`
----     `inv.call_file` (true today for hello_joypad — the C unit is the LAST unit, and atom-side file indices fit 1-based).
 ---   * Per spec, the `.debug_line_str` section (DWARF5 §7.5.6) holds the strings referenced by `DW_FORM_line_strp`.
 ---     The legacy DWARF3 format embeds strings directly with null terminators. This helper handles BOTH.
----   * File entries may have multiple forms (gcc -gdwarf-5 with `DW_LNCT_directory_index`
+---   * File entries may have multiple forms (gcc -gdwarf-5 with `DW_LNCT_directory_index` emits 2 forms: path + dir_index). 
----     emits 2 forms: path + dir_index). The helper supports:
+---     The helper supports:
 ---     - DW_FORM_line_strp (DWARF5; offset into .debug_line_str)
 ---     - DW_FORM_string (DWARF4-compat; inline null-terminated in .debug_line)
 ---     - DW_FORM_udata (ULEB128)
@@ -851,12 +852,9 @@ end
 ---
 --- Behavior on failure: writes to stderr and returns nil.
 --- Helpers consumed by `passes/dwarf_injection.lua::init_file_index_lookup(elf_path)` calls this once at pass start to populate the module-level `basename_to_index` map;
---- downstream `resolve_provenance_file_index(path)` consumers
+--- downstream `resolve_provenance_file_index(path)` consumers consult the map directly.
---- (which replaced the former hardcoded `ATOM_SOURCE_FILE_INDEX` + `PROVENANCE_BASENAME_TO_FILE_INDEX` table per `conductor/tracks/dwarf_file_index_lookup_20260731/`)
---- consult the map directly.
 ---
---- @param elf_path string  -- absolute path to the post-link ELF (typically the gcc-emitted `.elf` BEFORE dwarf_injector's splice;
+--- @param elf_path string  -- absolute path to the post-link ELF (typically the gcc-emitted `.elf` BEFORE dwarf_injector's splice; both shapes work since the splice preserves `.debug_line`)
----                            both shapes work since the splice preserves `.debug_line`)
 --- @return table|nil, table|nil, table|nil
 ---   basename_to_index: { [basename] = 1-based-per-unit-file-index, ... }
 ---   basenames:         { [1-based-per-unit-file-index] = basename, ... }
@@ -894,8 +892,7 @@ function M.read_line_unit_file_table(elf_path)
 			return nil, p + M.DWARF5_DEBUG_LINE.form_data16_bytes
 		else
 			-- Unsupported form in a directory/file-table entry: best-effort skip.
-			-- We do NOT stderr-write because the crt0.s DWARF5 line unit (gcc-as emitted) uses DW_FORM_addr (0x01) for what is effectively a path entry, 
+			-- We do NOT stderr-write because the crt0.s DWARF5 line unit (gcc-as emitted) uses DW_FORM_addr (0x01) for what is effectively a path entry, which is non-standard.
-			-- which is non-standard.
 			-- The C-unit's DWARF3 paths are read via the parallel DWARF3 path and never see this error.
 			-- Callers should consult `basename_to_index` for the paths they care about and ignore this unit if it produced none.
 			return nil, p

scripts/passes/annotation.lua

@@ -22,11 +22,11 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -45,28 +45,28 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 --- @field warnings table[]
 
 --- @class AtomAnnotation
---- @field line    integer       -- source line of the atom_info call
+--- @field line    integer       -- Source line of the atom_info call
---- @field macro   string        -- the macro name (always "atom_info" in the new shape)
+--- @field macro   string        -- Macro name (always "atom_info" in the new shape)
---- @field name    string        -- the atom name
+--- @field name    string        -- Atom name
---- @field kind    string        -- always "info"
+--- @field kind    string        -- Always "info"
 --- @field binds   string|nil    -- Binds_X name if any
 --- @field reads   string[]      -- R_* names (read targets)
 --- @field writes  string[]      -- R_* names (write targets)
---- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
+--- @field errors  string[]|nil  -- Parse-time errors from scan_source (atom_info body malformed)
 
---- @class DebugSkipMarker  -- sub-shape of scan_source.lua's @class DebugSkipMarker
+--- @class DebugSkipMarker  -- Sub-shape of scan_source.lua's @class DebugSkipMarker
---- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
+--- @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 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 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
 
 --- @class Finding
---- @field line integer -- source line (or 0 for pass-level)
+--- @field line integer -- Source line (or 0 for pass-level)
---- @field msg  string  -- finding message
+--- @field msg  string  -- Finding message
 
 --- @class Findings
 --- @field errors   Finding[]
@@ -74,14 +74,14 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 --- @field info     Finding[]
 
 --- @class PipeCtx
---- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
+--- @field atom_index     table<string, AtomAnnotation> -- Name -> AtomAnnotation (only kind=="atom")
---- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
+--- @field binds_index    table<string, BindsStruct>    -- Name -> BindsStruct
---- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
+--- @field annot_counts   table<string, integer>        -- Name -> annotation count (for unique_annotation check)
---- @field types          table<string, RegTypeDefault> -- from scan_source
+--- @field types          table<string, RegTypeDefault> -- From scan_source
---- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- From scan_source
---- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field seen_defaults  table<string, integer>        -- Duplicate atom_dbg_reg_default detection
 --- @field seen_field     table<string, integer>        -- Binds_* -> count of fields (set/checked by check_binds_no_duplicate_fields)
---- @field _scan          SourceScan                    -- full scan payload (typed-view sub-calls live here) 
+--- @field _scan          SourceScan                    -- Full scan payload (typed-view sub-calls live here) 
 
 --- @class AnnotatedResult
 --- @field atoms    AtomEntry[]
@@ -95,11 +95,10 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-check functions (the CHECK_RULES table's payload)
 -- ════════════════════════════════════════════════════════════════════════════
---
 --- The dispatcher in `validate()` routes each result by convention: existence checks write errors[] and shape checks write warnings[].
 --- `macro_word_drift` writes errors[] for missing or mismatched metadata and info[] for a match.
 
---- Check: every annotated atom must have a matching MipsAtom_(name) declaration.
+--- Check: Every annotated atom must have a matching MipsAtom_(name) declaration.
 --- @param a        AtomAnnotation
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
@@ -112,8 +111,8 @@ local function check_atom_decl_exists(a, pipe_ctx, findings)
 	end
 end
 
---- Check: every atom may have AT MOST ONE annotation.
+--- Check:     Every atom may have AT MOST ONE annotation.
---- Post-loop: needs full-corpus `annot_counts` from pipe_ctx.
+--- Post-loop: Needs full-corpus `annot_counts` from pipe_ctx.
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_unique_annotation(pipe_ctx, findings)
@@ -146,7 +145,7 @@ 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 wc table<string, integer> -- the shared word-count table (from ctx.shared.word_counts)
+--- @param wc table<string, integer> -- Shared word-count table (from ctx.shared.word_counts)
 --- @param findings Findings
 local function check_macro_word_drift(m, wc, findings)
 	local  declared = wc[m.name]
@@ -304,7 +303,7 @@ local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
 	end
 end
 
--- Check: debug-skip markers must satisfy shape + placement constraints.
+-- Check: Debug-skip markers must satisfy shape + placement constraints.
 --- 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. marker_kind ~= "atom_dbg_skip" -> legacy/renamed spelling (use `atom_dbg_skip`)
@@ -315,12 +314,11 @@ end
 ---   6. unsupported target_kind     -> marker precedes an unrelated declaration
 --- 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 _pipe_ctx PipeCtx     -- Unused; kept for consistency with per_annot // TODO(Ed): Remove?
 --- @param findings Findings
 local function check_skip_marker(marker, _pipe_ctx, findings)
 	local kind = marker.marker_kind
 	local line = marker.marker_line
-
 	-- 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
@@ -371,8 +369,6 @@ local function check_skip_marker(marker, _pipe_ctx, findings)
 end
 
 --- Warn when a source references an unregistered alias.
----
---- R_TapePtr, R_AtomJmp, R_PrimCursor, R_FaceCursor, R_VertBase, and R_OtBase opt in through `#define atom_reg` in lottes_tape.h.
 --- 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.
 --- @param _src     SourceFile
 --- @param pipe_ctx PipeCtx
@@ -426,8 +422,7 @@ local CHECK_RULES = {
 -- ════════════════════════════════════════════════════════════════════════════
 -- Validation
 -- ════════════════════════════════════════════════════════════════════════════
---
+-- Pure check: Read from src.scan, run validations, emit findings. The scan was done once upstream.
--- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
 
 --- 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.
@@ -473,7 +468,7 @@ 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.
+--- @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, corpus_pipe_ctx)
 	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
@@ -503,14 +498,8 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	end
 
 	-- 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`.
-	local seen_defaults = {}
+	local seen_defaults   = {}; for reg, _ in pairs (scan.types      or {}) do seen_defaults[reg] = (seen_defaults[reg] or 0) + 1 end
-	for reg, _ in pairs(scan.types or {}) do
+	local atom_infos_list = {}; for _, ai  in ipairs(scan.atom_infos or {}) do atom_infos_list[#atom_infos_list + 1] = ai         end
-		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
-	end
-	local atom_infos_list = {}
-	for _, ai in ipairs(scan.atom_infos or {}) do
-		atom_infos_list[#atom_infos_list + 1] = ai
-	end
 
 	local pipe_ctx = {
 		atom_index               = {},

scripts/passes/atoms_source_map.lua

@@ -1,8 +1,8 @@
 --- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
 ---
 --- Writer: this pass, given `atom.paths` (the per-atom mutable surface owned by `emission_model`). Readers:
---- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and the gdb-runtime
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and 
---- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
+--- the gdb-runtime wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
 ---
 --- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: 
 ---  one `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
@@ -28,7 +28,6 @@
 ---   ...
 ---   ENDATOM
 ---   ```
----
 --- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -95,8 +94,8 @@ end
 --- 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>` (component invocation)
 ---   `WORD N  CALL <src-path>:<src-line>  RAW` (raw `.word` outside any mac_* component)
---- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was declared in `corpus.word_counts` 
---- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
+--- (populated by word_count_eval + components passes).
 --- @param src  table
 --- @param atom table
 --- @param wc   table -- identity alias of corpus.word_counts
@@ -105,20 +104,17 @@ local function emit_provenance_stanza(src, atom, wc)
 	local lines          = {}
 	local rel_path       = src.path:gsub("\\\\", "/")
 	local entries, total = canonical_word_entries(atom)
-
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
 
 	for _, entry in ipairs(entries) do
 		local inv         = entry.invocation
 		local macro_count = inv and wc["mac_" .. inv.component_name]
 		if inv and macro_count ~= nil then
-			lines[#lines + 1] = string.format(
+			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d'
-				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				, entry.pos, rel_path, entry.line, inv.component_name
-				entry.pos, rel_path, entry.line, inv.component_name,
+				, inv.def_path or "", inv.def_line or 0, entry.body_line)
-				inv.def_path or "", inv.def_line or 0, entry.body_line)
 		else
-			lines[#lines + 1] = string.format(
+			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
-				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
 		end
 	end
 
@@ -248,9 +244,10 @@ local function build_atom_table(ctx)
 	return matched
 end
 
---- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`, the convenience
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`,
---- vars set in `emit_gdb_runtime` provide printf args, and each command is a static sequence of `printf` / `tbreak` /
+--- the convenience vars set in `emit_gdb_runtime` provide printf args, and 
---- `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.
+--- The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
 ---
 --- Why hardcoded per-atom: gdb's `$` substitution doesn't concat inside var names — `$__atom_name_$__i` in a `while`
 --- loop resolves to one literal identifier, not `name_i`. Compile-time emission is the only path.
@@ -395,8 +392,7 @@ local function append_gdb_commands(lines, matched)
 end
 
 --- Emit the gdb-runtime file (post-link). Pure gdb scripting — addresses come from `mipsel-none-elf-nm -S`, get embedded
---- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb 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

scripts/passes/components.lua

@@ -6,10 +6,9 @@
 --- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations,
 --- then resolves the function-args string from the preceding `FI_ Slice_MipsCode 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.
+--- Emits one `gen/macs.h` per *immediate source directory* with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
----
+--- All sources inside the same directory contribute to the same file (per-directory aggregation).
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- The directory itself is the namespace, so the filename does not repeat the module name.
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -41,29 +40,30 @@ local MAC_PREFIX_LEN  = 4
 local BYTE_NEWLINE    = 10
 local BYTE_SLASH      = 47
 
--- Source dir basename used as the output `.macs.h` filename.
+-- Output gen subdirectory + filename (per-directory aggregation; the directory name is the namespace).
 local GEN_SUBDIR      = "gen"
+local MACS_FILENAME   = "macs.h"
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
+--- @field sources            SourceFile[]           -- All source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table                  -- Cross-pass shared state
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
+--- @field project_root       string                 -- Project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
+--- @field upstream           table<string, table>   -- Per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @field verbose            boolean                -- log diagnostic info
+--- @field verbose            boolean                -- Log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -71,13 +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 -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @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
+--- @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)
@@ -231,7 +231,6 @@ end
 --
 -- Skips `//` sequences that are inside string or character literals
 -- (a rough heuristic — sufficient for component bodies which don't have those constructs).
---
 --- @param s string
 --- @return string
 local function convert_line_comments_to_block(s)
@@ -339,6 +338,116 @@ local function count_all_components(components, wc)
 	return counts
 end
 
+-- ═══════════════════════════════════════════
+-- Per-component metadata derivation (replaces the hardcoded `M.GP0_MACRO_CONTRIB` + `M.INSTRUCTION_LATENCY[mac_*]` tables that previously lived in `duffle.lua`).
+--
+-- Each `MipsAtomComp_(ac_X) { body }` definition in `code/duffle/lottes_tape.h` is the canonical source.
+-- The `mac_X(...)` macros are GENERATED from these definitions by `emit_component_macros_h` for tape-side composition;
+-- the metaprogram must NEVER walk the generated variants to derive metadata.
+-- Always walk the original `MipsAtomComp_` body via `cc.body_tokens`.
+-- ═══════════════════════════════════════════
+
+--- (internal) Recursive cycle-cost derivation. Sum `latency[ident]` per emitted instruction in the component body,
+--- recursing through nested `mac_*` calls (so `mac_format_g4_color`'s cost = 4 × `mac_pack_color_word`'s cost).
+--- Special rule: `mac_yield`'s cost = 0 (per `lottes_tape.h:125-130` "the runtime cost lands in the next atom's prologue").
+--- @param name         string  -- component bare name (e.g. "yield", "pack_color_word")
+--- @param comp_by_name table<string, Component>
+--- @param latency      table<string, integer>
+--- @param cache        table<string, integer>  -- shared memoization; `-1` sentinel detects cycles
+--- @return integer
+local function cycle_cost_rec(name, comp_by_name, latency, cache)
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1
+	local cc = comp_by_name[name]
+	local n
+	if cc then
+		if name == "yield" then
+			-- mac_yield's cost is 0 by convention (the runtime cost lands in the next atom's prologue).
+			n = 0
+		else
+			n = 0
+			local tokens = cc.body_tokens
+			for _, t in ipairs(tokens) do
+				local trimmed = t.tok
+				if trimmed ~= "" then
+					local ident = duffle.read_ident(trimmed, 1)
+					if ident and ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+						-- Nested `mac_X(...)` call: recurse.
+						local nested = ident:sub(MAC_PREFIX_LEN + 1)
+						n = n + cycle_cost_rec(nested, comp_by_name, latency, cache)
+					else
+						-- Leaf instruction or pseudo-macro. Look up in INSTRUCTION_LATENCY; default 1.
+						n = n + (latency[ident] or 1)
+					end
+				end
+			end
+		end
+	else
+		n = 1
+	end
+	cache[name] = n
+	return n
+end
+
+--- (internal) Recursive GP0 prim-buffer contribution. Count `store_word` / `store_half` / `store_byte` 
+--- calls in the component body that target `R_PrimCursor` (these are the
+--- RAM-side prim-buffer words the macro contributes), recursing through nested `mac_*` calls.
+--- Only `R_PrimCursor`-targeting stores count. Stores targeting other registers (e.g. `R_OtBase`, heap pointers) are not prim-buffer contributions.
+--- @param name         string
+--- @param comp_by_name table<string, Component>
+--- @param cache        table<string, integer>
+--- @return integer
+local function gp0_contrib_rec(name, comp_by_name, cache)
+	if cache[name] ~= nil then return cache[name] end
+	cache[name] = -1
+	local cc = comp_by_name[name]
+	local n
+	if cc then
+		n = 0
+		local tokens = cc.body_tokens
+		for _, t in ipairs(tokens) do
+			local trimmed = t.tok
+			if    trimmed ~= "" then
+				local ident = duffle.read_ident(trimmed, 1)
+				if    ident and ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+					-- Nested `mac_X(...)` call: recurse.
+					local nested = ident:sub(MAC_PREFIX_LEN + 1)
+					n = n + gp0_contrib_rec(nested, comp_by_name, cache)
+				elseif ident == "store_word" or ident == "store_half" or ident == "store_byte" then
+					if trimmed:find("R_PrimCursor", 1, true) then
+						n = n + 1
+					end
+				end
+			end
+		end
+	else
+		n = 0
+	end
+	cache[name] = n
+	return n
+end
+
+--- Compute `cycle_cost` + `gp0_contrib` for every component in `components` in a single pass.
+--- Memoization cache is built ONCE (per source) and shared across both helpers so that
+--- a nested `mac_Y` reference inside a `mac_X` body computes its values once.
+--- @param components Component[]
+--- @param latency    table<string, integer>
+--- @return table<string, {cycle_cost=integer, gp0_contrib=integer}>
+local function compute_components_metadata(components, latency)
+	local comp_by_name = {}
+	for _, cc in ipairs(components) do comp_by_name[cc.name] = cc end
+	local cc_cache  = {}
+	local gc_cache  = {}
+	local out       = {}
+	for _, c in ipairs(components) do
+		out[c.name] = {
+			cycle_cost = cycle_cost_rec(c.name, comp_by_name, latency, cc_cache),
+			gp0_contrib = gp0_contrib_rec(c.name, comp_by_name, gc_cache),
+		}
+	end
+	return out
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-component emit logic
 -- ════════════════════════════════════════════════════════════════════════════
@@ -445,9 +554,15 @@ end
 
