finished reviewing normalize_v3s4 for now

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

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

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

.gitignore

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

.vscode/launch.json

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

code/duffle/bios.h

@@ -0,0 +1,14 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	pragma once
+#endif
+enum {
+	bios_init_pad_2  = 0x12,
+	bios_start_pad_2 = 0x13,
+	bios_flushcache  = 0x44,
+	bios_table_addr  = 0xA0,
+	bios_btable_addr = 0xB0,
+};
+
+enum {
+	bios_pad_buffer_size = 0x22,
+};

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
 
 /* ============================================================================
@@ -71,11 +70,31 @@
 /* ----------------------------------------------------------------------------
  * atom_reg (per-enum opt-in marker for the DWARF register-alias registry)
  *
- * The bare `atom_reg` token adjacent to an enum entry in mips.h / lottes_tape.h flags that alias as debug-visible for scan_source's register_alias_registry.
+ * Bare `atom_reg` token adjacent to an enum entry that alias as debug-visible for scan_source's register_alias_registry.
- * The C preprocessor strips it to a comment so no runtime symbol is created; the Lua scanner reads the bare token.
+ * Lua scanner reads the bare token.
  * ----------------------------------------------------------------------------*/
 #define atom_reg /* atom_reg: opt the preceding enum entry into the DWARF registry */
 
+// ----------------------------------------------------------------------------
+// atom_auto_reg(atom, sym) — per-atom auto-allocated GPR binding.
+//   enum {
+//       atom_auto_reg(cube_g4_face, R_Fwdx),                // expands to: R_Fwdx = R_Fwdx_Code /* atom_auto_reg: cube_g4_face */,
+//       atom_auto_reg(cube_g4_face, R_Eye_z) atom_type(S4), // atom_type chains after
+//   };
+// (The macro IS the entire enum entry — no separate LHS=RHS. The `atom` scope is
+//  preserved in a trailing C-comment on the RHS so the Lua scanner can recover
+//  it after preprocessing strips the macro form. R_<Sym>_Code is resolved from gen/auto_reg.h which the .c file #include's before the enum declaration.)
+#define atom_auto_reg(atom, sym) sym = sym ## _Code /* atom_auto_reg: atom */
+
+// ----------------------------------------------------------------------------
+// phase_auto_reg(phase, sym) — per-phase auto-allocated GPR binding.
+//   enum {
+//       phase_auto_reg(cube_g4, R_Temp0),     // expands to: R_Temp0 = R_Temp0_Code /* phase_auto_reg: cube_g4 */,
+//       phase_auto_reg(cube_g4, R_Temp1),
+//   };
+// (Same macro-as-enum-entry form as atom_auto_reg above; the `phase` scope is preserved in a trailing C-comment on the RHS for the Lua scanner to recover.)
+#define phase_auto_reg(phase, sym) sym = sym ## _Code /* phase_auto_reg: phase */
+
 /* ============================================================================
  *   atom_info :
  *      MipsAtom_(cube_tri) atom_info(
@@ -148,12 +167,12 @@
  *   ... 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

@@ -3,7 +3,7 @@
 #	include "assert.h"
 #endif
 
-#define offset_of(type, member)  cast(U8,__builtin_offsetof(type,member))
+#define offset_of(type, member)  cast(U8,__builtin_offsetof(type,member)) // Compiler builtin version of O_
 #define static_assert            _Static_assert
 #define typeof                   __typeof__
 #define typeof_ptr(ptr)          typeof((ptr)[0])
@@ -28,8 +28,9 @@
 #define internal static // internal
 
 #define asm           __asm__
-#define align_(value) __attribute__((aligned (value)))             // for easy alignment
 
+#define A_(data)      (& data)
+#define align_(value) __attribute__((aligned (value)))             // for easy alignment
 #define align_(value) __attribute__((aligned (value)))             // for easy alignment
 #define C_(type,data) ((type)(data))                               // for enforced precedence
 #define expect_(x, y) __builtin_expect(x, y)                       // so compiler knows the common path
@@ -43,7 +44,9 @@
 
