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

@@ -17,3 +17,7 @@ toolchain/PSn00bSDK
 .vscode/settings.json
 toolchain/lfs
 toolchain/lpeg
+
+scratch
+toolchain/libpsn00b
+scripts/pcsx_debug_helper.zip

.vscode/launch.json

@@ -74,41 +74,7 @@
 			]
 		},
 		{
-			"name": "Debug: Hello GTE Psy-Q!",
-			"type": "gdb",
-			"request": "attach",
-			"target": "localhost:3333",
-			"remote": true,
-			"cwd": "${workspaceRoot}/build",
-			"valuesFormatting": "parseText",
-			"registerLimit": "1-32",
-			"frameFilters": false,
-			"showDevDebugOutput": false,
-			"printCalls": false,
-			"stopAtConnect": true,
-			"gdbpath": "gdb-multiarch",
-			"windows": {
-				"gdbpath": "gdb-multiarch.exe"
-			},
-			"osx": {
-				"gdbpath": "gdb"
-			},
-			"executable": "${workspaceRoot}/build/hello_gte.elf",
-			"setupCommands": [
-				{ "text": "set mi-async off" },
-				{ "text": "set remotetimeout 0" },
-				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
-				{ "text": "set logging redirect on" }
-			],
-			"autorun": [
-				"monitor reset shellhalt",
-				"load hello_gte.elf",
-				"tbreak main",
-				"continue"
-			]
-		},
-		{
-			"name": "Debug: Hello GTE Psy-Q! (atoms debug — DWARF-injected)",
+			"name": "Debug: Hello GTE!",
 			"type": "gdb",
 			"request": "attach",
 			"target": "localhost:3333",
@@ -138,7 +104,76 @@
 				"monitor reset shellhalt",
 				"load build/hello_gte.dwarf-injected.elf",
 				"source scripts/gdb/gdb_tape_atoms.gdb",
-				"source build/gen/hello_gte.gdbinit",
+				"tbreak main",
+				"continue"
+			]
+		},
+		{
+			"name": "Debug: Hello Joypad!",
+			"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_joypad.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_joypad.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_joypad.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"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.
@@ -139,10 +142,11 @@ typedef void Proc_(VoidFn) (void);
 #define mega(n)                (C_(U4, n) << 20)
 #define giga(n)                (C_(U4, n) << 30)
 #define tera(n)                (C_(U4, n) << 40)
+
 #define null                    C_(U4,    0)
 #define nullptr                 C_(void*, 0)
-#define O_(type, field) (C_(U4, & C_(type*,0)->field))
- 
+#define O_(type, field)         C_(U4, & C_(type*,0)->field)
+#define OA_(type, member, idx)  C_(U4, & C_(type*,0)->member[idx])
 #define OT_(field)              O_(typeof_ptr(& field), filed))
 #define S_(data)                C_(U4, sizeof(data))
 

code/duffle/gcc_asm.h

@@ -50,17 +50,13 @@
 #define asm_words(...) m_expand(glue(GCC_ASM_INL_, GCC_ASM_COUNT_ARGS(__VA_ARGS__))(__VA_ARGS__))
 // Very nasty macro expansion. See the Cruft pragma region after all the DSL defines
 
-/* reg_str(n) — Stringify an integer register id into the GCC asm
- * string form (e.g. 12 → "$12"). Use this anywhere GCC's parser
- * expects a literal string identifying a register: clobber lists,
- * asm templates, etc. The two-level macro is the standard preprocessor
- * idiom for forcing one level of expansion before stringify — without
- * it, `#n` would stringify the macro name `R_T4` to `"R_T4"` instead
- * of expanding `R_T4` to its value first.
+/* reg_str(n) — Stringify an integer register id into the GCC asm string form (e.g. 12 → "$12").
+ * Use this anywhere GCC's parser expects a literal string identifying a register: clobber lists,
+ * asm templates, etc. The two-level macro is the standard preprocessor idiom for forcing one level of expansion before stringify —
+ * without it, `#n` would stringify the macro name `R_T4` to `"R_T4"` instead of expanding `R_T4` to its value first.
  *
- * For declaring a register variable bound to a specific GPR, use the
- * `rgcc(n)` bundle from gcc_asm.h instead — it adds the `__asm__()`
- * qualifier around the string.
+ * For declaring a register variable bound to a specific GPR, use the `rgcc(n)` bundle from gcc_asm.h instead —
+ * it adds the `__asm__()` qualifier around the string.
  *
  *   register V3_S2* p0 __asm__(reg_str(R_T4)) = ...;        // verbose
  *   register V3_S2* p0 rgcc(R_T4) = ...;                    // bundled
@@ -85,7 +81,7 @@
  *  - The string "$12" is derived from it via reg_str, so they cannot drift apart.
  *  - Spelling `__asm__(reg_str(R_T4_Code))` at every call site is noise.
  *
- *  tmpl defined in dsl.h (the token-paste glue). 
+ *  tmpl defined in dsl.h (token-paste glue). 
  *  rgcc define here (gcc_asm.h) because the `__asm__` keyword is GCC-specific. 
  *  Anyone porting to a different compiler's asm dialect overrides rgcc,
  *  and the integer→string derivation in rlit can be retargeted in one place.
@@ -94,12 +90,10 @@
  * ------------------------------------------------------------------------ */
 #define rgcc(n) __asm__(rlit(n))
 
-/* rgcc_ref(n) — GCC operand-reference form "%N". Not currently used
- * by the placeholder-pun macros (the .word bodies are fully baked
- * at compile time and have no runtime operand references), but kept
- * here for completeness in case a future asm template needs to refer
- * to a runtime input by position. Mirror of rgcc but produces "%N"
- * instead of "$N". */
+/* rgcc_ref(n) — GCC operand-reference form "%N". Not currently used by the placeholder-pun macros
+ * (the .word bodies are fully baked at compile time and have no runtime operand references),
+ * but kept here for completeness in case a future asm template needs to refer to a runtime input by position.
+ * Mirror of rgcc but produces "%N" instead of "$N". */
 #define rgcc_ref_(n)   "%" #n
 #define rgcc_ref(n)    rgcc_ref_(n)
 
@@ -147,11 +141,9 @@
     9,  8,  7,  6,  5,  4,  3,  2,  1,  0))
 
 /* --- 2. String Concatenation Helpers --- *
- * NOTE: we use `%0`, `%1`, ... not `%c0`, `%c1`, ... because GCC's
- * asm-parser rejects `%cN` in this position with "invalid use of '%c'".
- * The `%cN` form is for printing *character* constants; for arbitrary
- * integer immediates (the only kind `"i"(...)` produces), the plain
- * `%N` form is the right one. Both expand to the bare immediate.
+ * NOTE: we use `%0`, `%1`, ... not `%c0`, `%c1`, ... because GCC's asm-parser rejects `%cN` in this position with "invalid use of '%c'".
+ * The `%cN` form is for printing *character* constants; for arbitrary integer immediates (the only kind `"i"(...)` produces),
+ * the plain `%N` form is the right one. Both expand to the bare immediate.
  */
 #define GCC_ASM_W1                  "%0"
 #define GCC_ASM_W2   GCC_ASM_W1   ", %1"

code/duffle/gen/duffle.macs.h

@@ -1,139 +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_post_rtpt(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
-WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
-
-#define mac_format_g4_color(r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
-,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
-,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
-,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
-WORD_COUNT(mac_format_g4_color, 12)
-
-/* 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).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-#define mac_gte_store_g4_p012_post_rtpt_pre_rtps(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
-WORD_COUNT(mac_gte_store_g4_p012_post_rtpt_pre_rtps, 3)
-
-/* 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 — DO NOT read SXY0 here, that's the bug this name
- * prevents).
- */
-#define mac_gte_store_g4_p3_post_rtps(...) \
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
-WORD_COUNT(mac_gte_store_g4_p3_post_rtps, 1)
-

code/duffle/gen/duffle.offsets.h

@@ -1,9 +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/gp.h

@@ -39,8 +39,8 @@
 /* ============================================================================
  *  Hardware MMIO Addresses
  *  ============================================================================
- *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2
- *  0x1F800000+. GP0 (offset 0x10) is the data port (commands + params).
+ *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2 0x1F800000+.
+ *  GP0 (offset 0x10) is the data port (commands + params).
  *  GP1 (offset 0x14) is the control port (status, ctrl writes).
  * ============================================================================ */
 /* IO base address (KSEG2 0x1F800000+ for the I/O register region).
@@ -73,9 +73,7 @@ enum {
  *  GP0 command byte constants + Layer 1 (GPU bitfield shifts)
  *  ============================================================================
  *  8-bit GP0 opcodes (the upper byte of a primitive's first word). These are the BYTE only.
- *  The layer-1 bitfield-layout constants live in the same enum block so the encoder can reference them by name. 
  *  NO macro body past this point uses a raw shift or raw mask.
- *  Every shift/width/mask is named here, named once.
  *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
  * ============================================================================ */
 enum {
@@ -176,6 +174,9 @@ enum {
 #define gp0_word_copy_vram()         enc_gp0_cmd_word(gp0_cmd_CopyVram)
 #define gp0_word_read_vram()         enc_gp0_cmd_word(gp0_cmd_ReadVram)
 
+/* NOP — bare-cmd word (no effect; used as DR_ENV padding). */
+#define gp0_word_nop()               enc_gp0_cmd_word(gp0_cmd_Nop)
+
 /* ============================================================================
  *  GP1 command byte constants + Layer 1 (display-mode + range + draw-area bitfield shifts)
  *  ============================================================================
@@ -199,8 +200,7 @@ enum {
 		* DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
 
 	/* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
-     * Bit positions match the encoder shifts below; values are the
-     * *payload* bits only (the cmd byte is OR'd in by enc_gp1_disp_mode_word). */
+		* Bit positions match the encoder shifts below; values are the *payload* bits only (cmd byte is OR'd in by enc_gp1_disp_mode_word). */
 	gp1_disp_HRes_256   = 0x0,
 	gp1_disp_HRes_320   = 0x1,
 	gp1_disp_HRes_512   = 0x2,
@@ -235,7 +235,7 @@ enum {
 #define enc_gp1_disp_hres(h)      (((h) & gp1_disp_hres_mask)      << gp1_disp_hres_shift)
 #define enc_gp1_disp_vres(v)      (((v) & gp1_disp_vres_mask)      << gp1_disp_vres_shift)
 #define enc_gp1_disp_color(c)     (((c) & gp1_disp_color_mask)     << gp1_disp_color_shift)
-#define enc_gp1_disp_interlace(i) (((i) & gp1_disp_interlace_mask << gp1_disp_interlace_shift)
+#define enc_gp1_disp_interlace(i) (((i) & gp1_disp_interlace_mask) << gp1_disp_interlace_shift)
 
 #define enc_gp1_hrange_x1(x1) (((x1) & gp1_hrange_x1_mask) << gp1_hrange_x1_shift)
 #define enc_gp1_hrange_x2(x2) (((x2) & gp1_hrange_x2_mask) << gp1_hrange_x2_shift)
@@ -255,6 +255,11 @@ enum {
 #define enc_gp0_draw_area_br_word(x, y)    (enc_gp0_cmd(gp0_cmd_SetDrawArea_BotRight) | enc_gp1_draw_x(x) | enc_gp1_draw_y(y))
 
 /* ---- Layer 3: GP1 semantic word builders ---- */
+#define gp1_word_Reset()             enc_gp0_cmd_word(gp1_cmd_Reset)
+#define gp1_word_ResetCmdBuffer()    enc_gp0_cmd_word(gp1_cmd_ResetCmdBuffer)
+#define gp1_word_AcknowledgeIRQ()    enc_gp0_cmd_word(gp1_cmd_AcknowledgeIRQ)
+#define gp1_word_StartDisplayArea()  enc_gp0_cmd_word(gp1_cmd_StartDisplayArea)
+
 #define gp1_word_display_enable(on)                         (enc_gp0_cmd(gp1_cmd_DisplayEnable) | ((on) & 1))
 #define gp1_word_display_disable()                          gp1_word_display_enable(0)
 #define gp1_word_display_mode_320x240_15bit_ntsc            enc_gp1_disp_mode_word(gp1_disp_HRes_320, gp1_disp_VRes_240, gp1_disp_Color15, 0)
@@ -279,6 +284,9 @@ enum {
 #define gp1_word_display_enabled   enc_gp0_cmd_word(gp1_cmd_DisplayEnable)
 #define gp1_word_display_disabled (enc_gp0_cmd_word(gp1_cmd_DisplayEnable) | 1)
 
+#define gp1_word_DisplayOn()       gp1_word_display_enable(0)
+#define gp1_word_DisplayOff()      gp1_word_display_enable(1)
+
 /* ---- DMA direction (2-bit payload on DMADirection cmd 0x04) ---- */
 enum {
 	gp1_dma_dir_Off            = 0,
@@ -287,6 +295,8 @@ enum {
 	gp1_dma_dir_GPUREAD_to_CPU = 3,
 };
 #define gp1_word_dma_direction(dir) (enc_gp0_cmd(gp1_cmd_DMADirection) | ((dir) & 0x3))
+#define gp1_word_dma_to_gpu()       gp1_word_dma_direction(gp1_dma_dir_CPU_to_GPU)
+#define gp1_word_dma_read_cpu()     gp1_word_dma_direction(gp1_dma_dir_GPUREAD_to_CPU)
 
 /* ---- Standard display ranges (NTSC + PAL pre-baked) ---- */
 /* Horizontal range values are in video clock units (8 units/pixel); vertical range values are scanline numbers. */
@@ -314,14 +324,49 @@ enum {
 /* ---- Draw-mode setting (TPage / draw-area allowance) ---- */
 /* The "drawing enabled" word is the standard post-init state. */
 enum {
-    gp0_DrawMode_DrawToDispBit = 10,
+	/* Per psx-spx, the standard 0xE1 layout has dfe at bit 10. But libpsyx's PutDrawEnv
+	 * uses bit 19 (in the "unused" 14-23 range) for dfe in the DR_ENV code[0] — and the
+	 * PSX hardware honors bit 19 in the DR_ENV context (not bit 10). So we need a
+	 * separate bit definition for the DR_ENV-specific DrawMode. */
+	gp0_DrawMode_DrawToDispBit         = 10,   // standard psx-spx bit 10 (dfe)
+	gp0_DrawMode_DR_ENV_DrawToDispBit  = 19,   // libpsyx DR_ENV code[0] (dfe in DR_ENV context)
+	gp0_DrawMode_DR_ENV_isbgBit        = 19,   // libpsyx uses bit 19 for isbg too
 };
 #define gp0_word_draw_mode_drawing_allowed (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit))
 
-/* ---- DrawArea pre-baked at origin (0,0) and full screen (320x240) ---- */
+/* DR_ENV-specific DrawMode variants (libpsyx SetDrawEnv layout).
+ * The DR_ENV is a 16-word packet emitted at boot by gp_screen_init's ac_put_draw_env_demo
+ * atom component. Within the DR_ENV, the 0xE1 command is reused in three different bit
+ * configurations:
+ *   code[0] = `gp0_word_draw_mode_drawing_allowed` (dfe=1; standard post-init state)
+ *   code[6] = `gp0_word_dr_env_bg_color_cmd(isbg, r, g, b)` (initial-bg-color path)
+ *   code[7] = `gp0_word_dr_env_draw_mode(isbg)` (isbg-flag path)
+ * Bits 0-23 of the 0xE1 word are the payload; bits 24-31 are the cmd byte (0xE1). */
+#define gp0_word_dr_env_bg_color_cmd(isbg, r, g, b) (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
+#define gp0_word_dr_env_draw_mode(isbg)             (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0))
+
+/* State-setter bare-cmd words (no immediate payload; the GPU uses the current state machine already programmed). */
+#define gp0_word_set_texture_window() enc_gp0_cmd_word(gp0_cmd_SetTextureWindow)
+#define gp0_word_set_draw_offset()    enc_gp0_cmd_word(gp0_cmd_SetDrawOffset)
+#define gp0_word_set_mask_bit()       enc_gp0_cmd_word(gp0_cmd_SetMaskBit)
+
+/* DR_ENV code[5] Mask (0xE6 cmd + isbg bit). The isbg bit is set so the GPU knows the auto-clear path is active (paired with code[6] + code[7]). */
+#define gp0_word_dr_env_mask() (gp0_word_set_mask_bit() | gp0_dr_env_isbg_bit)
+
+/* DR_ENV pre-baked constants (libpsyx PutDrawEnv layout).
+ * DR_ENV is a 16-word packet: tag = (length << 24) | addr, where length = 15 (15 code words follow) and addr = 0 (chain to nothing). */
+enum {
+	PolyTag_len_bits  =  8,
+	PolyTag_addr_bits = 24,
+
+	gp0_dr_env_tag      = (15 << 24) | 0x00FFFFFF,
+	gp0_dr_env_isbg_bit = (1 << gp0_DrawMode_DR_ENV_isbgBit),
+};
+
+/* ---- DrawArea at origin (0,0) and full screen (320x240) ---- */
 #define gp0_word_draw_area_top_left_origin      enc_gp0_draw_area_tl_word(0, 0)
-#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(320, 240)
-#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(640, 480)
+#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(319, 239)
+#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(639, 479)
 
 #pragma endregion GPU Ports & Commands
 
@@ -360,10 +405,10 @@ typedef Struct_(RGB8) { B1 r; B1 g; B1 b; };
 #define rgb8(r,g,b) ((RGB8){r,g,b})
 
 /* ---------- PolyTag (the OT-link header; 1 word) ---------- */
-enum {
-    PolyTag_len_bits  =  8,
-    PolyTag_addr_bits = 24,
-};
+// enum {
+// 	PolyTag_len_bits  =  8,
+// 	PolyTag_addr_bits = 24,
+// };
 typedef Struct_(PolyTag) {
 	union {
 		U4 code;
@@ -525,6 +570,9 @@ enum {
 	gp0_tpage_color_8bpp  = 0x1,
 	gp0_tpage_color_16bpp = 0x2,
 
+	/* Default TPage value libpsyx's SetDefDrawEnv writes (matches the `li v1, 10; sh v1, 20(v0)` sequence at C11_only.elf:0x8001273C). */
+	gp0_tpage_default    = 10,
+
 	/* TPage semi-transparency mode payload values (NOT bit positions). */
 	gp0_tpage_semi_trans_none  = 0x0,
 	gp0_tpage_semi_trans_alpha = 0x1,

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/gte.h

@@ -17,9 +17,8 @@
  *            gte_lw_v0_xy(base) (gte + lw + v0 + xy)
  *            load_upper_i       (load-upper + immediate, unique verb)
  *
- *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are
- *  NOT in this header. They live in the opt-in `gte_vendor_sym.h` for
- *  users who prefer the textbook MIPS assembly mnemonics.
+ *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are NOT in this header.
+ *  They are in the opt-in `gte_vendor_sym.h` for users who prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -34,20 +33,16 @@
  *  gte.h — Geometry Transformation Engine (COP2) for the PS1
  * ============================================================================
  *
- *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word`
- *  constants from C. No GCC inline-assembly string syntax in the code body.
+ *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word` constants from C.
+ *  No GCC inline-assembly string syntax in the code body.
  *
  *  STYLE NOTES
  *  -----------
- *  - Per-field encoders are named `enc_gte_<field>(value)` and each one
- *    self-masks its argument before shifting. Mirrors the `enc_op / enc_rs
- *    / enc_rt / ...` family in mips.h.
- *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of
- *    the per-field encoders, plus the COP2/CO base.
- *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined
- *    for the common cases so call sites read like assembly source.
- *  - All register/field values are enums (not `#define`s) so they show up
- *    in debugger symbol tables and IDE autocomplete.
+ *  - Per-field encoders are named `enc_gte_<field>(value)` and each one self-masks its argument before shifting.
+ *    Mirrors the `enc_op / enc_rs / enc_rt / ...` family in mips.h.
+ *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of the per-field encoders, plus the COP2/CO base.
+ *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined for the common cases so call sites read like assembly source.
+ *  - All register/field values are enums (not `#define`s) so they show up in debugger symbol tables and IDE autocomplete.
  *
  *  SEE ALSO
  *  --------
@@ -58,8 +53,7 @@
 
 /* --- GTE Data Registers (Coprocessor 2) ---
  * Preprocessor-visible integer ids for the COP2 data register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
  * Same pattern as the GPR `_Code` set in mips.h. */
 #define C2_VXY0_Code  0
 #define C2_VZ0_Code   1
@@ -192,10 +186,8 @@ enum {
 
 /* --- GTE Control Register Indices (for ctc2/cfc2) ---
  * Preprocessor-visible integer ids for the COP2 control register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
- * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23
- * are reserved/unused on real hardware, so there's a gap. */
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23 are reserved/unused on real hardware, so there's a gap. */
 #define gte_cr_RT11_Code  0
 #define gte_cr_RT12_Code  1  /* packed with RT13 in bits 16..31 */
 #define gte_cr_RT13_Code  2  /* packed with RT22 in bits 16..31 */
@@ -223,8 +215,9 @@ enum {
 #define gte_cr_RFC_Code  27
 #define gte_cr_GFC_Code  28
 #define gte_cr_BFC_Code  29
-#define gte_cr_OFX_Code  30
-#define gte_cr_OFY_Code  31
+#define gte_cr_OFX_Code  24
+#define gte_cr_OFY_Code  25
+#define gte_cr_H_Code    26
 
 enum {
 	gte_cr_RT11 = gte_cr_RT11_Code, gte_cr_RT12 = gte_cr_RT12_Code, gte_cr_RT13 = gte_cr_RT13_Code,
@@ -246,21 +239,16 @@ enum { _C2_OPS_ = 0
 
 /* COP2 transfer sub-opcodes (5-bit field in the `rs` slot of enc_gte_tx).
  *
- * Spans the 2x2 {From, To} × {Data, Control} register classes that the
- * GTE exposes:
- *
+ * Spans the 2x2 {From, To} × {Data, Control} register classes that the GTE exposes:
  *   bit 1 (0x02): register class — 0 = data,  1 = control
  *   bit 2 (0x04): direction      — 0 = read,  1 = write
  *
- * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit
- * numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
- * (which target the data register file on any coprocessor). They are
- * re-aliased here so the four-way table reads like the spec mnemonics
- * (MFC2 / CFC2 / MTC2 / CTC2) and so the encoding lives next to its
- * only consumer (this header).
+ * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
+ * (which target the data register file on any coprocessor).
+ * They are re-aliased here so the four-way table reads like the spec mnemonics (MFC2 / CFC2 / MTC2 / CTC2)
+ * and so the encoding lives next to its only consumer (this header).
  *
- * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2)
- * live in gte_vendor_sym.h. */
+ * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2) live in gte_vendor_sym.h. */
 enum { _C2_TX_SUBS_ = 0
 	, sub_mfc2 = 0x00 /* MFC2: Move From Coprocessor 2 data   reg */
 	, sub_cfc2 = 0x02 /* CFC2: Copy From Coprocessor 2 ctrl   reg */
@@ -314,8 +302,8 @@ enum { _C2_TX_SUBS_ = 0
  * `swc2` is redundant when we're already inside the `gte_` namespace.
  *   gte_lw rt, base, off  →  lwc2 rt, off(base)
  *   gte_sw rt, base, off  →  swc2 rt, off(base)
- * For the typical user-facing vector-level load (xy + z as two
- * instructions), use the higher-level `gte_load_vN` macros below. */
+ * For the typical user-facing vector-level load (xy + z as two instructions),
+ * use the higher-level `gte_load_vN` macros below. */
 #define gte_lw(rt, base, off) enc_gte_lw(rt, base, off)
 #define gte_sw(rt, base, off) enc_gte_sw(rt, base, off)
 
@@ -323,13 +311,12 @@ enum { _C2_TX_SUBS_ = 0
  * Opcode is always MIPS_OP_COP2, RS is always 1 (CO).
  * The lower 25 bits are the GTE-specific command payload.
  *
- * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs`
- * pattern in mips.h: each one self-masks and shifts its own field, so a
- * caller can build up a GTE command piece by piece (handy for state-driven
- * MVMVA emitters that vary one field at a time).
+ * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs` pattern in mips.h:
+ * Each one self-masks and shifts its own field, so a caller can build up a GTE command piece by piece
+ * (handy for state-driven MVMVA emitters that vary one field at a time).
  *
- * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command
- * word in one go. It just ORs the per-field encoders together. */
+ * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command word in one go.
+ * It just ORs the per-field encoders together. */
 #define gte_cmd_base (enc_op(op_cop2) | (1 << 25))
 
 /* Per-field encoders. Each one does (value & mask) << shift on its own. */
@@ -363,8 +350,9 @@ enum { _C2_TX_SUBS_ = 0
  * (standard rotation-matrix, no scaling factor, V0 vector, translation vector, no clamp),
  * so the only varying bits are the `cmd` field.
  *
- * Naming follows the file's convention: `gte_cmd_*` is the raw 6-bit `cmd` field id, `gte_cmdw_*`
- *  is the fully-encoded 32-bit instruction word ready to drop into a `.word` directive.
+ * Naming convention:
+ * - `gte_cmd_*` : Raw 6-bit `cmd` field id 
+ * - `gte_cmdw_* : 32-bit instruction word ready to drop into a `.word` directive.
  *
  * --------------------------------------------------------------------------
  *  PsyQ-compatibility note (RTPS/RTPT):
@@ -413,20 +401,16 @@ enum { _C2_TX_SUBS_ = 0
 
 /**
  * @brief Loads a single SVECTOR to GTE vector register V0
- *
  * @details Loads values from an SVECTOR struct to GTE data registers C2_VXY0
  * (XY at offset 0) and C2_VZ0 (Z at offset 4) using `lwc2`.
  *
- * Uses string-style GCC inline asm with `%0` substitution because the 
- * base register `r0` is a runtime GPR chosen by the compiler.
+ * Uses string-style GCC inline asm with `%0` substitution because the base register `r0` is a runtime GPR chosen by the compiler.
  * It cannot be encoded into a static `.word` constant.
  *
- * Usage:
- *   asm_gte_load_v0(svector_ptr);
+ * Usage: asm_gte_load_v0(svector_ptr);
  */
 