 --- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
 --- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
---- @param src SourceFile
+--- @param dir     string       -- the absolute source directory
+--- @param sources SourceFile[] -- sources contributing to this directory (for the header comment)
 --- @return string[]
-local function header_boilerplate(src)
+local function header_boilerplate(dir, sources)
+	local source_lines = { "// Directory: " .. duffle.to_absolute_path(dir) .. "/" }
+	for _, src in ipairs(sources) do
+		source_lines[#source_lines + 1] = "//   source: " .. duffle.to_absolute_path(src.path)
+	end
+	local source_blob = table.concat(source_lines, "\n")
 	return {
 		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
 		-- The build does manual unity includes (the user controls include order), so the pragma is only active for IDE/tooling.
@@ -455,7 +570,7 @@ local function header_boilerplate(src)
 		"#pragma once",
 		"#endif",
 		"// Auto-generated by ps1_meta.lua — DO NOT EDIT",
-		"// Source: " .. duffle.to_absolute_path(src.path),
+		source_blob,
 		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
 		"",
 		-- Self-contained: define WORD_COUNT if not already defined.
@@ -468,30 +583,30 @@ local function header_boilerplate(src)
 	}
 end
 
---- Compute the output path for one source's `.macs.h` file.
+--- Compute the per-directory output path for `.macs.h`.
---- The pre-rework convention uses the *directory* basename (not the source file basename)
+---  e.g. any source in `code/duffle/` produces `code/duffle/gen/macs.h` regardless of source filename.
----  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+--- The directory name is the namespace; the filename does not repeat it.
---- This matches what the C codebase #includes.
+--- @param dir string  -- the absolute source directory
---- @param src SourceFile
 --- @return string  -- the output directory
 --- @return string  -- the full output path
-local function compute_macs_h_path(src)
+local function compute_macs_h_path(dir)
-	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
+	local  out_dir  = dir .. "/" .. GEN_SUBDIR
-	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+	local  out_path = out_dir .. "/" .. MACS_FILENAME
 	return out_dir, out_path
 end
 
---- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
+--- Emit a per-directory `.macs.h` header with the aggregated `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).
 --- @param ctx        PassCtx
---- @param src        SourceFile
+--- @param dir        string                  -- the absolute source directory
---- @param components Component[]
+--- @param sources    SourceFile[]            -- sources contributing to this directory (for the header comment)
+--- @param components Component[]             -- aggregated components from all sources in this directory
 --- @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)
+local function emit_component_macros_h(ctx, dir, sources, components, counts)
 	if #components == 0 then return nil end
-	local out_dir, out_path = compute_macs_h_path(src)
+	local out_dir, out_path = compute_macs_h_path(dir)
-	local lines             = header_boilerplate(src)
+	local lines             = header_boilerplate(dir, sources)
 
 	for _, c in ipairs(components) do
 		for _, l in ipairs(build_component_lines(c, counts)) do
@@ -534,18 +649,20 @@ end
 
 --- (internal) Populate `corpus.components` with this source's components-by-name map.
 --- First declaration wins; later declarations of the same bare name are dropped and recorded as a collision via `corpus.collisions` (kind = "component").
---- The pass does NOT write to `ctx.shared.components` (ownership follows the canonical contract).
+--- The pass does NOT write to `ctx.shared.components`.
---- The `debug_skip` field mirrors the scanner-owned declaration record (`c.debug_skip`).
 --- No parallel skip map is built here; consumers that need the per-component skip state read `corpus.components[name].debug_skip` directly.
+--- The `cycle_cost` + `gp0_contrib` fields are populated from `metadata[c.name]` (computed by `compute_components_metadata` against the original `MipsAtomComp_` body).
 --- @param corpus     table   -- the corpus
 --- @param src        SourceFile
 --- @param components Component[]
-local function update_canonical_components(corpus, src, components)
+--- @param metadata    table<string, {cycle_cost=integer, gp0_contrib=integer}>
+local function update_canonical_components(corpus, src, components, metadata)
 	local rel_path = src.path:gsub("\\", "/")
 	for _, c in ipairs(components) do
 		-- Keyed by bare name (e.g. `yield`, `load_tri_indices`).
 		-- The atoms_source_map pass looks up components by bare name from the corpus;
 		-- `mac_` prefix lives at the call-site identifier and is stripped before lookup.
+		local m = metadata and metadata[c.name] or nil
 		if corpus.components[c.name] == nil then
 			corpus.components[c.name] = {
 				name        = c.name,
@@ -553,6 +670,8 @@ local function update_canonical_components(corpus, src, components)
 				path        = rel_path,
 				kind        = c.kind or "comp_bare",
 				debug_skip  = c.debug_skip == true,
+				cycle_cost  = m and m.cycle_cost  or nil,
+				gp0_contrib = m and m.gp0_contrib or nil,
 			}
 		else
 			-- A second declaration of the same bare name: record a typed collision so static-analysis + the report can surface it.
@@ -624,21 +743,39 @@ function M.run(ctx)
 	--   * `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
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/macs.h`.
-		-- project_components reads from src.scan + does backward lookups on src.text
+	-- The directory itself is the namespace. `corpus.sources_by_dir` preserves source-order within each bucket (matches `corpus.source_order`).
-		local components = project_components(src.text, src.scan)
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
-		if #components > 0 then
+	for dir, sources in pairs(sources_by_dir) do
-			-- Compute all component word counts once per source.
+		-- Aggregate components from every source in this directory.
-			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
+		-- `project_components` returns nil for sources with no `MipsAtomComp_` declarations; we skip those.
-			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
+		local aggregated_components = {}
-			local counts = count_all_components(components, corpus.word_counts)
+		local metadata_per_source   = {}
-			local macs_path = emit_component_macros_h(ctx, src, components, counts)
+		for _, src in ipairs(sources) do
+			local per_source = project_components(src.text, src.scan) or {}
+			for _, c in ipairs(per_source) do
+				aggregated_components[#aggregated_components + 1] = c
+			end
+			if #per_source > 0 then
+				metadata_per_source[src] = compute_components_metadata(per_source, duffle.INSTRUCTION_LATENCY)
+			end
+		end
+		if #aggregated_components > 0 then
+			-- Compute word counts across the aggregated set. `corpus.word_counts` carries the
+			-- same-source + prior-directory entries so the recursive lookup sees both.
+			local counts    = count_all_components(aggregated_components, corpus.word_counts)
+			local macs_path = emit_component_macros_h(ctx, dir, sources, aggregated_components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				-- Populate the projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
+				-- Populate the projections AFTER disk emission (byte-identical `.macs.h` contract).
-				update_canonical_word_counts(corpus, components, counts)
+				update_canonical_word_counts(corpus, aggregated_components, counts)
-				update_canonical_components(corpus, src, components)
+				for _, src in ipairs(sources) do
-				update_canonical_component_body_index(corpus, src, components, src.scan)
+					local per_source = project_components(src.text, src.scan) or {}
+					if #per_source > 0 then
+						update_canonical_components(corpus, src, per_source, metadata_per_source[src])
+						update_canonical_component_body_index(corpus, src, per_source, src.scan)
+					end
+				end
 			end
 		end
 	end

scripts/passes/dwarf_injection.lua

@@ -73,10 +73,6 @@ local DW_RLE_start_length  = DWARF5_RNGLISTS.start_length
 
 -- File-index lookup for the existing main line unit (Unit 2).
 -- Populated at pass start by `init_file_index_lookup(elf_path)` from the runtime ELF (see `elf_dwarf.read_line_unit_file_table`).
--- The hardcoded indices and the `PROVENANCE_BASENAME_TO_FILE_INDEX` table that previously lived here were retired in `conductor/tracks/dwarf_file_index_lookup_20260731/`
--- (red of the
--- `TODO(Ed): Remove this HARDCODE` from line 156); the runtime lookup reads the
--- actual gcc-emitted `.debug_line` file table instead.
 local _file_index_by_basename    = nil  -- [basename] = 1-based line-table file index
 local _file_path_by_index        = nil  -- [1-based index] = full source path (diagnostics / future consumers)
 local _default_atom_source_index = nil  -- any valid index used in opaque-row fallbacks
@@ -188,15 +184,17 @@ end
 --- Resolve an absolute provenance path to the line-unit file index used by the emitting line program.
 --- Normalizes mixed `/` and `\` separators to a basename and looks it up against the runtime-computed file table populated by `init_file_index_lookup`.
 ---
---- Fails loudly on an unknown provenance basename: adding a new component source file will produce a clear error message naming the missing basename and listing the .debug_line file table contents, 
+--- Returns 0 (the DWARF `set_file(0)` "no file change" sentinel) when the basename is not in the file table.
---- so the user can either confirm the gcc include order, the unity-root, or the `.debug_line` file table contents.
+--- This is a normal occurrence: the compiler only adds a file to the `.debug_line` file table when the file has line-numbered content (i.e., code).
---- Silent fallback would mask the new-file case by misattributing component rows to an arbitrary source file.
+--- Files containing only static-array data (e.g. `MipsAtomComp_` declarations in `gp.atom.c`, `psyq.atom.c`, `pad.atom.c` — the OT-tag inserts, etc.) produce no line numbers,
+--- so gcc omits them from the file table.
+--- The DWARF emitter then keeps the previous line-program file state instead of pointing at a file that has no entries to walk.
+--- A stderr warning is emitted per-miss so the user can audit which files the compiler dropped.
 --- @param path string  -- absolute provenance path (mixed slashes accepted)
---- @return integer  -- 1-based line-unit file index
+--- @return integer  -- 1-based line-unit file index, or 0 on miss (DWARF no-change sentinel)
 local function resolve_provenance_file_index(path)
 	if _file_index_by_basename == nil then
-		error("[dwarf_injection] resolve_provenance_file_index called before init_file_index_lookup. "
+		error("[dwarf_injection] resolve_provenance_file_index called before init_file_index_lookup. Is M.run being entered correctly (with --elf)?")
-			.. "Is M.run being entered correctly (with --elf)?")
 	end
 	if path == nil or path == "" then
 		error("[dwarf_injection] resolve_provenance_file_index: empty path")
@@ -211,13 +209,11 @@ local function resolve_provenance_file_index(path)
 	for i, p in pairs(_file_path_by_index) do
 		if p and p:gsub("\\", "/") == normalized then return i end
 	end
-	-- Build an error message listing the known basenames for fast diagnostics.
+	-- File is in the corpus but gcc omitted it from the .debug_line file table (data-only content).
-	local known = {}
+	-- Return 0 = DWARF `set_file(0)` no-change sentinel so the line program keeps its prior file state.
-	for k in pairs(_file_index_by_basename) do known[#known + 1] = k end
+	io.stderr:write(string.format("[dwarf_injection] line-table miss: '%s' (basename '%s') not in .debug_line file table; "
-	table.sort(known)
+		.. "falling back to set_file(0)\n", path, basename))
-	error(string.format("[dwarf_injection] resolve_provenance_file_index: unknown provenance basename '%s' (from '%s'). "
+	return 0
-		.. "Known basenames in the .debug_line file table (%d): %s"
-		, basename, path, #known, table.concat(known, ", ")))
 end
 
 local DW_FORM_addr           = 0x01
@@ -232,7 +228,6 @@ local DW_FORM_sec_offset     = 0x17   -- 4-byte section-relative offset (into .d
 
 -- DW_OP_reg0 + DW_OP_piece are declared above (lines 114-116) alongside the other DWARF5 §7.7.3 loclist opcodes.
 
-
 local DW_ATE_unsigned        = 0x07   -- DWARF5 §7.8.1: DW_ATE_unsigned (used for U4 base type)
 
 -- (DW_LANG_Mips_Assembler = 0x8001 was used in the, but we want this CU to look like a C TU so VSCode's Variables pane treats it as code.)
@@ -616,10 +611,8 @@ local function build_atom_sequence(atom)
 		-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call).
 		for ai, anc in ipairs(entry_1_ancestry) do
 			assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-			assert(anc.body_lines[1] ~= nil
+			assert(anc.body_lines[1] ~= nil,              "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
-				, "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
+			assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation "            .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
-			assert(anc.call_path and anc.call_path ~= ""
-				, "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
 			emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line,    true)
 			local is_outermost = (ai == 1)
 			if not (is_outermost and anc.debug_skip) then
@@ -644,8 +637,7 @@ local function build_atom_sequence(atom)
 			-- all OTHER ancestors emit body_lines[1] with is_stmt = not debug_skip.
 			--
 			-- This re-emits the outer ancestor's call-site + body rows at the inner's first word PC
-			-- for debugger context: source-level stepping now shows the outer body line
+			-- for debugger context: source-level stepping now shows the outer body line (not the inner body line) when stepping into the inner. PROBLEM B fix.
-			-- (not the inner body line) when stepping into the inner. PROBLEM B fix.
 			-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content line in the parent's source), 
 			-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call: 
 			-- gdb 12.1 picks the displayed line as the LAST row at the same PC in byte-stream order,
@@ -653,11 +645,8 @@ local function build_atom_sequence(atom)
 			local ancestry = ancestry_idx[idx]
 			for ai, anc in ipairs(ancestry) do
 				assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-				assert(anc.body_lines[1] ~= nil
+				assert(anc.body_lines[1] ~= nil, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d", anc.component_name, anc.start_pos, #(anc.body_lines or {})))
-					, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d",
+				assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
-						anc.component_name, anc.start_pos, #(anc.body_lines or {})))
-				assert(anc.call_path and anc.call_path ~= ""
-					, "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
 				emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line, true)
 				local is_outermost = (ai == 1)
 				if not (is_outermost and anc.debug_skip) then
@@ -673,9 +662,7 @@ local function build_atom_sequence(atom)
 			-- Marked invocations emit non-statement body rows at every body word; unmarked invocations emit statement body rows.
 			assert(inv.body_lines, "missing body_lines: emitter did not run emission-model")
 			local words_into = idx - inv.start_pos
-			assert(inv.body_lines[words_into] ~= nil
+			assert(inv.body_lines[words_into] ~= nil, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d", words_into, inv.component_name, inv.start_pos, #(inv.body_lines or {}), idx))
-				, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d",
-					words_into, inv.component_name, inv.start_pos, #(inv.body_lines or {}), idx))
 			emit_row(resolve_provenance_file_index(inv.def_path), inv.body_lines[words_into], not inv.debug_skip)
 		else
 			-- RAW word: single call-site row, always a statement target (the word itself is unmarked).
@@ -777,9 +764,8 @@ local function build_atom_table(corpus, addrs)
 
 	local out = {}
 	-- Walk every source's atom list (which preserves source order + per-source src_path).
-	-- Cross-ref with the nm symbol table; atoms absent from `addrs` are skipped (an atom
+	-- Cross-ref with the nm symbol table; atoms absent from `addrs` are skipped
-	-- declared in source but not emitted as a symbol is a metaprogram or atom-info bug, not
+	-- (an atom declared in source but not emitted as a symbol is a metaprogram or atom-info bug, not a source-correlation bug — emit_no_emit would catch it upstream).
-	-- a source-correlation bug — emit_no_emit would catch it upstream).
 	for _, src in ipairs((corpus and corpus.source_order) or {}) do
 		local src_path = src.path or ""
 		for _, atom_rec in ipairs(((src.scan or {}).atoms) or {}) do
@@ -840,12 +826,14 @@ end
 ---   load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
 ---   ...
 ---
+--- Also matches `load_half` / `load_half_u` / `load_byte` / `load_byte_u` (any MIPS load instruction with `(R_<reg>, R_<base>, O_(<Binds_X>, FieldName))` shape).
+--- Every field's `byte_size` + `offset` determine which load to emit; this function only records the (reg, field) pair.
+---
 --- The GPR for each `R_<reg>` is looked up in the merged register_alias_registry; aliases absent from the registry
 --- (no `atom_reg` opt-in) are silently skipped — the resulting rbind record will be incomplete and the atom will fail to bind a usable piece chain.
 --- This is intentional: silently falling back to a hardcoded GPR would mask the missing opt-in.
 ---
---- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level
+--- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level statements (each entry is a single `load_*` call or other statement).
---- statements (each entry is a single `load_word(...)` call or other statement).
 --- @param body_tokens table[]  -- the atom's pre-tokenized body statements (from atom.body_tokens)
 --- @param binds_name  string   -- expected Binds_X name (skip pairs with mismatching binds)
 --- @param registries  table    -- merged registries from collect_per_source_registries
@@ -853,14 +841,18 @@ end
 local function parse_body_load_pairs(body_tokens, binds_name, registries)
 	local pairs = {}
 	local reg_index_by_name = (registries and registries.register_alias_registry) or {}
+	-- One regex that matches any of: load_word, load_half, load_half_u, load_byte, load_byte_u, gte_lw, gte_lwc2.
+	-- The captured ident is `kind`; `inner` holds the parens body for arg parsing.
+	local load_pattern = "^(load_word|load_half|load_half_u|load_byte|load_byte_u|gte_lw|gte_lwc2)%s*%((.*)%)$"
 	for _, t in ipairs(body_tokens or {}) do
 		local tok         = duffle.trim(t.tok or "")
-		-- Match "load_word(...)" — the entire call is one body_tokens entry.
+		local kind, inner = tok:match(load_pattern)
-		local inner = tok:match("^load_word%s*%((.*)%)$")
+		if kind then
-		if inner then
 			local args = duffle.split_top_level_commas(inner)
-			-- Expected shape: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
+			-- Expected shape for an rbind piece-chain load: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
-			if #args >= 3 then
+			-- The second arg MUST be R_TapePtr — loads from other bases (e.g. `load_byte_u(R_RawStatus, R_PadRaw, 0)`)
+			-- are field-derivative loads that read already-bound tape values; they're NOT a new piece-chain.
+			if #args >= 3 and duffle.trim(args[2]) == "R_TapePtr" then
 				local reg_name  = duffle.trim(args[1])
 				local third_arg = duffle.trim(args[3])
 				-- Match O_(Binds_<X>, FieldName)
@@ -879,9 +871,7 @@ local function parse_body_load_pairs(body_tokens, binds_name, registries)
 end
 