 #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 +70,8 @@
 #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.
@@ -87,12 +91,13 @@
 #define PtrSet_(type)  TypeR_(type); typedef TypeV_(type)
 #define TSet_(type)    type; typedef PtrSet_(type)
 
-#define array_len(a)                   (U4)(sizeof(a) / sizeof(typeof((a)[0])))
+#define Array_len(a)                   (U4)(sizeof(a) / sizeof(typeof((a)[0])))
-#define array_decl(type, ...)          (type[]){__VA_ARGS__}
+#define Array_decl(type, ...)          (type[]){__VA_ARGS__}
 #define Array_sym(type,len)            A ## len ## _ ## type
 #define Array_expand(type,len)         type Array_sym(type, len)[len]; typedef PtrSet_(Array_sym(type, len))
 #define Array_(type,len)               Array_expand(type,len)
 #define Bit_(id,b)                     id = (1 << b), tmpl(id,pos) = b
+#define Bitmask_(b)                    (1u << b)
 #define Enum_(underlying_type, symbol) underlying_type TSet_(symbol); enum   symbol
 #define Proc_(symbol)                  symbol
 #define Relative_(symbol)              // Does nothing but annotate that a symbol is associated with another.
@@ -130,21 +135,20 @@ typedef __UINT32_TYPE__ TSet_(B4);
 #define u4_v(value) C_(U4 V_*, value)
 enum { false = 0, true  = 1, true_overflow, };
 
-#define u4_lo(value) ((value) & 0xFFFFU)
+#define u4_lo(value) (u4_(value) & 0xFFFFU)
-#define u4_hi(value) ((value) >> 12)
+#define u4_hi(value) (u4_(value) >> (S_(U2) * 8))
 
 typedef void Proc_(VoidFn) (void);
 
-#define kilo(n)                (C_(U4, n) << 10)
+#define Kilo_(n)                (C_(U4, n) << 10)
-#define mega(n)                (C_(U4, n) << 20)
+#define Mega_(n)                (C_(U4, n) << 20)
-#define giga(n)                (C_(U4, n) << 30)
+#define Giga_(n)                (C_(U4, n) << 30)
-#define tera(n)                (C_(U4, n) << 40)
+#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 OA_(type, member, idx)  C_(U4, & C_(type*,0)->member[idx])
+#define OT_(field)              O_(typeof_ptr(& field), field))
-#define OT_(field)              O_(typeof_ptr(& field), filed))
 #define S_(data)                C_(U4, sizeof(data))
 
 #define sop_1(op,a,b) C_(U1, s1_(a) op s1_(b))
@@ -165,6 +169,8 @@ def_signed_ops(le, <=)
 #undef def_signed_ops
 #undef def_signed_op
 
+// Unused, we arent' doing any C-like asm since we have the asm dsl. We'll keep the non-generics if we somehow do.
+#if 0
 #define def_generic_sop(op, a, ...) _Generic((a), U1:  op ## _s1, U2: op ## _s2, U4: op ## _s4) (a, __VA_ARGS__)
 #define add_s(a,b) def_generic_sop(add,a,b)
 #define sub_s(a,b) def_generic_sop(sub,a,b)
@@ -174,11 +180,12 @@ def_signed_ops(le, <=)
 #define ge_s(a,b)  def_generic_sop(ge, a,b)
 #define le_s(a,b)  def_generic_sop(le, a,b)
 #undef def_generic_sop
+#endif
 
 #define alignas                             _Alignas
 #define alignof                             _Alignof
 #define byte_pad(amount, ...)               B1 glue(_PAD_, __VA_ARGS__) [amount]
-#define pcast(type, data)                   (C_(type*, & (data)) [0])
+#define C_ptr(type, data)                   (C_(type*, & (data)) [0])
 
 #define dbg_args(...)                      __VA_ARGS__
 
@@ -193,6 +200,8 @@ def_signed_ops(le, <=)
 #define defer_info(type,expr, ...) for(type       info= {__VA_ARGS__}; info.once!=1;++info.once,(expr))    // Defer with tracked state
 
 #define do_while(cond) for (U8 once=0; once!=1 || (cond); ++once)
+
+#define Jmp_nZero_(cond,label) if (cond) goto label;
 #pragma endregion Control Flow & Iteration
 