 /* lwc2 encoding helpers parameterized on the base GPR.
- *
  * gte_lw_v0_xy(base)  → lwc2 $0,  0(base)   ; C2_VXY0
  * gte_lw_v0_z(base)   → lwc2 $1,  4(base)   ; C2_VZ0
  * gte_lw_v1_xy(base)  → lwc2 $2,  0(base)   ; C2_VXY1
@@ -435,8 +419,7 @@ enum { _C2_TX_SUBS_ = 0
  * gte_lw_v2_z(base)   → lwc2 $5,  4(base)   ; C2_VZ2
  *
  * `base` is the GPR number to bake into the .word constant's `rs` field.
- * These are pure compile-time integers; the C compiler constant-folds
- * them into .word directives. */
+ * These are pure compile-time integers; the C compiler constant-folds them into .word directives. */
 
 enum {
 	GTE_Z_Offset = 4
@@ -459,8 +442,8 @@ enum {
  *     gte_load_v0(p_in_12, R_T4);   // R_T4 = 12, base is $12
  *
  * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register `p_in_12` can live in),
- * which is exactly the register the .word constants expect. A `"$12"` clobber would conflict with the register-variable binding
- * ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
+ * which is exactly the register the .word constants expect.
+ * A `"$12"` clobber would conflict with the register-variable binding ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
  * The other ABI-clobbers ($2/$8/$9/$31) stay because the GTE instructions don't touch caller-saved GPRs but the kernel does treat them as volatile.
  *
  * WHICH REGISTER TO PICK
@@ -499,10 +482,8 @@ enum {
 
 /* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — prelude to gte_cmd_rtpt.
  *
- * Loads all three GTE input vectors (6 words) from three separate pointers,
- * one per GTE vector register, each loaded from its own base GPR.
- * Caller must bind each `pN` to `bN` via a register variable.
- *
+ * Loads all three GTE input vectors (6 words) from three separate pointers, one per GTE vector register,
+ * each loaded from its own base GPR. Caller must bind each `pN` to `bN` via a register variable.
  *   register V3_S2* p0 rgcc(R_T4) = verts[0].ptr;  // → __asm__("$12")
  *   register V3_S2* p1 rgcc(R_T5) = verts[1].ptr;  // → __asm__("$13")
  *   register V3_S2* p2 rgcc(R_T6) = verts[2].ptr;  // → __asm__("$14")
@@ -521,29 +502,20 @@ enum {
 
 /**
  * @brief Rotate, Translate and Perspective Triple (23 cycles)
- *
- * @details Performs rotation, translation and perspective calculation of three
- * vertices at once. The equation performed is the same as gte_rtps() only
- * repeated three times for each vertex. The result of the first vertex is
- * stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then
- * C2_SXY2.
+ * @details Performs rotation, translation and perspective calculation of three vertices at once.
+ * The equation performed is the same as gte_rtps() only repeated three times for each vertex.
+ * The result of the first vertex is stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then C2_SXY2.
  *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency — the GTE
- *      takes ~8 cycles per perspective divide, and the nops let any
- *      preceding lwc2/swc2 retire before RTPT starts reading its
- *      inputs from V0/V1/V2.
- *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the
- *      pre-baked encoders above) — `0x0280030` decoded as
- *      `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields
- *      zero (standard rotation, no scaling, V0 vector, translation
- *      vector, no clamp).
+ *   1. Two `nop` words fill the COP2 pipeline latency — the GTE takes ~8 cycles per perspective divide, 
+ *      and the nops let any preceding lwc2/swc2 retire before RTPT starts reading its inputs from V0/V1/V2.
+ *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the pre-baked encoders above) — 
+ *      `0x0280030` decoded as `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields zero
+ *      (standard rotation, no scaling, V0 vector, translation vector, no clamp).
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register — those have to be saved by the caller if
- * they need to survive across the call (RTPT writes SXY0..2, SZ0..3,
- * OTZ, MAC0..3, IR0..3, etc.).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI)
+ * plus the standard "memory" barrier. Does not clobber any COP2 data/control register — 
+ * those have to be saved by the caller if they need to survive across the call (RTPT writes SXY0..2, SZ0..3, OTZ, MAC0..3, IR0..3, etc.).
  */
 #define gte_rtpt()                        \
 	asm volatile(                           \
@@ -559,32 +531,24 @@ enum {
 
 /**
  * @brief Normal clipping (8 cycles)
- *
- * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for
- * backface culling. If the value of C2_MAC0 is negative, the coordinates are
- * inverted and thus the triangle is back facing.
+ * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for backface culling.
+ * If the value of C2_MAC0 is negative, the coordinates are inverted and thus the triangle is back facing.
  *
  * The following equation is performed when executing this GTE command:
- *
  *     MAC0 = SX0*SY1 + SX1*SY2 + SX2*SY0 - SX0*SY2 - SX1*SY0 - SX2*SY1
- *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency - the GTE
- *      pipeline takes a few cycles per op, and the nops let any
- *      preceding lwc2/swc2/RTPT retire before NCLIP starts reading
- *      its inputs from SXY0/SXY1/SXY2.
- *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the
- *      pre-baked encoders above) - `0x01400006` decoded as
- *      `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields
- *      zero. NCLIP is spec-clean in the original PsyQ source
- *      (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat`
- *      quirk), so `gte_cmdw_nclip` does NOT OR in any reserved bits.
+ *   1. Two `nop` words fill the COP2 pipeline latency
+ *      - the GTE pipeline takes a few cycles per op, and the nops let any preceding
+ *      lwc2/swc2/RTPT retire before NCLIP starts reading its inputs from SXY0/SXY1/SXY2.
+ *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the pre-baked encoders above)
+ *      - `0x01400006` decoded as  `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields zero.
+ *      NCLIP is spec-clean in the original PsyQ source (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat` quirk),
+ *      so `gte_cmdw_nclip` does NOT OR in any reserved bits.
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register - those have to be saved by the caller if
- * they need to survive across the call (NCLIP writes MAC0 only; it
- * is purely a sign-of-double-product computation on SXY0..2).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI) plus the standard "memory" barrier.
+ * Does not clobber any COP2 data/control register.
+ * Those have to be saved by the caller if they need to survive across the call (NCLIP writes MAC0 only;
+ * it is purely a sign-of-double-product computation on SXY0..2).
  */
 #define gte_nclip()                       \
 	asm volatile(                           \
@@ -610,13 +574,10 @@ enum {
 		"cop2 0x0158002D;")
 
 /* asm_gte_matrix_set_rotation(r0)
+ * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix control registers (RT11..RT22, indices 0..4) via ctc2.
  *
- * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix
- * control registers (RT11..RT22, indices 0..4) via ctc2.
- *
- * Memory layout at r0: five contiguous 32-bit words (offsets 0..16),
- * each holding two packed 16-bit matrix elements. The first 1.5 rows
- * of a standard PSX SDK MATRIX struct (where each row is laid out as
+ * Memory layout at r0: five contiguous 32-bit words (offsets 0..16), each holding two packed 16-bit matrix elements.
+ * The first 1.5 rows of a standard PSX SDK MATRIX struct (where each row is laid out as
  * [RT_xx, RT_xy] | [RT_xz, pad] | ...).
  *
  * Generated MIPS (mirrors the source macro):
@@ -631,27 +592,22 @@ enum {
  *   ctc2 $13,  $3         ; → C2_RT21
  *   ctc2 $14,  $4         ; → C2_RT22
  *
- * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a
- * register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
- * instructions to read from the right base. The `"r"(r0)` constraint
- * alone doesn't force a specific GPR — it just lets GCC pick one.
- * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field
- * of each lw, so the lw instructions will only do the right thing
- * if $12/$13/$14 hold the matrix base at runtime.
+ * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
+ * instructions to read from the right base. The `"r"(r0)` constraint alone doesn't force a specific GPR — it just lets GCC pick one.
+ * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field of each lw, so the lw instructions will 
+ * only do the right thing if $12 / $13 / $14 hold the matrix base at runtime.
  *
  *   M3_S2* m = ...;
  *   register M3_S2* m_in_12 rgcc(R_T4) = m;
  *   asm_gte_matrix_set_rotation(m_in_12);
  *
- * We clobber $12/$13/$14 (the ones we use as scratch inside the
- * inline asm) plus the system clobbers; we don't clobber `r0` because
- * the `rgcc` binding already says "this variable lives in $12".
+ * We clobber $12/$13/$14 (the ones we use as scratch inside the inline asm)
+ * plus the system clobbers; we don't clobber `r0` because the `rgcc` binding already says "this variable lives in $12".
  *
- * WARNING: Incomplete by design. The source macro only writes RT11..RT22
- * (5 of 9 rotation elements); RT23 and the entire RT3x row are left
- * untouched. Real libpsn00b SetRotMatrix writes all 9. Use only when the
- * GTE's remaining rotation entries are already correct, or you will
- * get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
+ * WARNING: Incomplete by design. The source macro only writes RT11..RT22 (5 of 9 rotation elements);
+ * RT23 and the entire RT3x row are left untouched.
+ * Real libpsn00b SetRotMatrix writes all 9. Use only when the GTE's remaining rotation entries are already correct,
+ * or you will get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
  */
 #define asm_gte_matrix_set_rotation(r0) \
 	asm volatile(                         \

code/duffle/lottes_tape.h

@@ -1,19 +1,122 @@
 #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 "gen/duffle.macs.h"
-#	include "gen/duffle.offsets.h"
+#	include "dsl.atom.h"
 #endif
 
-typedef U4 const MipsCode;
-typedef Slice_(MipsCode);
-typedef Slice_MipsCode MipsAtom;
+#pragma region Tape Drive
+/* -----------------------------------------------------------------------------
+ *  TAPE DRIVE ABI
+ * -----------------------------------------------------------------------------
+ * 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 
+ * or finalized conceptually.
+ * -----------------------------------------------------------------------------
+ * This ABI and its associated legos were directly inspired by researching
+ * the work of Timothy Lottes and Onat Türkçüoğlu; along with many others.
+ * 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.
+ * These eventually lead to a terminal atom for the tape which is defined
+ * below as "tape_exit".
+ * 
+ * This behaves as one of the simplest runtime harnesses ontop of a 
+ * host-enviornment's execution engine to author and compose programs with.
+ * From here various conventions can be further applied.
+ * 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
+ * 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 */
+/* Stringification codes for the GCC inline assembler clobber lists. */
+#define R_AtomJmp_Code R_T8_Code
+#define R_TapePtr_Code R_T9_Code
 
+// R_InCursor = R_T4,
+// #define R_InCursor_Code    R_T4_Code
+
+// Reserved Registers (Callee-saved):
+// - R_T9: Holds the Tape Ptr which we need to increment
+//         If we hit a wall with register allocations we can clobber V0 & V1 (return values), defering as opt-in by user.
+// - R_RA: Not sure??
+// Needed by ac_yield but can be used as atom scratch:
+// - R_T8: Will be used as the atom jump register.
+
+// All allocatable registers for mips atoms:
+	R_TScratchVolatile = R_AT, // This one is reserved for psuedo instructions, but you can technically use it.
+	R_TScratch0  = R_T0,
+	R_TScratch1  = R_T1,
+	R_TScratch2  = R_T2,
+	R_TScratch3  = R_T3,
+	R_TScratch4  = R_T4,
+	R_TScratch5  = R_T5,
+	R_TScratch6  = R_T6,
+	R_TScratch7  = R_T7,
+	R_TScratch8 =  R_T8,
+	R_TScratch10 = R_V0, // Tend to be used with gte DMAs
+	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.
+	// A 0-2
+	// S 0-7
+};
+
+typedef U4 const MipsCode; // Underlying type to mips asm words.
+typedef Slice_(MipsCode);
+
+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).
@@ -23,78 +126,85 @@ typedef Slice_MipsCode MipsAtom;
 #define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
 
 // Used for components with value-args (e.g., ac_format_f3_color).
-//   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+//   FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
 // expands to:
-//   FI_ MipsAtom ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+//   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); }
 
-// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
+/* 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.
 
-/* Register aliases */
-enum {
-	R_AtomJmp    = R_T9 atom_reg,  /* debug-visible; tape yield handshake scratch */
-	R_TapePtr    = R_T8 atom_reg,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,
+   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
 
-	R_PrimCursor = R_T7 atom_reg atom_type(U4 *),    /* VRAM output cursor (primitive buffer) */
-	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2 *), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
-	R_VertBase   = R_T5 atom_reg atom_type(V3_S2 *), /* Base address of the vertex array */
-	R_OtBase     = R_T6 atom_reg atom_type(U4 *),    /* Base address of the Ordering Table */
-
-/* Stringification codes for the GCC inline assembler clobber lists. */
-#define R_TapePtr_Code     R_T8_Code
-#define R_InCursor_Code    R_T4_Code
-
-#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
-};
-
-#pragma region Tape Drive
-/* ---------------------------------------------------------------------------
- *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
- * ---------------------------------------------------------------------------*/
+ typedef Slice_(MipsAtom); typedef Slice_MipsAtom Tape;
 
 /* The 'Exit' Atom */
 atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
-/* Generalized Tape Engine Runner */
-NI_ void tape_run(Slice_MipsCode tape) { register U4* tp rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
+// TODO(Ed): When we have a substantial workload/throughput, profile each of these to see impact at ABI boundaries.
+
+/* Tape Runner (Default) */
+FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
-			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
-		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
-		, load_word(  R_AtomJmp, R_TapePtr, 0)         /* Bootstrap the first jump */
-		, add_ui_self(R_TapePtr, S_(MipsCode))         /* Advance tape */
+		  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 */
-		, load_word(  R_RA, R_SP, 0)                   /* Restore $ra from stack */
-		, add_ui_self(R_SP, MipsStackAlignment)        /* Deallocate stack space */
 	)
-	asm_rpins, r_use(tp)
+	asm_rpins, r_use(tape_ptr)
 	asm_clobber: 
-			rlit(R_AT)
-		, rlit(R_V0), rlit(R_V1)
-		, rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3)
-		/* Tell GCC the tape engine owns and destroys the workspace registers */
-		, rlit(R_PrimCursor), rlit(R_FaceCursor), rlit(R_VertBase), rlit(R_OtBase)
-		, rlit(R_T9)
-		, clb_mem_drain 
+		rlit(R_AT),
+		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 }; }
 FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }; }
 
-#define tb_emit_(tb, atom) tb_emit(tb, atom)
 FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
+#define tb_emit_(atom)        tb_emit(& tb, atom)
+#define tb_data_(field, data) tb_data(& tb, u4_(data))
 
-FI_ Slice_MipsCode tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Slice_MipsCode){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Slice_MipsCode){ C_(U4*,tb.ptr),  tb.used }; }
+FI_ Tape tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Tape){ C_(U4*,tb->ptr), tb->used }; }
+FI_ Tape tb_slice(TapeBuilder  tb) {                        return (Tape){ C_(U4*,tb.ptr),  tb.used }; }
 #define tb_scope(tb)     for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,tape_exit))
 
+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
@@ -104,118 +214,30 @@ FI_ Slice_MipsCode tb_slice(TapeBuilder  tb) {                        return (Sl
  * ---------------------------------------------------------------------------*/
 
 // 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,
 };
 
-/* Words: 3; Loads 3 S2 indices from the face array */
-atom_dbg_skip MipsAtomComp_(ac_load_tri_indices) {
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
-	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
-	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+atom_dbg_skip MipsAtomComp_(ac_yield_load) {
+	load_word(R_AtomJmp, R_TapePtr, 0),
 };
 
-/* 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),
+atom_dbg_skip MipsAtomComp_(ac_yield_tail) {
+	add_ui_self(R_TapePtr, S_(MipsCode)),
+	jump_reg(  R_AtomJmp), nop,
 };
 
-/* 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_ MipsAtom 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_ MipsAtom 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_post_rtpt) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
-};
-
-/* Words: 12; Emits the four (code|color) words of a Poly_G4.
- * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-FI_ MipsAtom ac_format_g4_color(
-	U1 r0, U1 g0, U1 b0,
-  U1 r1, U1 g1, U1 b1,
-  U1 r2, U1 g2, U1 b2,
-  U1 r3, U1 g3, U1 b3)
-MipsAtomComp_Proc_(ac_format_g4_color, {
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
-	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1),
-	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2),
-	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3),
-})
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
- * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
- * three registers aligned with v0/v1/v2 you must store before RTPS).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
-};
-
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its
- * single-vertex result to SXY2; SXY0 still holds v0.screen from the
- * earlier RTPT — DO NOT read SXY0 here, that's the bug this name
- * prevents).
- */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
-
 #pragma endregion Macro Atom Components
 
 #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; };
 
@@ -237,120 +259,10 @@ FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
 	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
 }
 
-#define mipsatom_from_builder(ab) (MipsAtom){ab.start, ab.used}
-
+#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()
-};
-
-/* 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()
-};
-
 #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

@@ -11,6 +11,7 @@ enum {
 	v3s2_byteoff = 3, // log2(8), used with shift_left_logical op for index via byte offset.
 };
 
+typedef Array_(U1, 2);
 typedef Array_(U4, 2);
 typedef Array_(S2, 2);
 typedef Array_(S2, 3);
@@ -22,6 +23,7 @@ typedef S2 A3x3_S2[3][3];
 typedef Struct_(Extent2_S2) { S2 width; S2 height; };
 typedef Struct_(Extent2_S4) { S4 width; S4 height; };
 
+typedef Struct_(V2_U1)      { U1 x; U1 y; };
 typedef Struct_(V2_S2)      { S2 x; S2 y; };
 typedef Struct_(V2_S4)      { S4 x; S4 y; };
 typedef Struct_(V3_S2)      { S2 x; S2 y; S2 z; S2 pad; };
@@ -37,6 +39,7 @@ typedef Struct_(Rect_S4) { S4 x; S4 y; S4 width; S4 height; };
 
 typedef Struct_(M3_S2)      { A3x3_S2 m; A3_S4 t; };
 
+typedef Array_(V2_S2, 2);
 typedef Array_(V2_S2, 3);
 typedef Array_(V2_S2, 4);
 
@@ -58,10 +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) { 
-	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   (V3_S4_R out_a, V3_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));  }

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) --- */
@@ -444,18 +462,15 @@ enum { _BitOffsets = 0
 #define load_imm_1w_s0(rt, imm) add_si((rt)), R_0, (imm))
 
 /* load_imm_2w — unconditional 2-word `li` form: `lui` + (ori | addi).
- *
  * Granular companion to `load_imm`: skips the compile-time range checks and always emits 2 .words. Use this when:
  *   - you know `imm` is > 0xFFFF (otherwise you're wasting a word), OR
- *   - `imm` is not a compile-time constant and you want predictable
- *     2-word emission without the `__builtin_constant_p` branches.
+ *   - `imm` is not a compile-time constant and you want predictable 2-word emission without the `__builtin_constant_p` branches.
  *
  * The lo16 strategy is still chosen at expansion time on the lo half:
  *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui already cleared bits 15..0)
  *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the intended bit pattern)
  *
- * For situations where you need to bypass even this choice 
- * (e.g. to force a specific encoding for a known discontiguous high/low pair),
+ * For situations where you need to bypass even this choice (e.g. to force a specific encoding for a known discontiguous high/low pair),
  * see `load_imm_2w_ori_forced` and `load_imm_2w_addi_forced` below.
  * Statement-level (not expression-level): emits its own `asm volatile(...)`.
  */

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/pad.h

@@ -0,0 +1,73 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "dsl.h"
+#endif
+
+/* PSX button bit positions — 1:1 with PSX-SPX docs at docs/psx-spx/docs/controllersandmemorycards.md:405-421.
+ * Wire is active-low (0 = pressed).
+ * The decoder atom computes buttons = (~raw_buttons) & 0xFFFF; the active-low-to-active-high inversion is applied bit-by-bit. */
+enum {
+	Bit_(Pad_Select,  0),
+	Bit_(Pad_L3,      1),
+	Bit_(Pad_R3,      2),
+	Bit_(Pad_Start,   3),
+	Bit_(Pad_Up,      4),
+	Bit_(Pad_Right,   5),
+	Bit_(Pad_Down,    6),
+	Bit_(Pad_Left,    7),
+	Bit_(Pad_L2,      8),
+	Bit_(Pad_R2,      9),
+	Bit_(Pad_L1,     10),
+	Bit_(Pad_R1,     11),
+	Bit_(Pad_Triangle, 12),
+	Bit_(Pad_Circle,  13),
+	Bit_(Pad_Cross,   14),
+	Bit_(Pad_Square,  15),
+};
+
+enum {
+	PadId_Offset = 4,
+
+	Pad0 = 0 << PadId_Offset,
+	Pad1 = 1 << PadId_Offset,
+};
+
+#define pad0_(btn_id) (btn_id << Pad0)
+#define pad1_(btn_id) (btn_id << Pad1)
+
+/* ============================================================
+ * BIOS pad-buffer subsystem: docs/psx-spx/docs/kernelbios.md (B(12h) + B(13h))
+ * ============================================================ */
+
+enum {
+	PAD_BIOS_RAW_SIZE = 0x22,
+};
+typedef Struct_(PadBiosRaw) {
+	U1 bytes[PAD_BIOS_RAW_SIZE];
+};
+
+typedef Enum_(U4, PadStatus) {
+	PadStatus_Disconnected,
+	PadStatus_Digital,
+	PadStatus_AnalogStick,
+	PadStatus_AnalogPad,
+	PadStatus_Unsupported,
+	PadStatus_Pending,
+	PadStatus_Invalid,
+};
+
+/* PadState — per-port normalized runtime state.
+ * Field order is chosen so that the 4 axes (left_x, left_y, right_x, right_y)
+ * form a contiguous 4-byte block at offset 8, allowing a single `store_word` to clear-or-write all 4 axes in one MIPS instruction.
+ * The struct size stays 12 bytes (unchanged from the prior order,
+ * which left the C compiler to insert 1 byte of trailing pad to reach the 4-byte struct alignment). */
+typedef Struct_(PadState) {
+	PadStatus status;  /* offset 0,  size 4 (U4) */
+	U2        buttons; /* offset 4,  size 2 */
+	U1        id;      /* offset 6,  size 1 */
+	U1        pad;     /* offset 7,  size 1 — explicit pad to align the axes block */
+	U1        left_x;  /* offset 8,  size 1 — store_word target (4-byte aligned) */
+	U1        left_y;  /* offset 9,  size 1 */
+	U1        right_x; /* offset 10, size 1 */
+	U1        right_y; /* offset 11, size 1 */
+};

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

@@ -0,0 +1,103 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#	include "dsl.h"
+#	include "math.h"
+#	include "gp.h"
+#endif
+
+typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };
+typedef Struct_(DrawEnv) {
+	Rect_S2 clip_area;
+	V2_S2   drawing_offset[2];
+	Rect_S2 texture_window;
+	S2      texture_page;
+	B1      flag_dither;
+	B1      flag_draw_on_display;
+	B1      enable_auto_clear;
+	RGB8    initial_bg_color;
+	DrawEnv_Packed dr_env; // reserved
+};
+typedef Struct_(DisplayEnv) {
+	Rect_S2 display_area;
+	Rect_S2 screen;
+	B1      vinterlace;
+	B1      color24;
+	B1      pad0;
+	B1      pad1;
+};
+typedef Array_(DrawEnv,    2);
+typedef Array_(DisplayEnv, 2);
+
+typedef Struct_(DoubleBuffer) {
+	A2_DrawEnv    draw;
+	A2_DisplayEnv display;
+};
+
+DisplayEnv* displayenv_init(DisplayEnv* env, S4 x, S4 y, S4 w, S4 h) asm("SetDefDispEnv");
+DrawEnv*    drawenv_init   (DrawEnv*    env, S4 x, S4 y, S4 w, S4 h) asm("SetDefDrawEnv");
+
+DisplayEnv* displayenv_put(DisplayEnv* env) asm("PutDispEnv");
+DrawEnv*    drawenv_put   (DrawEnv*    env) asm("PutDrawEnv");
+
+U4   geom_init(void)             asm("InitGeom");
+void geom_set_offset(U4 x, U4 y) asm("SetGeomOffset");
+void geom_set_screen(U4 h)       asm("SetGeomScreen");
+
+U4* orderingtbl_clear_reverse(U4* ot, U4 len) asm("ClearOTagR");
+
+U4 reset_graph(U4 mode)           asm("ResetGraph");
+void set_display_enabled(U4 mask) asm("SetDispMask");
+
+U4 draw_sync(U4 mode) asm("DrawSync");
+U4 vsync(U4 mode)     asm("VSync");
+
+void draw_orderingtbl(U4* buf) asm("DrawOTag");
+
+typedef Struct_(Tile) {
+	U4      tag;
+	RGB8    color;
+	B1      code;
+	Rect_S2 rect;
+};
+
+/*
+	Linear Algebra
+*/
+
+M3_S2* m3s2_rotation   (V3_S2* vec, M3_S2* mat) asm("RotMatrix");
+M3_S2* m3s2_translation(M3_S2* mat, V3_S4* vec) asm("TransMatrix");
+M3_S2* m3s2_scale      (M3_S2* mat, V3_S4* vec) asm("ScaleMatrix");
+
+// Rotation, Translation, Perspective
+
+S4 rtp_v3s2_raw(V3_S2* vec, S4* xy, S4* pp, S4* flag) asm("RotTransPers");
+FI_ S4 rtp_v3s2(V3_S2* vec, V2_S2* xy, A2_S2* pp, S4* flag) { return rtp_v3s2_raw(vec, C_(S4*R_, & xy->x), C_(S4*R_, pp), r_(flag)); }
+
+S4 rtp_avg_nclip_a3_v3s2_raw(V3_S2* v0, V3_S2* v1, V3_S2* v2, S4* xy1, S4* xy2, S4* xy3, S4* pp, S4* otz, S4* flag) asm("RotAverageNclip3");
+FI_  S4 rtp_avg_nclip_a3_v3s2(
+	V3_S2* v0,  V3_S2* v1,  V3_S2* v2, 
+	V2_S2* xy0, V2_S2* xy1, V2_S2* xy2, 
+	A2_S2* pp, S4* otz, S4* flag
+){
+	return rtp_avg_nclip_a3_v3s2_raw(
+		v0, v1, v2, 
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
+	);
+}
+
+S4 rtp_avg_nclip_a4_v3s2_raw(V3_S2* v0, V3_S2* v1, V3_S2* v2, V3_S2* v3, S4* xy1, S4* xy2, S4* xy3, S4* xy4, S4* pp, S4* otz, S4* flag) asm("RotAverageNclip4");
+FI_ S4 rtp_avg_nclip_a4_v3s2(
+	V3_S2* v0,  V3_S2* v1,  V3_S2* v2,  V3_S2* v3,
+	V2_S2* xy0, V2_S2* xy1, V2_S2* xy2, V2_S2* xy3,
+	A2_S2* pp,  S4* otz,    S4* flag
+){
+	return rtp_avg_nclip_a4_v3s2_raw(
+		v0, v1, v2, v3,
+		C_(S4*R_, xy0), C_(S4*R_, xy1), C_(S4*R_, xy2), C_(S4*R_, xy3),
+		C_(S4*R_, pp),  C_(S4*R_, otz), C_(S4*R_, flag)
+	);
+}
+
+void gte_matrix_set_rotation   (M3_S2* mat) asm("SetRotMatrix");
+void gte_matrix_set_translation(M3_S2* mat) asm("SetTransMatrix");