 --- Collect every rbind atom + the matching Binds_X struct + (reg, field) pairs.
----
 --- Inputs come from the dep-closed `scan-source` pass (the per-source `src.scan` payload is preserved on each `corpus.source_order` entry).
----
 --- Returns:
 ---   rbind_atoms     = {[atom_name] = {binds, fields, regs, byte_size, info_line}}
 ---   rbind_structs   = {[binds_name] = {byte_size, fields, atom_names}}
@@ -981,7 +971,8 @@ end
 --- (the final unit, referenced by the main CU's DW_AT_stmt_list).
 ---
 --- This builder extends the main compilation unit.
---- A detached synthetic line unit has no DW_AT_stmt_list referencing it, so gdb ignored it (a previous experiment); byte 13 is the first special opcode, not the extended-opcode marker.
+--- A detached synthetic line unit has no DW_AT_stmt_list referencing it, so gdb ignored it (a previous experiment);
+--- byte 13 is the first special opcode, not the extended-opcode marker.
 --- The existing final unit already contains hello_gte_tape.c as file index 11 and ends with a valid end_sequence.
 --- We preserve its bytes, append independent atom sequences, and increase only that unit's DWARF32 unit_length.
 --- @param existing   string -- existing section bytes, byte-for-byte
@@ -992,9 +983,7 @@ local function build_dwarf_line_section(existing, atom_table)
 
 	-- Build the sequences.
 	local sequences = {}
-	for _, atom in ipairs(atom_table) do
+	for _, atom in ipairs(atom_table) do sequences[#sequences + 1] = build_atom_sequence(atom) end
-		sequences[#sequences + 1] = build_atom_sequence(atom)
-	end
 	local appended = table.concat(sequences)
 
 	-- Walk DWARF32 line units and retain the final unit's bounds.
@@ -1003,7 +992,7 @@ local function build_dwarf_line_section(existing, atom_table)
 	while unit_pos < #existing do
 		if unit_pos + 4 > #existing then return existing end
 		local unit_length =  elf_dwarf.read_u32_le(existing, unit_pos)
-		if    unit_length == elf_dwarf.ELF32.dw_dwarf32_terminator then return existing end
+		if    unit_length == elf_dwarf.dw_dwarf32_terminator then return existing end
 		local unit_end_excl = unit_pos + 4 + unit_length
 		if unit_end_excl > #existing then return existing end
 		last_pos, last_length, last_end = unit_pos, unit_length, unit_end_excl
@@ -1064,7 +1053,7 @@ local function build_dwarf_aranges_section(existing, atom_table)
 	while i < #existing do
 		-- Read this unit's length.
 		local ul = elf_dwarf.read_u32_le(existing, i)
-		if    ul == elf_dwarf.ELF32.dw_dwarf32_terminator then
+		if    ul == elf_dwarf.dw_dwarf32_terminator then
 			-- DWARF64 marker - not supported.
 			io.stderr:write("[dwarf_injection] WARN: .debug_aranges contains a DWARF64 marker (0xFFFFFFFF); the 64-bit extension is not supported by this metaprogram; passing through unchanged\n")
 			return existing
@@ -1528,13 +1517,12 @@ end
 ---        DW_AT_location  = piece-chain (DW_FORM_exprloc)
 ---        DW_AT_type      = ref4 → structure_type DIE
 ---
---- This function does NOT emit the final 0 byte (root terminator). build_debug_info_section splices bytes ahead of the root terminator
+--- This function does NOT emit the final 0 byte (root terminator).
---- and preserves existing DIE bytes exactly.
+--- build_debug_info_section splices bytes ahead of the root terminator and preserves existing DIE bytes exactly.
 ---
 --- ref4 basis: DW_FORM_ref4 is CU-relative (offset from the first byte of the CU header).
 --- Our inserted DIEs live in the main CU, so every ref4 = (target section offset) - main_cu_offset.
 --- Per-die section offsets are tracked via the running `next_offset` cursor (= section offset of the NEXT byte to emit).
----
 --- @param main_cu_offset integer    -- 0-based section offset of the main CU's unit_length field
 --- @param main_cu_end_excl integer  -- 0-based section offset of the first byte AFTER the main CU
 --- @param atom_table table[]        -- atoms (with atom.rbind set if rbind; atom.invocations set if mac_X(...) calls)
@@ -1654,6 +1642,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	--
 	-- The table is small + explicit — the prototype principle treats the typed-view struct layout as data, not derived state.
 	local STRUCT_MEMBER_TABLE = {
+		-- TODO(Ed): This hardcoding is brittle...
+		-- TODO(Ed): Better to just have a table for the fundamental types in duffle/dsl.h, we can derive the rest via typedef parsing...
 		-- 2-element signed short vector (rare; placeholder for future use).
 		V2_S2 = { byte_size = 4, members = {
 			{ name = "x", offset = 0, byte_size = 2 },
@@ -1906,6 +1896,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		end
 		local atom_view = (registries.atom_views or {})[atom.name]
 		-- Build the atom-name lookup table once (cheap; O(atom_table)) so step (b) and step (d) can resolve rbind_atom names.
+		-- TODO(Ed): Bad assignment?
 		local atom_by_name = atom_by_name or (function() local m = {}; for _, a in ipairs(atom_table) do if a.name then m[a.name] = a end end; return m end)()
 		-- step (b) inputs: this atom's `atom_ctx(<rbind_atom>)` (resolved from the registries' atom_ctxs)
 		local this_ctx = registries.atom_ctxs and registries.atom_ctxs[atom.name]
@@ -1986,6 +1977,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			end,
 			-- (e) enum-site atom_type(<T>) default on the registry entry.
 			function(r_name, alias_code)
+				-- TODO(Ed): Bad definition?
 				if alias and alias.default_type and alias.default_depth and alias.default_depth > 0 then
 					return type_chain_offsets[alias.default_type .. "|" .. alias.default_depth]
 				end
@@ -2098,7 +2090,6 @@ local function build_debug_str_section(existing, atom_table, registries)
 end
 
 --- Build the new .debug_info: SPLICE inserted DIEs into the MAIN CU as children.
----
 --- This implementation:
 ---   1. Builds the inserted-children bytes (base_type, struct_types, subprograms with their RR_* + bind_args children) via build_inserted_children.
 ---   2. Patches the main CU's `unit_length` field to account for the inserted bytes.
@@ -2217,7 +2208,8 @@ function M.run(ctx)
 
 	-- Read the existing DWARF sections directly (no subprocess; io.open + manual ELF32 section-header walk).
 	-- We need all 8 sections: .debug_line / .debug_aranges / .debug_rnglists get extended (additional rows appended to the existing unit),
-	-- and .debug_info / .debug_abbrev / .debug_str / .debug_loc / .debug_loclists get spliced (the main CU's unit_length is patched; no new compile unit is appended; .debug_loc/.debug_loclists may not exist in the source ELF so we add-section them on splice).
+	-- and .debug_info / .debug_abbrev / .debug_str / .debug_loc / .debug_loclists get spliced
+	-- (the main CU's unit_length is patched; no new compile unit is appended; .debug_loc/.debug_loclists may not exist in the source ELF so we add-section them on splice).
 	-- The per-section dispatch is inlined in the writers loop below.
 	local existing_sections = elf_dwarf.read_elf_sections(elf_path, {
 		".debug_line", ".debug_aranges", ".debug_rnglists",
@@ -2234,8 +2226,7 @@ function M.run(ctx)
 	-- `corpus` is the sole canonical source projection.
 	local corpus     = (ctx.shared and ctx.shared.corpus) or {}
 	local registries = collect_per_source_registries(corpus)
-	-- Read nm symbols (the ONLY disk-side input to the atom table) and join
+	-- Read nm symbols (the ONLY disk-side input to the atom table) and join them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
-	-- them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
 	-- Disk source-map/provenance text is not consulted (those are diagnostic artifacts; semantic inputs are in memory).
 	local addrs      = elf_dwarf.read_nm(ctx.flags.elf_path)
 	local atom_table = build_atom_table(corpus, addrs)

scripts/passes/emission_model.lua

@@ -41,7 +41,6 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- 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.
@@ -59,8 +58,7 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	-- `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)
+	local root_body_line  = root_line_of(atom_record.body_off - 1) or atom_record.line or 0
-		or atom_record.line or 0
 	local component_index = corpus.component_body_index or {}
 	local word_items      = {}
 

scripts/passes/offsets.lua

@@ -3,12 +3,11 @@
 --- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
 --- for `MipsAtom_(name)` and `MipsCode code_<name>` declarations, computes the word offset
 --- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
---- `<dir_basename>.offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+--- `gen/offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+--- Per-directory aggregation: every source in the same directory contributes to the same `gen/offsets.h`.
+--- The directory itself is the namespace; the filename does not repeat the module name.
 ---
 --- The offset is `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding: branch_offset = relative_pc_in_words - 1).
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -16,8 +15,7 @@
 
 -- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
@@ -39,17 +37,17 @@ local OFFSET_MACRO_COL = 44
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table -- Cross-pass shared state
---- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared.corpus      table -- Corpus projection
 --- @field shared.word_counts table
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string -- Output root (e.g. "build/gen")
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -57,22 +55,24 @@ 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      -- 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      -- Target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field branch_word integer  -- branch 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 per consuming instruction (see `compute_offsets`)
+--- @field consuming_encoder string|nil  -- Instruction consuming the offset (e.g. "branch_le_zero", "jump", "call_addr")
+--- @field consuming_arg_pos integer|nil -- 1-based arg position within the consuming instruction's arg list
 
 --- @class AtomData
---- @field name        string         -- atom name
+--- @field name        string         -- Atom name
---- @field total_words integer        -- total word count of the atom body
+--- @field total_words integer        -- Total word count of the atom body
---- @field offsets     BranchOffset[] -- per-branch offset list
+--- @field offsets     BranchOffset[] -- Per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Canonical marker projection
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- MARKER_PROJECTORS is the marker-kind data table.
--- The emission-model pass already records marker word positions;
+-- The emission-model pass already records marker word positions + consuming-instruction context;
 -- this pass only projects those records into the label/branch lookup shape needed by offset computation.
 local MARKER_PROJECTORS = {
 	label = function(state, marker)
@@ -83,6 +83,8 @@ local MARKER_PROJECTORS = {
 			tag                = marker.name,
 			target             = marker.target,
 			branch_word        = marker.word_index,
+			consuming_encoder  = marker.consuming_encoder,
+			consuming_arg_pos  = marker.consuming_arg_pos,
 		}
 	end,
 }
@@ -104,7 +106,18 @@ end
 -- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
+--- Compute branch offsets per consuming instruction.
+--- Disposition table:
+---   `branch_*`           -> relative offset: `target_word - branch_word - 1` (MIPS branch-immediate encoding).
+---   `jump` / `call_addr` -> same value as `branch_*` (a relative word offset). 
+---     The duffle headers' `enc_i` macro truncates the value to the immediate-field width (16 bits for branches, 26 bits for jumps).
+---     For tape-atom bodies within a single module, this works for `j`/`jal` because the linker's symbol resolution produces the correct 26-bit absolute target via standard `j` relocations.
+---     For cross-module `j`/`jal` (atom body in one module, target in another), the linker emits a `R_MIPS_26` relocation against the lower 26 bits; the upper 4 bits come from the PC of the delay slot following the `j`.
+---     The metaprogram doesn't know either at compile time, so the emitted value is the relative word offset that the duffle `enc_i` macro places in the immediate field; the toolchain handles the rest.
+---   `jump_reg` / `call_reg` / `jump_link` -> ERROR. Register-form jumps have no offset field; `atom_offset` is invalid.
+---
+--- Top-level `atom_offset(F, T)` markers (where the marker is the entire token — `consuming_encoder` == nil) default to `branch_*` behavior (relative offset).
+--- This preserves backward compatibility for any top-level marker that may exist outside a control-transfer instruction.
 --- @param labels   table<string, integer>
 --- @param branches table[]
 --- @return BranchOffset[]
@@ -115,11 +128,23 @@ local function compute_offsets(labels, branches)
 		if not target then
 			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")")
 		end
+		local consuming = br.consuming_encoder
+		local offset
+		if consuming == "jump_reg" or consuming == "call_reg" or consuming == "jump_link" then
+			-- Register-form jumps have no offset field. `atom_offset` cannot be used here.
+			error("atom_offset cannot be used with " .. consuming
+				.. " (register-form jumps have no offset field); at word " .. br.branch_word)
+		end
+		-- All other consuming instructions (including `branch_*`, `jump`, `call_addr`, and nil for top-level markers) use the same relative offset value.
+		-- The MIPS encoding differs per opcode but the duffle `enc_i` macro handles the truncation to the immediate-field width.
+		offset = target - br.branch_word - 1
 		results[#results + 1] = {
 			target             = br.target,
 			tag                = br.tag,
 			branch_word        = br.branch_word,
-			offset      = target - br.branch_word - 1,
+			offset             = offset,
+			consuming_encoder  = br.consuming_encoder,
+			consuming_arg_pos  = br.consuming_arg_pos,
 		}
 	end
 	return results
@@ -167,41 +192,45 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
---- Generate the per-source .offsets.h header.
+--- Generate the per-directory .offsets.h header.
---- @param source_path string
+--- @param dir          string     -- the absolute source directory
+--- @param sources      table[]    -- sources contributing to this directory (for the header comment)
 --- @param atoms_data   AtomData[]
 --- @return string
-local function generate_header(source_path, atoms_data)
+local function generate_header(dir, sources, atoms_data)
-	local basename = duffle.basename_no_ext(source_path)
+	local dir_basename = duffle.basename_no_ext(dir)
 
 	local lines = {}
 	local function add(s) lines[#lines + 1] = s end
 
 	add("// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT")
-	add("// Source: " .. source_path)
+	add("// Directory: " .. dir:gsub("/", "\\") .. "\\")
+	for _, src in ipairs(sources) do
+		add("//   source: " .. src.path:gsub("/", "\\"))
+	end
 	add("#pragma once")
 	add("")
-	add("#pragma region " .. basename)
+	add("#pragma region " .. dir_basename)
 	add("")
 	add("")
 	for _, atom in ipairs(atoms_data) do
 		emit_atom_offsets(add, atom)
 	end
-	add("#pragma endregion " .. basename)
+	add("#pragma endregion " .. dir_basename)
 	add("")
 	return table.concat(lines, "\n") .. "\n"
 end
 
 local M = {}
 
---- (internal) Process one source: render offsets from canonical atom paths.
+--- (internal) Aggregate atoms from every source in one directory, render the per-directory `offsets.h`.
 --- Returns the offsets_h path if a header was written, or nil.
 --- @param ctx     PassCtx
---- @param src SourceFile
+--- @param dir     string       -- the absolute source directory
+--- @param sources SourceFile[] -- sources in this directory
 --- @return string|nil  -- the offsets_h path
-local function process_source(ctx, src)
+local function process_directory(ctx, dir, sources)
 	local atoms_data = {}
-	local scan = src.scan or {}
 
 	local function append_atom(atom)
 		local  paths = atom and atom.paths
@@ -214,19 +243,22 @@ local function process_source(ctx, src)
 		}
 	end
 
+	for _, src in ipairs(sources) do
+		local scan = src.scan or {}
 		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
+	end
 	if #atoms_data == 0 then return nil end
 
-	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
+	local out_path = dir .. "/gen/offsets.h"
 	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(dir, sources, atoms_data))
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
+--- For each canonical source-directory, emits a per-directory `gen/offsets.h`
---- containing constants for every marker recorded in atom.paths.
+--- containing constants for every marker recorded in atom.paths across every source in that directory.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -235,12 +267,17 @@ function M.run(ctx)
 	local warnings = {}
 
 	local corpus = ctx.shared and ctx.shared.corpus
-	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+	if type(corpus) ~= "table" then
-		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+		error("offsets.run requires ctx.shared.corpus", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order.", 0)
 	end
 
-	for _, src in ipairs(corpus.source_order) do
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`.
-		local out_path = process_source(ctx, src)
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
+	for dir, sources in pairs(sources_by_dir) do
+		local out_path = process_directory(ctx, dir, sources)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }
 		end

scripts/passes/report.lua

@@ -1,22 +1,21 @@
---- passes/report.lua — Per-MODULE annotation report renderer +
+--- passes/report.lua — Per-MODULE annotation report renderer + project-wide summary writer.
---- project-wide summary writer.
 ---
 --- Two output files per build:
 ---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
 ---   - `build/gen/annotation_validation.txt` — the project summary.
 ---
 --- The annotation pass emits `errors.h` files per module and the canonical `corpus.sources_by_dir` projection groups sources by directory.
---- This pass iterates the canonical dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
+--- This pass iterates the dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- 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: See `ps1_meta.lua` for the rationale.
--- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- 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).
+-- 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")
@@ -59,20 +58,20 @@ local PASS_NAME = "report"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field sources            SourceFile[]         -- all source files in the build
+--- @field sources            SourceFile[]         -- All source files in the build
---- @field metadata_path      string               -- path to word_count.metadata.h
+--- @field metadata_path      string               -- Path to word_count.metadata.h
---- @field shared             table                -- cross-pass shared state
+--- @field shared             table                -- Cross-pass shared state
---- @field out_root           string               -- output root (e.g. "build/gen")
+--- @field out_root           string               -- Output root (e.g. "build/gen")
---- @field project_root       string               -- project root (e.g. "code/")
+--- @field project_root       string               -- Project root (e.g. "code/")
---- @field upstream           table<string, table> -- per-pass upstream outputs
+--- @field upstream           table<string, table> -- Per-pass upstream outputs
 --- @field flags              table                -- CLI flags + per-pass stash
---- @field verbose            boolean              -- if true, log diagnostic info
+--- @field verbose            boolean              -- If true, log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -82,60 +81,60 @@ local PASS_NAME = "report"
 -- Shapes produced by `passes/annotation.lua`'s `M.validate()`.
 