 #define span_iter(type, iter, m_begin, op, m_end) ( \
@@ -209,16 +218,16 @@ def_signed_ops(le, <=)
 typedef Span_(S4);
 typedef Span_(U4);
 
-#if 0
 #pragma region Debug
-#define debug_trap() __builtin_debugtrap()
+#define debug_trap() __builtin_trap()
 #if BUILD_DEBUG
-IA_ void assert(U8 cond) { if(cond){return;} else{debug_trap(); ms_exit_process(1);} }
+	#define assert(cond) if(cond == false){debug_trap();}
 #else
-#define assert(cond)
+#	ifndef assert
+#		include <assert.h>
+#	endif
 #endif
 #pragma endregion Debug
-#endif
 
 #define GCC_OPTIMIZATION_DISABLE _Pragma("GCC push_options") _Pragma("GCC optimize(\"O0\")")
 #define GCC_OPTIMIZATION_ENABLE  _Pragma("GCC pop_options")

code/duffle/gen/duffle.macs.h

@@ -1,150 +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).
-//   - 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 */
-/* Words: 3; Loads 3 S2 indices from the face array */
-#define mac_load_tri_indices(...) \
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)) \
-,	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)) \
-,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
-WORD_COUNT(mac_load_tri_indices, 3)
-
-/* atom_dbg_skip */
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-#define mac_gte_load_tri_verts(...) \
-	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY0) \
-,	gte_mv_to_data_r(R_V1, C2_VZ0) \
-,	shift_lleft(R_AT, R_T1, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY1) \
-,	gte_mv_to_data_r(R_V1, C2_VZ1) \
-,	shift_lleft(R_AT, R_T2, v3s2_byteoff) \
-,	add_u_self(R_AT, R_VertBase) \
-,	load_word(R_V0, R_AT, O_(V3_S2,x)) \
-,	load_word(R_V1, R_AT, O_(V3_S2,z)) \
-,	gte_mv_to_data_r(R_V0, C2_VXY2) \
-,	gte_mv_to_data_r(R_V1, C2_VZ2)
-WORD_COUNT(mac_gte_load_tri_verts, 18)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-#define mac_insert_ot_tag_f3(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (5 - 1) << 24 = 4 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_f3, 11)
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-#define mac_insert_ot_tag_g4(...) \
-	shift_lleft( R_T1, R_T1, S_(U4)/2)                        /* T1 = otz * S_(U4) (otz arg is implicit R_T1) */ \
-,	add_u_self(  R_T1, R_OtBase)                              /* T1 = & OrderingTable[OTZ] */ \
-,	load_word(   R_AT, R_T1,         O_(PolyTag,code))        /* AT = old_ot_head */ \
-,	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) /* V0 = (9 - 1) << 24 = 8 << 24 */ \
-,	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits))    /* Strip upper 8 bits (length from prev cell) → keep only low 24 */ \
-,	or_u(        R_AT, R_AT, R_V0)                            /* Merge length */ \
-,	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code))        /* prim->tag = packed(prim_length, old_addr) */ \
-,	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits))    /* AT = (prim_length << 24) | old_addr */ \
-,	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)) \
-,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
-WORD_COUNT(mac_insert_ot_tag_g4, 11)
-
-/* atom_dbg_skip */
-#define mac_pack_color_word(off, cmd, r, g, b) \
-	load_upper_i(R_AT, (cmd) << 8  | (b)) \
-,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
-,	store_word(  R_AT, R_PrimCursor, (off))
-WORD_COUNT(mac_pack_color_word, 3)
-
-/* atom_dbg_skip */
-#define mac_format_f3_color(r, g, b) \
-	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
-WORD_COUNT(mac_format_f3_color, 3)
-
-/* atom_dbg_skip */
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
- * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-#define mac_gte_store_f3(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2))
-WORD_COUNT(mac_gte_store_f3, 3)
-
-#define mac_format_g4_color(r0, g0, b0, r1, g1, b1, r2, g2, b2, r3, g3, b3) \
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0) \
-,	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1) \
-,	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2) \
-,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
-WORD_COUNT(mac_format_g4_color, 12)
-
-/* atom_dbg_skip */
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
- * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
-#define mac_gte_store_g4_p012(...) \
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)) \
-,	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)) \
-,	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2))
-WORD_COUNT(mac_gte_store_g4_p012, 3)
-
-/* atom_dbg_skip */
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its single-vertex result to SXY2;
- * SXY0 still holds v0.screen from the earlier RTPT.
- */
-#define mac_gte_store_g4_p3(...) \
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
-WORD_COUNT(mac_gte_store_g4_p3, 1)
-