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.
 //
@@ -22,6 +22,7 @@ WORD_COUNT(call_reg,                         1)
 WORD_COUNT(call_addr,                        1)
 WORD_COUNT(branch_le_zero,                   1)
 WORD_COUNT(branch_equal,                     1)
+WORD_COUNT(branch_ne,                        1)
 WORD_COUNT(add_ui,                           1)
 WORD_COUNT(set_lt_u,                         1)
 WORD_COUNT(set_lt_s,                         1)
@@ -29,7 +30,9 @@ WORD_COUNT(set_lt_si,                        1)
 WORD_COUNT(set_lt_ui,                        1)
 WORD_COUNT(load_word,                        1)
 WORD_COUNT(load_half_u,                      1)
+WORD_COUNT(load_byte_u,                      1)
 WORD_COUNT(store_word,                       1)
+WORD_COUNT(store_byte,                       1)
 WORD_COUNT(add_ui_self,                      1)
 WORD_COUNT(add_u_self,                       1)
 WORD_COUNT(add_u,                            1)
@@ -37,6 +40,7 @@ WORD_COUNT(or_i,                             1)
 WORD_COUNT(or_i_self,                        1)
 WORD_COUNT(or_u,                             1)
 WORD_COUNT(or_u_self,                        1)
+WORD_COUNT(nor_u,                            1)
 WORD_COUNT(shift_lleft,                      1)
 WORD_COUNT(shift_lleft_self,                 1)
 WORD_COUNT(shift_lright,                     1)

code/gte_hello/hello_gte_tape.c

@@ -1,156 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
-#	include "duffle/gen/duffle.offsets.h"
-#	include "duffle/atom_dsl.h"
-#	include "duffle/lottes_tape.h"
-#	include "duffle/word_count.metadata.h"
-#	include "gen/gte_hello.offsets.h"
-#	include "hello_gte.h"
-#endif
-
-#pragma region MACs (Mips Atom components)
-
-
-
-#pragma endregion MACs
-
-#pragma region Baked Atoms
-
-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)
-){
-	/* 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_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-	nop2, gte_cmdw_nclip,
-
-	nop2,  gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
-
-	store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
-	mac_format_g4_color(
-		/* c0 magenta */ 0xFF, 0x00, 0xFF,
-		/* c1 yellow  */ 0xFF, 0xFF, 0x00,
-		/* c2 cyan    */ 0x00, 0xFF, 0xFF,
-		/* c3 green   */ 0x00, 0xFF, 0x00),
-	mac_gte_store_g4_p012_post_rtpt_pre_rtps(),
-
-	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),
-
-	nop2, gte_cmdw_rotate_translate_perspective_single,
-	mac_gte_store_g4_p3_post_rtps(),
-
-	nop2, gte_cmdw_avg_sort_z4,
-	nop2, 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(),
-
-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)
-){
-	/* 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_FaceCursr)
-) {
-	mac_load_tri_indices(  R_T0, R_T1, R_T2),
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple, // 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_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-	mac_gte_store_f3_post_rtpt(),
-
-	/* 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,
-	/* Insert into Ordering Table Linked List */
-	mac_insert_ot_tag_f3(),
-	add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
-		// Note(Ed): No bounds checking, should be checked before atom runs.
-
-/* Advance Input Cursor & Yield (Both branch targets land here) */
-atom_label(floor_f3_face_exit)
-	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/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_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
+
+#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_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/gen/gte_hello.offsets.h

@@ -5,10 +5,10 @@
 #pragma region hello_gte_tape
 
 
-// --- atom: cube_g4_face (87 words) ---
+// --- atom: cube_g4_face (77 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           48
-#define _atom_offset_bounds_chk_cube_g4_face_exit     12
+#define _atom_offset_cull_cube_g4_face_exit           42
+#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,
@@ -18,7 +18,7 @@ enum {
 // --- atom: floor_f3_face (58 words) ---
 
 #define _atom_offset_culling_floor_f3_face_exit       25
-#define _atom_offset_bounds_chk_floor_f3_face_exit    13
+#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,

code/hello_gte/gen/hello_gte.offsets.h

@@ -0,0 +1,29 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\hello_gte\hello_gte.tape.c
+#pragma once
+
+#pragma region hello_gte.tape
+
+
+// --- atom: cube_g4_face (77 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           42
+#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_gte.tape
+

code/hello_gte/hello_gte.c

@@ -14,16 +14,16 @@
 #include "duffle/gp.h"
 #include "duffle/gte.h"
 
-#	include "duffle/gen/duffle.macs.h"
-#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
 #include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
 #include "duffle/word_count.metadata.h"
 
-#	include "gen/gte_hello.offsets.h"
+#	include "gen/offsets.h"
 #include "hello_gte.h"
 
-#include "hello_gte_tape.c"
+#include "hello_gte.tape.c"
 
 typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
 typedef Array_(OrderingTable_Buffer, 2);
@@ -122,7 +122,7 @@ global SMemory smem;
 extern SMemory smem;
 
 // TODO(Ed):
-FI_ U4* spad_warm(MipsAtom atom) {
+FI_ U4* spad_warm(Slice_MipsCode atom) {
 	return nullptr;
 }
 

code/hello_gte/hello_gte.tape.c

@@ -0,0 +1,218 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+#	include "duffle/atom_dsl.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/word_count.metadata.h"
+#	include "gen/offsets.h"
+#	include "hello_gte.h"
+#endif
+
+#pragma region MACs (Mips Atom components)
+
+
+
+#pragma endregion MACs
+
+#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()
+};
+
+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)
+){
+	/* 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_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,
+		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(),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(),
+
+		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(
+				/* 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)
+){
+	/* 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_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(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(),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* 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,
+			mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(),                 /* 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_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

@@ -0,0 +1,76 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.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
+#pragma once
+
+#pragma region hello_joypad
+
+
+// --- 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,
+};
+
+// --- 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,
+};
+
+// --- 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,
+};
+
+#pragma endregion hello_joypad
+

code/hello_joypad/hello_joypad.atom.c

@@ -0,0 +1,638 @@
+#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_joypad.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(),
+};
+
+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()
+};
+
+/* ----- 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(),
+};
+
+/* ----- 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(),
+};
+
+#pragma endregion Baked Atoms

code/hello_joypad/hello_joypad.c

@@ -0,0 +1,441 @@
+#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/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Joypade Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
+#include "hello_joypad.h"
+#pragma region Joypad Headers
+
+#pragma region Hello Joypad TUs
+#include "hello_joypad.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 (0) // Pad Input (dead — kept for the source-as-written record; references the deleted `pad_state` field)
+	{
+		(void)Pad_Left; (void)Pad_Right; /* suppress unused-token warnings */
+		if (false) {
+			smem.cube.rot.y  += 30;
+			smem.floor.rot.y += 5;
+		}
+		if (false) {
+			smem.cube.rot.y  -= 30;
+			smem.floor.rot.y -= 5;
+		}
+	}
+	if (1) // Pad Input (Tape version)
+	{
+		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 (0)
+	{
+		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);
+		for (U4 face_id = 0; face_id < Cube_num_faces; face_id += 1)
+		{
+			Poly_G4* quad = prim_alloc(Poly_G4); set_poly_g4(quad);
+			quad->c0 = rgb8(255,   0, 255);
+			quad->c1 = rgb8(255, 255,   0);
+			quad->c2 = rgb8(  0, 255, 255);
+			quad->c3 = rgb8(  0, 255,   0);
+
+			V4_S2* face = & smem.cube.faces[face_id];
+			V3_S2* p0   = & smem.cube.verts[face->x];
+			V3_S2* p1   = & smem.cube.verts[face->y];
+			V3_S2* p2   = & smem.cube.verts[face->z];
+			V3_S2* p3   = & smem.cube.verts[face->w];
+
+			nclip = rtp_avg_nclip_a4_v3s2(
+				p0, p1, p2, p3,
+				& quad->p0, & quad->p1, & quad->p2, & quad->p3,
+				& p, & orderingtbl_z, & flag
+			);
+			if (nclip <= 0) {
+				continue;
+			}
+
+			if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+				orderingtbl_add_primitive(ordering_buf[orderingtbl_z], quad);
+			}
+		}
+		// smem.cube.rot.x +=  6;
+		// smem.cube.rot.y +=  8;
+		// smem.cube.rot.z += 12;
+		smem.cube.rot.y += 30;
+	}
+	// Draw cube (tape method) - two triangles per face
+	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 (0)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+		for (U4 face_id = 0; face_id < Floor_num_faces; face_id += 1)
+		{
+			Poly_F3* tri = prim_alloc(Poly_F3); set_poly_f3(tri);
+			tri->color   = rgb8(255, 255, 255);
+
+			V3_S2* face = & smem.floor.faces[face_id];
+			register V3_S2* p0 rgcc(R_T4) = & smem.floor.verts[face->x];
+			register V3_S2* p1 rgcc(R_T5) = & smem.floor.verts[face->y];
+			register V3_S2* p2 rgcc(R_T6) = & smem.floor.verts[face->z];
+
+			gte_load_v0(p0, R_T4);
+			/*
+			asm volatile( ".word " "%0" ", %1" : : 
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_xy & REG_MASK) << RT_SHIFT) | (0            & IMM_MASK)),
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_z  & REG_MASK) << RT_SHIFT) | (GTE_Z_Offset & IMM_MASK)), 
+				"r"(p0) : 
+				"$2", "$8", "$9", "$31", "memory" 
+			);
+			*/
+			gte_load_v1(p1, R_T5);
+			gte_load_v2(p2, R_T6);
+
+			gte_rtpt();
+			gte_nclip();
+			gte_stotz(& nclip);
+
+			// nclip = rtp_avg_nclip_a3_v3s2(p0, p1, p2
+			// 	, & tri->p0, & tri->p1, & tri->p2
+			// 	, & p, & orderingtbl_z, & flag
+			// );
+			// if (nclip <= 0) {
+			// 	continue;
+			// }
+
+			if (nclip > 0 ) {
+				gte_stsxy3(& tri->p0, & tri->p1, & tri->p2);
+				gte_avsz3();
+				gte_stotz(& orderingtbl_z);
+
+				if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+					orderingtbl_add_primitive(ordering_buf[orderingtbl_z], tri);
+				}
+			}
+		}
+		smem.floor.rot.y += 5;
+	}
+	// Draw floor tape method
+	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_joypad/hello_joypad.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_joypad/sio.tape.c

@@ -0,0 +1,659 @@
+
+#if 0 /* ac_pad_sio_write_pad_state — superseded by pad_bios_snapshot */
+
+/* ============================================================
+ * raw_sio_pad_poll_20260802 — superseded by bios_pad_buffer_snapshot_20260803.
+ * The doomed raw-SIO production atoms (ac_pad_sio_write_pad_state,
+ * pad_sio_init, pad_sio_step, pad_sio_diag_pin, pad_sio_diag_byte_exchange)
+ * reference symbols that were removed from code/duffle/pad.h during
+ * Phase 1. Each is wrapped in a narrow `#if 0` so the C compile skips
+ * the body while the source-as-written text stays in place for the
+ * Phase 5.1 deletion pass. The wrap is removed (and the bodies are
+ * deleted) by Phase 5.1 of this track.
+ * ============================================================ */
+
+ * Writes the per-port PadState in 5 instructions plus 4 store_word calls (status,
+ * buttons, left_x/y/right_x/right_y packed, attempt). The provisional decode publishes
+ * 0x0000FFFF buttons + centered axes on every path until response-byte decode lands.
+ *
+ * Args:
+ *   status_val    - the PadSioStatus enum value to publish
+ *   state_ptr_reg - the PadState* base (R_PadState at the call site)
+ *   scratch_reg   - scratch register for the value being stored (e.g., R_T0)
+ *
+ * Emits 9 instructions (status/buttons/axes/attempt stores plus the
+ * two-instruction zero-extended buttons load).
+ */
+FI_ Slice_MipsCode ac_pad_sio_write_pad_state(U4 status_val, U4 state_ptr_reg, U4 scratch_reg)
+MipsAtomComp_Proc_(ac_pad_sio_write_pad_state, {
+	add_ui(scratch_reg, R_0, status_val),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,status)),
+	/* FIX 2026-08-02: buttons = 0x0000FFFF = "no buttons pressed" in
+	 * libetc convention. Build it with LUI + ORI so addiu does not
+	 * sign-extend 0xFFFF to 0xFFFFFFFF. */
+	load_upper_i(scratch_reg, 0x0000),
+	or_i(scratch_reg, scratch_reg, 0xFFFF),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,buttons)),
+	add_ui(scratch_reg, R_0, 0x80808080),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,left_x)),
+	add_ui(scratch_reg, R_0, 0),
+	store_word(scratch_reg, state_ptr_reg, O_(PadState,attempt))
+})
+#endif /* end ac_pad_sio_write_pad_state wrap */
+
+/* ----- pad_sio_init -----
+ * Boot-time SIO0 init. Caller pins R_T6 = sio_base_addr0.
+ * Issues SIO CTRL=0x0040 (reset), MODE=0x000D, BAUD=0x0088.
+ * (Phase 2 fills the body.)
+ */
+#if 0 /* pad_sio_init — superseded by pad_bios_init_start (Phase 1.3) */
+internal MipsAtom_(pad_sio_init) atom_info(atom_phase(pad_init)
+, atom_reads(R_T5, R_T6)
+, atom_writes(R_T5, R_T6)
+) {
+	/* FIX 2026-08-02: explicitly load the KSEG1 base into R_T6 at the top of
+	 * the atom body. The rgcc(R_PadSioBase) binding in main() pins R_T6 = base
+	 * when main() runs, but $12 is caller-saved per the O32 ABI — when tape_run
+	 * is invoked, R_T6 is fair game. The atom body cannot rely on the value. */
+	load_upper_i(R_T6, pad_IO_KSEG1_BASE >> 16),   /* R_T6 high 16 = 0xBF80 */
+	or_i(R_T6, R_T6, pad_IO_KSEG1_BASE & 0xFFFF),   /* R_T6 = 0xBF800000 */
+
+	/* SIO CTRL = 0x0040 (reset) */
+	add_ui(R_T5, R_0, pad_SIO_CTRL_RESET),
+	store_half(R_T5, R_T6, pad_SIO_CTRL_OFFSET),
+	/* SIO MODE = 0x000D (MUL1, 8-bit, no parity, idle-high) */
+	add_ui(R_T5, R_0, pad_SIO_MODE_INIT),
+	store_half(R_T5, R_T6, pad_SIO_MODE_OFFSET),
+	/* SIO BAUD = 0x0088 (~250 kHz) */
+	add_ui(R_T5, R_0, pad_SIO_BAUD_INIT),
+	store_half(R_T5, R_T6, pad_SIO_BAUD_OFFSET),
+	mac_yield(),
+};
+#endif /* end pad_sio_init wrap */
+
+/* ----- pad_sio_step -----
+ * Per-frame bounded raw-SIO transaction. Reads PadState pointers + SIO
+ * base addresses from Binds_PadSioStep; writes per-port status +
+ * buttons + axes into smem.pad[0..1].
+ * Body shape (per spec §"Transaction model (per port, per pad_sio_step)"):
+ *   port 0: CTRL=CLEANUP → settle → CTRL=port-select → settle → exchange 5
+ *           bytes (addr + 0x42 0x00 0x00 0x00) → decode → write PadState[0]
+ *           → CTRL=CLEANUP.
+ *   port 1: swap scratch regs (sio_base_addr1 → R_PadSioBase, state1 →
+ *           R_PadState) → mirror port 0 sequence.
+ *
+ * Bounded-loop semantics: every countdown is wrapped in
+ *   add_ui_self(R_T1, -1) + branch_ne(R_T1, R_0, ...)
+ * with a known maximum (pad_SIO_SETTLE_BEFORE_TX=1000, pad_SIO_SETTLE_AFTER_TX=2000,
+ * pad_SIO_WAIT_BUDGET=4096). The static-analysis pass currently reports
+ * has_loops = true; the follow-up metaprogram track that learns modeled-bounded
+ * loops is out of scope here (per spec §"Risks").
+ *
+ * Scratch register strategy:
+ *   R_PadStatus    = R_T4 — RESERVED for port-1 swap (holds state1)
+ *   R_PadCountdown = R_T5 — RESERVED for port-1 swap (holds sio_base_addr1)
+ *   R_T0           — byte-exchange value + STAT read (clobbered freely)
+ *   R_T1           — countdown budget (clobbered freely)
+ *   R_PadState     = R_T7 — PadState* (preserved for PadState writes)
+ *   R_PadSioBase   = R_T6 — SIO base (preserved through the port)
+ *
+ * Response decode (Task 3.1 teaching scope):
+ *   - status   = PadSioStatus_Digital (hardcoded)
+ *   - buttons  = 0xFFFF (no buttons pressed in the provisional libetc
+ *                         convention; full response-byte decode is follow-up)
+ *   - axes     = 0x80808080 (centered: left_x=0x80, left_y=0x80,
+ *                             right_x=0x80, right_y=0x80)
+ *   - attempt  = 0
+ *   - DualShock handshake (0x43 0x01 → 0x44 0x01 0x03 → 0x43 0x00) is
+ *     follow-up scope; the hardcoded digital decode is a placeholder.
+ *
+ * Both ports raise /CS (CTRL = pad_SIO_CTRL_CLEANUP) before exit. Both ports
+ * treat response timeout as PadSioStatus_Disconnected per the spec §"Failure
+ * handling" + the canonical per-port timeout semantics.
+ */
+#if 0 /* pad_sio_step — superseded by pad_bios_snapshot (Phase 2.1) */
+internal MipsAtom_(pad_sio_step) atom_info(atom_bind(Binds_PadSioStep)
+, atom_reads(R_TapePtr, R_PadSioBase, R_PadState, R_PadStatus, R_PadCountdown)
+, atom_writes(R_PadStatus, R_PadCountdown)
+) {
+	/* FIX 2026-08-02: explicitly load KSEG1 base into R_PadSioBase (R_T6) at the
+	 * top. The rgcc() binding in main() does NOT survive the tape_run call
+	 * because R_T6 is caller-saved per the O32 ABI. The pad_sio_init atom
+	 * (also in the per-frame tape) reloads R_T6 separately. */
+	load_upper_i(R_PadSioBase, pad_IO_KSEG1_BASE >> 16),
+	or_i(R_PadSioBase, R_PadSioBase, pad_IO_KSEG1_BASE & 0xFFFF),
+
+	/* Pop Binds from tape (in Binds_PadSioStep declaration order) */
+	load_word(R_PadState,     R_TapePtr, O_(Binds_PadSioStep,state0)),
+	load_word(R_PadStatus,    R_TapePtr, O_(Binds_PadSioStep,state1)),    /* reserved for port-1 swap */
+	load_word(R_PadSioBase,   R_TapePtr, O_(Binds_PadSioStep,sio_base_addr0)),
+	load_word(R_PadCountdown, R_TapePtr, O_(Binds_PadSioStep,sio_base_addr1)),  /* reserved for port-1 swap */
+	add_ui_self(R_TapePtr, S_(Binds_PadSioStep)),
+
+	/* ============== PORT 0 TRANSACTION ============== */
+	/* Use R_T0 (byte value / STAT read) + R_T1 (countdown) as scratch.
+	 * R_PadStatus (state1) + R_PadCountdown (sio_base_addr1) are preserved
+	 * through the port-0 body and swapped into R_PadSioBase + R_PadState
+	 * at atom_offset(port1_start, ...) below. */
+
+	/* 1. Cleanup: CTRL = 0x0010 (raise /CS, clear stale status) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_BEFORE_TX = 1000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_BEFORE_TX),
+atom_label(settle_pre_port0)
+	nop,  /* BD slot */
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_pre_port0, settle_pre_port0)),
+
+	/* 2. Port-select: CTRL = 0x0003 (TX enable + DTR /CS) for port 0 */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),  /* set /CS line low */
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_AFTER_TX = 2000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_AFTER_TX),
+atom_label(settle_post_port0)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_post_port0, settle_post_port0)),
+
+	/* 3. Address byte (0x01) — send + RX-ready wait + read response + RX-drain confirmation */
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack0_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack0_port0, ack0_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack0_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack0_port0, wait_ack0_port0)),
+	/* RX timeout → mark disconnected; skip to port 1 */
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack0)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack0, port1_start)),
+
+atom_label(ack0_received_port0)
+	/* Read open-bus response byte 0 — discard per docs/psx-spx §controllersandmemorycards.md */
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Confirm RX FIFO drained before sending byte 1. Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel0_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel0_port0, ack_released_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel0_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel0_port0, wait_ackrel0_port0)),
+	/* RX-drain timeout → disconnected; skip to port 1 */
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel0)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel0, port1_start)),
+
+atom_label(ack_released_port0)
+
+	/* === Byte 1 (port 0): send 0x42 (cmd read) + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack1_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack1_port0, ack1_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack1_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack1_port0, wait_ack1_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack1)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack1, port1_start)),
+
+atom_label(ack1_received_port0)
+	/* Read response ID byte — discarded for teaching scope (decode hardcoded). */
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* RX FIFO drain wait. Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel1_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel1_port0, ack_released1_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel1_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel1_port0, wait_ackrel1_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel1)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel1, port1_start)),
+
+atom_label(ack_released1_port0)
+
+	/* === Byte 2 (port 0): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack2_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack2_port0, ack2_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack2_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack2_port0, wait_ack2_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack2)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack2, port1_start)),
+
+atom_label(ack2_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel2_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel2_port0, ack_released2_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel2_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel2_port0, wait_ackrel2_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel2)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel2, port1_start)),
+
+atom_label(ack_released2_port0)
+
+	/* === Byte 3 (port 0): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack3_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack3_port0, ack3_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack3_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack3_port0, wait_ack3_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ack3)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ack3, port1_start)),
+
+atom_label(ack3_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel3_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel3_port0, ack_released3_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel3_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel3_port0, wait_ackrel3_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_ackrel3)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_ackrel3, port1_start)),
+
+atom_label(ack_released3_port0)
+
+	/* === Byte 4 (FINAL, port 0): send 0x00 + RX-not-empty wait + read final byte === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_rx4_port0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_rx4_port0, rx4_received_port0)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_rx4_port0)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_rx4_port0, wait_rx4_port0)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port0_from_rx4)
+	branch_equal(R_0, R_0, atom_offset(skip_port0_from_rx4, port1_start)),
+
+atom_label(rx4_received_port0)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),  /* discard final byte */
+
+	/* === RESPONSE DECODE (hardcoded for teaching scope) ===
+	 * Per the plan §"Phase 3 task 3.1" + spec §"Architecture":
+	 *   - Full decode (buttons/axes from response bytes) is follow-up scope.
+	 *   - Teaching scope: hardcode digital poll response.
+	 *     status = PadSioStatus_Digital
+	 *     buttons = 0x0000FFFF (no buttons pressed — placeholder)
+	 *     axes = 0x80808080 (left_x=0x80, left_y=0x80, right_x=0x80, right_y=0x80)
+	 *     attempt = 0
+	 */
+atom_label(decode_port0)
+	mac_pad_sio_write_pad_state(PadSioStatus_Digital, R_PadState, R_T0),
+
+	/* /CS cleanup: raise /CS, clear stale status before exiting port 0. */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+	/* ============== PORT 1 SETUP ============== */
+	/* Swap: R_PadCountdown holds sio_base_addr1; R_PadStatus holds state1. */
+atom_label(port1_start)
+	add_u(R_PadSioBase, R_0, R_PadCountdown),  /* sio_base_addr1 → R_PadSioBase */
+	add_u(R_PadState,   R_0, R_PadStatus),     /* state1 → R_PadState */
+
+	/* ============== PORT 1 TRANSACTION (mirror of port 0) ============== */
+	/* R_PadStatus + R_PadCountdown are no longer reserved (port 1 is the
+	 * last transaction); we still use R_T0/R_T1 as scratch to match port 0. */
+
+	/* 1. Cleanup: CTRL = 0x0010 (raise /CS, clear stale status) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_BEFORE_TX = 1000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_BEFORE_TX),
+atom_label(settle_pre_port1)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_pre_port1, settle_pre_port1)),
+
+	/* 2. Port-select: CTRL = 0x0003 | (1 << 13) (port 1 select) */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),
+	or_i(R_T0, R_T0, 1 << 13),  /* port 1 select bit (CTRL bit 13 = port select) */
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	/* Bounded by pad_SIO_SETTLE_AFTER_TX = 2000 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_SETTLE_AFTER_TX),
+atom_label(settle_post_port1)
+	nop,
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(settle_post_port1, settle_post_port1)),
+
+	/* 3. Address byte (0x01) — send + RX-ready wait + read response + RX-drain confirmation */
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack0_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack0_port1, ack0_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack0_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack0_port1, wait_ack0_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack0)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack0, end_atom)),
+
+atom_label(ack0_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel0_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel0_port1, ack_released_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel0_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel0_port1, wait_ackrel0_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel0)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel0, end_atom)),
+
+atom_label(ack_released_port1)
+
+	/* === Byte 1 (port 1): send 0x42 (cmd read) + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack1_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack1_port1, ack1_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack1_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack1_port1, wait_ack1_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack1)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack1, end_atom)),
+
+atom_label(ack1_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel1_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel1_port1, ack_released1_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel1_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel1_port1, wait_ackrel1_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel1)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel1, end_atom)),
+
+atom_label(ack_released1_port1)
+
+	/* === Byte 2 (port 1): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack2_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack2_port1, ack2_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack2_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack2_port1, wait_ack2_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack2)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack2, end_atom)),
+
+atom_label(ack2_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel2_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel2_port1, ack_released2_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel2_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel2_port1, wait_ackrel2_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel2)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel2, end_atom)),
+
+atom_label(ack_released2_port1)
+
+	/* === Byte 3 (port 1): send 0x00 + RX-ready wait + read response + RX-drain confirmation === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ack3_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_ack3_port1, ack3_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ack3_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ack3_port1, wait_ack3_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ack3)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ack3, end_atom)),
+
+atom_label(ack3_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_ackrel3_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_equal(R_T0, R_0, atom_offset(wait_ackrel3_port1, ack_released3_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_ackrel3_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_ackrel3_port1, wait_ackrel3_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_ackrel3)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_ackrel3, end_atom)),
+
+atom_label(ack_released3_port1)
+
+	/* === Byte 4 (FINAL, port 1): send 0x00 + RX-not-empty wait + read final byte === */
+	/* Bounded by pad_SIO_WAIT_BUDGET = 4096 iterations. */
+	add_ui(R_T0, R_0, 0x00),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(wait_rx4_port1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(wait_rx4_port1, rx4_received_port1)),
+	add_ui_self(R_T1, -1),
+atom_label(continue_wait_rx4_port1)
+	branch_ne(R_T1, R_0, atom_offset(continue_wait_rx4_port1, wait_rx4_port1)),
+	mac_pad_sio_write_pad_state(PadSioStatus_Disconnected, R_PadState, R_T0),
+atom_label(skip_port1_from_rx4)
+	branch_equal(R_0, R_0, atom_offset(skip_port1_from_rx4, end_atom)),
+
+atom_label(rx4_received_port1)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),  /* discard final byte */
+
+	/* === RESPONSE DECODE (port 1) === */
+atom_label(decode_port1)
+	mac_pad_sio_write_pad_state(PadSioStatus_Digital, R_PadState, R_T0),
+
+	/* /CS cleanup: raise /CS, clear stale status before exiting port 1. */
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+atom_label(end_atom)
+	mac_yield(),
+};
+#endif /* end pad_sio_step wrap */
+
+/* ----- pad_sio_diag_pin -----
+ * Per-frame diagnostic counter. The caller binds R_DiagPinScratch to
+ * scratch_for_atom_diag_pin for temporary gdb verification.
+ */
+#if 0 /* pad_sio_diag_pin — superseded (raw-SIO phase removed) */
+internal MipsAtom_(pad_sio_diag_pin) atom_info(atom_phase(pad_init)
+, atom_reads(R_T0, R_T1, R_DiagPinScratch)
+, atom_writes(R_T0, R_T1, R_DiagPinScratch)
+) {
+	/* FIX 2026-08-02: explicitly reload R_DiagPinScratch (R_T3 = $t3). Caller-saved
+	 * per O32 ABI; the rgcc binding in main() does not survive tape_run. */
+	load_upper_i(R_DiagPinScratch, 0x8001),
+	or_i(R_DiagPinScratch, R_DiagPinScratch, 0xC800),
+
+	/* High half = 0xD1A6; low half increments once per atom invocation. */
+	load_word(R_T1, R_DiagPinScratch, 0),
+	nop,
+	add_ui(R_T1, R_T1, 1),
+	and_i(R_T0, R_T1, 0xFFFF),
+	load_upper_i(R_T1, 0xD1A6),
+	or_i(R_T1, R_T1, 0),
+	or_u(R_T1, R_T1, R_T0),
+	store_word(R_T1, R_DiagPinScratch, 0),
+	mac_yield(),
+};
+#endif /* end pad_sio_diag_pin wrap */
+
+/* ----- pad_sio_diag_byte_exchange -----
+ * Temporary two-byte wire probe: sends 0x01 and 0x42, then stores the
+ * open-bus byte and response ID in scratch_for_atom_diag_pin.
+ */
+#if 0 /* pad_sio_diag_byte_exchange — superseded (raw-SIO phase removed) */
+internal MipsAtom_(pad_sio_diag_byte_exchange) atom_info(atom_phase(pad_init)
+, atom_reads(R_T0, R_T1, R_T2, R_PadSioBase, R_DiagPinScratch)
+, atom_writes(R_T0, R_T1, R_T2, R_PadSioBase, R_DiagPinScratch)
+) {
+	/* FIX 2026-08-02: explicitly reload R_DiagPinScratch (R_T3 = $t3). Caller-saved
+	 * per O32 ABI; the rgcc binding in main() does not survive tape_run. */
+	load_upper_i(R_DiagPinScratch, 0x8001),
+	or_i(R_DiagPinScratch, R_DiagPinScratch, 0xC800),
+
+	/* FIX 2026-08-02: explicitly load KSEG1 base into R_PadSioBase (R_T6) at the
+	 * top. The rgcc() binding in main() does NOT survive the tape_run call
+	 * because R_T6 is caller-saved per the O32 ABI. */
+	load_upper_i(R_PadSioBase, pad_IO_KSEG1_BASE >> 16),
+	or_i(R_PadSioBase, R_PadSioBase, pad_IO_KSEG1_BASE & 0xFFFF),
+
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	add_ui(R_T0, R_0, pad_SIO_CTRL_TX_ENABLE),
+	or_i(R_T0, R_T0, pad_SIO_CTRL_DTR_CS),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+
+	add_ui(R_T0, R_0, pad_PROTO_ADDR),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(diag_wait_ack0)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(diag_wait_ack0, diag_ack0_done)),
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(diag_wait_ack0, diag_wait_ack0)),
+	add_ui(R_T0, R_0, 0xDEADAC01),
+	store_word(R_T0, R_DiagPinScratch, 0),
+	branch_equal(R_0, R_0, atom_offset(diag_timeout_ack0, diag_timeout)),
+atom_label(diag_ack0_done)
+	load_byte_u(R_T2, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T0, R_0, pad_PROTO_CMD_READ),
+	store_byte(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	add_ui(R_T1, R_0, pad_SIO_WAIT_BUDGET),
+atom_label(diag_wait_ack1)
+	load_half_u(R_T0, R_PadSioBase, pad_SIO_STAT_OFFSET),
+	nop,
+	and_i(R_T0, R_T0, pad_SIO_STAT_RX_NOT_EMPTY),
+	branch_ne(R_T0, R_0, atom_offset(diag_wait_ack1, diag_ack1_done)),
+	add_ui_self(R_T1, -1),
+	branch_ne(R_T1, R_0, atom_offset(diag_wait_ack1, diag_wait_ack1)),
+	add_ui(R_T0, R_0, 0xDEADAC02),
+	store_word(R_T0, R_DiagPinScratch, 0),
+	branch_equal(R_0, R_0, atom_offset(diag_timeout_ack1, diag_timeout)),
+atom_label(diag_ack1_done)
+	load_byte_u(R_T0, R_PadSioBase, pad_SIO_DATA_OFFSET),
+	nop,
+	shift_lleft(R_T0, R_T0, 8),
+	or_u(R_T2, R_T2, R_T0),
+	store_word(R_T2, R_DiagPinScratch, 0),
+atom_label(diag_success)
+	branch_equal(R_0, R_0, atom_offset(diag_success, diag_done)),
+	nop,
+atom_label(diag_timeout_ack0)
+	add_ui(R_T0, R_0, 0xDEADAC01),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_timeout_ack1)
+	add_ui(R_T0, R_0, 0xDEADAC02),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_timeout)
+	add_ui(R_T0, R_0, 0xDEADACFF),
+	store_word(R_T0, R_DiagPinScratch, 0),
+atom_label(diag_done)
+	add_ui(R_T0, R_0, pad_SIO_CTRL_CLEANUP),
+	store_half(R_T0, R_PadSioBase, pad_SIO_CTRL_OFFSET),
+	mac_yield(),
+};
+#endif /* end pad_sio_diag_byte_exchange wrap */