 --- @class AtomEntry
---- @field name string   -- atom name (e.g. "cube_g4_face")
+--- @field name string   -- Atom name (e.g. "cube_g4_face")
---- @field line integer  -- source line of the atom declaration
+--- @field line integer  -- Source line of the atom declaration
 
 --- @class AnnotEntry
---- @field line    integer  -- source line
+--- @field line    integer     -- Source line
---- @field macro   string   -- the macro name (e.g. "atom_reads")
+--- @field macro   string      -- Macro name (e.g. "atom_reads")
---- @field name    string   -- the atom name (if a `name(...)` was given)
+--- @field name    string      -- Atom name (if a `name(...)` was given)
 --- @field kind    string      -- "atom_info" | "atom_bind" | ...
 --- @field binds   string|nil  -- Binds_X name if any
 --- @field reads   string[]    -- R_* names (read targets)
 --- @field writes  string[]    -- R_* names (write targets)
---- @field error   string|nil  -- error message if annotation was malformed
+--- @field error   string|nil  -- Error message if annotation was malformed
 
 --- @class BindsField
---- @field name   string   -- field name
+--- @field name   string   -- Field name
---- @field offset integer  -- byte offset within the Binds_X struct
+--- @field offset integer  -- Byte offset within the Binds_X struct
 
 --- @class BindsStruct
---- @field name   string         -- struct name (e.g. "Binds_Floor")
+--- @field name   string         -- Struct name (e.g. "Binds_Floor")
---- @field line   integer        -- source line of the typedef
+--- @field line   integer        -- Source line of the typedef
---- @field bytes  integer        -- total byte size
+--- @field bytes  integer        -- Total byte size
---- @field fields BindsField[]   -- the field list
+--- @field fields BindsField[]   -- The field list
 
 --- @class MacroEntry
---- @field name  string   -- macro name (e.g. "WORD_COUNT(my_macro, 4)")
+--- @field name  string   -- Macro name (e.g. "WORD_COUNT(my_macro, 4)")
---- @field line  integer  -- source line
+--- @field line  integer  -- Source line
---- @field words integer  -- declared word count
+--- @field words integer  -- Declared word count
 
 --- @class Finding
---- @field line integer  -- source line
+--- @field line integer  -- Source line
---- @field msg  string   -- finding message
+--- @field msg  string   -- Finding message
 
 --- @class AnnotationResult
---- @field source   string         -- set by this pass; original source path
+--- @field source   string         -- Set by this pass; original source path
---- @field atoms    AtomEntry[]    -- atom declarations in this source
+--- @field atoms    AtomEntry[]    -- Atom declarations in this source
---- @field annots   AnnotEntry[]   -- annotation entries
+--- @field annots   AnnotEntry[]   -- Annotation entries
---- @field macros   MacroEntry[]   -- macro word-count declarations
+--- @field macros   MacroEntry[]   -- Macro word-count declarations
 --- @field binds    BindsStruct[]  -- Binds_* struct declarations
---- @field errors   Finding[]      -- errors from validation
+--- @field errors   Finding[]      -- Errors from validation
---- @field warnings Finding[]      -- warnings from validation
+--- @field warnings Finding[]      -- Warnings from validation
---- @field info     table          -- info summary (not rendered here)
+--- @field info     table          -- Info summary (not rendered here)
 
 --- @class ModuleEntry
---- @field dir          string   -- absolute directory path
+--- @field dir          string   -- Absolute directory path
---- @field dir_basename string   -- basename (e.g. "duffle", "gte_hello")
+--- @field dir_basename string   -- Basename (e.g. "duffle", "gte_hello")
---- @field atoms_count  integer  -- pre-counted atoms for filtering
+--- @field atoms_count  integer  -- Pre-counted atoms for filtering
 
 --- @class ModuleReport
---- @field dir       string             -- module directory
+--- @field dir       string             -- Module directory
---- @field sources   SourceFile[]       -- sources in this module
+--- @field sources   SourceFile[]       -- Sources in this module
---- @field results   AnnotationResult[] -- per-source validate() results
+--- @field results   AnnotationResult[] -- Per-source validate() results
 
 --- @class ProjectReport
---- @field results AnnotationResult[]  -- all per-source results
+--- @field results AnnotationResult[]  -- All per-source results
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-MODULE annotation report (aggregated across all sources in a dir)
@@ -153,9 +152,16 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- Render the thin project-wide summary (`build/atom_meta_report.summary.md`).
---- @param all_results { module:string, atoms:integer, annots:integer, binds:integer,
+--- @param all_results { 
----                       macros:integer, findings:integer, errors:integer,
+--- 	module:string,
----                       warnings:integer, info:integer }[]
+--- 	atoms:integer, 
+--- 	annots:integer, 
+--- 	binds:integer,
+---		macros:integer,   
+---		findings:integer,
+---		errors:integer,
+--- 	warnings:integer,
+--- 	info:integer }[]
 --- @return string
 local function render_project_summary(all_results)
 	local lines = {
@@ -165,13 +171,10 @@ local function render_project_summary(all_results)
 		"| module | atoms | annots | binds | macros | findings | errors | warnings | info |",
 		"|--------|-------|--------|-------|--------|----------|--------|----------|------|",
 	}
-	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0,
+	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0, findings = 0, errors = 0, warnings = 0, info = 0 }
-		findings = 0, errors = 0, warnings = 0, info = 0 }
 	for _, e in ipairs(all_results) do
-		lines[#lines + 1] = string.format(
+		lines[#lines + 1] = string.format("| %s | %d | %d | %d | %d | %d | %d | %d | %d |"
-			"| %s | %d | %d | %d | %d | %d | %d | %d | %d |",
+			, e.module, e.atoms, e.annots, e.binds, e.macros, e.findings, e.errors, e.warnings, e.info)
-			e.module, e.atoms, e.annots, e.binds, e.macros,
-			e.findings, e.errors, e.warnings, e.info)
 		totals.atoms    = totals.atoms    + e.atoms
 		totals.annots   = totals.annots   + e.annots
 		totals.binds    = totals.binds    + e.binds
@@ -181,10 +184,8 @@ local function render_project_summary(all_results)
 		totals.warnings = totals.warnings + e.warnings
 		totals.info     = totals.info     + e.info
 	end
-	lines[#lines + 1] = string.format(
+	lines[#lines + 1] = string.format("| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |"
-		"| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |",
+		, totals.atoms, totals.annots, totals.binds, totals.macros, totals.findings, totals.errors, totals.warnings, totals.info)
-		totals.atoms, totals.annots, totals.binds, totals.macros,
-		totals.findings, totals.errors, totals.warnings, totals.info)
 	return table.concat(lines, "\n") .. "\n"
 end
 
@@ -326,8 +327,8 @@ local function render_module_meta_report(dir, dir_sources, annot_results, sa_res
 				local binds  = a.binds  or "—"
 				local reads  = (#a.reads  > 0 and table.concat(a.reads,  ",")) or "—"
 				local writes = (#a.writes > 0 and table.concat(a.writes, ",")) or "—"
-				add(string.format("| %s | %d | %s | %s | %s | %s |",
+				add(string.format("| %s | %d | %s | %s | %s | %s |"
-					src_name, a.line, a.name, binds, reads, writes))
+					, src_name, a.line, a.name, binds, reads, writes))
 			end
 		end
 	end

scripts/passes/scan_source.lua

@@ -2,7 +2,6 @@
 ---
 --- Single source-walk pass that produces the fat `SourceScan` payload consumed by all downstream passes. Walks each corpus source record once,
 --- extracting every construct type the metaprograms need:
----
 ---   MipsAtom_              (kind = "atom", with optional atom_info inner)
 ---   MipsAtomComp_          (kind = "comp_bare")
 ---   MipsAtomComp_Proc_     (kind = "comp_proc", body inside last {})
@@ -14,8 +13,6 @@
 --- The result is attached to each `src.scan` so downstream passes can read from `src.scan.atoms` / `src.scan.binds` / etc. without re-walking the source.
 --- This is the first pass in the dep graph (no deps).
 --- Every other pass that reads source structure depends on this one — see `ps1_meta.lua :: PASSES`.
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible
 
 -- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
@@ -50,12 +47,12 @@ local parse_enum_int_literal
 --- @field line_of           fun(pos: integer): integer       -- shared LineIndex closure
 
 --- @class DebugSkipMarker
---- @field marker_kind    string      -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive; any other ident reaches the unrelated fallback and is never associated with a declaration.
+--- @field marker_kind    string      -- Exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive; any other ident reaches the unrelated fallback and is never associated with a declaration.
---- @field marker_line    integer     -- line of the marker ident start
+--- @field marker_line    integer     -- Line of the marker ident start
---- @field marker_pos     integer     -- byte position of the marker ident start (the comment walker anchors here)
+--- @field marker_pos     integer     -- Byte position of the marker ident start (the comment walker anchors here)
 --- @field is_bare        boolean     -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
 --- @field has_parens     boolean     -- true iff a `(...)` follows the marker ident (diagnostic-only)
---- @field args           string|nil  -- trimmed args inside the `(...)` (nil when has_parens is false)
+--- @field args           string|nil  -- Trimmed args inside the `(...)` (nil when has_parens is false)
 --- @field pending        boolean     -- true while awaiting the following declaration
 --- @field superseded_by_marker_line integer|nil -- set when a newer marker bumped this one out of the pending slot
 --- @field target_kind    string|nil  -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed (nil if no declaration ever followed)
@@ -71,15 +68,15 @@ local parse_enum_int_literal
 --- @field reg           string  -- "R_T0"
 --- @field type_name     string
 --- @field pointer_depth integer
---- @field source_line   integer -- line of the call site (callsite or enum-site)
+--- @field source_line   integer -- Line of the call site (callsite or enum-site)
 
 --- @class AtomCtxEntry
---- @field rbind_atom string  -- the rbind atom ident that this consumer should propagate types from
+--- @field rbind_atom string  -- The rbind atom ident that this consumer should propagate types from
 --- @field info_line  integer
---- @field source     string  -- absolute path of the source file
+--- @field source     string  -- Absolute path of the source file
 
 --- @class AtomPhaseGroup
---- @field atoms string[]  -- atom names tagged with this phase label (source-order)
+--- @field atoms string[]  -- Atom names tagged with this phase label (source-order)
 
 --- @class AtomViewEntry
 --- @field atom_name        string                 -- e.g. "red_cube_g4_face"
@@ -88,11 +85,11 @@ local parse_enum_int_literal
 --- @field info_line        integer                -- line of the atom_info call
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (set by this pass)
+--- @field scan     table   -- Pre-scanned SourceScan payload (set by this pass)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -111,15 +108,15 @@ local parse_enum_int_literal
 
 --- @class AtomEntry
 --- @field line              integer
---- @field name              string     -- atom name (for components: without ac_ prefix)
+--- @field name              string       -- Atom name (for components: without ac_ prefix)
---- @field body              string     -- brace-delimited body (without the braces)
+--- @field body              string       -- Brace-delimited body (without the braces)
---- @field body_off          integer    -- char offset of body[1] in source
+--- @field body_off          integer      -- Char offset of body[1] in source
 --- @field kind              string       -- "atom" | "comp_bare" | "comp_proc" | "raw_atom"
---- @field raw_name          string     -- un-stripped name (for components: with ac_ prefix)
+--- @field raw_name          string       -- Un-stripped name (for components: with ac_ prefix)
---- @field ident_pos         integer    -- position of the MipsAtom_/MipsAtomComp_ ident start
+--- @field ident_pos         integer      -- Position of the MipsAtom_/MipsAtomComp_ ident start
---- @field after_paren       integer    -- position past the closing paren
+--- @field after_paren       integer      -- Position past the closing paren
 --- @field debug_skip        boolean      -- true when an `atom_dbg_skip` bare marker immediately precedes this declaration (sole-owner stamp; see push_debug_skip_marker)
---- @field declaration_comment string|nil -- populated by the scanner (backward walk past the marker, captures contiguous `/* */` or `//` block)
+--- @field declaration_comment string|nil -- Populated by the scanner (backward walk past the marker, captures contiguous `/* */` or `//` block)
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Local helpers (shared by per-form parsers)
@@ -139,10 +136,10 @@ local QUALIFIER_KEYWORDS = {
 local AC_PREFIX     = "ac_"
 local AC_PREFIX_LEN = 3
 
--- Strip the "ac_" prefix from a component name.
+--- Strip the "ac_" prefix from a component name.
--- Returns the input unchanged if it doesn't start with the prefix.
+--- Returns the input unchanged if it doesn't start with the prefix.
--- @param raw_name string
+--- @param raw_name string
--- @return string
+--- @return string
 local function strip_ac_prefix(raw_name)
 	if #raw_name > AC_PREFIX_LEN and raw_name:sub(1, AC_PREFIX_LEN) == AC_PREFIX then
 		return raw_name:sub(AC_PREFIX_LEN + 1)
@@ -184,19 +181,19 @@ local function find_body_braces(source, after_paren, fallback)
 	return body, after_brace, brace + 1
 end
 
--- Walk backward from `start_pos` capturing contiguous `/* */` block(s) and
+--- Walk backward from `start_pos` capturing contiguous `/* */` block(s) and
--- `//` line(s) that immediately precede it. The caller (preceding_declaration_comment)
+--- `//` line(s) that immediately precede it. The caller (preceding_declaration_comment)
--- supplies `start_pos` so the walker does not need to detect marker shape or prelude layout.
+--- supplies `start_pos` so the walker does not need to detect marker shape or prelude layout.
--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
+--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
---
+---
--- The walker captures:
+--- The walker captures:
---   - Block comment close `*/` followed by walking back to `/*`.
+---   - Block comment close `*/` followed by walking back to `/*`.
---   - `//` line comments (the line containing the current non-ws position starts with `//`).
+---   - `//` line comments (the line containing the current non-ws position starts with `//`).
--- It stops at the first non-ws char that does not begin a comment block or line.
+--- It stops at the first non-ws char that does not begin a comment block or line.
--- Empty string if no comment is adjacent.
+--- Empty string if no comment is adjacent.
--- @param source string
+--- @param source    string
--- @param start_pos integer -- exclusive upper bound for the captured block
+--- @param start_pos integer -- exclusive upper bound for the captured block
--- @return string
+--- @return string
 local function preceding_comment_walk_backward(source, start_pos)
 	local pieces  = {}
 	local scan_pos = start_pos
@@ -249,13 +246,13 @@ local function preceding_comment_walk_backward(source, start_pos)
 	return table.concat(pieces, "\n")
 end
 
--- Resolve the start position for the declaration-comment walk.
+--- Resolve the start position for the declaration-comment walk.
--- When a debug-skip marker is pending, the walker must start from the position immediately before the marker ident
+--- When a debug-skip marker is pending, the walker must start from the position immediately before the marker ident
--- (so it walks backward past the marker text and any `FI_ MipsAtom ac_X(args)` proc-prelude layout — neither of which is visible if we start from the declaration ident_pos).
+--- (so it walks backward past the marker text and any `FI_ MipsAtom ac_X(args)` proc-prelude layout — neither of which is visible if we start from the declaration ident_pos).
--- When no marker is pending, the walker starts from the declaration ident_pos directly.
+--- When no marker is pending, the walker starts from the declaration ident_pos directly.
--- @param pending_marker DebugSkipMarker|nil
+--- @param pending_marker DebugSkipMarker|nil
--- @param ident_pos integer -- declaration ident position
+--- @param ident_pos      integer -- declaration ident position
--- @return integer
+--- @return integer
 local function comment_walk_start(pending_marker, ident_pos)
 	if pending_marker then
 		return pending_marker.marker_pos - 1
@@ -263,16 +260,16 @@ local function comment_walk_start(pending_marker, ident_pos)
 	return ident_pos - 1
 end
 