code/duffle/gen/duffle.offsets.h

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

code/duffle/gen/macs.h

@@ -0,0 +1,308 @@
+#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\bios.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.c
+//   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, offset + O_(V3_S2,x)) \
+,	load_half( rs_y, r_base, offset + 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_load_v3s4(rs_x, rs_y, rs_z, r_base, offset) \
+	load_word( rs_x, r_base, offset + O_(V3_S4,x)) \
+,	load_word( rs_y, r_base, offset + O_(V3_S4,y)) \
+,	load_word( rs_z, r_base, offset + O_(V3_S4,z))
+WORD_COUNT(mac_load_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_store_v3s4(rt_x, rt_y, rt_z, base, offset) \
+	store_word(rt_x, base, offset + O_(V3_S4,x)) \
+,	store_word(rt_y, base, offset + O_(V3_S4,y)) \
+,	store_word(rt_z, base, offset + O_(V3_S4,z))
+WORD_COUNT(mac_store_v3s4, 3)
+
+/* atom_dbg_skip */
+#define mac_sub_v3s4(rds_x, rds_y, rds_z, rt_x, rt_y, rt_z) \
+	sub_s(rds_x, rds_x, rt_x) \
+,	sub_s(rds_y, rds_y, rt_y) \
+,	sub_s(rds_z, rds_z, rt_z)
+WORD_COUNT(mac_sub_v3s4, 3)
+
+/* 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_word_imm(dst, imm) \
+	load_upper_i(dst, u4_hi(imm)) \
+,	or_i_self(   dst, u4_lo(imm))
+WORD_COUNT(mac_load_word_imm, 2)
+
+/* 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)
+
+/* atom_dbg_skip */
+#define mac_gte_sqr_v3(r_sx, r_sy, r_sz, r_sq_x, r_sq_y, r_sq_z) \
+	mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, nop) \
+,	gte_mv_from_data_r(r_sq_x, C2_MAC1) \
+,	gte_mv_from_data_r(r_sq_y, C2_MAC2) \
+,	gte_mv_from_data_r(r_sq_z, C2_MAC3)
+WORD_COUNT(mac_gte_sqr_v3, 8)
+
+/* atom_dbg_skip */
+#define mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, nop_slot) \
+	gte_mv_to_data_r(r_sx, C2_IR1) \
+,	gte_mv_to_data_r(r_sy, C2_IR2) \
+,	gte_mv_to_data_r(r_sz, C2_IR3) \
+,	nop_slot \
+,	gte_cmdw_sqr
+WORD_COUNT(mac_gte_sqr_v3s4, 5)
+
+/* atom_dbg_skip */
+#define mac_gte_gpf_scale(r_sx, r_sy, r_sz, r_recip_est, r_shift, r_dx, r_dy, r_dz) \
+	gte_mv_to_data_r(r_recip_est, C2_IR0) \
+,	gte_mv_to_data_r(r_sx,        C2_IR1) \
+,	gte_mv_to_data_r(r_sy,        C2_IR2) \
+,	gte_mv_to_data_r(r_sz,        C2_IR3) \
+,	nop2 /* retire IR0..IR3 → GPF input pre-fill (matches libgte 0x80016134..0x80016138) */ \
+,	gte_cmdw_gpf \
+,	gte_mv_from_data_r(r_dx, C2_MAC1) \
+,	gte_mv_from_data_r(r_dy, C2_MAC2) \
+,	gte_mv_from_data_r(r_dz, C2_MAC3) \
+,	shift_aright_var(r_dx, r_dx, r_shift) \
+,	shift_aright_var(r_dy, r_dy, r_shift) \
+,	shift_aright_var(r_dz, r_dz, r_shift)
+WORD_COUNT(mac_gte_gpf_scale, 13)
+
+#define mac_trans_mt3s3s4(r_mtx, r_off, r_t0, r_t1, r_t2) \
+	load_word(r_t0, r_off, O_(V3_S4,x)) \
+,	load_word(r_t1, r_off, O_(V3_S4,y)) \
+,	load_word(r_t2, r_off, O_(V3_S4,z)) \
+,	store_word(r_t0, r_mtx, O_(MT3_S2S4,t[0])) \
+,	store_word(r_t1, r_mtx, O_(MT3_S2S4,t[1])) \
+,	store_word(r_t2, r_mtx, O_(MT3_S2S4,t[2]))
+WORD_COUNT(mac_trans_mt3s3s4, 6)
+
+/* atom_dbg_skip */
+#define mac_lzcr_round_even_half_shift(r_shift, r_mag_sq, r_mag_sq_copy) \