readme.md

@@ -7,12 +7,12 @@ A rest from the usual.
 ## Dependencies
 
 I will be programming from a Windows 11 machine (may eventually try this on the Steam Deck...):
-![system_info](./docs/assets/system_info.png)
 
 [armips](https://github.com/Kingcom/armips)
 
 * Supports doing bare-metal assembly for the ps1
 * `scoop install armips` or just clone and build..
+* Was used early in the course. Now I just use an macro asm dsl in C11.
 
 [luajit-2.1](https://github.com/LuaJIT/LuaJIT.git)
 
@@ -73,3 +73,9 @@ scoop install luajit
 ![hello_psyq!](./docs/assets/pcsx-redux_2025-08-05_23-01-19.png)
 ![cube!](./docs/assets/pcsx-redux_2025-10-11_03-04-01.png)
 ![cube and floor!](./docs/assets/pcsx-redux_2026-07-10_22-47-02.png)
+
+Win 11 machine:
+
+![system_info](./docs/assets/system_info.png)
+
+Still haven't gotten around to trying this on linux...

scripts/build_psyq.ps1

@@ -180,29 +180,15 @@ 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)
+	# 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math, mcrd, mcx, press, sio, snd, spu, tap)
+	# had LOAD lines in the map but ZERO .o files pulled in — they were unused.
+	# 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.).
 	$libraries = @(
 		"api",
 		"c",
-		"c2",
-		"card",
-		"cd",
-		"comb",
-		"ds",
 		"etc",
 		"gpu",
-		"gs",
-		"gte",
-		"gun",
-		"hmd",
-		"math",
-		"mcrd",
-		"mcx",
-		"pad",
-		"press",
-		"sio",
-		"snd",
-		"spu",
-		"tap"
+		"gte"
 	)
 	foreach ($lib in $libraries) {
 		$link_args += ($f_link_lib + $lib)
@@ -238,9 +224,7 @@ function ps1-meta { param(
 		[string[]]$passes = @('--pre-link'),
 		[string[]]$extra_args = @()
 	)
-	# `--unity-root` and `--source` are
-	# mutually exclusive. Exactly one of `$unity_root` / `$sources` must
-	# be supplied; the other must be absent.
+	# `--unity-root` and `--source` are mutually exclusive. Exactly one of `$unity_root` / `$sources` must be supplied; the other must be absent.
 	if ($null -ne $unity_root -and $unity_root -ne '') 
 	{
 		if ($null -ne $sources -and $sources.Count -gt 0) {
@@ -276,6 +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 += @()
 
@@ -350,10 +401,10 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function build-gte_hello {
+function build-hello_gte {
 	$includes += @()
 
-	$path_module        = join-path $path_code   'gte_hello'
+	$path_module        = join-path $path_code   'hello_gte'
 	$path_duffle        = join-path $path_code   'duffle'
 	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
 	$path_build_gen     = join-path $path_build   'gen'
@@ -402,64 +453,110 @@ function build-gte_hello {
 	# 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'
-	$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"
+	inject-dwarf $elf $path_build_gen
+}
+# build-hello_gte
+
+function build-hello_joypad {
+	$includes += @()
+
+	$path_module        = join-path $path_code   'hello_joypad'
+	$path_duffle        = join-path $path_code   'duffle'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build  'gen'
+
+	$src_c = join-path $path_module 'hello_joypad.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
+
+	$assemble_args = @()
+	$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_joypad_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_joypad.elf'
+	$exe = join-path $path_build 'hello_joypad.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
 	)
-		& $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;
-		}
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
 
-		$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"
+	# 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)
+
+	inject-dwarf $elf $path_build_gen
+}
+# build-hello_joypad
+
+function build-hello_camera {
+	$includes += @()
+
+	$path_module        = join-path $path_code   'hello_camera'
+	$path_duffle        = join-path $path_code   'duffle'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build  'gen'
+
+	$src_c = join-path $path_module 'hello_camera.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--pre-link')
+
+	$assemble_args = @()
+	$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
 	)
-		& $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;
-		}
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
 
-		# 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-gte_hello
+	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
+}
+build-hello_camera
 
 # 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.
----
---- One ownership statement, then the rest is signal:
----   * **Character classification** (`is_space`, `is_alpha`, `is_alnum`, `is_digit`, plus the byte-fast `_byte` variants).
----   * **String / path primitives** (`trim`, `dirname`, `basename_no_ext`, `normalize_path`, `canonical_path_key`, `find_byte`).
----   * **I/O primitives** (`read_file`, `write_file`, `ensure_dir`).
----   * **Corpus resolution** (`parse_direct_quoted_includes`, `resolve_source_corpus`).
----   * **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`).
----   * **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.
+---   * Character classification: `is_space`, `is_alpha`, `is_alnum`, `is_digit`, plus the byte-fast `_byte` variants.
+---   * String / path primitives: `trim`, `dirname`, `basename_no_ext`, `normalize_path`, `canonical_path_key`, `find_byte`.
+---   * I/O primitives:           `read_file`, `write_file`, `ensure_dir`.
+---   * Corpus resolution:        `parse_direct_quoted_includes`, `resolve_source_corpus`.
+---   * 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`.
+---   * 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`, `INSTRUCTION_LATENCY`.
 
 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 text     string    -- the full source text
---- @field dir      string    -- the directory containing the source
---- @field basename string    -- filename without extension
+--- @field path     Path      -- Absolute path to the source file
+--- @field text     string    -- Full source text
+--- @field dir      string    -- Directory containing the source
+--- @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('"')
--- Char literal: '...' with backslash escapes.
-local lpeg_chr_pat = P("'") * (P(1) - S("'\\") + P('\\') * P(1))^0 * P("'")
--- Line comment: // ... to end-of-line.
-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
+local lpeg_str_pat        = P('"') * (P(1) - S('"\\') + P('\\') * P(1))^0 * P('"') -- String literal: "..." with backslash escapes.
+local lpeg_chr_pat        = P("'") * (P(1) - S("'\\") + P('\\') * P(1))^0 * P("'") -- Char literal: '...' with backslash escapes.
+local lpeg_line_cmt_pat   = P("//") * (P(1) - S("\n"))^0                           -- Line comment: // ... to end-of-line.
+local lpeg_block_cmt_pat  = P("/*") * (P(1) - P("*/"))^0 * P("*/")                 -- Block comment: /* ... */ (no nesting per C standard).
+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).
 
 -- 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.
--- We accept any single char except the target.
+-- Used by scan_to_char for non-ident / non-bracket chars. 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;
--- the byte versions are what the hot loops should call).
+-- String-based wrappers (kept for callers that already have a single-char string; 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.
@@ -920,18 +897,25 @@ 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, 
-			-- so we stop the scan when we hit any of them.
+			-- These also appear as separators between argument lists inside the parens/braces/brackets, 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
+			-- Line-comment '// ... \n' (0x2F 0x2F): skip to (and past) the next newline, or to end-of-body.
+			if     c == BYTE_SLASH and body:byte(scan + 1) == BYTE_SLASH then
+				local nl = M.find_byte(body, BYTE_NEWLINE, scan)
+				scan = nl and (nl + 1) or (len + 1)
+			-- Block-comment '/* ... */' (0x2F 0x2A): skip to (and past) the matching '*/', or to end-of-body.
+			elseif c == BYTE_SLASH and body:byte(scan + 1) == BYTE_STAR then
+				local close = body:find("*/", scan + 2, true)
+				scan = close and (close + 2) or (len + 1)
 			-- Group opener bytes (consume the balanced group via the matching reader): '(' = 0x28, '{' = 0x7B, '[' = 0x5B.
-			if     c == BYTE_OPEN_PAREN then local _, a = M.read_parens   (body, scan); scan = a
+			elseif c == BYTE_OPEN_PAREN then local _, a = M.read_parens   (body, scan); scan = a
 			elseif c == BYTE_OPEN_BRACE then local _, a = M.read_braces   (body, scan); scan = a
 			elseif c == BYTE_OPEN_BRACK then local _, a = M.read_brackets (body, scan); scan = a
 			-- String-literal byte ('"' = 0x22 or '\'' = 0x27): skip past the quoted region in one shot.
 			elseif c == BYTE_DQUOTE or c == BYTE_SQUOTE then
-				scan = M.skip_str_or_cmt(body, scan) + 1
+				scan = (M.skip_str_or_cmt(body, scan) or scan) + 1
 			else
 				scan = scan + 1
 			end
@@ -1080,8 +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
--- phenomenon modeled from the command side), the values are:
+-- 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), 
+-- 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)
@@ -1091,7 +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 = {
@@ -1155,7 +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).
@@ -1180,10 +1163,9 @@ M.GTE_COMMAND_INPUTS = {
 --   * "mac_result"       : generic MAC output (nclip, op, mvmva)
 --
 -- Consumers:
---   * passes/static_analysis.lua::analyze_hardware_relations (the walker reads this after a GTE command to update
---     `forward_state.post_command_roles` for `gte_result_position`).
---   * passes/static_analysis.lua::check_gte_result_position (per-atom CHECK_RULES reader; renders role mismatches).
--- This table is consumed by the hardware-relation analyzer and result-position check.
+--   * passes/static_analysis.lua::analyze_hardware_relations (the walker reads this after a GTE command to update `forward_state.post_command_roles` for `gte_role_mismatch`).
+--   * passes/static_analysis.lua::check_gte_role_mismatch (per-atom CHECK_RULES reader; renders role mismatches).
+-- This table is consumed by the hardware-relation analyzer and the gte_role_mismatch check.
 M.GTE_COMMAND_OUTPUTS = {
 	-- RTPS: writes one screen coordinate (the perspective-divide result) into C2_SXY2.
 	-- The FIFO side effects leave SXY0 / SXY1 untouched, so `latest_screen_xy` is C2_SXY2.
@@ -1235,16 +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
--- producer step of `analyze_hardware_relations`.
+-- It is the inverse of the MTC2 -> command input propagation (register -> command direction), which is staged by the producer step of `analyze_hardware_relations`.
 --
--- The schema mirrors the producer-side relations (`direction`, `evidence`, `violation_kind`); `required` counts the
--- emitted words strictly between the command's last output word and the overwrite.
+-- The schema mirrors the producer-side relations (`direction`, `evidence`, `violation_kind`); 
+-- `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
--- command's outputs latch into the pipeline before a later MTC2/CTC2 overwrites them.
+-- They describe when a recent MTC2 / CTC2 must retire before the command issues. 
+-- 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).
@@ -1259,9 +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
-	-- newest one; the earlier SXY0 / SXY1 entries are part of the
-	-- batched triple).
+	-- 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).
 	["gte_cmdw_rtpt"] = {
 		{ register = "C2_SXY0", required = 4 },
 		{ register = "C2_SXY1", required = 4 },
@@ -1293,35 +1272,7 @@ M.GTE_COMMAND_LATCH_WINDOWS = {
 	},
 }
 
--- GTE component result contracts (immutable; keyed by bare component name).
---
--- Register-role claims that the `_post_<cmd>` suffix alone cannot infer live here.
--- The bare name (the component name stripped of the `_post_<cmd>` suffix) is the key; the row carries the expected
--- command, the expected role, and the expected C2 register.
---
--- Known rows:
---   * `gte_store_g4_p3_post_rtps`: post-RTPS polygon-emit slot reads the newest projected screen coordinate from C2_SXY2.
---     C2_SXY0 is wrong (C2_SXY0 is an older FIFO entry, never the newest post-RTPS result).
---
--- Unknown `_post_<cmd>` components (a `<name>_post_<cmd>`-suffixed component whose bare `<name>` is not a row key) emit one
--- `table_gap` info finding so downstream consumers can detect when the contract table is incomplete for an authored atom body.
---
--- Consumers:
---   * passes/static_analysis.lua::check_gte_result_position (renders result-position findings).
---   * passes/static_analysis.lua::emit_table_gap_warning    (called once per atom body; surfaces the missing-row diagnostic).
--- This table is consumed by the result-position check.
-M.GTE_COMPONENT_RESULT_CONTRACTS = {
-	-- Post-RTPS g4 p3 store contract: writes the latest screen XY (C2_SXY2) into the primitive's p3 slot.
-	-- Reading from C2_SXY0 is a semantic mismatch — C2_SXY0 is the oldest post-RTPS SXY, not the newest one.
-	["gte_store_g4_p3_post_rtps"] = {
-		command  = "gte_cmdw_rtps",
-		role     = "latest_screen_xy",
-		register = "C2_SXY2",
-	},
-}
-
 -- 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.
 --
@@ -1341,13 +1292,13 @@ M.OPERAND_READ_POSITIONS = {
 	["sub_s"]                  = {1, 2, 3},
 	["sub_u"]                  = {1, 2, 3},
 	["and_i"]                  = {1, 2},
-	["and_u"]                  = {1, 2, 3},
+	["and"]                    = {1, 2, 3},
 	["or_i"]                   = {1, 2},
 	["or_i_self"]              = {1},
-	["or_u"]                   = {1, 2, 3},
-	["or_u_self"]              = {1, 2},
+	["or"]                     = {1, 2, 3},
+	["or_self"]                = {1, 2},
 	["xor_i"]                  = {1, 2},
-	["xor_u"]                  = {1, 2, 3},
+	["xor"]                    = {1, 2, 3},
 	["slt_s"]                  = {1, 2, 3},
 	["slt_u"]                  = {1, 2, 3},
 	["slt_si"]                 = {1, 2},
@@ -1447,33 +1398,7 @@ M.GP0_CMD_BY_SHAPE = {
 	["g4"]  = 0x38, ["gt4"] = 0x3C,
 }
 
--- 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_post_rtpt"]                  = 3,
-	["mac_gte_store_g3_post_rtpt"]                  = 3,
-	["mac_gte_store_g4_p012_post_rtpt_pre_rtps"]    = 3,
-	["mac_gte_store_g4_p3_post_rtps"]               = 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.
---
+-- Per-instruction cycle cost (best-case, no stalls). Used by the static-analysis pass to emit per-atom cycle budgets.
 -- 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)
@@ -1487,8 +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
--- safe to clobber after 0-4 cycles).
+-- `docs/psx-spx/docs/gtepipelinetimings.md` for the hardware-verified input-latch boundaries
+-- (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,
@@ -1497,7 +1429,7 @@ M.INSTRUCTION_LATENCY = {
 	["add_s"]               = 1,  ["add_si"]          = 1,
 	["add_u"]               = 1,  ["add_u_self"]      = 1,
 	["sub_u"]               = 1,  ["sub_s"]           = 1,
-	["and_i"]               = 1,  ["and_u"]             = 1,
+	["and_i"]               = 1,  ["and"]             = 1,
 	["or_i"]                = 1,  ["or_i_self"]       = 1,
 	["or_u"]                = 1,  ["or_u_self"]       = 1,
 	["xor_i"]               = 1,  ["xor_u"]           = 1,
@@ -1533,17 +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
-	-- counted as part of the branch's cost)
+	-- Branches (branch + BD slot nop = 2 cycles; the BD slot's nop is 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 latency is usually absorbed by subsequent nops or by the next
+	-- COP2 transfers (mtc2/mfc2/ctc2/cfc2 = 1 cycle + COP2 latency; 
+	-- 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,
@@ -1583,20 +1517,7 @@ M.INSTRUCTION_LATENCY = {
 	["gte_load_v1"]         = 2,
 	["gte_load_v2"]         = 2,
 	["gte_load_v0v1v2"]     = 6,
-	-- 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_post_rtpt"]               = 3,
-	["mac_gte_store_g3_post_rtpt"]               = 3,
-	["mac_gte_store_g4_p012_post_rtpt_pre_rtps"] = 3,
-	["mac_gte_store_g4_p3_post_rtps"]            = 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,
@@ -1613,8 +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,
--- matches its `encoder` against `row.token`, and:
+-- The forward walker in `passes/static_analysis.lua::analyze_hardware_relations` reads every emitted word_event, 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`.
 --
@@ -1634,8 +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
---     measurement that the vendor caveats may still classify as `"conservative"` even when the underlying timing is numerically known.
+--   * `evidence.confidence` is one of `"exact"`, `"conservative"`, `"unknown"`. The severity comes from `violation_kind`;
+--     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:
@@ -1745,22 +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",
-		semantic       = "LWC2",
+		id             = "lwc2_to_gte_command",
+		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.
 	{
@@ -1973,7 +1913,7 @@ M.GPR_VALUE_RULES = {
 	-- Present register-form self variants. They are included here so a
 	-- known value is not needlessly lost when these encoders are used.
 	add_u_self       = { op = "add_u",       dest = 1, sources = {1, 2}, },
-	or_u_self        = { op = "or_u",        dest = 1, sources = {1, 2}, },
+	or_u_self        = { op = "or",          dest = 1, sources = {1, 2}, },
 	shift_lleft_self = { op = "shift_lleft", dest = 1, source = 1, immediate = 2, },
 }
 
@@ -2092,14 +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.
----   * **`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.
----   * **Known `mac_X(...)` calls**: recursively expand the indexed component body, including nested components. Every event from the expansion carries:
+---   * 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.
+---   * 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:
 ---     - `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.
----   * **Marker tokens** (`atom_label(...)` / `atom_offset(...)`): zero events (they are pure metaprogram hints, not emitted machine words).
+---   * 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).
 ---
 --- 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).
 ---
@@ -2123,12 +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 word_events table[] -- dense view of items where kind == "word"
---- @field markers     table[] -- dense view of items where kind == "label"|"offset"
+--- @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 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 warnings    table[] -- opaque warnings (e.g. unknown uncounted macro)
+--- @field errors      table[] -- Token-resolution failures surfaced without fail-loud
+--- @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
@@ -2136,20 +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 call_text       string   -- the 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 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 def_path        string   -- source path of the component definition
---- @field def_line        integer  -- source line of the component declaration
+--- @field component_name  string   -- Bare component name without the `mac_` prefix
+--- @field call_text       string   -- Immediate `mac_X(...)` token text (or root call text for the outermost entry)
+--- @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_line       integer  -- Source line of the call site
+--- @field def_path        string   -- Source path of the component definition
+--- @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.
@@ -2225,10 +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,
@@ -2238,6 +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,
@@ -2245,39 +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) 
-			else                          emit_marker("offset", args[1] or "", args[2] or "", 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, nil, nil, consuming_encoder, arg_pos)
 			end
 			pos = after_paren
 			::continue_loop::
@@ -2289,7 +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]`
@@ -2360,7 +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
@@ -2384,10 +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_root_call_text` is the 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.
-	-- 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.
+	-- walk_parent_inv_id:       Invocation ID of the enclosing call (0 for the root call).
+	-- walk_root_call_text:      Outermost `mac_X(...)` token text (preserved across recursion).
+	-- walk_immediate_call_text: IMMEDIATE outer `mac_X(...)` token text for words emitted in this body — nil for the root atom body.
+	-- 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 {}
@@ -2404,8 +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
@@ -2415,7 +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
@@ -2430,14 +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({
@@ -2452,7 +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]
@@ -2478,8 +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.
-			-- resolve_count may emit a warning if the count is unresolved.
+			-- Direct encoder, or mac_X-without-component: resolve count + emit n words.
+			-- 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,
---- and section-relative values are zero-based wire offsets. Only Lua string APIs receive
---- a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors, and section-relative values are zero-based wire offsets. 
+--- Only Lua string APIs receive a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- ELF/DWARF field offsets are expressed in hex so they map directly to the zero-based byte positions in the binary file.
 ---
---- ELF/DWARF field offsets are expressed in hex so they map directly to the
---- zero-based byte positions in the binary file.
+--- The ELF32 header / section / sym layout tables are within scripts/elf32.lua.
+--- The metaprogram re-exports the DWARF32 initial-length terminator.
 