--- Attach the pending marker to the next declaration.
+--- Attach the pending marker to the next declaration.
--- The declaration form disambiguates whole atoms from components; the resolved `debug_skip` is stamped directly on the declaration record
+--- The declaration form disambiguates whole atoms from components; the resolved `debug_skip` is stamped directly on the declaration record
--- (sole-owner discipline; see push_debug_skip_marker).
+--- (sole-owner discipline; see push_debug_skip_marker).
---
+---
--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
+--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
---   marker_kind == "atom_dbg_skip" AND is_bare == true
+---   marker_kind == "atom_dbg_skip" AND is_bare == true
--- Any other spelling or shape (parenthesized form, legacy name) is recorded as a raw marker for annotation validation but never stamps `debug_skip`.
+--- Any other spelling or shape (parenthesized form, legacy name) is recorded as a raw marker for annotation validation but never stamps `debug_skip`.
--- @param out SourceScan
+--- @param out         SourceScan
--- @param target_kind string|nil -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
+--- @param target_kind string|nil -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
--- @return boolean|nil -- true iff the marker is the positive bare form
+--- @return boolean|nil -- true iff the marker is the positive bare form
 local function attach_debug_skip_marker(out, target_kind)
 	local markers = out.debug_skip_markers
 	local marker  = markers[#markers]
@@ -1280,7 +1277,6 @@ local function parse_mips_atom_comp_proc(source, pos, ident_end, line_of, out)
 	local name     = strip_ac_prefix(raw_name)
 	-- Position of body[1] in source = open_paren + 1 (start of inner) + last_brace_pos + 1 (past '{').
 	local body_off = open_paren + 2 + last_brace_pos
-
 	register_atom(out, "comp_proc", line_of(pos), name, body, body_off, raw_name, pos, after_paren, source)
 
 	return after_paren
@@ -1396,7 +1392,7 @@ end
 ---   1. `typedef Struct_(<name>) { <body> } <alias>;` adds to type_name_registry (kind="struct"). 
 ---       Binds_* aliases also land in out.binds[].
 ---   2. `typedef Enum_(<underlying>, <name>) { <body> } <alias>;` 
----       adds to type_name_registry (kind="enum"). 
+---       Adds to type_name_registry (kind="enum"). 
 ---   3. `typedef <type> <alias>;` simple typedef alias.
 ---       Adds to type_name_registry (kind="typedef").
 ---   4. `typedef <type> TSet_(<name>);` duffle TSet_ convention.
@@ -1472,8 +1468,7 @@ local function parse_typedef_binds(source, pos, ident_end, line_of, out)
 		local  inner, after_paren = read_parens_after(source, id2_end, id2_end)
 		if not inner then return id2_end end
 		local tset_name = duffle.trim(inner)
-		-- Empty underlying span is acceptable; the TSet_ wrapper itself
+		-- Empty underlying span is acceptable; the TSet_ wrapper itself encodes the alias identity (per the duffle TSet_ convention).
-		-- encodes the alias identity (per the duffle TSet_ convention).
 		register_typedef_alias("", tset_name, pos, line_of, out)
 		attach_debug_skip_marker(out, "unrelated")
 		return after_paren

scripts/passes/static_analysis.lua

@@ -1,20 +1,17 @@
 --- passes/static_analysis.lua — Per-atom static-analysis checks.
----
+--- Ownership: `ctx.shared.corpus` canonical merged registry; per-source fallback synthesis is rejected.
---- Ownership: `ctx.shared.corpus` is the canonical merged registry; per-source fallback synthesis is rejected.
 --- `atom.paths` supplies the emitted and analysis projections consumed by this pass.
 ---
 --- Per-atom rules:
 ---   1. transfer_hazards: A single forward walker (`analyze_hardware_relations`) reads `atom.paths.word_events` once per atom.
 ---      For each emitted word event it (a) inspects pending CPU / COP0 / COP2 / GTE relations against the event as CONSUMER 
 ---      (recording a hazard on `atom.paths.hazards` when the producer→consumer gap is below the required retire-slot count), 
----      (b) applies the event's GPR value effects (`duffle.INSTRUCTION_GPR_EFFECTS`) to `atom.paths.forward_state.gpr_values`, 
+---      (b) Applies the event's GPR value effects (`duffle.INSTRUCTION_GPR_EFFECTS`) to `atom.paths.forward_state.gpr_values`, 
 ---      applies bounded constant propagation, and stages matching relation rows as PRODUCERS (with `destination_match` filters, e.g. for the IRGB fan-out).
 ---      The `transfer_hazards` CHECK_RULES reader projects `atom.paths.hazards` into per-atom findings.
----      The reader does NOT re-walk source; this is the per-check purity contract.
+---      The reader does NOT re-walk source. The walker runs once per atom before the per-atom dispatch; the reader runs inside the same dispatch.
----      The walker runs once per atom before the per-atom dispatch; the reader runs inside the same dispatch.
 ---   2. control_transfer_delay_slot_use: For every emitted branch/jump/call encoder in `duffle.CONTROL_TRANSFER_DELAY_SLOT_POLICIES` 
----      (the six `branch_*` encoders plus `jump` / `jump_reg` / `jump_link` / `call_reg` / `call_addr`),
+---      (the six `branch_*` encoders plus `jump` / `jump_reg` / `jump_link` / `call_reg` / `call_addr`), inspect the next emitted event in `atom.paths.word_events`.
----      inspect the next emitted event in `atom.paths.word_events`.
 ---      Emit an `info`-severity finding when the successor is `nop` or absent (the next emitted word IS the hardware delay slot).
 ---      `jump_reg(R_AtomJmp)` is suppressed by policy (the fixed `mac_yield()` handshake).
 ---      `nop2` needs no special case: emission-model emits two `nop` events for it, so the first expansion is the hardware delay slot.
@@ -23,7 +20,9 @@
 ---   4. Binding handoff: Every `atom_bind(Binds_X)` must reference a `typedef Struct_(Binds_X) { ... }` declaration.
 ---   5. GPU Port-Store Shape: Per-shape (`f3`/`f4`/`g4`/etc.) the sum of `mac_format_X_color` + `mac_gte_store_X_*` + `mac_insert_ot_tag_X` words
 ---      must equal the GP0 cmd's expected packet size.
----   6. Per-Atom Cycle Budget: Sum each atom body's instruction latencies (per `duffle.INSTRUCTION_LATENCY`); report total.
+---   6. Per-Atom Cycle Budget: Sum each atom body's instruction latencies — non-`mac_*` tokens look up `duffle.INSTRUCTION_LATENCY[ident]`; 
+---      `mac_*` tokens look up `pipe_ctx.components_by_name[bare_name].cycle_cost` (auto-derived from the original `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`).
+---      Report total.
 ---
 --- Per-source rules (registry-driven):
 ---   8. enum_alias_membership: Every `R_X` referenced from `atom_dbg_reg_default`, `atom_reg_types`, `atom_type(...)`, `atom_reads`, or `atom_writes`
@@ -99,10 +98,10 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -119,43 +118,42 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 --- @field outputs  table[]
 --- @field errors   table[]
 --- @field warnings table[]
---- @field info     table[]   -- finding-level info (kind == "info"); distinct from per-source scanned/cycles summary rows
+--- @field info     table[]   -- Finding-level info (kind == "info"); distinct from per-source scanned/cycles summary rows
 
 --- @alias AtomName    string  -- lower_snake_case atom nameMacroName   string  -- lower_snake_case macro identifier
---- @alias CheckName   string  -- "transfer_hazards" | "control_transfer_delay_slot_use" | "mac_yield_uniformity" | "abi_handoff" | "gpu_portstore_shape" | "per_atom_cycle_budget" | "enum_alias_membership" | "atom_type_consistency" | "binds_no_substruct_deref"
+--- @alias CheckName   string  -- "transfer_hazards" | "control_transfer_delay_slot_use" | "mac_yield_uniformity" | "yield_load_tail_pairing" | "abi_handoff" | "gpu_portstore_shape" | "per_atom_cycle_budget" | "enum_alias_membership" | "atom_type_consistency" | "binds_no_substruct_deref"
 
 --- @class AtomBody
---- @field line     integer  -- source line of the atom declaration
+--- @field line     integer  -- Source line of the atom declaration
---- @field name     AtomName -- atom name (e.g. "cube_g4_face")
+--- @field name     AtomName -- Atom name (e.g. "cube_g4_face")
---- @field body     string   -- the brace-delimited body (without the braces)
+--- @field body     string   -- Brace-delimited body (without the braces)
 --- @field body_off integer  -- char offset of body[1] in source
 --- @field kind     string   -- "atom" | "comp_bare" | "comp_proc"
 
 --- @class Token
---- @field tok      string     -- the raw token text (trimmed)
+--- @field tok      string     -- Raw token text (trimmed)
---- @field line     integer    -- source line of the token's start
+--- @field line     integer    -- Source line of the token's start
---- @field ident    string|nil -- the leading ident of the token (if any)
+--- @field ident    string|nil -- Leading ident of the token (if any)
 --- @field kind     string     -- "n_words" | "mac_yield" | "gte_cmdw" | "mac_format" | "mac_gte_store" | "mac_insert_ot_tag" | "atom_label" | "atom_offset" | "other"
 
 --- @class Finding 
---- @field line     integer   -- source line of the finding
+--- @field line     integer   -- Source line of the finding
---- @field atom     AtomName  -- the atom this finding is for (or "")
+--- @field atom     AtomName  -- Atom this finding is for (or "")
---- @field check    CheckName -- the check identifier
+--- @field check    CheckName -- Check identifier
 --- @field kind     string    -- "error" | "warning" | "info"
---- @field msg      string    -- the finding message
+--- @field msg      string    -- Finding message
 
 --- @class AtomAnalysis
 --- @field atom       AtomBody
---- @field tokens     Token[]   -- the tokens in the atom body, annotated
+--- @field tokens     Token[]   -- Tokens in the atom body, annotated
---- @field findings   Finding[] -- findings for this atom
+--- @field findings   Finding[] -- Findings for this atom
---- @field total_cycles integer -- sum of token cycle costs
+--- @field total_cycles integer -- Sum of token cycle costs
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-word-event helpers
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Pick the source-line field that best represents "where in the user's source file is this word?".
---
 -- `word_events` (populated by `passes/emission_model.lua::stamp_root_provenance`) carry four line fields:
 --   * `call_line`  — physical line in the ROOT atom's source (the line of the `mac_X(...)` call site that triggered this emission, or `body_line` for direct words in the atom body)
 --   * `body_line`  — physical line in the body containing the emitted word (the atom body for direct words; the component body for words expanded inside `mac_X(...)`)
@@ -167,8 +165,7 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 -- The user editing their atom body expects the line to point at THEIR source — i.e. the line where `mac_yield()`
 -- was called (e.g. `hello_gte_tape.c:35`). That line is `call_line`.
 --
--- For direct words in the atom body (no invocation wrapping them), `call_line == body_line` already,
+-- For direct words in the atom body (no invocation wrapping them), `call_line == body_line` already, so `call_line` works for both cases.
--- so `call_line` works for both cases.
 local function line_for_word_event(ev)
     if ev == nil then return 0 end
     return ev.call_line or ev.body_line or ev.line or ev.def_line or 0
@@ -193,36 +190,56 @@ end
 -- Each entry has:
 --   ident                 — the leading identifier (e.g. "load_word", "gte_cmdw_rtpt", "nop", "mac_yield")
 --   nop_words             — 0 / 1 / 2 (for "nop" / "nop2" / anything else)
---   nop_prefix    — consecutive nop words ending just BEFORE this token (forward-pass pre-compute;
+--   nop_prefix            — consecutive nop words ending just BEFORE this token (forward-pass pre-compute; makes preceding-nop lookup O(N))
---                   makes preceding-nop lookup O(N))
 --   is_yield              — true if this token is `mac_yield` or `mac_yield(...)`
 --   is_atom_label         — true if this token is `atom_label(name)`; label_name has the name
---   is_branch     — true if this token is `branch_*(...)`; branch_label has the label or false
+--   is_branch             — true if this token is `branch_*(...)` OR an unconditional-jump-with-offset (`jump(off)` / `call_addr(off)`); branch_label has the target label or false
---   is_load_word  — true if this token starts with `load_word(`
+--   is_unconditional_jump — true if this token is `jump` or `call_addr` (BD slot + single successor — taken only; no fall-through).
+--                           Mutually exclusive with the conditional-branch semantics; combined with `is_branch` above.
+--   is_terminal_jump      — true if this token is `jump_reg` / `call_reg` / `jump_link` (transfers control OUT of the current atom; the `mac_yield()` handshake ends in `jump_reg(R_AtomJmp), nop`). 
+--                           No offset field — `atom_offset` is invalid here. Terminates the current path in the CFG.
+--   is_load               — true if this token starts with any of: load_word, load_half, load_half_u, load_byte,
+--                           load_byte_u, gte_lw, gte_lwc2. These all have MIPS load-delay semantics (the destination register is volatile for 1 word after the load).
 --   is_store_word         — true if this token starts with `store_word(`
 --
 -- Checks that need the leading ident use `tok_class.ident` instead of re-matching the token string.
 -- Checks that need "how many nops before token i" use `tok_class.nop_prefix` instead of walking backwards.
 
 --- @class TokClass
---- @field ident              string          -- leading identifier
+--- @field ident              string           -- lLading identifier
 --- @field nop_words          integer          -- 0 / 1 / 2
---- @field nop_prefix         integer         -- consecutive nop words before this token
+--- @field nop_prefix         integer          -- Consecutive nop words before this token
 --- @field is_yield           boolean
 --- @field is_atom_label      boolean
---- @field label_name         string|nil      -- for atom_label(name)
+--- @field label_name         string|nil       -- For atom_label(name)
---- @field is_branch          boolean
+--- @field is_branch          boolean          -- Conditional branch OR unconditional-jump-with-offset
---- @field branch_label       string|false|nil -- for branch_*(..., atom_offset(F, label))
+--- @field is_unconditional_jump boolean       -- `jump` / `call_addr` only
---- @field is_load_word       boolean
+--- @field is_terminal_jump   boolean          -- `jump_reg` / `call_reg` / `jump_link` only
+--- @field branch_label       string|false|nil -- For branch_*(..., atom_offset(F, label)) OR jump/call_addr
+--- @field is_load            boolean          -- load_word | load_half | load_half_u | load_byte | load_byte_u | gte_lw | gte_lwc2
 --- @field is_store_word      boolean
 --- @field mac_format_shape   string|nil       -- "f3" / "g4" etc. for mac_format_X_color; nil otherwise
---- @field is_gte_store       boolean         -- ident matches `mac_gte_store_<shape>`
+--- @field is_gte_store       boolean          -- Ident matches `mac_gte_store_<shape>`
---- @field is_ot_tag          boolean         -- ident matches `mac_insert_ot_tag_<shape>`
+--- @field is_ot_tag          boolean          -- Ident matches `mac_insert_ot_tag_<shape>`
 --- @field writes_r_prim_cursor boolean        -- store_word targeting R_PrimCursor
---- @field reads_r_tape_ptr   boolean         -- any token referencing R_TapePtr
+--- @field reads_r_tape_ptr   boolean          -- Any token referencing R_TapePtr
---- @field o_arg1             string|nil      -- first arg of O_(<a>, <b>) captures; nil for non-O_ tokens
+--- @field o_arg1             string|nil       -- First arg of O_(<a>, <b>) captures; nil for non-O_ tokens
---- @field o_arg2             string|nil      -- second arg of O_(<a>, <b>) captures
+--- @field o_arg2             string|nil       -- Second arg of O_(<a>, <b>) captures
---- @field s_arg1             string|nil      -- arg of S_(<a>) captures; nil for non-S_ tokens
+--- @field s_arg1             string|nil       -- Arg of S_(<a>) captures; nil for non-S_ tokens
+
+-- The set of MIPS instruction idents that have a load-delay slot.
+-- Per MIPS I R3000A: `lw`, `lh`, `lhu`, `lb`, `lbu`, `lwc2` (gte_lw).
+-- Note: `lui` (load_upper_i) does NOT have a load delay on MIPS I — it's an ALU op, not a load.
+-- The `load_imm_*` macros are lui + ori sequences with no per-component load delay either.
+local LOAD_INSTRUCTION_IDENTS = {
+	load_word   = true,
+	load_half   = true,
+	load_half_u = true,
+	load_byte   = true,
+	load_byte_u = true,
+	gte_lw      = true,
+	gte_lwc2    = true,
+}
 
 -- Patterns for O_(<arg1>, <arg2>) and S_(<arg>) captures.
 -- UNANCHORED, the substring can appea anywhere in the token (e.g., `load_word(R_T0, R_TapePtr, O_(Binds_X, field))` matches at position ~24).
@@ -230,6 +247,18 @@ end
 local O_PATTERN = "O_%(([%w_]+),%s*([%w_]+)%s*%)"
 local S_PATTERN = "S_%(([%w_]+)%s*%)"
 
+-- Ident patterns for control-transfer instruction kinds:
+--   * `branch_*` (conditional): `branch_equal`, `branch_ne`, `branch_lt_zero`, `branch_ge_zero`, `branch_le_zero`, `branch_gt_zero`.
+--   * `jump` / `call_addr` (unconditional absolute): one immediate offset field; can carry `atom_offset(F, T)`.
+--   * `jump_reg` / `call_reg` / `jump_link` (register-form): no offset field; `atom_offset` is invalid; transfers OUT of the current atom.
+local BRANCH_PATTERN          = "^branch_[%w_]+%s*%("
+-- `jump_rel(off)` is an ergonomic alias for `branch_equal(R_0, R_0, off)` (the within-atom-safe unconditional jump — see `code/duffle/mips.h`).
+-- The C preprocessor expands it BEFORE the metaprogram sees the source, but for source-level metadata consistency we still match it here and classify it as a branch_equal.
+-- This keeps `consuming_encoder` canonical for any downstream tooling that consults the metadata field.
+local JUMP_REL_PATTERN      = "^jump_rel%s*%("
+local UNCOND_JUMP_PATTERN   = "^%f[%w](jump|call_addr)%f[%W]"
+local TERMINAL_JUMP_PATTERN = "^%f[%w](jump_reg|call_reg|jump_link)%f[%W]"
+
 local function classify_tokens(tokens)
 	local n       = #tokens
 	local tc      = {}
@@ -241,12 +270,17 @@ local function classify_tokens(tokens)
 		if     ident == "nop"  then nop_words = 1
 		elseif ident == "nop2" then nop_words = 2 end
 