+	and_i(r_shift, r_shift, gte_lzcr_even_mask) \
+,	or_u(r_mag_sq_copy, r_mag_sq, 0) \
+,	li_s(r_mag_sq, 31) \
+,	sub_s(r_mag_sq, r_mag_sq, r_shift) \
+,	shift_aright(r_mag_sq, r_mag_sq, 1)
+WORD_COUNT(mac_lzcr_round_even_half_shift, 5)
+
+#define mac_shift_aright_var_v3(rd_v0, rd_v1, rd_v2, rs_v0, rs_v1, rs_v2, r_shift) \
+	shift_aright_var(rd_v0, rs_v0, r_shift) \
+,	shift_aright_var(rd_v1, rs_v1, r_shift) \
+,	shift_aright_var(rd_v2, rs_v2, r_shift)
+WORD_COUNT(mac_shift_aright_var_v3, 3)
+
+#define mac_shift_aright_var_v3_self(rds_v0, rds_v1, rds_v2, r_shift) \
+	shift_aright_var(rds_v0, rds_v0, r_shift) \
+,	shift_aright_var(rds_v1, rds_v1, r_shift) \
+,	shift_aright_var(rds_v2, rds_v2, r_shift)
+WORD_COUNT(mac_shift_aright_var_v3_self, 3)
+
+#define mac_gte_general_purpose_interopolation(to_ir0, to_ir1, to_ir2, to_ir3, fr_mac1, fr_mac2, fr_mac3, nop_slot1, nop_slot2) \
+	gte_mv_to_data_r(to_ir0, C2_IR0) \
+,	gte_mv_to_data_r(to_ir1, C2_IR1) /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */ \
+,	gte_mv_to_data_r(to_ir2, C2_IR2) \
+,	gte_mv_to_data_r(to_ir3, C2_IR3) /* IR3 = src.z (reloaded) */ \
+,	LdSlot_ nop_slot1 \
+,	LdSlot_ nop_slot2 \
+,	gte_cmdw_gpf \
+,	gte_mv_from_data_r(fr_mac1, C2_MAC1) \
+,	gte_mv_from_data_r(fr_mac2, C2_MAC2) \
+,	gte_mv_from_data_r(fr_mac3, C2_MAC3)
+WORD_COUNT(mac_gte_general_purpose_interopolation, 10)
+
+#define mac_gte_mv_from_data_r_mac123(fr_mac1, fr_mac2, fr_mac3) \
+	gte_mv_from_data_r(fr_mac1, C2_MAC1) \
+,	gte_mv_from_data_r(fr_mac2, C2_MAC2) \
+,	gte_mv_from_data_r(fr_mac3, C2_MAC3)
+WORD_COUNT(mac_gte_mv_from_data_r_mac123, 3)
+
+/* atom_dbg_skip */
+#define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
+	mac_load_word_imm(reg_transfer, cmd) \
+,	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)
+
+/* atom_dbg_skip */
+#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(r_ot_base, r_prim_cursor, poly_size) \
+	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, (poly_size/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits) \
+,	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, 11)
+
+/* atom_dbg_skip */
+#define mac_pad_set_centered_axes(state, scratch) \
+	load_upper_i(scratch, (PadAxis_Centered >> 16) & 0xFFFF) \
+,	or_i_self(   scratch,  PadAxis_Centered        & 0xFFFF)		/* mac_load_word_imm(scratch, PadAxis_Centered), */ \
+,	store_word(       scratch, state, O_(PadState,axes))
+WORD_COUNT(mac_pad_set_centered_axes, 3)
+
+/* atom_dbg_skip */
+#define mac_pad_set_id_byte(state, r_id, id_value) \
+	add_ui(    r_id, R_0, id_value) \
+,	store_byte(r_id, state, O_(PadState,id))
+WORD_COUNT(mac_pad_set_id_byte, 2)
+
+/* atom_dbg_skip */
+#define mac_pad_set_status(r_tmp, r_state, pad_status) \
+	add_ui(    r_tmp, R_0, pad_status) \
+,	store_word(r_tmp, r_state, O_(PadState,status))
+WORD_COUNT(mac_pad_set_status, 2)
+
+/* atom_dbg_skip */
+#define mac_pad_store_inverted_buttons(r_buttons, r_pad_state) \
+	nor_u(       r_buttons, r_buttons, R_0) \
+,	store_half(  r_buttons, r_pad_state, O_(PadState,buttons))
+WORD_COUNT(mac_pad_store_inverted_buttons, 2)
+