-
---- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
-M.ELF32 = {
-	magic_offset          = 0x00,         -- 4-byte magic "\127ELF" at file offset 0x00
-	magic                 = "\127ELF",
-	class_offset          = 0x04,         -- 1-byte; 1 = ELF32, 2 = ELF64
-	class_elf32           = 1,
-	endian_offset         = 0x05,         -- 1-byte; 1 = little-endian, 2 = big-endian
-	endian_little         = 1,
-	header_bytes          = 0x34,         -- spec: gABI v1.2 §"ELF Header" — ELF32 header is 52 bytes total
-	e_shoff_offset        = 0x20,         -- 4-byte LE; section-header table file offset
-	e_shentsize_offset    = 0x2E,         -- 2-byte LE; section-header entry size in bytes
-	e_shnum_offset        = 0x30,         -- 2-byte LE; number of section headers
-	e_shstrndx_offset     = 0x32,         -- 2-byte LE; index of section-name string table
-	sh_size_bytes         = 0x28,         -- spec: gABI v1.2 §"Section Header Table" — each entry is 40 bytes
-	sh_name_offset        = 0x00,         -- 4-byte LE; offset into .shstrtab
-	sh_type_offset        = 0x04,         -- 4-byte LE; section type (SHT_*)
-	sh_offset_offset      = 0x10,         -- 4-byte LE; section's file offset
-	sh_size_offset        = 0x14,         -- 4-byte LE; section's size in bytes
-	dw_dwarf32_terminator = 0xFFFFFFFF,   -- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
-}
+--- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
+M.dw_dwarf32_terminator = E.dw_dwarf32_terminator
+-- TODO(Ed): Remove re-export.
 
 -- ----------------------------------------------------------------------------
 -- 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
-	-- directly to their position in the line-program header byte sequence.
+	-- opcode_base + line_range are 1-byte header fields; hex so they map 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,
@@ -204,6 +185,44 @@ M.DWARF_LINE_OPS = {
 	set_address_payload_size  = 0x05,   -- size = sub_opcode(1) + addr(4)
 }
 
+-- ----------------------------------------------------------------------------
+-- DWARF5 .debug_line (per DWARF5 spec §6.2.4 — Line Number Program Header)
+-- ----------------------------------------------------------------------------
+-- All offsets are zero-based wire offsets from the start of the unit body
+-- (i.e. AFTER unit_length has been read and unit_length bytes skipped past unit_length's 4 bytes).
+--
+-- The DWARF3/4 line-program format differs:
+--   - It omits `address_size` (DWARF3 §6.2.4) + `segment_selector_size` (DWARF5 §6.2.4).
+--   - It uses null-terminated string lists for `include_directories` + `file_names`
+--     (vs. DWARF5's format_count + fields-list shape).
+-- These are documented inline at each parse site in read_line_unit_file_table below.
+
+--- spec: DWARF5 spec §6.2.4 (Line Number Program Header — version >= 5)
+M.DWARF5_DEBUG_LINE = {
+	-- Header fields (zero-based, AFTER unit_length has been read).
+	version_offset_post_il = 0x00,    -- 2-byte LE; expected = 5
+	addr_size_offset       = 0x02,    -- 1 byte;  expected = 4
+	seg_size_offset        = 0x03,    -- 1 byte;  expected = 0
+	header_length_offset   = 0x04,    -- 4-byte LE; length of program-header content that follows
+	program_header_start   = 0x08,    -- first byte of program-header content (after the 8 fixed bytes)
+
+	-- Per-form byte widths (used when reading directory / file-name entries).
+	form_addr_bytes        = 0x04,    -- DW_FORM_addr (32-bit) | DW_FORM_data4
+	form_strp_bytes        = 0x04,    -- DW_FORM_line_strp / DW_FORM_strp / DW_FORM_strp_sup
+	form_data16_bytes      = 0x10,    -- DW_FORM_data16 (MD5)
+
+	-- DWARF5 form codes (subset used in line-program directory + file tables).
+	form_line_strp         = 0x1A,    -- DWARF5 §7.5.6 — DW_FORM_line_strp (4-byte offset into .debug_line_str)
+	form_string            = 0x08,    -- DWARF4-compatible fallback (inline null-terminated; not in .debug_line_str)
+	form_udata             = 0x0F,    -- DW_FORM_udata (ULEB)
+	form_data16            = 0x18,    -- DW_FORM_data16 (16-byte MD5; gcc emits this for split debug info)
+
+	-- DWARF5 content-tag codes (DW_LNCT_* from §6.2.4.1 + §6.2.4.2).
+	lnct_path              = 0x01,
+	lnct_directory_index   = 0x02,
+	lnct_md5               = 0x05,    -- gcc with MD5 in file name table (rare)
+}
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- I/O helpers: little-endian byte read/write
 -- ════════════════════════════════════════════════════════════════════════════
@@ -211,37 +230,31 @@ M.DWARF_LINE_OPS = {
 --- Read a 4-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- Equivalent to `string.unpack("<I4", buf, off + 1)` but avoids the table-return shape + works under LuaJIT 2.1
 --- (which has partial `string.unpack` coverage).
----
 --- **Convention:** `off` is a zero-based wire offset; `+ 1` is applied only at the `string.byte` boundary.
 ---
---- **Byte weights** are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
---- byte 0 contributes its value directly; byte 1 is shifted left by 8
---- (= 0x100); byte 2 by 16 (= 0x10000); byte 3 by 24 (= 0x1000000).
+--- Thin forwarder: the canonical implementation lives in scripts/elf32.lua.
+--- The "second caller lifts" pattern keeps the metaprogram side fluent
+--- (`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 buf:byte(byte_off)
-		+ buf:byte(byte_off + 0x01) * 0x00000100
-		+ buf:byte(byte_off + 0x02) * 0x00010000
-		+ buf:byte(byte_off + 0x03) * 0x01000000
+	return E.read_u32_le(buf, off)
 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 buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
+	return E.read_u16_le(buf, off)
 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
--- can import them as file-scope locals per the 2nd-caller lift precedent
+-- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua can import them as file-scope locals per the 2nd-caller lift precedent
 -- (the uleb128 + sleb128 encoders were promoted the same way).
 function M.read_uleb128_at(buf, pos)
 	local value, shift = 0, 0
@@ -389,9 +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
-		-- by signature. The low 4 bytes (LE) are the type signature (content hash);
-		-- the high 4 bytes (LE) are a CU-relative offset into the matching type unit.
+		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): An 8-byte value identifying a type by signature.
+		-- The low  4 bytes (LE) are the type signature (content hash);
+		-- The high 4 bytes (LE) are a CU-relative offset into the matching type unit.
 		-- Consumers use the low  4 to look up the type unit (see M.find_type_unit_by_signature)
 		--          then the high 4 to resolve the specific type within it.
 		-- Return the low  4 as the primary value to preserve the (value, next_pos) shape;
@@ -417,20 +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).
--- Walk all units in `info` and return the 0-based offset of the first unit
--- whose `DW_AT_type_signature` (8-byte value at the end of the unit header) equals `target_sig`.
--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
---
--- Unit header layout (from pos 0):
---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
--- @param info string  -- the .debug_info section bytes
--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
+--- DWARF5 §7.5.6 (Type Entries).
+--- Walk all units in `info` and return the 0-based offset of the first unit whose `DW_AT_type_signature`
+--- (8-byte value at the end of the unit header) equals `target_sig`.
+--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
+--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
+---
+--- Unit header layout (from pos 0):
+---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
+---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
+--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
+--- @param info          string   -- the .debug_info section bytes
+--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
+--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
+--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
 function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
 	local pos         = 0
 	local section_len = #info
@@ -539,69 +552,58 @@ function M.read_elf_sections(elf_path, section_names)
 		return result
 	end
 
-	-- Read the ELF32 header.
-	local header = f:read(M.ELF32.header_bytes)
-	if not header or #header < M.ELF32.header_bytes then
-		io.stderr:write("[elf_dwarf.read_elf_sections] ELF too small for ELF32 header\n")
+	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 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.
-	if header:sub(M.ELF32.magic_offset + 1, M.ELF32.magic_offset + 0x04) ~= M.ELF32.magic then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not an ELF file\n")
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.class_offset + 1) ~= M.ELF32.class_elf32 then
-		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] not ELF32 (class=%d)\n", header:byte(M.ELF32.class_offset + 1)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.endian_offset + 1) ~= M.ELF32.endian_little then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not little-endian; unsupported\n")
+	local sections, walk_err = E.walk_sections(adapter, hdr)
+	if not sections then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] section walk failed: %s\n", tostring(walk_err)))
 		f:close()
 		return result
 	end
 
-	-- Parse section-header table location + dimensions from the header.
-	local e_shoff     = M.read_u32_le(header, M.ELF32.e_shoff_offset)
-	local e_shentsize = M.read_u16_le(header, M.ELF32.e_shentsize_offset)
-	local e_shnum     = M.read_u16_le(header, M.ELF32.e_shnum_offset)
-	local e_shstrndx  = M.read_u16_le(header, M.ELF32.e_shstrndx_offset)
-
-	-- Read the section-header string table (.shstrtab) so we can resolve section names from their `sh_name` offsets.
-	f:seek("set", e_shoff + e_shstrndx * e_shentsize)
-	local strtab_hdr = f:read(e_shentsize)
-	if not strtab_hdr or #strtab_hdr < e_shentsize then
-		io.stderr:write("[elf_dwarf.read_elf_sections] could not read .shstrtab header\n")
-		f:close()
-		return result
-	end
-	local strtab_offset = M.read_u32_le(strtab_hdr, M.ELF32.sh_offset_offset)
-	local strtab_size   = M.read_u32_le(strtab_hdr, M.ELF32.sh_size_offset)
-	f:seek("set", strtab_offset)
-	local strtab = f:read(strtab_size) or ""
-
-	-- Walk all section headers; collect (offset, size) for the wanted names.
-	local function read_section_bytes(sh_offset, sh_size)
-		f:seek("set", sh_offset)
-		return f:read(sh_size) or ""
-	end
-
-	for sh_idx = 0, e_shnum - 1 do
-		f:seek("set", e_shoff + sh_idx * e_shentsize)
-		local sh = f:read(e_shentsize)
-		if not sh or #sh < e_shentsize then break end
-		local sh_name   = M.read_u32_le(sh, M.ELF32.sh_name_offset)
-		local sh_offset = M.read_u32_le(sh, M.ELF32.sh_offset_offset)
-		local sh_size   = M.read_u32_le(sh, M.ELF32.sh_size_offset)
-
-		-- Extract the name (null-terminated C string in strtab).
-		local name_end = strtab:find("\0", sh_name + 1, true) or (sh_name + 1)
-		local name     = strtab:sub(sh_name + 1, name_end - 1)
-		if wanted[name] then
-			result[name] = read_section_bytes(sh_offset, sh_size)
+	-- Resolve the requested sections.
+	for _, s in ipairs(sections) do
+		if wanted[s.name] then
+			local bytes = E.read_section_bytes(adapter, s)
+			if bytes then result[s.name] = bytes end
 		end
 	end
 
@@ -618,49 +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).
-	local sections = M.read_elf_sections(elf_path, {".symtab", ".strtab"})
-	local symtab   = sections[".symtab"]
-	local strtab   = sections[".strtab"]
-	if not symtab or not strtab or #symtab == 0 or #strtab == 0 then
-		-- No symbol table (e.g. stripped ELF). Return empty.
+	-- Existence check first; an empty or missing ELF returns an empty map.
+	if lfs.attributes(elf_path, "mode") ~= "file" then
 		return addrs
 	end
 
-	-- Iterate the 16-byte ELF32 symtab entries.
-	-- Each entry (zero-based): st_name at 0, st_value at 4, st_size at 8, st_info at 12, st_other at 13, st_shndx at 14.
-	local SYM_ENTRY_BYTES = 0x10
-	local SYM_ST_NAME     = 0x00
-	local SYM_ST_VALUE    = 0x04
-	local SYM_ST_SIZE     = 0x08
-	local SYM_ST_INFO     = 0x0C
-	local n_syms = #symtab / SYM_ENTRY_BYTES
-	for i = 0, n_syms - 1 do
-		local entry_off  = i * SYM_ENTRY_BYTES
-		local st_info    = symtab:byte(entry_off + SYM_ST_INFO + 1)
-		-- High nibble = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
-		-- Use math.floor(/16) instead of bit.rshift for LuaJIT 2.1 compat
-		-- (LuaJIT's `>>` is 5.3+, but math.floor(x/16) works on all versions).
-		local binding    = math.floor(st_info / 16)
-		if binding == 0 or binding == 1 then            -- STB_LOCAL or STB_GLOBAL
-			local st_size   = M.read_u32_le(symtab, entry_off + SYM_ST_SIZE)
-			if st_size > 0 then
-				local st_name_off = M.read_u32_le(symtab, entry_off + SYM_ST_NAME)
-				-- Extract the name from .strtab (null-terminated C string).
-				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
-				local name     = strtab:sub(st_name_off + 1, name_end - 1)
-				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — 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 }
+	local f = io.open(elf_path, "rb")
+	if not f then
+		return addrs
 	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
 
@@ -787,6 +827,224 @@ function M.sleb128_size(n)
 	return bytes
 end
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- DWARF5 line-program file-table reader
+-- ════════════════════════════════════════════
+
+--- Read every line-program unit in `.debug_line` and produce one entry per file across all units.
+--- Returns three parallel maps keyed by 1-based file index.
+---
+--- Wire format notes:
+---   * 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 `inv.call_file`
+---   * 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` emits 2 forms: path + dir_index). 
+---     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)
+---     - DW_FORM_data16 (16-byte MD5; ignored — skip the form's bytes)
+---   * Symlink-canonicalisation: each path's `paths[i]` is stored verbatim from the wire
+---     (mixed `/` and `\` accepted; the basename is taken via the last path separator). Caller normalises as needed.
+---
+--- 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 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; 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, ... }
+---   paths:             { [1-based-per-unit-file-index] = full path (mixed slashes), ... }
+function M.read_line_unit_file_table(elf_path)
+	local sections = M.read_elf_sections(elf_path, { ".debug_line", ".debug_line_str" })
+	local line     = sections[".debug_line"]
+	local lstr     = sections[".debug_line_str"] or ""
+	if not line or line == "" then
+		io.stderr:write("[elf_dwarf.read_line_unit_file_table] no .debug_line section in: " .. tostring(elf_path) .. "\n")
+		return nil
+	end
+
+	local basenames         = {}
+	local basename_to_index = {}
+	local paths             = {}
+
+	--- Read one form-code's bytes from `buf` at position `p` according to `form`.
+	--- Returns (value, after) where `value` is:
+	---   * the resolved string (DW_FORM_line_strp / DW_FORM_string)
+	---   * the ULEB128 number  (DW_FORM_udata)
+	---   * nil + skip-bytes    (DW_FORM_data16; we don't surface the MD5)
+	local function read_form(buf, lstr_buf, p, form)
+		if form == M.DWARF5_DEBUG_LINE.form_line_strp then
+			local strp    = M.read_u32_le(buf, p)
+			local end_pos = lstr_buf:find("\0", strp + 1, true) or (#lstr_buf + 1)
+			return lstr_buf:sub(strp + 1, end_pos - 1), p + M.DWARF5_DEBUG_LINE.form_strp_bytes
+		elseif form == M.DWARF5_DEBUG_LINE.form_string then
+			local nul = buf:find("\0", p + 1, true) or (#buf + 1)
+			return buf:sub(p + 1, nul - 1), nul
+		elseif form == M.DWARF5_DEBUG_LINE.form_udata then
+			local  v, after = M.read_uleb128_at(buf, p)
+			return v, after
+		elseif form == M.DWARF5_DEBUG_LINE.form_data16 then
+			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, 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
+		end
+	end
+
+	--- Parse one DWARF-version-3-style unit (DWARF3/4 line program; gcc default in the PS1 toolchain still emits DWARF3 for line programs in `-g` mode).
+	--- Layout: null-terminated directory list, then path(null) + dir_idx(ULEB) + time(ULEB) + size(ULEB) file entries terminated by an empty null.
+	--- `content_start` = zero-based wire offset of the first byte of program-header content (after version + header_length fields).
+	--- @return unit_basenames { [idx_in_unit_1_based] = basename }
+	--- @return unit_paths     { [idx_in_unit_1_based] = full path }
+	local function parse_dwarf3_unit(buf, content_start, body_end)
+		local up = content_start
+		-- 5 fixed bytes: min_insn, default_is, line_base (signed), line_range, opcode_base
+		up = up + 5
+		local opcode_base = buf:byte(content_start + 5)
+		up = up + (opcode_base - 1)  -- std_opcode_lengths
+		local dirs = {}
+		while up < body_end do
+			local nul = buf:find("\0", up + 1, true) or (body_end + 1)
+			if    nul > body_end then break end
+			local len = nul - up - 1
+			if    len == 0 then up = nul break end
+			dirs[#dirs + 1] = buf:sub(up + 1, nul - 1)
+			up = nul
+		end
+		local unit_basenames = {}
+		local unit_paths = {}
+		while up < body_end do
+			local nul = buf:find("\0", up + 1, true) or (body_end + 1)
+			if    nul > body_end or nul == up + 1 then up = nul break end
+			local path = buf:sub(up + 1, nul - 1)
+			up = nul
+			local didx,  up_next  = M.read_uleb128_at(buf, up); up = up_next
+			local _time, up_next2 = M.read_uleb128_at(buf, up); up = up_next2
+			local _size, up_next3 = M.read_uleb128_at(buf, up); up = up_next3
+			local idx = #unit_basenames + 1
+			local bs  = path:match("[^/\\]+$") or path
+			unit_paths[idx]     = path
+			unit_basenames[idx] = bs
+			dirs[1] = dirs[1] or ""  -- safety: gcc emits "" sentinel dir at 0
+			if didx > 0 and dirs[didx] then
+				unit_paths[idx] = dirs[didx] .. "/" .. path
+			end
+		end
+		return unit_basenames, unit_paths
+	end
+
+	--- Parse one DWARF-version-5-style unit (DWARF5 line program; used by modern gcc with `-gdwarf-5`).
+	--- `content_start` is the first byte of program-header content (after the 8 fixed bytes version+addr_size+seg_size+header_length).
+	--- @return same shape as parse_dwarf3_unit
+	local function parse_dwarf5_unit(buf, lstr_buf, content_start, body_end)
+		local up = content_start
+		-- 6 fixed bytes: min_insn, max_ops_per_insn, default_is, line_base, line_range, opcode_base
+		up = up + 6
+		local opcode_base = buf:byte(content_start + 6)
+		up = up + (opcode_base - 1)  -- std_opcode_lengths
+		-- directories
+		local dir_format_count, after = M.read_uleb128_at(buf, up); up = after
+		local dir_formats = {}
+		for i = 1, dir_format_count do
+			local f, a2 = M.read_uleb128_at(buf, up); up = a2
+			dir_formats[i] = f
+		end
+		local dir_count, a3 = M.read_uleb128_at(buf, up); up = a3
+		local dirs = {}
+		for i = 1, dir_count do
+			local combined = ""
+			for j = 1, dir_format_count do
+				local v, a4 = read_form(buf, lstr_buf, up, dir_formats[j])
+				up = a4
+				if j == 1 and type(v) == "string" then combined = v end
+			end
+			dirs[i] = combined
+		end
+		-- file names
+		local file_format_count, after2 = M.read_uleb128_at(buf, up); up = after2
+		local file_formats = {}
+		for i = 1, file_format_count do
+			local f, a2 = M.read_uleb128_at(buf, up); up = a2
+			file_formats[i] = f
+		end
+		local file_count, a3 = M.read_uleb128_at(buf, up); up = a3
+		local unit_basenames = {}
+		local unit_paths     = {}
+		for i = 1, file_count do
+			local combined = ""
+			local didx = 0
+			for j = 1, file_format_count do
+				local v, a4 = read_form(buf, lstr_buf, up, file_formats[j])
+				up = a4
+				if j == 1 and type(v) == "string" then combined = v end
+				if j == 2 and type(v) == "number"  then didx  = v end
+			end
+			local idx = #unit_basenames + 1
+			local bs  = combined:match("[^/\\]+$") or combined
+			unit_paths[idx]     = combined
+			unit_basenames[idx] = bs
+			if didx > 0 and dirs[didx] then
+				unit_paths[idx] = dirs[didx] .. "/" .. combined
+			end
+		end
+		return unit_basenames, unit_paths
+	end
+
+	--- Walk every line-program unit in the section.
+	local p = 0
+	local section_end = #line
+	while p + 4 <= section_end do
+		local unit_length = M.read_u32_le(line, p)
+		if    unit_length == 0xFFFFFFFF then
+			io.stderr:write("[elf_dwarf.read_line_unit_file_table] 64-bit DWARF (initial-length 0xFFFFFFFF); not supported\n")
+			return nil
+		end
+		local body_start = p + 4
+		local body_end   = p + 4 + unit_length
+		if body_end > section_end then break end
+		local version = M.read_u16_le(line, body_start)
+		local unit_basenames, unit_paths
+		if version >= 5 then
+			-- DWARF5 header: version(2) + addr_size(1) + seg_size(1) + header_length(4) + content
+			local header_length_offset = body_start + 6  -- past version(2) + addr_size(1) + seg_size(1) - wait that's wrong; past hdr len is at +6
+			local content_start = body_start + 8  -- past version(2) + addr_size(1) + seg_size(1) + header_length(4)
+			unit_basenames, unit_paths = parse_dwarf5_unit(line, lstr, content_start, body_end)
+		elseif version >= 2 then
+			-- DWARF2/3/4 header: version(2) + header_length(4) + content
+			local content_start = body_start + 6  -- past version(2) + header_length(4)
+			unit_basenames, unit_paths = parse_dwarf3_unit(line, content_start, body_end)
+		else
+			io.stderr:write(string.format("[elf_dwarf.read_line_unit_file_table] unsupported DWARF version %d (offset 0x%x)\n", version, p))
+			p = body_end
+			goto continue
+		end
+		-- Per-unit 1-based file indices are aligned with `inv.call_file` values because the metaprogram emits `DW_LNS_set_file` with the per-unit index.
+		-- When multiple units are present (crt0.s + C unit), the per-unit index in each unit matches the metaprogram's intent (gcc always sets file in unit-local terms).
+		-- We therefore store directly without global re-indexing; the caller is responsible for knowing which unit the file-index applies to.
+		-- For DWARF3 (C unit is the unit that matters for atom line tables), this matches.
+		-- For DWARF5 (crt0.s + C unit), each carries its own per-unit file-table map;
+		-- the atom-side DW_LNS_set_file(N) refers to the C unit's indices, NOT crt0.s's.
+		-- Since the C unit is the one with full include_directories + 12 entries, we can use it directly.
+		for idx, bs in pairs(unit_basenames) do
+			basenames[idx] = bs
+			paths[idx] = unit_paths[idx]
+			basename_to_index[bs] = idx
+		end
+		p = body_end
+		::continue::
+	end
+
+	return basename_to_index, basenames, paths
+end
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- I/O helpers: atoms source-map + native directory glob
 -- ════════════════════════════════════════════════════════════════════════════

scripts/gdb/gdb_tape_atoms.gdb

@@ -3,16 +3,16 @@
 # Wrapper for the tape-atom step-debug helpers.
 # The 9 user commands are defined here as STUBS (degraded-state messages). 
 # The real implementations + the per-atom data tables are emitted by `passes/atoms_source_map.lua` 
-# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/gen/gdb_tape_atoms_runtime.gdb`.
+# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/gdb_tape_atoms_runtime.gdb`.
 # Sourcing that file RE-DEFINES the commands with real implementations.
 #
-# If `build/gen/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
+# If `build/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
 # The user just needs to re-run `build_psyq.ps1` to regenerate.
 
-# ── Stub commands (defined here so they're always present, even if the runtime file is missing). The runtime file overrides these if sourced. ──
+# ?? Stub commands (defined here so they're always present, even if the runtime file is missing). The runtime file overrides these if sourced. ??
 
 define tape_atoms
-  echo "[gdb_tape_atoms] STUB: runtime file build/gen/gdb_tape_atoms_runtime.gdb not found."
+  echo "[gdb_tape_atoms] STUB: runtime file build/gdb_tape_atoms_runtime.gdb not found."
   echo "[gdb_tape_atoms] STUB: run .\\build_psyq.ps1 to regenerate, then re-source this file."
 end
 document tape_atoms
@@ -21,35 +21,35 @@ document tape_atoms
 end
 
 define break_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document break_atom
   Set a breakpoint at the start of tape atom <name>. STUB state.
 end
 
 define step_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document step_atom
   Resume execution until the next atom boundary. STUB state.
 end
 
 define next_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document next_atom
   Alias for step_atom. STUB state.
 end
 
 define where_in_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document where_in_atom
   Report current atom name, .rodata addr, word offset, and source line (if known). STUB state.
 end
 
 define stepi_inside_atom
-  echo "[gdb_tape_atoms] STUB: build/gen/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
 end
 document stepi_inside_atom
   One MIPS-instruction step, then where_in_atom. STUB state.
@@ -89,17 +89,17 @@ document wave_ctx
 end
 
 
-# ── Source the runtime file (re-defines commands with real impls + data). ──
+# ?? Source the runtime file (re-defines commands with real impls + data). ??
 