-		local is_yield      = ident == "mac_yield"
+		-- `mac_yield_tail` is the canonical end-of-atom terminator when paired with `mac_yield_load`
+		-- in a preceding branch's BD-slot. It runs `addiu_self R_TapePtr; jr R_AtomJmp; nop` — the
+		-- "tail" half of the lego split. Treat it as a yield for the `mac_yield_uniformity` check.
+		local is_yield      = ident == "mac_yield" or ident == "mac_yield_tail"
 		local is_atom_label = false
 		local label_name    = nil
 		local is_branch     = false
+		local is_unconditional_jump = false
+		local is_terminal_jump   = false
 		local branch_label  = nil
-		local is_load_word  = ident == "load_word"
+		local is_load       = LOAD_INSTRUCTION_IDENTS[ident] == true
 		local is_store_word = ident == "store_word"
 
 		-- Per-check pre-computes (R3 lift).
@@ -262,9 +296,21 @@ local function classify_tokens(tokens)
 		if ident == "atom_label" then
 			is_atom_label = true
 			label_name    = tok:match("^atom_label%s*%(%s*([%w_]+)%s*%)")
-		elseif tok:match("^branch_[%w_]+%s*%(") then
+		elseif tok:match(BRANCH_PATTERN) or tok:match(JUMP_REL_PATTERN) then
+			-- Conditional branch OR `jump_rel` (the within-atom-safe unconditional jump alias).
+			-- Both encode a 16-bit signed relative word offset.
 			is_branch    = true
 			branch_label = tok:match("atom_offset%s*%([^,]+,%s*([%w_]+)%s*%)") or false
+		elseif tok:match(UNCOND_JUMP_PATTERN) then
+			-- Unconditional absolute jump / call: `jump(off)` / `call_addr(off)`.
+			-- One immediate offset field; can carry an `atom_offset(F, T)` marker (the offsets pass dispatches on `consuming_encoder` — see `passes/offsets.lua::compute_offsets`).
+			is_branch              = true
+			is_unconditional_jump = true
+			branch_label           = tok:match("atom_offset%s*%([^,]+,%s*([%w_]+)%s*%)") or false
+		elseif tok:match(TERMINAL_JUMP_PATTERN) then
+			-- Register-form jump / call: no offset field; `atom_offset` is invalid here (the offsets pass will error if one is supplied).
+			-- Transfers control OUT of the current atom — the CFG treats this as a path terminator.
+			is_terminal_jump = true
 		end
 
 		-- mac_format_X_color / mac_gte_store_<shape> / mac_insert_ot_tag_<shape> (used by check_gpu_portstore_shape).
@@ -290,8 +336,10 @@ local function classify_tokens(tokens)
 			is_atom_label          = is_atom_label,
 			label_name             = label_name,
 			is_branch              = is_branch,
+			is_unconditional_jump  = is_unconditional_jump,
+			is_terminal_jump       = is_terminal_jump,
 			branch_label           = branch_label,
-			is_load_word         = is_load_word,
+			is_load                = is_load,
 			is_store_word          = is_store_word,
 			mac_format_shape       = mac_format_shape,
 			is_gte_store           = is_gte_store,
@@ -346,6 +394,16 @@ end
 -- therefore counts ONLY words strictly between the producer and the consumer.
 -- ─────────────────────────────────────────────────────────────────────────
 
+-- True iff `consumer_event` is a GTE command (gte_cmdw_* or one of the human-readable aliases mapped in `duffle.GTE_COMMAND_ALIASES`).
+-- Used by the LWC2 retirement-regime dispatch in the forward walker: a GTE-command consumer can read the LWC2 result in the very next slot
+-- (the GTE pipeline latches the LWC2 data); any other consumer must observe the standard MIPS load delay (gap >= 1).
+local function is_gte_command(consumer_event)
+	local tok = consumer_event.encoder or consumer_event.ident or ""
+	if tok:sub(1, 9) == "gte_cmdw_" then return true end
+	local  aliases = duffle.GTE_COMMAND_ALIASES or {}
+	return aliases[tok] ~= nil
+end
+
 -- True iff `consumer_word` falls inside the COP2 command's input set OR inside the producer's `fanout_to` set (for IRGB writes).
 -- Used by the consumer-match step of the forward walker.
 local function is_cop2_consumer_of(consumer_event, destination, producer_rel)
@@ -381,8 +439,7 @@ local function is_cop2_consumer_of(consumer_event, destination, producer_rel)
 end
 
 -- True iff `consumer_event` reads the GPR operand at any position the destination register occupies.
--- The read-position lookup consults `duffle.OPERAND_READ_POSITIONS` 
+-- The read-position lookup consults `duffle.OPERAND_READ_POSITIONS` for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
--- for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
 local function is_gpr_consumer_of(consumer_event, destination)
 	local consumer_token = consumer_event.encoder or consumer_event.ident
 	local read_pos       = duffle.OPERAND_READ_POSITIONS or {}
@@ -449,8 +506,8 @@ local function shift_left_u4(value, amount)
 	return wrap_u4(value * (2 ^ amount))
 end
 
--- Resolve only a standalone integer literal. Compound C expressions remain
+-- Resolve only a standalone integer literal.
--- unknown by design; the analyzer must not pretend to be a C evaluator.
+-- Compound C expressions remain unknown by design; the analyzer must not pretend to be a C evaluator.
 local function parse_integer_literal(raw)
 	if type(raw) ~= "string" then return nil end
 	raw = duffle.trim(raw)
@@ -686,8 +743,7 @@ local function consume_cu2_transition(atom, event, ev_word, forward)
 	else
 		append_cu2_finding(atom, event, forward, transition, gap,
 			"error", "exact",
-			string.format(
+			string.format("%s at line %d: COP2 unavailable after SR.CU2 was disabled"
-				"%s at line %d: COP2 unavailable after SR.CU2 was disabled"
 				.. " (gap=%d, required=%d)",
 				atom.name, event_line,
 				gap, transition.required))
@@ -763,9 +819,17 @@ local function analyze_hardware_relations(atom)
 			local relation = prod.relation
 			local semantic = relation.semantic
 			local is_match = false
-			if semantic == "MTC2" or semantic == "CTC2" or semantic == "LWC2" then
+			if semantic == "MTC2" or semantic == "CTC2" or semantic == "LWC2_to_GTE" or semantic == "LWC2_to_other" then
 				-- Consumer is a GTE command whose input set contains the producer's COP2 destination (or a fan-out target).
+				-- LWC2_to_GTE   — GTE-command consumer: gap = 0 OK (the pipeline latches the LWC2 result).
+				-- LWC2_to_other — non-GTE consumer:     standard load delay applies.
+				if relation.id == "lwc2_to_gte_command" then
+					is_match = is_gte_command(ev) and is_cop2_consumer_of(ev, prod.destination, relation)
+				elseif relation.id == "lwc2_to_other_consumer" then
+					is_match = (not is_gte_command(ev)) and is_cop2_consumer_of(ev, prod.destination, relation)
+				else
 					is_match = is_cop2_consumer_of(ev, prod.destination, relation)
+				end
 			elseif semantic == "MFC2" or semantic == "CFC2" or semantic == "MFC0" then
 				-- Consumer is any encoder that reads the producer's GPR destination as an operand.
 				is_match = is_gpr_consumer_of(ev, prod.destination)
@@ -1024,7 +1088,9 @@ end
 -- A subsequent MFC2 (or any encoder that reads a C2 register) that picks the WRONG register for the active role emits a `result_role_mismatch` warning.
 -- For example, reading `C2_SXY0` after RTPS is wrong: the `latest_screen_xy` role is `C2_SXY2`.
 --
--- Note: the OLD `gte_result_position` check also emitted table-gap info findings for `_post_<cmd>` components missing a row in `duffle.GTE_COMPONENT_RESULT_CONTRACTS`. That table-gap check was based on the `_post_<cmd>` NAMING convention rather than hardware truth, and was removed (the user did not want naming to encode ordering semantics; a proper `atom_info` directive for ordering semantics is a future TODO).
+-- Note: the OLD `gte_result_position` check also emitted table-gap info findings for `_post_<cmd>` components missing a row in `duffle.GTE_COMPONENT_RESULT_CONTRACTS`.
+-- That table-gap check was based on the `_post_<cmd>` NAMING convention rather than hardware truth, and was removed
+-- (the user did not want naming to encode ordering semantics; A proper `atom_info` directive for ordering semantics is a future TODO).
 --
 -- The first `transfer_hazards` reader comment above records the projection contract.
 -- ─────────────────────────────────────────────────────────────────────────
@@ -1107,6 +1173,8 @@ end
 local function check_hazard_nop_use(atom, _pipe_ctx, findings)
 	local forward = atom.paths and atom.paths.forward_state
 	local events  = atom.paths.word_events or {}
+	-- GPR effects table used to resolve load destinations when classifying load-delay-slot nops.
+	local gpr_effects = duffle.INSTRUCTION_GPR_EFFECTS or {}
 	if not events or #events == 0 then return end
 	-- Runtime-helper atoms / components (e.g. tape_exit, ac_yield) carry `debug_skip = true` from the bare
 	-- `atom_dbg_skip` marker; their structural nops are part of the fixed handshake and not author choices.
@@ -1124,9 +1192,10 @@ local function check_hazard_nop_use(atom, _pipe_ctx, findings)
 
 		-- Classify the nop BEFORE its event is applied to the pending state.
 		if ev_ident == "nop" and prev_ev ~= nil then
-			-- Skip BD-slot nops: they are exclusively owned by control_transfer_delay_slot_use.
+			-- Skip BD-slot nops that are exclusively owned by control_transfer_delay_slot_use
+			-- (the nop after a branch/jump — covered by that check separately).
 			-- Every BD-slot nop is structural; this check never reports on it.
-			-- (The earlier `if not suppressed then is_bd_slot = true end` form inverted the suppression — the `mac_yield()` handshake's `jump_reg(R_AtomJmp)` was incorrectly flagged.)
+			-- (The earlier `if not suppressed then is_bd_slot = true end` form inverted the suppression - the `mac_yield()` handshake's `jump_reg(R_AtomJmp)` was incorrectly flagged.)
 			local prev_ident   = prev_ev.encoder or ""
 			local bd_policies  = duffle.CONTROL_TRANSFER_DELAY_SLOT_POLICIES or {}
 			local is_bd_slot   = bd_policies[prev_ident] ~= nil
@@ -1187,6 +1256,42 @@ local function check_hazard_nop_use(atom, _pipe_ctx, findings)
 				else
 					-- Track the slot_kind so the BD-separation case can assert the mac_yield handshake is still suppressed.
 					local slot_kind = "plain"
+					-- MIPS load-delay slot: a `load_*` wrote a register in the previous slot, and the result is unavailable for 1 cycle.
+					-- This `nop` is structurally required; classifying it as `modeled-required` is the correct signal
+					-- (removing it would make the following instruction read the OLD value of the loaded register, a load-use hazard).
+					-- The `load_delay_violations` check (Concern 3) catches the actual read-side error; here we suppress the `modeled-redundant` misclassification.
+					-- The set of load instructions mirrors the LOAD_INSTRUCTION_IDENTS in `check_load_delay_slots`.
+					local load_idents = {
+						 load_word   = true,
+						 load_half   = true,
+						 load_half_u = true,
+					   load_byte   = true,
+						 load_byte_u = true,
+						 gte_lw      = true,
+						 gte_lwc2    = true 
+					}
+					local is_load_delay = load_idents[prev_ident] == true
+					if    is_load_delay then
+						-- Determine the destination register from the load's `writes` field.
+						local prev_writes = gpr_effects[prev_ident] and gpr_effects[prev_ident].writes or {}
+						local prev_args   = prev_ev.args or {}
+						local load_dest   = prev_writes[1] and prev_args[prev_writes[1]] or "<load-destination>"
+						findings[#findings + 1] = {
+							check                = "hazard_nop_use",
+							kind                 = "info",
+							atom                 = atom.name,
+							line                 = ev_line,
+							source               = ev.def_path or ev.source or "",
+							nop_classification   = "modeled-required",
+							nop_word_index       = ev_word,
+							retired_relation     = "load_delay_slot",
+							producer_destination = load_dest,
+							consumer_token       = "<would-be-consumer>",
+							msg = string.format("%s at line %d: nop at word %d is modeled-required (load-delay slot for %s)"
+								, atom.name, ev_line, ev_word, load_dest
+							),
+						}
+					else
 						findings[#findings + 1] = {
 							check              = "hazard_nop_use",
 							kind               = "info",
@@ -1204,6 +1309,7 @@ local function check_hazard_nop_use(atom, _pipe_ctx, findings)
 					end
 				end
 			end
+		end
 