 # Try to source from project-root-relative path first (the typical case).
 # If the user is in a different CWD, the source will fail and stubs remain.
-# The runtime file path is computed relative to the ELF's source map convention (build/gen/gdb_tape_atoms_runtime.gdb).
+# The runtime file path is computed relative to the ELF's source map convention (build/gdb_tape_atoms_runtime.gdb).
 echo [gdb_tape_atoms] Wrapper loaded. Sourcing runtime file...
 # Suppress the "Redefine command" prompts that would otherwise appear when the runtime file overrides the 9 stub commands defined above.
-# The runtime's `define` blocks are intended to overwrite — there's no ambiguity to confirm.
+# The runtime's `define` blocks are intended to overwrite ? there's no ambiguity to confirm.
 set confirm off
 
 # Source the runtime file (re-defines commands with real impls + data).
-source build/gen/gdb_tape_atoms_runtime.gdb
+source build/gdb_tape_atoms_runtime.gdb
 set confirm on
 echo [gdb_tape_atoms] Runtime sourced successfully (9 commands now have real implementations).

scripts/passes/annotation.lua

@@ -7,18 +7,11 @@
 ---
 --- Ownership: the canonical `ctx.shared.corpus` supplies cross-source registries, while each `src.scan` supplies its source's declarations and bodies.
 --- A context without `ctx.shared.corpus` is rejected with an explicit canonical-corpus message.
----
---- Writes `<ctx.out_root>/<dir_basename>.errors.h` once per module, with `#error` directives for findings that the C compile surfaces.
---- `passes/report.lua` renders annotations.txt from `corpus.sources_by_dir`, re-validating each source through `M.validate()`.
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
 
 -- Bootstrap follows the entry scripts; `scripts/duffle_paths.lua` sets package.path and package.cpath. See `ps1_meta.lua` for the rationale.
 -- `debug.getinfo(1, "S").source` locates this file for standalone and orchestrated runs, then `duffle_paths.lua` returns the loaded `duffle` module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local write_file     = duffle.write_file
-local ensure_dir     = duffle.ensure_dir
 
 -- The annotation pass reads the source-derived registries from scan_source:
 --   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
@@ -29,11 +22,11 @@ local ensure_dir     = duffle.ensure_dir
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -52,28 +45,28 @@ local ensure_dir     = duffle.ensure_dir
 --- @field warnings table[]
 
 --- @class AtomAnnotation
---- @field line    integer       -- source line of the atom_info call
---- @field macro   string        -- the macro name (always "atom_info" in the new shape)
---- @field name    string        -- the atom name
---- @field kind    string        -- always "info"
+--- @field line    integer       -- Source line of the atom_info call
+--- @field macro   string        -- Macro name (always "atom_info" in the new shape)
+--- @field name    string        -- Atom name
+--- @field kind    string        -- Always "info"
 --- @field binds   string|nil    -- Binds_X name if any
 --- @field reads   string[]      -- R_* names (read targets)
 --- @field writes  string[]      -- R_* names (write targets)
---- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
+--- @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
---- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
+--- @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_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 msg  string  -- finding message
+--- @field line integer -- Source line (or 0 for pass-level)
+--- @field msg  string  -- Finding message
 
 --- @class Findings
 --- @field errors   Finding[]
@@ -81,14 +74,14 @@ local ensure_dir     = duffle.ensure_dir
 --- @field info     Finding[]
 
 --- @class PipeCtx
---- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
---- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
---- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
---- @field types          table<string, RegTypeDefault> -- from scan_source
---- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
---- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field atom_index     table<string, AtomAnnotation> -- Name -> AtomAnnotation (only kind=="atom")
+--- @field binds_index    table<string, BindsStruct>    -- Name -> BindsStruct
+--- @field annot_counts   table<string, integer>        -- Name -> annotation count (for unique_annotation check)
+--- @field types          table<string, RegTypeDefault> -- From scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- From scan_source
+--- @field seen_defaults  table<string, integer>        -- Duplicate atom_dbg_reg_default detection
 --- @field seen_field     table<string, integer>        -- Binds_* -> count of fields (set/checked by check_binds_no_duplicate_fields)
---- @field _scan          SourceScan                    -- full scan payload (typed-view sub-calls live here) 
+--- @field _scan          SourceScan                    -- Full scan payload (typed-view sub-calls live here) 
 
 --- @class AnnotatedResult
 --- @field atoms    AtomEntry[]
@@ -102,11 +95,10 @@ local ensure_dir     = duffle.ensure_dir
 -- ════════════════════════════════════════════════════════════════════════════
 -- 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
@@ -119,8 +111,8 @@ local function check_atom_decl_exists(a, pipe_ctx, findings)
 	end
 end
 
---- Check: every atom may have AT MOST ONE annotation.
---- Post-loop: needs full-corpus `annot_counts` from pipe_ctx.
+--- Check:     Every atom may have AT MOST ONE annotation.
+--- Post-loop: Needs full-corpus `annot_counts` from pipe_ctx.
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_unique_annotation(pipe_ctx, findings)
@@ -153,7 +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]
@@ -311,7 +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`)
@@ -322,12 +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
@@ -378,8 +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
@@ -433,8 +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.
@@ -480,7 +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)
@@ -510,14 +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 = {}
-	for reg, _ in pairs(scan.types or {}) do
-		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
-	end
-	local atom_infos_list = {}
-	for _, ai in ipairs(scan.atom_infos or {}) do
-		atom_infos_list[#atom_infos_list + 1] = ai
-	end
+	local seen_defaults   = {}; for reg, _ in pairs (scan.types      or {}) do seen_defaults[reg] = (seen_defaults[reg] or 0) + 1 end
+	local atom_infos_list = {}; for _, ai  in ipairs(scan.atom_infos or {}) do atom_infos_list[#atom_infos_list + 1] = ai         end
 
 	local pipe_ctx = {
 		atom_index               = {},
@@ -605,40 +587,6 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	}
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Per-DIRECTORY (per-module) output: errors.h + annotations.txt
--- ════════════════════════════════════════════════════════════════════════════
-
---- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
---- Empty directories (no errors, no atoms) produce no file.
-local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
-	if atoms_count == 0 and #errors == 0 then
-		return nil
-	end
-	local out_path = ctx.out_root .. "/" .. dir_basename .. ".errors.h"
-	local lines    = {
-		"// Auto-generated by ps1_meta.lua (passes/annotation.lua) — DO NOT EDIT",
-		string.format("// Module: %s   Sources: %d", dir_basename, #sources),
-		"#pragma once",
-		"",
-	}
-	if #errors == 0 then
-		lines[#lines + 1] = "// annotation pass OK"
-	else
-		for _, e in ipairs(errors) do
-			local src_tag = ""
-			if e.source then
-				local src_name = e.source:match("([^/\\]+)$") or e.source
-				src_tag = src_name .. ": "
-			end
-			lines[#lines + 1] = string.format('#error "%s%s (line %d)"', src_tag, e.msg, e.line)
-		end
-	end
-	ensure_dir(ctx.out_root)
-	write_file(out_path, table.concat(lines, "\n") .. "\n")
-	return out_path
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- M.run — orchestrator entry
 -- ════════════════════════════════════════════════════════════════════════════
@@ -683,11 +631,6 @@ function M.run(ctx)
 				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
 			end
 		end
-
-		local err_path = emit_module_errors_h(ctx, dir_basename, dir_atoms, dir_errors, dir_sources)
-		if err_path then
-			table.insert(outputs, { errors_h = err_path })
-		end
 	end
 
 	return { outputs = outputs, errors = errors, warnings = warnings }

scripts/passes/atoms_source_map.lua

@@ -1,23 +1,21 @@
 --- 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
---- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and 
+--- 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.
+--- 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.
 ---
---- **Two output forms** (per the workspace's per-emission-form pattern from `guide_metaprogram_ssdl.md`):
----   1. **Sourcemap.txt form** — `<out_root>/<basename>.atoms.sourcemap.txt`. Format-version-tagged for forward-compat.
----      Lives in `<out_root>/` (build/gen). Mirrors the convention used by `annotation.lua`
----      (`<out_root>/<basename>.errors.h`) and `static_analysis.lua` (`<out_root>/<basename>.static_analysis.txt`).
+--- Two output forms:
+---   1. Markdown form: Handled by `passes/report.lua` (writes `<module>.atoms.md`).
+---      The render functions `render_source_map` + `render_provenance` are exported for `report.lua` to call directly.
 ---      Compile artifacts (`*.macs.h`, `*.offsets.h`) stay in `<source_dir>/gen/`.
----   2. **gdb-runtime form** — `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`. A pure gdb command script — addresses come
----      from `nm`, the 9 user commands are static `define ... end` blocks. Emitted when `ctx.flags.gdb_runtime` is true
----      AND `ctx.flags.elf_path` points to an existing ELF. Useful for `gdb-multiarch --without-python` users
----      (the common case on Windows MinGW builds) — `source <path>` loads it with no Python / Tcl / Guile required.
+---   2. `gdb_tape_atoms_runtime.gdb`: Post-link opt-in (`ctx.flags.gdb_runtime`),
+---      so the gdb wrapper script + the generated runtime script share the same canonical location.
+---      Triggered by `--post-link` or `--gdb-runtime`.
 ---
---- **Output format** (sourcemap.txt form):
+--- Output forma (sourcemap.txt form):
 ---   ```
 ---   # FORMAT_VERSION 1
 ---   # auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT
@@ -30,10 +28,7 @@
 ---   ...
 ---   ENDATOM
 ---   ```
----
 --- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
----
---- **Conventions:** tabs (1/level), EmmyLua annotations, no regex, Lua 5.3 compatible.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -99,8 +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
---- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was declared in `corpus.word_counts` 
+--- (populated by word_count_eval + components passes).
 --- @param src  table
 --- @param atom table
 --- @param wc   table -- identity alias of corpus.word_counts
@@ -109,20 +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(
-				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
-				entry.pos, rel_path, entry.line, inv.component_name,
-				inv.def_path or "", inv.def_line or 0, entry.body_line)
+			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d'
+				, entry.pos, rel_path, entry.line, inv.component_name
+				, inv.def_path or "", inv.def_line or 0, entry.body_line)
 		else
-			lines[#lines + 1] = string.format(
-				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
 		end
 	end
 
@@ -252,9 +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
---- vars set in `emit_gdb_runtime` provide printf args, and 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.
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`,
+--- the convenience vars set in `emit_gdb_runtime` provide printf args, and 
+--- 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.
@@ -399,8 +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
---- source-time.
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb source-time.
 --- @param ctx PassCtx
 local function emit_gdb_runtime(ctx)
 	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
@@ -458,7 +450,22 @@ local function emit_gdb_runtime(ctx)
 	-- Confirmation line for the source operator.
 	lines[#lines + 1] = 'printf "[gdb_tape_atoms] runtime loaded %d atoms from %s\\n", $__atom_count, $__elf_path'
 
-	local out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	local out_path
+	-- Move out of `<out_root>/gdb_tape_atoms_runtime.gdb` to `<out_root>/../gdb_tape_atoms_runtime.gdb` when the conventional `<out_root>` is `<build>/gen`
+	-- (any equivalent spelling — relative, absolute backslash, absolute forward-slash, trailing-separator variants).
+	-- This puts the gdb runtime alongside the ELF at `build/` rather than under the report subdir.
+	local function ends_with_gen_dir(p)
+		if type(p) ~= "string" then return false end
+		return p:match("[/\\]gen[/\\]?$") ~= nil or p == "build/gen" or p == "build\\gen"
+	end
+	if ends_with_gen_dir(ctx.out_root) then
+		-- Strip the trailing `/gen` segment, then write the runtime script under `build/`.
+		-- e.g. "C:/projects/Pikuma/ps1/build/gen" -> "C:/projects/Pikuma/ps1/build".
+		local parent = ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		out_path = parent .. "/gdb_tape_atoms_runtime.gdb"
+	else
+		out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	end
 	duffle.ensure_dir(duffle.dirname(out_path))
 	duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
 	-- io.stderr:write(string.format("[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
@@ -470,10 +477,62 @@ end
 
 local M = {}
 
---- Pass entry. For each source that declares at least one `MipsAtom_(name)` / `MipsCode code_<name>`, emit two files
---- in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
---- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation). When `ctx.flags.gdb_runtime`
---- is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
+-- Expose the pure render functions so `report.lua` and the focused tests can call them directly without triggering the file-emit path.
+M.render_source_map = render_source_map
+M.render_provenance = render_provenance
+
+--- Render ONE atom's sourcemap stanza.
+--- @param atom table   -- atom record (must have `atom.paths` populated)
+--- @return string
+function M.render_atom_source_map(atom)
+	assert(type(atom) == "table", "render_atom_source_map: atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_source_map: atom.paths must be a table")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+	end
+	lines[#lines + 1] = "ENDATOM"
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render ONE atom's provenance stanza — no per-file format header, no enumeration of other atoms.
+---
+--- `rel_path` is the source path (forward-slashes) embedded in every `CALL` line.
+--- The .md caller (report.lua) is expected to derive this once per `## <source>` heading and pass it down for each atom in that source.
+--- @param atom      table   -- atom record (must have `atom.paths` populated)
+--- @param wc        table   -- identity alias of `corpus.word_counts`
+--- @param rel_path  string  -- source path (forward-slashes) for `CALL` fields
+--- @return string
+function M.render_atom_provenance(atom, wc, rel_path)
+	assert(type(atom)       == "table", "render_atom_provenance:  atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_provenance:  atom.paths must be a table")
+	assert(type(rel_path)   == "string", "render_atom_provenance: rel_path must be a string")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		local inv = entry.invocation
+		local macro_count = inv and wc and wc["mac_" .. inv.component_name]
+		if inv and macro_count ~= nil then
+			lines[#lines + 1] = string.format(
+				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				entry.pos, rel_path, entry.line, inv.component_name,
+				inv.def_path or "", inv.def_line or 0, entry.body_line)
+		else
+			lines[#lines + 1] = string.format(
+				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Pass entry. For each source that declares at least one `MipsAtom_(name)` / `MipsCode code_<name>`,
+--- emit two files in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
+--- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation).
+--- When `ctx.flags.gdb_runtime` is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -495,34 +554,8 @@ function M.run(ctx)
 		}
 	end
 
-	-- Always emit the text form (per-source).
-	for _, src in ipairs(corpus.source_order) do
-		local has_projection = false
-		for _, atom in ipairs((src.scan or {}).atoms or {}) do
-			if (atom.kind == "atom" or atom.kind == "raw_atom") and atom.paths then
-				has_projection = true; break
-			end
-		end
-		if not has_projection then
-			for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
-				if atom.paths then has_projection = true; break end
-			end
-		end
-		if has_projection then
-			local basename = duffle.basename_no_ext(src.path)
-			-- (1) atoms.sourcemap.txt — format-1 per-word call-site map.
-			local sourcemap_path = ctx.out_root .. "/" .. basename .. ".atoms.sourcemap.txt"
-			local sourcemap_body = render_source_map(src)
-			-- (2) atoms.provenance.txt — format-1 per-word definition/body map.
-			local prov_path = ctx.out_root .. "/" .. basename .. ".atoms.provenance.txt"
-			local prov_body = render_provenance(src, wc)
-			duffle.ensure_dir(duffle.dirname(sourcemap_path))
-			duffle.write_file_lf(sourcemap_path, sourcemap_body)
-			duffle.write_file_lf(prov_path,    prov_body)
-			outputs[#outputs + 1] = { kind = "report", path = sourcemap_path }
-			outputs[#outputs + 1] = { kind = "report", path = prov_path }
-		end
-	end
+	-- atoms.sourcemap.txt + atoms.provenance.txt content moved to report.lua via `<module>.atoms.md` markdown file.
+	-- This pass emits only the post-link gdb_runtime artifact (see emit_gdb_runtime below).
 
 	-- Optionally emit the gdb-runtime form (post-link, one file per build).
 	if ctx.flags and ctx.flags.gdb_runtime then

scripts/passes/components.lua

@@ -4,12 +4,11 @@
 --- Scanner owns `declaration_comment` and `debug_skip` on each declaration record; this pass projects both forward.
 ---
 --- 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_ MipsAtom ac_X(...)` declaration via a backward walk.
+--- 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.
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
+--- 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).
+--- The directory itself is the namespace, so the filename does not repeat the module name.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
@@ -29,7 +28,7 @@ local duffle          = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- Atom component declaration identifiers.
 local ATOM_COMP_PROC  = "MipsAtomComp_Proc_"
-local MIPS_ATOM       = "MipsAtom"  -- prefix on the function declaration that wraps an AtomComp_Proc_
+local MIPS_ATOM       = "Slice_MipsCode"  -- prefix on the function declaration that wraps an AtomComp_Proc_
 
 -- Component-name prefixes.
 local AC_PREFIX       = "ac_"  -- arg to MipsAtomComp_(ac_X); the X is the atom name
@@ -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 text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
---- @field out_root           string                 -- output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
+--- @field sources            SourceFile[]           -- All source files in the build
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
+--- @field shared             table                  -- Cross-pass shared state
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
+--- @field project_root       string                 -- Project root (e.g. "code/")
+--- @field upstream           table<string, table>   -- Per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @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 body       string     -- brace-delimited body (without the braces)
---- @field args       string|nil -- function-args string (function form only)
---- @field line       integer    -- source line of the declaration
---- @field comment    string|nil -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @field name       string     -- Atom name (without `ac_` prefix)
+--- @field body       string     -- Brace-delimited body (without the braces)
+--- @field args       string|nil -- Function-args string (function form only)
+--- @field line       integer    -- Source line of the declaration
+--- @field comment    string|nil -- 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)
@@ -97,9 +97,9 @@ local M = {}
 --- Returns the args string (e.g., `"U4 off, U4 code, U1 r, U1 g, U1 b"`) or nil if no function declaration is found.
 ---
 --- Convention: function form is
----   `FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
+---   `FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
 --- We find the LAST occurrence of `"ac_X("` before `before_pos` and extract the args from inside the parens.
---- We then verify the preceding context ends with `MipsAtom`
+--- We then verify the preceding context ends with `Slice_MipsCode`
 --- (the function-decl keyword with possible qualifiers between).
 ---
 --- @param source     string
@@ -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.
---- The pre-rework convention uses the *directory* basename (not the source file basename)
----  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
---- This matches what the C codebase #includes.
---- @param src SourceFile
+--- Compute the per-directory output path for `.macs.h`.
+---  e.g. any source in `code/duffle/` produces `code/duffle/gen/macs.h` regardless of source filename.
+--- The directory name is the namespace; the filename does not repeat it.
+--- @param dir string  -- the absolute source directory
 --- @return string  -- the output directory
 --- @return string  -- the full output path
-local function compute_macs_h_path(src)
-	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
-	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+local function compute_macs_h_path(dir)
+	local  out_dir  = dir .. "/" .. GEN_SUBDIR
+	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 components Component[]
+--- @param dir        string                  -- the absolute source directory
+--- @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 lines             = header_boilerplate(src)
+	local out_dir, out_path = compute_macs_h_path(dir)
+	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 `debug_skip` field mirrors the scanner-owned declaration record (`c.debug_skip`).
+--- The pass does NOT write to `ctx.shared.components`.
 --- 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
-		-- project_components reads from src.scan + does backward lookups on src.text
-		local components = project_components(src.text, src.scan)
-		if #components > 0 then
-			-- Compute all component word counts once per source.
-			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
-			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
-			local counts = count_all_components(components, corpus.word_counts)
-			local macs_path = emit_component_macros_h(ctx, src, components, counts)
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/macs.h`.
+	-- The directory itself is the namespace. `corpus.sources_by_dir` preserves source-order within each bucket (matches `corpus.source_order`).
+	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
+		-- Aggregate components from every source in this directory.
+		-- `project_components` returns nil for sources with no `MipsAtomComp_` declarations; we skip those.
+		local aggregated_components = {}
+		local metadata_per_source   = {}
+		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).
-				update_canonical_word_counts(corpus, components, counts)
-				update_canonical_components(corpus, src, components)
-				update_canonical_component_body_index(corpus, src, components, src.scan)
+				-- Populate the projections AFTER disk emission (byte-identical `.macs.h` contract).
+				update_canonical_word_counts(corpus, aggregated_components, counts)
+				for _, src in ipairs(sources) do
+					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

@@ -71,9 +71,11 @@ local DW_LNE_set_address  = DWARF_LINE_OPS.DW_LNE_set_address
 local DW_RLE_end_of_list   = DWARF5_RNGLISTS.end_of_list
 local DW_RLE_start_length  = DWARF5_RNGLISTS.start_length
 
--- File index 11 in the existing main line unit is hello_gte_tape.c.
--- The injector extends that unit rather than appending an unreferenced unit.
-local ATOM_SOURCE_FILE_INDEX = 11
+-- 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`).
+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
 
 -- RR_<R_Name> debug-visible variables come from the merged register_alias_registry filtered to aliases whose code is a valid MIPS GPR 0..31
 -- (see collect_per_source_registries + by_alias in build_inserted_children).
@@ -98,9 +100,8 @@ local ABBREV_INLINED_SUBROUTINE  = 0x6C   -- 108: DW_TAG_inlined_subroutine with
 -- (each field transitions from tape memory to GPR at load_pc + 8 = MIPS I load-delay slot boundary).
 local ABBREV_BIND_VAR_LOCLIST = 0x6D   -- 109: DW_TAG_variable no children + DW_AT_type = ref4 + DW_AT_location = sec_offset
 -- Typed-view pointer_type (for the synthetic V4_S2* / V3_S2* / U4* / void* chains).
--- MUST be a fresh abbrev code in the appended table — emitting uleb128(9) collides with GCC's
--- existing abbrev 9 (a pointer_type that carries DW_AT_byte_size + DW_AT_type), so gdb misparses
--- our 4-byte ref4 as (byte_size, type[0..2]) and lands the cursor mid-attribute.
+-- MUST be a fresh abbrev code in the appended table — emitting uleb128(9) collides with GCC's existing abbrev 9
+-- (a pointer_type that carries DW_AT_byte_size + DW_AT_type), so gdb misparses our 4-byte ref4 as (byte_size, type[0..2]) and lands the cursor mid-attribute.
 local ABBREV_TYPED_VIEW_POINTER = 0x6E   -- 110: DW_TAG_pointer_type no children + DW_AT_type = ref4 (typed-view / U4 / void chain)
 
 -- DWARF5 §7.7.3 loclist opcodes.
@@ -153,44 +154,66 @@ local DW_AT_inline           = 0x20   -- DWARF5 §7.7.1: DW_AT_inline (used by a
 local DW_AT_decl_file        = 0x3A   -- DWARF5 §7.7.1: DW_AT_decl_file (1-based file index into the CU's file table)
 local DW_AT_decl_line        = 0x3B   -- DWARF5 §7.7.1: DW_AT_decl_line
 
--- File index lookup table for the existing main line unit (Unit 2).
--- Provenance paths come back with mixed slashes; we normalize to basename and look up against the line unit's actual file table.
--- Current scope has two provenance basenames: hello_gte_tape.c (the atom's call site) and lottes_tape.h (the component definition).
--- Both live in the existing gcc-generated line unit; their 1-based indices are stable across rebuilds because the include order
--- in code/gte_hello/hello_gte.c determines the unit's file table.
--- gcc only adds a file to the line table when it has actual line-number entries;
--- headers that are pure macros/typedefs (dsl.h, memory.h, math.h, mips.h, gp.h, gte.h, etc.) never appear.
--- lottes_tape.h is the FIRST include that emits line entries (MipsAtomComp_ declarations), so it is the FIRST entry after the primary file.
-local PROVENANCE_BASENAME_TO_FILE_INDEX = {
-	["hello_gte_tape.c"] = ATOM_SOURCE_FILE_INDEX,  -- = 11
-	["lottes_tape.h"]    = 2,
-}
+-- Replaced the hardcoded `ATOM_SOURCE_FILE_INDEX = 11` and the `PROVENANCE_BASENAME_TO_FILE_INDEX` table below with a runtime lookup
+-- (`init_file_index_lookup` + `resolve_provenance_file_index`) that reads the actual `.debug_line` file table from the post-link ELF.
+
+--- Populate the module-level file-index lookup table from the `.debug_line` section of the post-link ELF pointed at by `elf_path`.
+--- This MUST be called exactly once at pass start (from `M.run`) before any `resolve_provenance_file_index` invocation;
+--- downstream callers handle a nil table as "no file info available; fall back to errors".
+---
+--- The lookup uses `elf_dwarf.read_line_unit_file_table` (which parses both DWARF3 and DWARF5 line-program units — 
+--- the crt0.s assembler-side DWARF5 unit may emit non-standard form codes for paths and is intentionally skipped).
+--- @param elf_path string|nil
+local function init_file_index_lookup(elf_path)
+	if not elf_path or elf_path == "" then return end
+	local b2i, _basenames, paths = elf_dwarf.read_line_unit_file_table(elf_path)
+	if type(b2i) ~= "table" or type(paths) ~= "table" then
+		io.stderr:write("[dwarf_injection] read_line_unit_file_table returned no file table for: " .. tostring(elf_path) .. "\n")
+		return
+	end
+	_file_index_by_basename = b2i
+	_file_path_by_index    = paths
+	-- Pick any valid index for the opaque-row fallbacks at lines 466 + 570
+	-- (both sites legitimately want "any file index"; gdb resolves whatever index we emit to whatever that file's line happens to be).
+	for idx in pairs(paths) do
+		_default_atom_source_index = idx
+		break
+	end
+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 known file table.
+--- 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 requires extending
---- `PROVENANCE_BASENAME_TO_FILE_INDEX` so the line-program emission contract stays explicit.
---- Silent fallback to ATOM_SOURCE_FILE_INDEX would mask the new-file case by misattributing component rows to the atom's source file.
---- @param path string  -- absolute provenance path (e.g. "C:/.../lottes_thttps://www.youtube.com/watch?v=ORM4yLkdKx8ape.h" or "C:\\...\\lottes_tape.h")
---- @return integer  -- 1-based line-unit file index
+--- Returns 0 (the DWARF `set_file(0)` "no file change" sentinel) when the basename is not in the file table.
+--- 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).
+--- 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, 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. Is M.run being entered correctly (with --elf)?")
+	end
 	if path == nil or path == "" then
 		error("[dwarf_injection] resolve_provenance_file_index: empty path")
 	end
-	-- Normalize backslashes → forward slashes (paths arrive with mixed separators from the provenance file: forward slashes from Lua's io.lines;
-	-- backslashes if the input ever round-trips through Windows shell expansion).
+	-- Normalize backslashes → forward slashes (paths arrive with mixed separators from the provenance file).
 	local normalized = path:gsub("\\", "/")
 	-- Take the last path component (the basename).
 	local basename = normalized:match("([^/]+)$") or normalized
-	local idx      = PROVENANCE_BASENAME_TO_FILE_INDEX[basename]
-	if idx == nil then
-		error(string.format(
-			"[dwarf_injection] resolve_provenance_file_index: unknown provenance basename '%s' (from '%s'). "
-			.. "Extend PROVENANCE_BASENAME_TO_FILE_INDEX in passes/dwarf_injection.lua.",
-			basename, path))
+	local idx      = _file_index_by_basename[basename]
+	if idx ~= nil then return idx end
+	-- Last-resort exact-path match (handles paths that don't reduce to a known basename).
+	for i, p in pairs(_file_path_by_index) do
+		if p and p:gsub("\\", "/") == normalized then return i end
 	end
-	return idx
+	-- File is in the corpus but gcc omitted it from the .debug_line file table (data-only content).
+	-- Return 0 = DWARF `set_file(0)` no-change sentinel so the line program keeps its prior file state.
+	io.stderr:write(string.format("[dwarf_injection] line-table miss: '%s' (basename '%s') not in .debug_line file table; "
+		.. "falling back to set_file(0)\n", path, basename))
+	return 0
 end
 
 local DW_FORM_addr           = 0x01
@@ -205,7 +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.)
@@ -461,7 +483,7 @@ local function build_atom_sequence(atom)
 	if atom.debug_skip then
 		return table.concat({
 			set_address(atom.addr),
-			set_file(ATOM_SOURCE_FILE_INDEX),
+			set_file(resolve_provenance_file_index(atom.src_path)),
 			advance_line(atom.entries[1].line - 1),
 			negate_stmt(),
 			copy_op(),
@@ -508,10 +530,9 @@ local function build_atom_sequence(atom)
 	--   * If the invocation's body has any NESTED invocations (parent_id == top_inv.id), the body's
 	--     first content is the call_line of the earliest nested invocation (by start_pos).
 	--   * Otherwise (only RAW words in the body), it's the line of the first raw word = body_lines[1].
-	-- This is the value the multi-row PC's body_lines[1] row must reference for source-order display:
-	-- `anc.body_lines[1]` is the line of the FIRST WORD (which for an outer whose body starts with a
-	-- nested expansion is inside the inner's expansion = wrong for display purposes); `anc.body_first_line`
-	-- is the body's first content line in the parent's source (= correct for display).
+	-- This is the value the multi-row PC's body_lines[1] row must reference for source-order display: `anc.body_lines[1]` is the line of the FIRST WORD 
+	-- (which for an outer whose body starts with a nested expansion is inside the inner's expansion = wrong for display purposes);
+	-- `anc.body_first_line` is the body's first content line in the parent's source (= correct for display).
 	local body_first_line_of = {}
 	for _, top_inv in ipairs(invs) do
 		local earliest_nested_call_line = nil
@@ -565,7 +586,7 @@ local function build_atom_sequence(atom)
 		parts[#parts + 1] = copy_op()
 	end
 
-	local call_file_idx = ATOM_SOURCE_FILE_INDEX
+	local call_file_idx = resolve_provenance_file_index(atom.src_path)
 
 	-- --- Atom entry (idx 1) -------------------------------------------------
 	local entry_1         = atom.entries[1]
@@ -574,13 +595,11 @@ local function build_atom_sequence(atom)
 	-- If atom entry 1 starts inside an invocation, walk the ancestry and emit a call-site row + (when applicable)
 	-- a body_lines[1] row for every active ancestor. For a non-nested invocation this is just the one pair;
 	-- for nested invocations this emits the outer call-site + body_lines[1] rows BEFORE the inner pair so the debugger displays
-	-- the outer body line at the inner's first word
-	-- (PROBLEM B fix).
+	-- the outer body line at the inner's first word (PROBLEM B fix).
 	--
-	-- A marked OUTERMOST ancestor's body_lines[1] row is suppressed at this PC (the existing full-skip
-	-- contract is preserved for the marked outer range); its call-site row IS still emitted as a
-	-- statement. Marked INNER ancestors always emit their body_lines[1] row with is_stmt=false
-	-- (the per-invocation `want_body = not inv.debug_skip` predicate).
+	-- A marked OUTERMOST ancestor's body_lines[1] row is suppressed at this PC (the existing full-skip contract is preserved for the marked outer range);
+	-- Its call-site row IS still emitted as a statement.
+	-- Marked INNER ancestors always emit their body_lines[1] row with is_stmt=false (the per-invocation `want_body = not inv.debug_skip` predicate).
 	if #entry_1_ancestry == 0 then
 		-- RAW word at atom entry: single call-site row, always a statement target.
 		emit_row(call_file_idx, entry_1.line, true)
@@ -588,15 +607,12 @@ local function build_atom_sequence(atom)
 		-- Atom starts in an invocation. Walk the ancestry outermost-first.
 		-- Each ancestor emits one call-site row (statement) and one body_lines[1] row
 		-- (statement iff unmarked; suppressed for marked outermost).
-		-- 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).
+		-- 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).
 		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
-				, "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.body_lines[1] ~= nil,              "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.")
 			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
@@ -615,28 +631,22 @@ local function build_atom_sequence(atom)
 
 		if inv and idx == inv.start_pos + 1 then
 			-- First word of the innermost active invocation (PROBLEM B fix — nested-display rule).
-			-- Walk the active ancestry outermost-first; for each ancestor emit a call-site row
-			-- (statement) + a body_lines[1] row. The inner-most invocation's call-site + body pair
-			-- become the LAST two rows in the sequence. Marked outermost ancestors suppress their
-			-- body_lines[1] row at this PC (the existing full-skip contract is preserved for the
-			-- marked outer range); all OTHER ancestors emit body_lines[1] with is_stmt = not debug_skip.
+			-- Walk the active ancestry outermost-first; for each ancestor emit a call-site row (statement) + a body_lines[1] row.
+			-- The inner-most invocation's call-site + body pair become the LAST two rows in the sequence.
+			-- Marked outermost ancestors suppress their body_lines[1] row at this PC (the existing full-skip contract is preserved for the marked outer range);
+			-- 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 (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, so the disc=1 row's
-			-- value matters for what's shown when stepping into the nested case).
+			-- 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.
+			-- 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,
+			-- so the disc=1 row's value matters for what's shown when stepping into the nested case).
 			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
-					, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d",
-						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.")
+				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 {})))
+				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
@@ -647,15 +657,12 @@ local function build_atom_sequence(atom)
 			-- Subsequent body word of the innermost active invocation: `body_lines[k]` is indexed by the 1-based offset of this word inside the invocation.
 			-- Both `idx` (1-based DWARF entry index) and `inv.start_pos` (0-based emitted-word position stamped at `emit_invoke_begin`) come from the same
 			-- monotonic counter, so `idx - inv.start_pos` is exactly the 1-based k (the first word of the invocation has `idx == inv.start_pos + 1`, hence `k == 1`).
-			-- atom_dbg_step_ux_20260725: `want_body = not inv.debug_skip`. The previous `want = not marked_idx[idx]`
-			-- (which suppressed ALL body rows when any ancestor was marked) is replaced by the per-invocation
-			-- predicate. Marked invocations emit non-statement body rows at every body word; unmarked
-			-- invocations emit statement body rows.
+			-- atom_dbg_step_ux_20260725: `want_body = not inv.debug_skip`.
+			-- The previous `want = not marked_idx[idx]` (which suppressed ALL body rows when any ancestor was marked) is replaced by the per-invocation predicate.
+			-- 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
-				, 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))
+			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))
 			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).
@@ -704,7 +711,9 @@ local function build_atom_table(corpus, addrs)
 	local atoms_by_name = corpus.atoms_by_name or {}
 
 	-- Per-atom ingest. Returns nil if the atom is absent from the corpus; the caller skips it via the `if atom then ...` guard.
-	local function ingest_atom(name, info)
+	-- `src_path` is the absolute source path that declared this atom; the build_atom_table iteration below threads `src.path` through.
+	-- This is consumed by `build_atom_sequence::set_file(...)` for opaque-row fallbacks + raw-word rows (atoms where no invocation ancestry exists).
+	local function ingest_atom(name, info, src_path)
 		local atom_record = atoms_by_name[name]
 		if not atom_record then return nil end
 
@@ -730,6 +739,7 @@ local function build_atom_table(corpus, addrs)
 			words      = #word_events,
 			entries    = entries,
 			debug_skip = atom_record.debug_skip == true,
+			src_path   = src_path or "",
 		}
 
 		-- Consume invocation records from `atom.paths.invocations`. It is the single producer of per-invocation body_lines, per-invocation debug_skip,
@@ -753,10 +763,26 @@ local function build_atom_table(corpus, addrs)
 	end
 
 	local out = {}
-	for name, info in pairs(addrs) do
-		local atom = ingest_atom(name, info)
+	-- 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 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).
+	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
+			local info = addrs[atom_rec.name or atom_rec.raw_name]
+			if info then
+				local atom = ingest_atom(atom_rec.name or atom_rec.raw_name, info, src_path)
 				if atom then out[#out + 1] = atom end
 			end
+		end
+		for _, atom_rec in ipairs(((src.scan or {}).raw_atoms) or {}) do
+			local info = addrs[atom_rec.name or atom_rec.raw_name]
+			if info then
+				local atom = ingest_atom(atom_rec.name or atom_rec.raw_name, info, src_path)
+				if atom then out[#out + 1] = atom end
+			end
+		end
+	end
 	table.sort(out, function(a, b) return a.addr < b.addr end)
 	return out
 end
@@ -800,12 +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
---- statements (each entry is a single `load_word(...)` call or other statement).
+--- 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).
 --- @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
@@ -813,14 +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*%((.*)%)$"
 	for _, t in ipairs(body_tokens or {}) do
 		local tok         = duffle.trim(t.tok or "")
-		-- Match "load_word(...)" — the entire call is one body_tokens entry.
-		local inner = tok:match("^load_word%s*%((.*)%)$")
-		if inner then
+		local kind, inner = tok:match(load_pattern)
+		if kind then
 			local args = duffle.split_top_level_commas(inner)
-			-- Expected shape: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
-			if #args >= 3 then
+			-- Expected shape for an rbind piece-chain load: (R_<reg>, R_TapePtr, O_(Binds_<X>, FieldName))
+			-- 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)
@@ -839,9 +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}}
@@ -941,7 +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
@@ -952,9 +983,7 @@ local function build_dwarf_line_section(existing, atom_table)
 
 	-- Build the sequences.
 	local sequences = {}
-	for _, atom in ipairs(atom_table) do
-		sequences[#sequences + 1] = build_atom_sequence(atom)
-	end
+	for _, atom in ipairs(atom_table) do sequences[#sequences + 1] = build_atom_sequence(atom) end
 	local appended = table.concat(sequences)
 
 	-- Walk DWARF32 line units and retain the final unit's bounds.
@@ -963,7 +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
@@ -1024,7 +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
@@ -1488,13 +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
---- and preserves existing DIE bytes exactly.
+--- This function does NOT emit the final 0 byte (root terminator).
+--- 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)
@@ -1614,6 +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 },
@@ -1866,6 +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]
@@ -1946,6 +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
@@ -2058,7 +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.
@@ -2177,7 +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",
@@ -2186,12 +2218,15 @@ function M.run(ctx)
 		-- reading them just returns "" which is the "missing" case the builder handles.
 		".debug_loc", ".debug_loclists",
 	})
+	-- Resolve the per-file line-table indices from the same .debug_line bytes;
+	-- this MUST run before any atom sequence is emitted (build_atom_sequence below
+	-- calls resolve_provenance_file_index when populating call-site / body rows).
+	init_file_index_lookup(elf_path)
 	-- Skip state lives in `corpus.atoms_by_name[*].debug_skip` (whole-atom) and `atom.paths.invocations[*].debug_skip` (per-invocation).
 	-- `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
-	-- them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
+	-- 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.
 	-- 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)
-		or atom_record.line or 0
+	local root_body_line  = root_line_of(atom_record.body_off - 1) or atom_record.line or 0
 	local component_index = corpus.component_body_index or {}
 	local word_items      = {}
 

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
--- both standalone and when require'd from the orchestrator.
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
@@ -39,17 +37,17 @@ local OFFSET_MACRO_COL = 44
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field shared             table                  -- cross-pass shared state
---- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared             table -- Cross-pass shared state
+--- @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 target      string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field branch_word integer  -- branch word position within the atom body
---- @field offset      integer  -- computed `target_word - branch_word - 1`
+--- @field tag               string      -- Marker tag (e.g. "F" 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 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 total_words integer        -- total word count of the atom body
---- @field offsets     BranchOffset[] -- per-branch offset list
+--- @field name        string         -- Atom name
+--- @field total_words integer        -- Total word count of the atom body
+--- @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.
---- @param source_path string
+--- Generate the per-directory .offsets.h header.
+--- @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 basename = duffle.basename_no_ext(source_path)
+local function generate_header(dir, sources, atoms_data)
+	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`
---- containing constants for every marker recorded in atom.paths.
+--- For each canonical source-directory, emits a per-directory `gen/offsets.h`
+--- 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
-		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	if type(corpus) ~= "table" then
+		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
-		local out_path = process_source(ctx, src)
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`.
+	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

@@ -1,26 +1,21 @@
---- passes/report.lua — Per-MODULE annotation report renderer +
---- project-wide summary writer.
+--- passes/report.lua — Per-MODULE annotation report renderer + project-wide summary writer.
 ---
 --- Two output files per build:
 ---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
 ---   - `build/gen/annotation_validation.txt` — the project summary.
 ---
 --- The annotation pass 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.
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
+--- 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: 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: See `ps1_meta.lua` for the rationale.
+-- Bootstrap: Load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
+-- Bootstrap: Load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
@@ -30,6 +25,11 @@ local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- that the report pass renders into the per-module `<dir_basename>.annotations.txt` output.
 local annotation = dofile(_bootstrap_dir .. "annotation.lua")
 
+-- Load atoms_source_map for the `render_source_map` / `render_provenance` module functions (used by `render_module_atoms_md` to produce `<module>.atoms.md` without re-walking source tokens).
+-- The pass itself emits no per-source files anymore; we only consume the two pure renderers here.
+-- Defined BEFORE the renderer functions below so their upvalues resolve to this local (not the global `atoms_source_map`, which is nil).
+local atoms_source_map = dofile(_bootstrap_dir .. "atoms_source_map.lua")
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
@@ -44,8 +44,7 @@ local SECTION_HEADER_MACROS    = "── Macro word-count declarations ───
 local SECTION_HEADER_ERRORS    = "── Errors ──────────────────────────────────────────────"
 local SECTION_HEADER_WARNINGS  = "── Warnings ────────────────────────────────────────────"
 
--- Lua pattern that captures the basename (last path segment) of a
--- forward- or back-slash separated path.
+-- Lua pattern that captures the basename (last path segment) of a forward- or back-slash separated path.
 local BASENAME_PATTERN = "([^/\\]+)$"
 
 -- Debug flag name — set to truthy in `_G` to enable verbose logging.
@@ -59,20 +58,20 @@ local PASS_NAME = "report"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field sources            SourceFile[]         -- all source files in the build
---- @field metadata_path      string               -- path to word_count.metadata.h
---- @field shared             table                -- cross-pass shared state
---- @field out_root           string               -- output root (e.g. "build/gen")
---- @field project_root       string               -- project root (e.g. "code/")
---- @field upstream           table<string, table> -- per-pass upstream outputs
+--- @field sources            SourceFile[]         -- All source files in the build
+--- @field metadata_path      string               -- Path to word_count.metadata.h
+--- @field shared             table                -- Cross-pass shared state
+--- @field out_root           string               -- Output root (e.g. "build/gen")
+--- @field project_root       string               -- Project root (e.g. "code/")
+--- @field upstream           table<string, table> -- Per-pass upstream outputs
 --- @field flags              table                -- CLI flags + per-pass stash
---- @field 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 line integer  -- source line of the atom declaration
+--- @field name string   -- Atom name (e.g. "cube_g4_face")
+--- @field line integer  -- Source line of the atom declaration
 
 --- @class AnnotEntry
---- @field line    integer  -- source line
---- @field macro   string   -- the macro name (e.g. "atom_reads")
---- @field name    string   -- the atom name (if a `name(...)` was given)
+--- @field line    integer     -- Source line
+--- @field macro   string      -- Macro name (e.g. "atom_reads")
+--- @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 offset integer  -- byte offset within the Binds_X struct
+--- @field name   string   -- Field name
+--- @field offset integer  -- Byte offset within the Binds_X struct
 
 --- @class BindsStruct
---- @field name   string         -- struct name (e.g. "Binds_Floor")
---- @field line   integer        -- source line of the typedef
---- @field bytes  integer        -- total byte size
---- @field fields BindsField[]   -- the field list
+--- @field name   string         -- Struct name (e.g. "Binds_Floor")
+--- @field line   integer        -- Source line of the typedef
+--- @field bytes  integer        -- Total byte size
+--- @field fields BindsField[]   -- The field list
 
 --- @class MacroEntry
---- @field name  string   -- macro name (e.g. "WORD_COUNT(my_macro, 4)")
---- @field line  integer  -- source line
---- @field words integer  -- declared word count
+--- @field name  string   -- Macro name (e.g. "WORD_COUNT(my_macro, 4)")
+--- @field line  integer  -- Source line
+--- @field words integer  -- Declared word count
 
 --- @class Finding
---- @field line integer  -- source line
---- @field msg  string   -- finding message
+--- @field line integer  -- Source line
+--- @field msg  string   -- Finding message
 
 --- @class AnnotationResult
---- @field source   string         -- set by this pass; original source path
---- @field atoms    AtomEntry[]    -- atom declarations in this source
---- @field annots   AnnotEntry[]   -- annotation entries
---- @field macros   MacroEntry[]   -- macro word-count declarations
+--- @field source   string         -- Set by this pass; original source path
+--- @field atoms    AtomEntry[]    -- Atom declarations in this source
+--- @field annots   AnnotEntry[]   -- Annotation entries
+--- @field macros   MacroEntry[]   -- Macro word-count declarations
 --- @field binds    BindsStruct[]  -- Binds_* struct declarations
---- @field errors   Finding[]      -- errors from validation
---- @field warnings Finding[]      -- warnings from validation
---- @field info     table          -- info summary (not rendered here)
+--- @field errors   Finding[]      -- Errors from validation
+--- @field warnings Finding[]      -- Warnings from validation
+--- @field info     table          -- Info summary (not rendered here)
 
 --- @class ModuleEntry
---- @field dir          string   -- absolute directory path
---- @field dir_basename string   -- basename (e.g. "duffle", "gte_hello")
---- @field atoms_count  integer  -- pre-counted atoms for filtering
+--- @field dir          string   -- Absolute directory path
+--- @field dir_basename string   -- Basename (e.g. "duffle", "gte_hello")
+--- @field atoms_count  integer  -- Pre-counted atoms for filtering
 
 --- @class ModuleReport
---- @field dir       string             -- module directory
---- @field sources   SourceFile[]       -- sources in this module
---- @field results   AnnotationResult[] -- per-source validate() results
+--- @field dir       string             -- Module directory
+--- @field sources   SourceFile[]       -- Sources in this module
+--- @field results   AnnotationResult[] -- Per-source validate() results
 
 --- @class ProjectReport
---- @field results AnnotationResult[]  -- all per-source results
+--- @field results AnnotationResult[]  -- All per-source results
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-MODULE annotation report (aggregated across all sources in a dir)
@@ -148,328 +147,453 @@ local function source_basename(path)
 	return path:match(BASENAME_PATTERN) or path
 end
 
---- (internal) Format a single annotation entry as one rendered line.
---- @param a AnnotEntry
---- @param src_name string
+-- ════════════════════════════════════════════════════════════════════════════
+-- Markdown renderers (consolidated-report-files refactor, 2026-07-26)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render the thin project-wide summary (`build/atom_meta_report.summary.md`).
+--- @param all_results { 
+--- 	module:string,
+--- 	atoms:integer, 
+--- 	annots:integer, 
+--- 	binds:integer,
+---		macros:integer,   
+---		findings:integer,
+---		errors:integer,
+--- 	warnings:integer,
+--- 	info:integer }[]
 --- @return string
-local function format_annot_line(a, src_name)
-	if a.error then
-		return string.format("  ✗ line %d  %s  [ERROR: %s]  [%s]", a.line, a.macro or "?", a.error, src_name)
+local function render_project_summary(all_results)
+	local lines = {
+		"# Project summary",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua).",
+		"",
+		"| module | atoms | annots | binds | macros | findings | errors | warnings | info |",
+		"|--------|-------|--------|-------|--------|----------|--------|----------|------|",
+	}
+	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0, findings = 0, errors = 0, warnings = 0, info = 0 }
+	for _, e in ipairs(all_results) do
+		lines[#lines + 1] = string.format("| %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)
+		totals.atoms    = totals.atoms    + e.atoms
+		totals.annots   = totals.annots   + e.annots
+		totals.binds    = totals.binds    + e.binds
+		totals.macros   = totals.macros   + e.macros
+		totals.findings = totals.findings + e.findings
+		totals.errors   = totals.errors   + e.errors
+		totals.warnings = totals.warnings + e.warnings
+		totals.info     = totals.info     + e.info
 	end
-	local line = string.format("  ●  line %d  %s  [%s]", a.line, a.name, src_name)
-	if a.binds       then line = line .. "  binds="   .. a.binds end
-	if #a.reads  > 0 then line = line .. "  reads={"  .. table.concat(a.reads,  ",") .. "}" end
-	if #a.writes > 0 then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
-	return line
+	lines[#lines + 1] = string.format("| **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)
+	return table.concat(lines, "\n") .. "\n"
 end
 
---- (internal) Tally totals across all results in a module.
---- @param results AnnotationResult[]
---- @return integer, integer, integer, integer, integer, integer
-local function tally_module_totals(results)
-	local total_atoms, total_annots, total_binds, total_macros = 0, 0, 0, 0
-	local total_errors, total_warnings = 0, 0
-	for _, r in ipairs(results) do
-		total_atoms    = total_atoms    + #r.atoms
-		total_annots   = total_annots   + #r.annots
-		total_binds    = total_binds    + #r.binds
-		total_macros   = total_macros   + #r.macros
-		total_errors   = total_errors   + #r.errors
-		total_warnings = total_warnings + #r.warnings
+--- Render the per-module verbose source-map markdown (`build/<module>.atoms.md`).
+--- Per-source sub-section, per-atom stanza with sourcemap + provenance rows.
+--- Pulls sourcemap + provenance from `atoms_source_map` (no second source walk).
+--- @param dir         string
+--- @param dir_sources SourceFile[]
+--- @param wc          table<string, integer>
+--- @return string
+local function render_module_atoms_md(dir, dir_sources, wc)
+	local dir_basename = source_basename(dir)
+	local lines        = {
+		"# " .. dir_basename .. " — atoms (verbose source map)",
+		"> Per-word call-site + provenance. Auto-generated.",
+		"",
+	}
+	for _, src in ipairs(dir_sources) do
+		local src_name = source_basename(src.path)
+		lines[#lines + 1] = "## " .. src_name
+		lines[#lines + 1] = ""
+		-- For each atom with a projection, render its sourcemap + provenance.
+		local atoms_list = {}
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
 		end
-	return total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
+		end
+		if #atoms_list == 0 then
+			lines[#lines + 1] = "_(no atom projections)_"
+			lines[#lines + 1] = ""
+		else
+			-- Per-source forward-slash path (same one `emit_atom_stanza` / `emit_provenance_stanza` would derive;
+			-- computed once per `## <source>` heading and reused by each atom's `WORD N CALL ...` field).
+			local rel_path = src.path:gsub("\\\\", "/")
+			for _, atom in ipairs(atoms_list) do
+				lines[#lines + 1] = string.format(
+					"### atom: %s (line %d, %d words)",
+					atom.name, atom.line or 0, #(atom.paths.items or {}))
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Sourcemap** — per-word call site:"
+				lines[#lines + 1] = "```"
+				-- Per-atom invariant: call the per-atom renderers, NOT the per-source ones.
+				-- The per-source renderers enumerate every atom in `src`;
+				-- calling them in a per-atom loop would repeat the whole source under every `### atom:` heading.
+				lines[#lines + 1] = atoms_source_map.render_atom_source_map(atom):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Provenance** — per-word definition + body:"
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = atoms_source_map.render_atom_provenance(atom, wc, rel_path):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+			end
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
 end
 
--- (internal) Section renderer: per-source atom declarations.
-local function render_module_atoms_section(add, results)
-	add(SECTION_HEADER_ATOMS)
-	for _, r in ipairs(results) do
+--- Render the consolidated per-module markdown (`build/<module>.atom_meta_report.md`).
+--- Aggregates annotation + static-analysis content across all sources in `dir`.
+--- Annotations come from re-running `annotation.validate()` per source (the existing pattern);
+--- static-analysis comes from `corpus.static_analysis_results[dir_basename]` (populated by `static_analysis.lua` — no second corpus_pipe_ctx build).
+--- @param dir           string
+--- @param dir_sources   SourceFile[]
+--- @param annot_results AnnotationResult[]
+--- @param sa_results    table -- corpus.static_analysis_results[dir_basename]
+--- @return string
+local function render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+	local dir_basename = source_basename(dir)
+	local lines        = {
+		"# " .. dir_basename .. " — atom meta report",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua). Do not edit.",
+		"",
+	}
+	local function add(s) lines[#lines + 1] = s end
+
+	-- Module summary table.
+	local n_atoms        = 0
+	local n_annot        = 0
+	local n_binds        = 0
+	local n_macros       = 0
+	local n_bare, n_proc = 0, 0
+	for _, r in ipairs(annot_results) do
+		n_atoms  = n_atoms  + #r.atoms
+		n_annot  = n_annot  + #r.annots
+		n_binds  = n_binds  + #r.binds
+		n_macros = n_macros + #r.macros
+	end
+	for _, a in ipairs(sa_results.atoms or {}) do
+		if     a.kind == "comp_bare" then n_bare = n_bare + 1
+		elseif a.kind == "comp_proc" then n_proc = n_proc + 1
+		end
+	end
+
+	add("## Module summary"); add("")
+	add("| metric | value |"); add("|--------|-------|")
+	add(string.format("| sources | %d |", #dir_sources))
+	add(string.format("| atoms | %d (atoms: %d, comp_bare: %d, comp_proc: %d) |",
+		#(sa_results.atoms or {}),
+		#(sa_results.atoms or {}) - n_bare - n_proc, n_bare, n_proc))
+	add(string.format("| annotations | %d |", n_annot))
+	add(string.format("| binds structs | %d |", n_binds))
+	add(string.format("| macro decls | %d |", n_macros))
+	add(string.format("| findings | %d (errors: %d, warnings: %d, info: %d) |",
+		#(sa_results.findings or {}),
+		#(sa_results.errors   or {}),
+		#(sa_results.warnings or {}),
+		#(sa_results.info     or {})))
+	add("")
+
+	-- Sources
+	add("## Sources"); add("")
+	for _, s in ipairs(dir_sources) do add("- `" .. s.path .. "`") end
+	add("")
+
+	-- Atoms (annotation)
+	add("## Atoms"); add("")
+	add("| kind | name | source | line |"); add("|------|------|--------|------|")
+	for _, r in ipairs(annot_results) do
 		local src_name = source_basename(r.source)
 		for _, a in ipairs(r.atoms) do
-			add(string.format("  MipsAtom_(%s)   line %d   [%s]", a.name, a.line, src_name))
+			add(string.format("| atom | %s | %s | %d |", a.name, src_name, a.line))
 		end
 	end
 	add("")
-end
 
--- (internal) Section renderer: per-source annotation entries.
-local function render_module_annots_section(add, results)
-	add(SECTION_HEADER_ANNOTS)
-	for _, r in ipairs(results) do
+	-- Annotations
+	add("## Annotations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | name | binds | reads | writes |")
+		add("|--------|------|------|-------|-------|--------|")
+		for _, r in ipairs(annot_results) do
 			local src_name = source_basename(r.source)
 			for _, a in ipairs(r.annots) do
-			add(format_annot_line(a, src_name))
+				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 |"
+					, src_name, a.line, a.name, binds, reads, writes))
+			end
 		end
 	end
 	add("")
-end
 