 		-- Update the pending snapshot for the next iteration.
 		-- The replay is observation-only; we mirror the walker's staging
@@ -1315,6 +1421,107 @@ local function check_control_transfer_delay_slot_use(atom, pipe_ctx, findings)
 	end
 end
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- Check #1d: load-delay slot violations (per-atom)
+-- �═══════════════════════════════════════════════════════════════════════════
+
+--- Walk every emitted word event of one atom. For each `is_load` event (lw / lh / lhu / lb / lbu / lwc2),
+--- mark the destination register as "volatile through" the NEXT emitted slot — MIPS I R3000A load-delay
+--- semantics. If any subsequent event in that 1-slot window reads the volatile register, emit a `load_delay_violation`
+--- finding (severity: error — the load result is unavailable in the delay slot).
+---
+--- The register becomes non-volatile again at word N+2 (the load has retired), OR sooner if a non-load instruction overwrites the register
+--- (the overwriter's write is the fresh producer; the load's value is shadowed and never observed by any reader).
+---
+--- Runtime-helper atoms / components (`debug_skip == true`) are exempt: their internal load-then-use sequences
+--- are part of the fixed handshake (e.g. `ac_load_tri_indices` loads into R_T0..R_T2, but those are caller-supplied).
+---
+--- The walker reads `duffle.OPERAND_READ_POSITIONS[event.encoder]` to determine which args are read-source
+--- (the destination of a load is in `writes`, not `reads` — see `duffle.INSTRUCTION_GPR_EFFECTS`).
+--- The check is purely structural; it does not consult the GPR-value lattice (no constant propagation needed for load-delay detection — the volatility window is unconditional).
+local function check_load_delay_slots(atom, pipe_ctx, findings)
+	if atom.kind ~= "atom" then return end
+	local events = atom.paths.word_events or {}
+	if   #events == 0 then return end
+	if is_runtime_helper(atom) then return end
+
+	local gpr_effects = duffle.INSTRUCTION_GPR_EFFECTS or {}
+	local read_positions = duffle.OPERAND_READ_POSITIONS or {}
+	-- volatile_until[reg] = 1-based word_events index; the slot AFTER which the register is safe.
+	-- `nil` means "not currently volatile".
+	local volatile_until = {}
+
+	-- Compute the "net reads" of an event: read-positions MINUS write-positions.
+	-- A position that is BOTH read and written (e.g. `add_ui rt, rs, imm` where the duffle table lists position 1 as both.
+	-- See `duffle.OPERAND_READ_POSITIONS["add_ui"] = {1, 2}` and `INSTRUCTION_GPR_EFFECTS["add_ui"].writes = {1}` — 
+	-- and for genuine RMW ops like `add rt, rs, rt` where position 1 IS both read+written) is not a "read" for load-delay purposes:
+	-- The write shadows whatever value the register previously held. Only positions that are reads WITHOUT a co-occurring write to the same register count as net reads.
+	local function net_reads(event_ident, args)
+		local effect = gpr_effects[event_ident]
+		local positions = read_positions[event_ident]
+		if not positions then return {} end
+		local writes_set = {}
+		if effect and effect.writes then
+			for _, pos in ipairs(effect.writes) do writes_set[pos] = true end
+		end
+		local net = {}
+		for _, pos in ipairs(positions) do
+			if not writes_set[pos] then net[#net + 1] = pos end
+		end
+		return net
+	end
+
+	for event_idx, event in ipairs(events) do
+		local event_ident = event.encoder or event.ident
+		local args        = event.args or {}
+		local is_load     = LOAD_INSTRUCTION_IDENTS[event_ident] == true
+
+		-- (1) Is this event reading a register that's still volatile from a previous load?
+		-- Skip the load instruction itself (the load's own argument list may "read" its destination via `OPERAND_READ_POSITIONS`:
+		-- e.g. `addiu rt, rs, imm` lists position 1 (rt) as a "read", but rt is the destination; the within-load argument list is not a separate consumer).
+		-- Use `net_reads` to ignore RMW positions (write shadows read within the same instruction).
+		if not is_load then
+			for _, pos in ipairs(net_reads(event_ident, args)) do
+				local reg = args[pos]
+				if type(reg) == "string" and reg:sub(1, 2) == "R_" then
+					local until_idx = volatile_until[reg]
+					if until_idx and event_idx <= until_idx then
+						local ev_line = line_for_word_event(event)
+						findings[#findings + 1] = {
+							atom  = atom.name,
+							line  = ev_line,
+							check = "load_delay_violation",
+							kind  = "error",
+							msg   = string.format("%s at line %d reads %s at word %d, but a prior load's "
+								.. "delay slot is not over until word %d; insert a `nop` between the "
+								.. "load and this instruction.",
+								atom.name, ev_line, reg, event_idx, until_idx),
+						}
+					end
+				end
+			end
+		end
+
+		-- (2) Update the volatile set based on what this event writes.
+		local effect = gpr_effects[event_ident]
+		if effect and effect.writes then
+			for _, pos in ipairs(effect.writes) do
+				local reg = args[pos]
+				if type(reg) == "string" and reg:sub(1, 2) == "R_" then
+					if is_load then
+						-- Load: destination volatile for exactly 1 slot (the delay slot).
+						volatile_until[reg] = event_idx + 1
+					else
+						-- Non-load write to this register: overwrites shadow the load; the volatile state ends.
+						-- If another reader comes later, it sees the overwriter's value (or unknown), not the stale load value.
+						volatile_until[reg] = nil
+					end
+				end
+			end
+		end
+	end
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Check #2: mac_yield uniformity
 -- ════════════════════════════════════════════════════════════════════════════
@@ -1420,7 +1627,127 @@ local function check_mac_yield_uniformity(atom, pipe_ctx, findings)
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Check #3: Binding handoff discipline
+-- Check #3: mac_yield_load / mac_yield_tail pairing (the lego split)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- The lego split — `mac_yield_load()` in a branch BD-slot + `mac_yield_tail()` at the branch's target
+--- label — must be used as a pair. Otherwise `R_AtomJmp` is not loaded for the tail's `jr R_AtomJmp`,
+--- and the tape runtime would jump to garbage.
+---
+--- Rules:
+---   1. Every `mac_yield_load()` must be in a branch BD-slot (the immediately preceding token must be a branch).
+---   2. Every `mac_yield_tail()` must be the first instruction after an `atom_label()`, AND
+---      at least one branch targeting that label must have `mac_yield_load()` in its BD-slot.
+---   3. `mac_yield_tail()` as the atom-end terminator (last token) is a WARNING, not an error
+---      (the safe default for atom-endings is `mac_yield()` which re-loads `R_AtomJmp`).
+---
+--- Per-atom. Runtime-helper atoms (`debug_skip`) are exempt.
+--- Takes `(atom, pipe_ctx, findings)`; `pipe_ctx` is unused.
+local function check_yield_load_tail_pairing(atom, _pipe_ctx, findings)
+	if atom.kind ~= "atom" then return end
+	if is_runtime_helper(atom) then return end
+
+	local tokens       = atom.paths.tokens
+	local line_in_body = atom.paths.line_in_body
+	local tc           = atom.paths.tok_class
+	local n            = #tokens
+
+	local function line_for(idx)
+		return atom.line + line_in_body[tokens[idx].rel]
+	end
+
+	-- ── Rule 1: every `mac_yield_load()` must be in a branch BD-slot.
+	for tok_idx = 1, n do
+		local c = tc[tok_idx]
+		if c.ident == "mac_yield_load" then
+			if tok_idx < 2 or not tc[tok_idx - 1].is_branch then
+				local prev_ident = (tok_idx >= 2) and (tc[tok_idx - 1].ident or "?") or "<none>"
+				findings[#findings + 1] = {
+					atom  = atom.name,
+					line  = tok_idx >= 2 and line_for(tok_idx) or atom.line,
+					check = "yield_load_tail_pairing",
+					kind  = "error",
+					msg   = string.format(
+						"%s at line %d has `mac_yield_load()` at word %d but the previous token is `%s`, not a branch — `mac_yield_load()` must fill a branch BD-slot."
+						, atom.name, tok_idx >= 2 and line_for(tok_idx) or atom.line, tok_idx, prev_ident),
+				}
+			end
+		end
+	end
+
+	-- ── Rule 2: every `mac_yield_tail()` must be at a labeled target whose branch BD-slot is `mac_yield_load()`.
+	for tok_idx = 1, n do
+		local c = tc[tok_idx]
+		if c.ident ~= "mac_yield_tail" then goto continue end
+
+		-- The immediately preceding token must be an `atom_label()` (no instructions between them).
+		local prev_idx = tok_idx - 1
+		if prev_idx < 1 or not tc[prev_idx].is_atom_label then
+			if tok_idx == n then
+				-- Atom-ending case: last token is `mac_yield_tail()` without a preceding label. WARNING.
+				findings[#findings + 1] = {
+					atom  = atom.name,
+					line  = line_for(tok_idx),
+					check = "yield_load_tail_pairing",
+					kind  = "warning",
+					msg   = string.format(
+						"%s at line %d has `mac_yield_tail()` as the atom-end terminator. The safe default for atom-endings is `mac_yield()` (which re-loads R_AtomJmp). The split is for BD-slot fill, not atom-endings."
+						, atom.name, line_for(tok_idx)),
+				}
+			else
+				findings[#findings + 1] = {
+					atom  = atom.name,
+					line  = line_for(tok_idx),
+					check = "yield_load_tail_pairing",
+					kind  = "error",
+					msg   = string.format(
+						"%s at line %d has `mac_yield_tail()` at word %d but it's not the first instruction after an `atom_label()` — `mac_yield_tail()` must be the first token of its target label's body."
+						, atom.name, line_for(tok_idx), tok_idx),
+				}
+			end
+			goto continue
+		end
+
+		local label_name = tc[prev_idx].label_name
+		-- Find at least one branch targeting `label_name` whose BD-slot is `mac_yield_load()`.
+		local found_pairing = false
+		for branch_idx = 1, n do
+			local bt = tc[branch_idx]
+			if bt.is_branch and bt.branch_label == label_name then
+				local bd_idx   = branch_idx + 1
+				local bd_tc   = bd_idx <= n and tc[bd_idx] or nil
+				if bd_tc and bd_tc.ident == "mac_yield_load" then
+					found_pairing = true
+				else
+					findings[#findings + 1] = {
+						atom  = atom.name,
+						line  = line_for(branch_idx),
+						check = "yield_load_tail_pairing",
+						kind  = "error",
+						msg   = string.format(
+							"%s at line %d has `mac_yield_tail()` at label `%s` (word %d) but the branch targeting it (at word %d) has BD-slot `%s` instead of `mac_yield_load()`."
+							, atom.name, line_for(branch_idx), label_name, tok_idx, branch_idx, bd_tc and bd_tc.ident or "?"),
+					}
+				end
+			end
+		end
+		if not found_pairing then
+			findings[#findings + 1] = {
+				atom  = atom.name,
+				line  = line_for(tok_idx),
+				check = "yield_load_tail_pairing",
+				kind  = "error",
+				msg   = string.format(
+					"%s at line %d has `mac_yield_tail()` at label `%s` but no branch in the body targets this label with `mac_yield_load()` in its BD-slot — R_AtomJmp would not be loaded."
+					, atom.name, line_for(tok_idx), label_name),
+			}
+		end
+		::continue::
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Check #4: Binding handoff discipline
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- For every atom with `atom_bind(Binds_X)`, verify the atom body reads every field of `Binds_X` from R_TapePtr (in any order) 
@@ -1463,7 +1790,7 @@ local function check_abi_handoff(atom, pipe_ctx, findings)
 	for tok_idx = 1, #tokens do
 		local tc_entry = tc[tok_idx]
 		-- scan: load_word(R_*, R_TapePtr, O_(<Binds_X>, <field>))
-		if tc_entry.is_load_word and tc_entry.reads_r_tape_ptr and tc_entry.o_arg1 == binds_name then
+		if tc_entry.is_load and tc_entry.reads_r_tape_ptr and tc_entry.o_arg1 == binds_name then
 			local field = tc_entry.o_arg2
 			if field then
 				found_field_set[field] = true
@@ -1515,7 +1842,8 @@ end
 --- Soft behavior (warnings):
 ---   - Atoms emitting a primitive via raw `store_word(R_PrimCursor, ...)` (no `mac_format_X_color` call) emit a "manual packet assembly" advisory.
 ---     Cannot auto-validate.
----   - Atoms containing a `mac_<name>(...)` call whose name is not in duffle.GP0_MACRO_CONTRIB emit a "new macro; update duffle.GP0_MACRO_CONTRIB" advisory.
+---   - Atoms containing a `mac_<name>(...)` call whose `name` is not registered in `pipe_ctx.components_by_name` emit a "new macro;
+---     Not in corpus.components" advisory — the auto-derivation returned nil for that name.
 ---
 --- Applies only to `kind = "atom"` (baked atoms). Components don't emit full primitives.
 local function check_gpu_portstore_shape(atom, pipe_ctx, findings)
@@ -1540,15 +1868,23 @@ local function check_gpu_portstore_shape(atom, pipe_ctx, findings)
 				cmd_line = atom.line + line_in_body[tokens[tok_idx].rel]
 			end
 			saw_format = true
-			local n = duffle.GP0_MACRO_CONTRIB["mac_format_" .. shape .. "_color"]
+			-- gp0_contrib is auto-derived from the original `MipsAtomComp_(ac_format_<shape>_color) { body }` body
+			-- in `passes/components.lua::compute_components_metadata` and stored on `corpus.components`.
+			local comp = pipe_ctx.components_by_name["format_" .. shape .. "_color"]
+			local n    = comp and comp.gp0_contrib
 			if    n then contrib = contrib + n end
 		end
 		if tc_entry.is_gte_store then
-			local n = duffle.GP0_MACRO_CONTRIB[tc_entry.ident]
+			-- `tc_entry.ident` is the macro-variant form (`mac_gte_store_f3`); strip the `mac_` prefix for the bare-name corpus lookup.
+			local bare    = tc_entry.ident:sub(#"mac_" + 1)
+			local comp    = pipe_ctx.components_by_name[bare]
+			local n       = comp and comp.gp0_contrib
 			if    n then contrib = contrib + n end
 		end
 		if tc_entry.is_ot_tag then
-			local n = duffle.GP0_MACRO_CONTRIB[tc_entry.ident]
+			local bare    = tc_entry.ident:sub(#"mac_" + 1)
+			local comp    = pipe_ctx.components_by_name[bare]
+			local n       = comp and comp.gp0_contrib
 			if    n then contrib = contrib + n end
 		end
 		if tc_entry.writes_r_prim_cursor then
@@ -1585,23 +1921,30 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- Walk all paths through an atom body and return per-path cycle sums.
---- Builds a tiny CFG: each token has a "next" pointer; branches have two (fall-through + taken).
+--- Builds a tiny CFG: each token has a "next" pointer. Three control-transfer kinds are recognized (set by `classify_tokens`):
---- The BD-slot nop after a branch is absorbed into the branch's cost (MIPS-accurate: BD slot always runs), 
+---   * `branch_*` (conditional): 2 successors — fall-through (BD slot absorbed) + taken (if `atom_offset` target known).
---- and is SKIPPED when continuing down the fall-through path (otherwise we'd double-count it).
+---   * `jump` / `call_addr` (unconditional absolute): 1 successor — taken only (BD slot absorbed into the cost).
+---   * `jump_reg` / `call_reg` / `jump_link` (register-form): terminator — transfers control OUT of the current atom (e.g. `mac_yield()` ends in `jump_reg(R_AtomJmp), nop`).
+---
+--- The BD-slot nop after ANY of these (conditional branch, unconditional jump, terminal jump) is absorbed into the control-transfer's cost
+--- (MIPS-accurate: the BD slot always runs) and is SKIPPED in the successor list (otherwise we'd double-count it).
 ---
 --- Returns:
 ---   cycles_min     - shortest path through the body (sum of token costs)
 ---   cycles_max     - longest path through the body
----   branches       - number of branches in the body
+---   branches       - number of branches in the body (conditional + unconditional-with-offset)
----   paths          - number of distinct paths reached (terminated at mac_yield or end-of-body)
+---   paths          - number of distinct paths reached (terminated at mac_yield / terminal_jump / end-of-body)
 ---   has_loops      - true iff a path re-entered a token it had visited (warning; loop bodies aren't supported)
----   unknown_macros - list of unique macro names not in duffle.INSTRUCTION_LATENCY
+---   unknown_macros - list of unique ident names with no cost lookup: non-`mac_*` idents not in `duffle.INSTRUCTION_LATENCY`,
-local function analyze_atom_paths(atom)
+---                    plus `mac_*` idents whose bare name is missing from `pipe_ctx.components_by_name` (i.e. no `MipsAtomComp_` for it).
+local function analyze_atom_paths(atom, pipe_ctx)
 	local tokens = atom.paths.tokens or duffle.tokenize_body(atom.body)
 	local tc     = atom.paths.tok_class or classify_tokens(tokens)
 	local n      = #tokens
 
 	-- Build label + branch maps from the pre-computed classification (no re-scan).
+	-- `branches` keys both `branch_*` (conditional) and `jump`/`call_addr` (unconditional absolute);
+	-- The latter resolve via `tc[tok_idx].branch_label` the same way (the offsets pass produces a valid relative offset for both).
 	local labels   = {}
 	local branches = {}
 	for tok_idx = 1, n do
@@ -1615,23 +1958,46 @@ local function analyze_atom_paths(atom)
 	end
 
 	-- Pre-compute per-token cycle costs from the pre-computed ident (no re-match).
+	-- For non-`mac_*` tokens: lookup `duffle.INSTRUCTION_LATENCY[c.ident]` directly.
+	-- For `mac_*` tokens: lookup `pipe_ctx.components_by_name[bare_name].cycle_cost`, which `passes/components.lua::compute_components_metadata` derived from the originals
+	-- `MipsAtomComp_(ac_X) { body }` definition (sum of `INSTRUCTION_LATENCY` per emitted instruction,
+	-- recursing through nested `mac_*` calls). `mac_yield` is special-cased to 0 by `compute_components_metadata` (the runtime cost lands in the next atom's prologue).
 	local costs       = {}
 	local unknown_set = {}
 	for tok_idx = 1, n do
 		local c      = tc[tok_idx]
-		local cost   = duffle.INSTRUCTION_LATENCY[c.ident]
+		local ident  = c.ident
+		local cost
+		if ident:sub(1, #"mac_") == "mac_" then
+			-- `mac_*` token: lookup corpus.components[bare_name].cycle_cost.
+			local bare = ident:sub(#"mac_" + 1)
+			local comp = pipe_ctx.components_by_name and pipe_ctx.components_by_name[bare]
+			if comp and comp.cycle_cost ~= nil then
+				cost = comp.cycle_cost
+			else
+				cost = duffle.UNKNOWN_INSTRUCTION_CYCLES
+				unknown_set[ident] = true
+			end
+		else
+			cost = duffle.INSTRUCTION_LATENCY[ident]
 			if cost == nil then
 				cost = duffle.UNKNOWN_INSTRUCTION_CYCLES
-			unknown_set[c.ident] = true
+				unknown_set[ident] = true
+			end
 		end
 		costs[tok_idx] = cost
 	end
 
-	-- A token is a terminator if it's `mac_yield`.
+	-- Three control-transfer predicates (set by `classify_tokens`):
-	local function is_terminator(tok_idx) return tc[tok_idx].is_yield end
+	--   is_terminator — path ends here (`mac_yield` or register-form jump); empty successors.
-
+	--   is_branch — has an immediate offset (`branch_*`, `jump`, `call_addr`); 1-2 successors depending on unconditional_jump.
-	-- A token is a "branch" if the classification says so.
+	--   is_unconditional_jump — when is_branch is also true: skip fall-through (target only).
+	local function is_terminator(tok_idx)
+		local c = tc[tok_idx]
+		return c.is_yield or c.is_terminal_jump
+	end
 	local function is_branch(tok_idx) return tc[tok_idx].is_branch end
+	local function is_unconditional_jump(tok_idx) return tc[tok_idx].is_unconditional_jump end
 	local function successors(tok_idx)
 		local tok = tokens[tok_idx].tok
 		if is_terminator(tok_idx) then
@@ -1640,11 +2006,22 @@ local function analyze_atom_paths(atom)
 		if is_branch(tok_idx) then
 			local label = branches[tok_idx]  -- may be false for literal-offset branches
 			local succ  = {}
-			-- Fall-through: skip the BD slot (tok_idx+1). Use tok_idx+2.
+			if is_unconditional_jump(tok_idx) then
+				-- Unconditional absolute jump / call: BD slot absorbed; single successor — the taken path.
+				-- The instruction word after the BD slot is unreachable in this atom's execution.
+				if label then
+					local label_pos = labels[label]
+					if label_pos and label_pos + 1 <= n then
+						succ[#succ + 1] = label_pos + 1
+					end
+				end
+				-- For literal-offset jumps (label == false), the target is a non-tracked address; conservatively omit.
+				return succ, nil
+			end
+			-- Conditional branch: BD slot absorbed; two successors — fall-through (tok_idx+2) + taken (if known).
 			if tok_idx + 2 <= n then
 				succ[#succ + 1] = tok_idx + 2
 			end
-			-- Taken: only if the branch has a known atom_offset target.
 			if label then
 				local label_pos = labels[label]
 				if label_pos and label_pos + 1 <= n then
@@ -1680,11 +2057,11 @@ local function analyze_atom_paths(atom)
 			return
 		end
 
-		-- Add this token's cost. For a branch, ADD the BD-slot cost too
+		-- Add this token's cost. For ANY control-transfer (conditional branch, unconditional jump, terminal jump),
-		-- (and skip the BD slot in the successor list — already done in `successors` above for fall-through; 
+		-- ADD the BD-slot cost too — MIPS-accurate: the BD slot always runs. Skip the BD slot in the successor list (already done in `successors` above;
-		-- for taken path the BD slot was at tok_idx+1 which is now skipped entirely).
+		-- for the taken path the BD slot was at tok_idx+1 which is now skipped entirely).
 		local cost = costs[tok_idx]
-		if is_branch(tok_idx) and tok_idx + 1 <= n then
+		if (is_branch(tok_idx) or is_terminator(tok_idx)) and tok_idx + 1 <= n then
 			cost = cost + costs[tok_idx + 1]
 		end
 		local new_acc = acc + cost
@@ -1715,7 +2092,7 @@ local function analyze_atom_paths(atom)
 	for macro_name in pairs(unknown_set) do unknown_list[#unknown_list + 1] = macro_name end
 	table.sort(unknown_list)
 
-	-- branch_count: number of `branch_*(...)` tokens.
+	-- branch_count: number of control-transfer tokens with an immediate offset (`branch_*` + `jump` + `call_addr`).
 	local branch_count = 0
 	for _ in pairs(branches) do branch_count = branch_count + 1 end
 
@@ -1733,9 +2110,9 @@ local function analyze_atom_paths(atom)
 end
 
 --- Per-source check that emits one finding per unknown macro seen
---- (deduplicated across atoms so the warning section doesn't get spammed with N copies of "macro X not in duffle.INSTRUCTION_LATENCY").
+--- (deduplicated across atoms so the warning section doesn't get spammed with N copies of the same diagnostic).
 --- Per-atom: emit one finding per unknown macro seen, deduplicated across atoms
---- (so the warning section doesn't get spammed with N copies of "macro X not in duffle.INSTRUCTION_LATENCY").
+--- (so the warning section doesn't get spammed with N copies of the same diagnostic).
 --- Reuses `analyze_atom_paths`'s per-atom unknown_macros discovery, which walks tokens and computes per-token cycle costs.
 local function check_per_atom_cycle_budget(atom, pipe_ctx, findings)
 	local p = atom.paths or {}
@@ -1745,8 +2122,11 @@ local function check_per_atom_cycle_budget(atom, pipe_ctx, findings)
 			findings[#findings + 1] = {
 				atom  = atom.name, line = atom.line,
 				check = "per_atom_cycle_budget", kind = "warning",
-				msg   = string.format("%s at line %d uses macro `%s` which is not in duffle.INSTRUCTION_LATENCY; "
+				msg   = string.format("%s at line %d uses macro `%s` with no cycle_cost lookup; "
-					.. "cycle count will be +%d per call (best-case). Add an entry to duffle.INSTRUCTION_LATENCY."
+					.. "cycle count will be +%d per call (best-case). For `mac_*` idents, ensure the "
+					.. "corresponding `MipsAtomComp_(ac_X)` is in scope of the build so "
+					.. "`passes/components.lua::compute_components_metadata` can derive its cost; "
+					.. "for non-`mac_*` idents, add an entry to `duffle.INSTRUCTION_LATENCY`."
 					, atom.name, atom.line, name, duffle.UNKNOWN_INSTRUCTION_CYCLES),
 			}
 		end
@@ -1910,7 +2290,7 @@ local function check_binds_no_substruct_deref(_src, pipe_ctx, findings)
 		local line_in_body = a.paths and a.paths.line_in_body or {}
 		for ti = 1, #tokens do
 			local tc_entry = tc[ti]
-			if (tc_entry.is_load_word or tc_entry.is_store_word)
+			if (tc_entry.is_load or tc_entry.is_store_word)
 				and tc_entry.o_arg1 and tc_entry.o_arg2 then
 				local type_name  = tc_entry.o_arg1
 				local field_name = tc_entry.o_arg2
@@ -1969,7 +2349,9 @@ local CHECK_RULES = {
 	{ name = "gte_role_mismatch",               per_atom   = check_gte_role_mismatch               },
 	{ name = "hazard_nop_use",                  per_atom   = check_hazard_nop_use                  },
 	{ name = "control_transfer_delay_slot_use", per_atom   = check_control_transfer_delay_slot_use },
+	{ name = "load_delay_violation",            per_atom   = check_load_delay_slots                },
 	{ name = "mac_yield_uniformity",            per_atom   = check_mac_yield_uniformity            },
+	{ name = "yield_load_tail_pairing",         per_atom   = check_yield_load_tail_pairing         },
 	{ name = "abi_handoff",                     per_atom   = check_abi_handoff                     },
 	{ name = "gpu_portstore_shape",             per_atom   = check_gpu_portstore_shape             },
 	{ name = "per_atom_cycle_budget",           per_atom   = check_per_atom_cycle_budget           },
@@ -2008,6 +2390,10 @@ local function build_corpus_pipe_ctx(ctx)
 		atom_phases             = corpus.atom_phases             or {},
 		binds_by_name           = corpus.binds_by_name           or {},
 		atoms_by_name           = corpus.atoms_by_name           or {},
+		-- Per-component metadata (cycle_cost + gp0_contrib) auto-derived from the original
+		-- `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`.
+		-- Keyed by bare name (e.g. `format_f3_color`, `gte_store_f3`); the `mac_` prefix at call sites is stripped before lookup. 
+		components_by_name      = corpus.components              or {},
 		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
 		atom_infos_list         = corpus.atom_infos              or {},
 		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
@@ -2059,15 +2445,16 @@ local function validate(ctx, src, corpus_pipe_ctx)
 		atom_infos_list         = atom_infos or {},
 		register_alias_registry = corpus_pipe_ctx.register_alias_registry,
 		type_name_registry      = corpus_pipe_ctx.type_name_registry,
+		-- Per-component metadata (cycle_cost + gp0_contrib) auto-derived from the original `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`.
+		components_by_name      = corpus_pipe_ctx.components_by_name,
 	}
-	-- Shared cross-source component-body index is owned by the corpus
+	-- Shared cross-source component-body index is owned by the corpus (`corpus.component_body_index`, populated by `passes/components.lua`).
-	-- (`corpus.component_body_index`, populated by `passes/components.lua`).
 	-- Per-atom checks consume the corpus-owned index directly.
 	pipe_ctx.component_body_index = (corpus and corpus.component_body_index) or {}
 
 	--- Per-atom pipeline. ONE iteration of atoms; the 5 check_* functions + analyze_atom_paths all run here, sharing a single tokenize_body + build_body_line_index per body.
 	--- Every piece of state derived from an atom body lives on `atom.paths` (per-atom mega-struct);
-	--- readers (analyze_atom_paths, the 5 checks, the renderers) all consume `atom.paths`, not the raw `atoms` list.
+	--- readers (analyze_atom_paths, the 5 checks, the renderers) all consume `atom.paths`.
 	--- Each `check_*` function accepts one atom and its shared context.
 	--- Per-source rules run once after this loop completes (no parallel dispatch table).
 	---
@@ -2093,7 +2480,7 @@ local function validate(ctx, src, corpus_pipe_ctx)
 		a.paths.tok_class = classify_tokens(a.paths.tokens)
 
 		-- analyze_atom_paths fills the *cycles / branches / has_loops / unknown_macros* fields of a.paths.
-		analyze_atom_paths(a)
+		analyze_atom_paths(a, pipe_ctx)
 
 		-- Run the single forward walker for transfer-hazard policy.
 		-- Runs once per atom BEFORE the CHECK_RULES per-atom dispatch so the `transfer_hazards` reader (`check_transfer_hazards`) can
@@ -2148,9 +2535,8 @@ local function validate(ctx, src, corpus_pipe_ctx)
 			evidence_confidence  = f.evidence_confidence,
 			evidence_source      = f.evidence_source,
 		}
-		-- Preserve relation fields such as target_state and status_register,
+		-- Preserve relation fields such as target_state and status_register status_value,
-		-- status_value, and future policy metadata) without making the binner
+		-- and future policy metadata) without making the binner another semantic walker.
-		-- another semantic walker.
 		for key, value in pairs(f) do
 			if payload[key] == nil then payload[key] = value end
 		end

scripts/passes/word_count_eval.lua

@@ -93,7 +93,7 @@ end
 --- Load the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts`.
 --- 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.
+--- after the components pass iterates `corpus.source_order` and writes each source-directory's `gen/macs.h` file.
 ---
 --- Contract:
 ---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).

scripts/ps1_meta.lua

@@ -16,7 +16,7 @@
 
 -- Bootstrap: load `duffle_paths.lua` via this script's own path.
 -- Use `arg[0]` when this file is the entry script (`arg[0]` ends in "ps1_meta.lua");
--- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path, not ours).
+-- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path).
 -- That single statement: (a) sets `package.path` + `package.cpath`, (b) at the bottom returns `require("duffle")`.
 -- So the dofile's return value is the duffle module.
 local _is_entry_script = arg and arg[0] and arg[0]:match("ps1_meta%.lua$") ~= nil
@@ -54,45 +54,44 @@ local PASS_FLAG_DISPATCH_KEY   = "__pass__"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class PassDescriptor
---- @field module string       -- module name passed to require()
+--- @field module string       -- Module name passed to require()
 --- @field kind   string       -- "shared" | "header-output" | "validation" | "diagnostic" | "report"
 ---                            -- Report severity is independent from process exit policy (see PASS_KIND_STOP_ON_ERROR).
---- @field deps   string[]     -- names of upstream passes
+--- @field deps   string[]     -- Names of upstream passes
---- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of
+--- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
----                            -- (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field metadata_path      string                 -- path to word_count.metadata.h
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table                  -- Cross-pass shared state
---- @field shared.corpus      table                  -- canonical authored-source/project projection
+--- @field shared.corpus      table                  -- Authored-source/project projection
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
 --- @field project_root       string                 -- PS1 repository root
 --- @field flags              table                  -- CLI flags + per-pass stash
---- @field verbose            boolean                -- if true, log diagnostic info
+--- @field verbose            boolean                -- If true, log diagnostic info
 
 --- @class Finding
---- @field line integer  -- source line (or 0 for pass-level)
+--- @field line integer  -- Source line (or 0 for pass-level)
---- @field msg  string   -- finding message
+--- @field msg  string   -- Finding message
 
 --- @class PassResult
---- @field outputs  PassOutputEntry[] -- emitted file paths
+--- @field outputs  PassOutputEntry[] -- Emitted file paths
---- @field errors   Finding[]         -- build-stops (per-pass kind policy)
+--- @field errors   Finding[]         -- Build-stops (per-pass kind policy)
---- @field warnings Finding[]         -- informational
+--- @field warnings Finding[]         -- Informational
 
 --- @class ParsedArgs
---- @field requested_set string[]   -- pass names to run (explicit --all expanded)
+--- @field requested_set string[]   -- Pass names to run (explicit --all expanded)
---- @field sources       string[]   -- exact --source values, retained in CLI order
+--- @field sources       string[]   -- Exact --source values, retained in CLI order
 --- @field unity_root    string|nil -- --unity-root value; mutually exclusive with sources
 --- @field metadata      string     -- --metadata value
 --- @field out_root      string     -- --out-root value (default "build/gen")
 --- @field project_root  string     -- PS1 repository root (derived from metadata by default)
---- @field verbose       boolean    -- if true, log diagnostic info
+--- @field verbose       boolean    -- If true, log diagnostic info
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- PASSES Table
@@ -138,7 +137,7 @@ local PASSES = {
 	["static-analysis"] = {
 		module = "passes.static_analysis",
 		-- "diagnostic" — every `error`/`warning` finding is written to the report file;
-		-- the orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
+		-- The orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
 		-- Report severity is independent from process exit policy.
 		kind   = "diagnostic",
 		deps   = {"scan-source", "word-counts", "components", "emission-model"},
@@ -163,12 +162,12 @@ local PASSES = {
 }
 
 -- ────────────────────────────────────────────────────────────────────────────
--- Phase-root selection: derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
+-- Phase-root selection: Derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
 -- topo_sort closes the transitive deps from there; dispatch_passes runs every resolved pass without phase-filtering.
 -- ────────────────────────────────────────────────────────────────────────────
 
---- @param group_name string  -- the build-phase group ("pre-link" | "post-link")
+--- @param group_name string  -- Build-phase group ("pre-link" | "post-link")
---- @return string[]          -- sorted root pass names belonging to that group
+--- @return string[]          -- Sorted root pass names belonging to that group
 local function roots_for_group(group_name)
 	local names = {}
 	for name, pass in pairs(PASSES) do
@@ -206,7 +205,7 @@ end
 -- Report severity is independent from process exit policy.
 -- A "diagnostic" pass still writes every `error`/`warning` finding into its report file,
 -- but `report_validation_errors` returns early for non-stopping kinds, so nothing is printed to stderr and the orchestrator does not exit non-zero.
--- Adding a new pass kind requires listing it here explicitly; an unknown kind must not silently fall back to "true".
+-- Adding a new pass kind requires listing it here explicitly; An unknown kind must not silently fall back to "true".
 local PASS_KIND_STOP_ON_ERROR = {
 	["shared"]        = false,
 	["header-output"] = true,
@@ -216,8 +215,7 @@ local PASS_KIND_STOP_ON_ERROR = {
 }
 
 -- Closed set of CLI flags -> pass names.
--- Per-pass flags (e.g. --word-counts) live here; phase flags (--pre-link, --post-link, --all)
+-- Per-pass flags (e.g. --word-counts); phase flags (--pre-link, --post-link, --all) are within FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
--- live in FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
 -- dwarf-injection is *also* a per-pass opt-in flag, but its selection + opt-in state are both owned by the explicit FLAG_HANDLERS entry below
 -- (it sets args.flags.dwarf_injection and appends "dwarf-injection" to requested_set), so it is intentionally absent from this table.
 local PASS_FLAG_TO_NAME = {
@@ -246,7 +244,6 @@ end
 
 -- Per-flag handlers. Each handler takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it). 
 -- Returning nil + os.exit() handles termination flags (--help). 
-
 local FLAG_HANDLERS = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -272,9 +269,9 @@ PASS_FLAGS:
   Or pick any subset:
     --scan-source         Scan sources into the fat SourceScan payload
     --word-counts         Load metadata.h + scan for existing .macs.h
-    --components          Generate <module>/gen/<basename>.macs.h
+    --components          Generate <srcdir>/gen/macs.h (per-directory aggregation)
     --validate            Run atom annotation DSL validation
-    --offsets             Generate <module>/gen/<basename>.offsets.h
+    --offsets             Generate <srcdir>/gen/offsets.h (per-directory aggregation)
     --atoms-source-map    Generate <basename>.atoms.sourcemap.txt per source
     --dwarf-injection     [opt-in] Select the post-link dwarf-injection pass + set the opt-in flag. Requires --elf.
     --static-analysis     Static analysis: GTE pipeline-fill, mac_yield, ABI handoff, cycle budget
@@ -317,8 +314,7 @@ local function require_flag_value(argv, arg_idx, flag)
 	local next_known = type(value) == "string"
 		and (FLAG_HANDLERS[value] ~= nil or PASS_FLAG_TO_NAME[value] ~= nil)
 	if value == nil or next_known then
-		io.stderr:write("ps1_meta: " .. flag .. " requires "
+		io.stderr:write("ps1_meta: " .. flag .. " requires " .. FLAG_VALUE_NAMES[flag] .. "\n")
-			.. FLAG_VALUE_NAMES[flag] .. "\n")
 		os.exit(EXIT_INTERNAL_ERROR)
 	end
 	return value, arg_idx + 1
@@ -328,12 +324,10 @@ end
 -- Termination flags like --help call os.exit() instead.
 -- Populated AFTER print_help so the --help handler can reference it as an upvalue (Lua resolves locals at closure-call time, 
 -- but if the closure is defined before the local, it falls back to _G). 
-FLAG_HANDLERS["--help"] = function(args)
-	print_help()
-	os.exit(0)
-end
 
+FLAG_HANDLERS["--help"]    = function(args) print_help(); os.exit(0) end
 FLAG_HANDLERS["--verbose"] = function(args) args.verbose = true      end
+
 FLAG_HANDLERS["--source"] = function(args, argv, arg_idx)
 	local value, value_idx = require_flag_value(argv, arg_idx, "--source")
 	args.sources[#args.sources + 1] = value
@@ -496,8 +490,7 @@ local function build_ctx(args)
 		or normalized_project_root:sub(1, 2) == "//"
 		or normalized_project_root:sub(1, 1) == "/"
 	if not project_root_is_absolute then
-		-- canonical_path_key validates ordinary relative paths and rejects
+		-- canonical_path_key validates ordinary relative paths and rejects drive-relative paths before the absolute-path rewrite is performed.
-		-- drive-relative paths before the absolute-path rewrite is performed.
 		duffle.canonical_path_key(normalized_project_root)
 		project_root = duffle.normalize_path(duffle.to_absolute_path(normalized_project_root))
 	else
@@ -511,8 +504,7 @@ local function build_ctx(args)
 			project_root = project_root,
 		})
 		if not ok_resolve then
-			io.stderr:write("ps1_meta: cannot resolve --unity-root "
+			io.stderr:write("ps1_meta: cannot resolve --unity-root " .. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
-				.. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
 			os.exit(EXIT_INTERNAL_ERROR)
 		end
 		resolution = resolved
@@ -528,8 +520,7 @@ local function build_ctx(args)
 			local path = duffle.normalize_path(input_path)
 			local key_ok, key_or_error = pcall(duffle.canonical_path_key, path)
 			if not key_ok then
-				error("ps1_meta: invalid --source " .. input_path .. ": "
+				error("ps1_meta: invalid --source " .. input_path .. ": " .. tostring(key_or_error), 0)
-					.. tostring(key_or_error), 0)
 			end
 			local file = io.open(path, "r")
 			if not file then
@@ -627,9 +618,7 @@ local function topo_sort(passes, requested_set)
 		changed = false
 		for name, _ in pairs(needed) do
 			local pass = passes[name]
-			if not pass then
+			if not pass then error("unknown pass '" .. name .. "' requested") end
-				error("unknown pass '" .. name .. "' requested")
-			end
 			for _, dep in ipairs(pass.deps) do
 				if not needed[dep] then
 					needed[dep] = true