--- (internal) Section renderer: per-source Binds_* struct declarations.
-local function render_module_binds_section(add, results)
-	add(SECTION_HEADER_BINDS)
-	for _, r in ipairs(results) do
+	-- Binds_* structs
+	add("## Binds_* structs"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		for _, r in ipairs(annot_results) do
 			local src_name = source_basename(r.source)
 			for _, b in ipairs(r.binds) do
-			add(string.format("  %s   line %d   %d bytes   [%s]", b.name, b.line, b.bytes, src_name))
+				add(string.format("### %s (%s:%d, %d bytes)",
+					b.name, src_name, b.line, b.bytes))
 				for _, f in ipairs(b.fields) do
-				add(string.format("    +%2d: %s", f.offset, f.name))
-			end
-		end
+					add(string.format("- `+%d %s`", f.offset, f.name))
 				end
 				add("")
 			end
+		end
+	end
 
--- (internal) Section renderer: per-source macro word-count declarations.
-local function render_module_macros_section(add, results)
-	add(SECTION_HEADER_MACROS)
-	for _, r in ipairs(results) do
+	-- Macro decls
+	add("## Macro word-count declarations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | macro declaration |")
+		add("|--------|------|-------------------|")
+		for _, r in ipairs(annot_results) do
 			local src_name = source_basename(r.source)
 			for _, m in ipairs(r.macros) do
-			add(string.format("  %s  line %d  words=%d  [%s]", m.name, m.line, m.words, src_name))
+				add(string.format("| %s | %d | %s |",
+					src_name, m.line, m.name))
+			end
 		end
 	end
 	add("")
-end
 
--- (internal) Section renderer: per-source errors (one-line + "(none)" if empty).
-local function render_module_errors_section(add, results, total_errors)
-	add(SECTION_HEADER_ERRORS)
-	if total_errors == 0 then
-		add("  (none)")
+	-- Findings by atom (static-analysis)
+	add("## Static analysis — findings by atom"); add("")
+	local by_atom = {}
+	for _, f in ipairs(sa_results.findings or {}) do
+		by_atom[f.atom] = by_atom[f.atom] or {}
+		by_atom[f.atom][#by_atom[f.atom] + 1] = f
+	end
+	if next(by_atom) == nil then
+		add("_(no findings)_")
 	else
-		for _, r in ipairs(results) do
-			local src_name = source_basename(r.source)
-			for _, e in ipairs(r.errors) do
-				add(string.format("  ✗ line %d  %s  [%s]", e.line, e.msg, src_name))
-			end
-		end
+		for _, a in ipairs(sa_results.atoms or {}) do
+			local fs = by_atom[a.name]
+			if fs then
+				add(string.format("### %s", a.name))
+				for _, f in ipairs(fs) do
+					add(string.format("- `[%s] %s`", f.check, f.msg))
 				end
 				add("")
 			end
+		end
+	end
 
--- (internal) Section renderer: per-source warnings (one-line + "(none)" if empty).
-local function render_module_warnings_section(add, results, total_warnings)
-	add(SECTION_HEADER_WARNINGS)
-	if total_warnings == 0 then
-		add("  (none)")
+	-- Errors / Warnings / Info
+	local function add_findings(label, entries)
+		add(string.format("## %s", label))
+		if #entries == 0 then
+			add("_(none)_")
 		else
-		for _, r in ipairs(results) do
-			local src_name = source_basename(r.source)
-			for _, w in ipairs(r.warnings) do
-				add(string.format("  ⚠ line %d  %s  [%s]", w.line, w.msg, src_name))
-			end
+			for _, e in ipairs(entries) do
+				add(string.format("- line %d  %s", e.line, e.msg))
 			end
 		end
 		add("")
 	end
+	add_findings("Errors",   sa_results.errors   or {})
+	add_findings("Warnings", sa_results.warnings or {})
+	add_findings("Info",     sa_results.info     or {})
 
--- ════════════════════════════════════════════════════════════════════════════
--- SECTION_RENDERERS — data-driven section dispatch (the plex pattern)
--- ════════════════════════════════════════════════════════════════════════════
---
--- Each entry maps a section to its (header, render_fn). The render_fn signature:
---   render_fn(add, results, totals)
---     add    -- the `add(line)` closure from the surrounding report renderer
---     results -- AnnotationResult[] (per-source results)
---     totals -- {atoms, annots, binds, macros, errors, warnings} counts
---
--- Sections that need to render "(none)" vs iterate use totals.errors / totals.warnings;
--- other sections ignore the totals arg.
--- Adding a new section = 1 row here + 1 render_<thing>_section function.
-local SECTION_RENDERERS = {
-	{ header = SECTION_HEADER_ATOMS,    render = render_module_atoms_section     },
-	{ header = SECTION_HEADER_ANNOTS,   render = render_module_annots_section    },
-	{ header = SECTION_HEADER_BINDS,    render = render_module_binds_section     },
-	{ header = SECTION_HEADER_MACROS,   render = render_module_macros_section    },
-	{ header = SECTION_HEADER_ERRORS,   render = function(add, results, totals) return render_module_errors_section(add, results, totals.errors)   end },
-	{ header = SECTION_HEADER_WARNINGS, render = function(add, results, totals) return render_module_warnings_section(add, results, totals.warnings) end },
-}
-
---- Render the per-MODULE annotation report (one `<dir_basename>.annotations.txt`).
---- @param dir string                  -- module directory path
---- @param sources SourceFile[]        -- sources in this module
---- @param results AnnotationResult[]  -- per-source validate() results
---- @return string                     -- the rendered report text
-local function render_module_report(dir, sources, results)
-	local lines = {}
-	local function add(s) lines[#lines + 1] = s end
-
-	add(RULE_THICK)
-	add("ANNOTATION PASS — module " .. source_basename(dir))
-	add(RULE_THICK)
-	add(string.format("Sources: %d", #sources))
-	for _, s in ipairs(sources) do add("  " .. s.path) end
-	add("")
-
-	local total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings = tally_module_totals(results)
-	add(string.format("Atoms: %d   Annotations: %d   Binds structs: %d   Macro decls: %d",
-		total_atoms, total_annots, total_binds, total_macros))
-	add("")
-
-	-- Bundle the totals so the section renderers don't need separate parameter lists.
-	-- Errors/warnings sections need their total count to decide "(none)" vs iterate.
-	-- Sections without totals (atoms/annots/binds/macros) ignore this arg.
-	local totals = {
-		atoms    = total_atoms,  annots = total_annots,  binds  = total_binds,
-		macros   = total_macros, errors = total_errors,  warnings = total_warnings,
-	}
-
-	-- THE per-section dispatch. ONE loop over SECTION_RENDERERS.
-	-- Each renderer writes its header + content via the `add` closure (pre-bound above).
-	-- Adding a new section = 1 row here + 1 render_<thing>_section function.
-	for _, section in ipairs(SECTION_RENDERERS) do
-		section.render(add, results, totals)
+	-- Per-atom cycle counts (path-aware)
+	add("## Per-atom cycle counts (path-aware, best case, no stalls)"); add("")
+	add("| atom | source | min | max | branches | paths | notes |")
+	add("|------|--------|-----|-----|----------|-------|-------|")
+	local sorted = {}
+	for _, a in ipairs(sa_results.atoms or {}) do sorted[#sorted + 1] = a end
+	table.sort(sorted, function(x, y)
+		return ((x.paths or {}).cycles_max or 0) > ((y.paths or {}).cycles_max or 0)
+	end)
+	for _, a in ipairs(sorted) do
+		local p        = a.paths or {}
+		local src_name = a.source_path and source_basename(a.source_path) or ""
+		local notes    = ""
+		if p.has_loops then notes = notes .. " [loop!]" end
+		if p.unknown_macros and #p.unknown_macros > 0 then
+			notes = notes .. " [unknown: " .. table.concat(p.unknown_macros, ", ") .. "]"
 		end
+		add(string.format("| %s | %s | %d | %d | %d | %d | %s |",
+			a.name, src_name,
+			p.cycles_min or 0, p.cycles_max or 0,
+			p.branches   or 0, p.paths      or 0, notes))
+	end
+	add("")
+
+	-- Per-source scan summary
+	add("## Per-source scan summary"); add("")
+	for _, src in ipairs(dir_sources) do
+		local src_atoms = {}
+		for _, a in ipairs(sa_results.atoms or {}) do
+			if a.source_path == src.path then src_atoms[#src_atoms + 1] = a end
+		end
+		if #src_atoms > 0 then
+			local mn, mx = math.huge, -1
+			for _, a in ipairs(src_atoms) do
+				local p = a.paths or {}
+				if (p.cycles_min or 0) < mn then mn = p.cycles_min or 0 end
+				if (p.cycles_max or 0) > mx then mx = p.cycles_max or 0 end
+			end
+			local path_str
+			if mx > 0 then
+				path_str = string.format("  cycles=%d..%d", mn, mx)
+			else
+				path_str = string.format("  %d cycles", mn)
+			end
+			add(string.format("- `%s` — %d atom%s%s",
+				src.basename, #src_atoms,
+				#src_atoms == 1 and "" or "s", path_str))
+		end
+	end
+	add("")
 
 	return table.concat(lines, "\n") .. "\n"
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Per-project summary
+-- REPORT_RENDERERS — data-driven report dispatch (one row per file kind)
 -- ════════════════════════════════════════════════════════════════════════════
-
---- Render the per-project summary (`build/gen/annotation_validation.txt`).
---- Aggregates totals across all sources; lists per-source error counts if any source has errors.
---- @param all_results AnnotationResult[]
---- @return string
-local function render_project_report(all_results)
-	local lines = {}
-	local function add(s) lines[#lines + 1] = s end
-
-	local total_atoms, total_annots, total_macros, total_binds = 0, 0, 0, 0
-	local total_errors, total_warnings = 0, 0
-	for _, r in ipairs(all_results) do
-		total_atoms    = total_atoms    + #r.atoms
-		total_annots   = total_annots   + #r.annots
-		total_macros   = total_macros   + #r.macros
-		total_binds    = total_binds    + #r.binds
-		total_errors   = total_errors   + #r.errors
-		total_warnings = total_warnings + #r.warnings
-	end
-
-	add(RULE_THICK)
-	add("ANNOTATION VALIDATION — project summary")
-	add(RULE_THICK)
-	add("")
-	add(string.format("Atoms:        %d", total_atoms))
-	add(string.format("Annotations:  %d", total_annots))
-	add(string.format("Macros:       %d", total_macros))
-	add(string.format("Binds:        %d", total_binds))
-	add("")
-	add(string.format("Errors:       %d", total_errors))
-	add(string.format("Warnings:     %d", total_warnings))
-	add("")
-
-	if total_errors > 0 then
-		add("Per-source error counts:")
-		for _, r in ipairs(all_results) do
-			if #r.errors > 0 then
-				local src_name = source_basename(r.source)
-				add(string.format("  %s : %d error(s)", src_name, #r.errors))
-			end
-		end
-		add("")
-	end
-
-	return table.concat(lines, "\n") .. "\n"
-end
-
--- ════════════════════════════════════════════════════════════════════════════
--- Orchestration helpers
--- ════════════════════════════════════════════════════════════════════════════
-
---- (internal) Re-validate every source in a directory against the canonical corpus projection.
---- Calls `annotation.validate()` per source to produce the per-source AnnotationResult (atoms / annots / macros / binds / errors / warnings)
---- that the report renderer consumes. Eeach report pass run is reproducible from the corpus.
---- Returns the list of module results + the flat list of all results (for the project-wide summary).
---- @param ctx PassCtx
---- @param dir_sources SourceFile[]
---- @return AnnotationResult[], AnnotationResult[]
-local function lookup_module_results(ctx, dir_sources)
-	local module_results = {}
-	local all_results    = {}
+-- `once = true` means render once at the project level (not per-module).
+-- `basename(dir_basename)` yields the file's basename for that kind.
+-- `gather(ctx, dir, dir_sources [, all_modules])` returns the rendered string.
+local REPORT_RENDERERS = {
+	{
+		name     = "atom_meta_report",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atom_meta_report" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			-- Annotations: re-run `annotation.validate()` per source (the existing pattern).
+			local annot_results = {}
 			for _, src in ipairs(dir_sources) do
 				if src.scan then
-			local result = annotation.validate(ctx, src, nil)
-			result.source = src.path                                    -- tag for downstream rendering
-			module_results[#module_results + 1] = result
-			all_results[#all_results + 1] = result
-		end
-	end
-	return module_results, all_results
-end
-
---- (internal) Does this module's results contain anything worth emitting?
---- @param module_results AnnotationResult[]
---- @return boolean
-local function module_has_content(module_results)
-	for _, r in ipairs(module_results) do
-		if #r.atoms  > 0 or #r.annots > 0 or #r.binds    > 0
-		or #r.macros > 0 or #r.errors > 0 or #r.warnings > 0 then
-			return true
-		end
-	end
-	return false
-end
-
---- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
---- @param fmt string
-local function debug_log(fmt, ...)
-	if _G[DEBUG_FLAG] then
-		io.stderr:write(string.format("[%s] " .. fmt, PASS_NAME, ...))
+					local r = annotation.validate(ctx, src, nil)
+					r.source = src.path
+					annot_results[#annot_results + 1] = r
 				end
 			end
+			-- Static-analysis: read stashed projection (no re-validate).
+			local dir_basename = dir:match("([^/\\]+)$") or dir
+			local sa_results   = (ctx.shared.corpus.static_analysis_results or {})[dir_basename] or {}
+			return render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+		end,
+	},
+	{
+		name     = "atoms",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atoms" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			return render_module_atoms_md(dir, dir_sources,
+				ctx.shared.corpus.word_counts or {})
+		end,
+	},
+	{
+		name     = "summary",
+		ext      = "md",
+		basename = function(_dir_basename) return "atom_meta_report.summary" end,
+		once     = true,
+		gather   = function(_ctx, _dir, _dir_sources, all_modules)
+			return render_project_summary(all_modules)
+		end,
+	},
+}
 
 -- ════════════════════════════════════════════════════════════════════════════
--- M — module exports
+-- M — public pass surface
 -- ════════════════════════════════════════════════════════════════════════════
 
 local M = {}
 
---- Run the report pass. 
---- Renders one `<dir_basename>.annotations.txt` per source-directory that has content, plus the project-wide `annotation_validation.txt` summary.
+--- Run the report pass. Emits 1 `atom_meta_report.summary.md` per build + 2 `atom_meta_report.md` + 2 `atoms.md` files per module (duffle + gte_hello).
+--- Reads `corpus.static_analysis_results` (added in Phase 1) to populate per-module findings without re-running validate().
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
 	local outputs = {}
-	local errors   = {}
-	local warnings = {}
-
-	-- Module grouping comes from `corpus.sources_by_dir` (the canonical projection).
-	-- Iterate it directly; no private cache, no per-pass stash.
 	local corpus  = ctx.shared and ctx.shared.corpus
 	local by_dir  = (corpus and corpus.sources_by_dir) or {}
 
-	duffle.ensure_dir(ctx.out_root)
+	-- `out_path_root`: when the conventional `out_root` is `build/gen` (any spelling — relative, absolute, separator variants).
+	-- Write the md files to `build/` (parent of `gen/`) instead of nested under `gen/`.
+	-- Mirrors the `gdb_tape_atoms_runtime.gdb` relocation.
+	local function ends_with_gen(p)
+		return type(p) == "string" and (p:match("[/\\]gen[/\\]?$") ~= nil
+			or p == "build/gen" or p == "build\\gen")
+	end
+	local out_root_effective = ends_with_gen(ctx.out_root)
+		and ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		or ctx.out_root
+
+	duffle.ensure_dir(out_root_effective)
+
+	-- Aggregator for the project-wide `once = true` summary renderer.
+	local all_modules = {}
 
-	local all_results_for_summary = {}
 	for dir, dir_sources in pairs(by_dir) do
 		local dir_basename = dir:match("([^/\\]+)$") or dir
-		debug_log("dir=%s basename=%s sources=%d\n", dir, dir_basename, #dir_sources)
 
-		if #dir_sources > 0 then
-			local module_results, all_results = lookup_module_results(ctx, dir_sources)
-			for _, r in ipairs(all_results) do
-				all_results_for_summary[#all_results_for_summary + 1] = r
-			end
-
-			if module_has_content(module_results) then
-				local out_path = ctx.out_root .. "/" .. dir_basename .. ".annotations.txt"
-				duffle.write_file(out_path, render_module_report(dir, dir_sources, module_results))
-				outputs[#outputs + 1] = { annotations_txt = out_path }
-			else
-				debug_log("  -> no content; skipping\n")
-			end
+		-- Per-renderer dispatch for the per-module renderers (once = false).
+		for _, renderer in ipairs(REPORT_RENDERERS) do
+			if not renderer.once then
+				local body     = renderer.gather(ctx, dir, dir_sources)
+				local out_path = out_root_effective .. "/" .. renderer.basename(dir_basename) .. "." .. renderer.ext
+				duffle.write_file(out_path, body)
+				outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
 			end
 		end
 
-	if #all_results_for_summary > 0 then
-		local summary_path = ctx.out_root .. "/annotation_validation.txt"
-		duffle.write_file(summary_path, render_project_report(all_results_for_summary))
-		outputs[#outputs + 1] = { summary_txt = summary_path }
+		-- For the summary, compute per-module totals once (re-validating annotations per source — same pattern as the meta_report renderer).
+		local annot_results = {}
+		for _, src in ipairs(dir_sources) do
+			if src.scan then
+				local r = annotation.validate(ctx, src, nil)
+				r.source = src.path
+				annot_results[#annot_results + 1] = r
+			end
+		end
+		local n_annot, n_binds, n_macros = 0, 0, 0
+		for _, r in ipairs(annot_results) do
+			n_annot  = n_annot  + #r.annots
+			n_binds  = n_binds  + #r.binds
+			n_macros = n_macros + #r.macros
+		end
+		local sa_results = (corpus.static_analysis_results or {})[dir_basename] or {}
+		all_modules[#all_modules + 1] = {
+			module   = dir_basename,
+			atoms    = #(sa_results.atoms or {}),
+			annots   = n_annot,
+			binds    = n_binds,
+			macros   = n_macros,
+			findings = #(sa_results.findings or {}),
+			errors   = #(sa_results.errors   or {}),
+			warnings = #(sa_results.warnings or {}),
+			info     = #(sa_results.info     or {}),
+		}
 	end
 
-	return { outputs = outputs, errors = errors, warnings = warnings }
+	-- Project-wide renderer (once = true): write the summary file.
+	for _, renderer in ipairs(REPORT_RENDERERS) do
+		if renderer.once then
+			local body     = renderer.gather(ctx, nil, nil, all_modules)
+			local out_path = out_root_effective .. "/" .. renderer.basename("") .. "." .. renderer.ext
+			duffle.write_file(out_path, body)
+			outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
+		end
+	end
+
+	return { outputs = outputs, errors = {}, warnings = {} }
 end
 
 return M

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_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_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_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 text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (set by this pass)
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
+--- @field scan     table   -- Pre-scanned SourceScan payload (set by this pass)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -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 body              string     -- brace-delimited body (without the braces)
---- @field body_off          integer    -- char offset of body[1] in source
+--- @field name              string       -- Atom name (for components: without ac_ prefix)
+--- @field body              string       -- Brace-delimited body (without the braces)
+--- @field body_off          integer      -- Char offset of body[1] in source
 --- @field kind              string       -- "atom" | "comp_bare" | "comp_proc" | "raw_atom"
---- @field raw_name          string     -- un-stripped name (for components: with ac_ prefix)
---- @field ident_pos         integer    -- position of the MipsAtom_/MipsAtomComp_ ident start
---- @field after_paren       integer    -- position past the closing paren
+--- @field raw_name          string       -- Un-stripped name (for components: with ac_ prefix)
+--- @field ident_pos         integer      -- Position of the MipsAtom_/MipsAtomComp_ ident start
+--- @field after_paren       integer      -- Position past the closing paren
 --- @field 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.
--- Returns the input unchanged if it doesn't start with the prefix.
--- @param raw_name string
--- @return string
+--- Strip the "ac_" prefix from a component name.
+--- Returns the input unchanged if it doesn't start with the prefix.
+--- @param raw_name 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
--- `//` 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.
--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
---
--- The walker captures:
---   - Block comment close `*/` followed by walking back to `/*`.
---   - `//` 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.
--- Empty string if no comment is adjacent.
--- @param source string
--- @param start_pos integer -- exclusive upper bound for the captured block
--- @return string
+--- Walk backward from `start_pos` capturing contiguous `/* */` block(s) and
+--- `//` 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.
+--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
+---
+--- The walker captures:
+---   - Block comment close `*/` followed by walking back to `/*`.
+---   - `//` 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.
+--- Empty string if no comment is adjacent.
+--- @param source    string
+--- @param start_pos integer -- exclusive upper bound for the captured block
+--- @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.
--- 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).
--- When no marker is pending, the walker starts from the declaration ident_pos directly.
--- @param pending_marker DebugSkipMarker|nil
--- @param ident_pos integer -- declaration ident position
--- @return integer
+--- 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
+--- (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.
+--- @param pending_marker DebugSkipMarker|nil
+--- @param ident_pos      integer -- declaration ident position
+--- @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.
--- 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).
---
--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
---   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`.
--- @param out SourceScan
--- @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
+--- 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
+--- (sole-owner discipline; see push_debug_skip_marker).
+---
+--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
+---   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`.
+--- @param out         SourceScan
+--- @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
 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
-		-- encodes the alias identity (per the duffle TSet_ convention).
+		-- Empty underlying span is acceptable; the TSet_ wrapper itself 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/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

@@ -2,24 +2,21 @@
 ---
 --- Dispatches to pass modules under `scripts/passes/`, resolving dependencies topologically (Kahn's algorithm + cycle detection).
 ---
---- **Architecture**:
----   - **PASSES table** — declarative dep graph (data, not code).
----   - **FLAG_HANDLERS table** — maps CLI flags to handlers.
----   - **parse_args** → **build_ctx** (resolves unity/direct includes or exact sources; no semantic scanning) → **topo_sort** → **dispatch_passes**.
+--- Architecture:
+---   - PASSES table: Declarative dep graph (data, not code).
+---   - FLAG_HANDLERS table: Maps CLI flags to handlers.
+---   - parse_args → build_ctx (resolves unity/direct includes or exact sources) → topo_sort → dispatch_passes.
 ---   - The first pass in the dep graph is `scan-source` (see `passes/scan_source.lua`).
 ---     It calls `duffle.scan_source` once per source to produce the fat `SourceScan` payload, which is attached to each `src.scan`. 
 ---     Every other pass that reads source structure depends on `scan-source` and consumes `src.scan` as a read-only. 
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
----
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Bootstrap: load `duffle_paths.lua` via this script's own path.
 -- Use `arg[0]` when this file is the entry script (`arg[0]` ends in "ps1_meta.lua");
--- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path, not ours).
+-- 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
@@ -57,45 +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 groups string[]?    -- OPTIONAL build-phase groups this pass is a root of
----                            -- (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
+--- @field deps   string[]     -- Names of upstream passes
+--- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
---- @field text     string  -- the full source text
---- @field dir      string  -- the directory containing the source
---- @field basename string  -- filename without extension
+--- @field path     string  -- Absolute path to the source file
+--- @field text     string  -- Full source text
+--- @field dir      string  -- Directory containing the source
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field metadata_path      string                 -- path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
---- @field shared.corpus      table                  -- canonical authored-source/project projection
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
+--- @field shared             table                  -- Cross-pass shared state
+--- @field shared.corpus      table                  -- Authored-source/project projection
+--- @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 msg  string   -- finding message
+--- @field line integer  -- Source line (or 0 for pass-level)
+--- @field msg  string   -- Finding message
 
 --- @class PassResult
---- @field outputs  PassOutputEntry[] -- emitted file paths
---- @field errors   Finding[]         -- build-stops (per-pass kind policy)
---- @field warnings Finding[]         -- informational
+--- @field outputs  PassOutputEntry[] -- Emitted file paths
+--- @field errors   Finding[]         -- Build-stops (per-pass kind policy)
+--- @field warnings Finding[]         -- Informational
 
 --- @class ParsedArgs
---- @field requested_set string[]   -- pass names to run (explicit --all expanded)
---- @field sources       string[]   -- exact --source values, retained in CLI order
+--- @field requested_set string[]   -- Pass names to run (explicit --all expanded)
+--- @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
@@ -141,7 +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"},
@@ -160,18 +156,18 @@ local PASSES = {
 	report = {
 		module = "passes.report",
 		kind   = "report",
-		deps   = {"annotation", "static-analysis"},
+		deps   = {"annotation", "static-analysis", "atoms-source-map"},  -- +atoms-source-map (consolidated-report-files refactor, 2026-07-26)
 		groups = { "pre-link" },
 	},
 }
 
 -- ────────────────────────────────────────────────────────────────────────────
--- 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")
---- @return string[]          -- sorted root pass names belonging to that group
+--- @param group_name string  -- Build-phase group ("pre-link" | "post-link")
+--- @return string[]          -- Sorted root pass names belonging to that group
 local function roots_for_group(group_name)
 	local names = {}
 	for name, pass in pairs(PASSES) do
@@ -206,9 +202,10 @@ end
 
 -- Pass-kind taxonomy: Which kinds stop the build on errors?
 --
--- Report severity is independent from process exit policy. A "diagnostic" pass still writes every `error`/`warning` finding into its report file,
+-- 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,
@@ -218,8 +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)
--- live in FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
+-- 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.
 -- 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 = {
@@ -248,7 +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 = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -274,9 +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
@@ -319,8 +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 "
-			.. FLAG_VALUE_NAMES[flag] .. "\n")
+		io.stderr:write("ps1_meta: " .. flag .. " requires " .. FLAG_VALUE_NAMES[flag] .. "\n")
 		os.exit(EXIT_INTERNAL_ERROR)
 	end
 	return value, arg_idx + 1
@@ -330,12 +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
@@ -498,8 +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
-		-- drive-relative paths before the absolute-path rewrite is performed.
+		-- canonical_path_key validates ordinary relative paths and rejects 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
@@ -513,8 +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 "
-				.. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
+			io.stderr:write("ps1_meta: cannot resolve --unity-root " .. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
 			os.exit(EXIT_INTERNAL_ERROR)
 		end
 		resolution = resolved
@@ -530,8 +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 .. ": "
-					.. tostring(key_or_error), 0)
+				error("ps1_meta: invalid --source " .. input_path .. ": " .. tostring(key_or_error), 0)
 			end
 			local file = io.open(path, "r")
 			if not file then
@@ -629,9 +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
-				error("unknown pass '" .. name .. "' requested")
-			end
+			if not pass then error("unknown pass '" .. name .. "' requested") end
 			for _, dep in ipairs(pass.deps) do
 				if not needed[dep] then
 					needed[dep] = true