code/duffle/gen/offsets.h

@@ -0,0 +1,65 @@
+// 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\bios.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\psyq.h
+//   source: C:\projects\Pikuma\ps1\code\duffle\pad.c
+//   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: normalize_v3s4 (56 words) ---
+
+#define _atom_offset_aligned_done_srav_path           3
+#define _atom_offset_srav_path_aligned_done           4
+
+enum {
+    atom_offset_aligned_done_srav_path = _atom_offset_aligned_done_srav_path,
+    atom_offset_srav_path_aligned_done = _atom_offset_srav_path_aligned_done,
+};
+
+// --- atom: pad_bios_snapshot (84 words) ---
+
+#define _atom_offset_snap_root_skip_disconnected      10
+#define _atom_offset_disconnected_snap_end            65
+#define _atom_offset_case_2_id_dispatch               9
+#define _atom_offset_pending_snap_end                 54
+#define _atom_offset_id_dispatch_try_analog_stick     12
+#define _atom_offset_id_dispatch_snap_end             40
+#define _atom_offset_try_analog_stick_try_analog_pad  13
+#define _atom_offset_analog_stick_snap_end            25
+#define _atom_offset_try_analog_pad_try_unsupported   12
+#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 +1,60 @@
+#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(AtomBuilder_R ab, U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_load_word_imm(reg_transfer, cmd),
+	store_word(  reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_store_rgb8(AtomBuilder_R ab, U1 rr, U1 rg, U1 rb, U4 base, U4 offset) 
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_byte(rr, base, offset + O_(RGB8,r)),
+	store_byte(rg, base, offset + O_(RGB8,g)),
+	store_byte(rb, base, offset + O_(RGB8,b)),
+})
+
+FI_ Slice_MipsCode ac_pack_color_word(AtomBuilder_R ab, U4 r_base, U4 off, U4 cmd, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_upper_i(R_AT, (cmd) << 8  | (b)),
+	or_i_self(   R_AT, ((g)  << 8) | (r)),
+	store_word(  R_AT, r_base, (off)),
+})
+
+FI_ Slice_MipsCode ac_format_f3_color(AtomBuilder_R ab, U4 r_base, U1 r, U1 g, U1 b)
+atom_dbg_skip MipsAtomComp_Proc_(ab, { mac_pack_color_word(r_base, O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+
+FI_ Slice_MipsCode ac_format_g4_color(AtomBuilder_R ab, 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)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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. */
+I_ Slice_MipsCode ac_insert_ot_tag(AtomBuilder_R ab, U4 r_ot_base, U4 r_prim_cursor, U4 poly_size) MipsAtomComp_Proc_(ab, {
+	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, (poly_size/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
+	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

@@ -21,7 +21,7 @@
  *    4. Semantic encoders      gp0_word_poly_f3(r,g,b)
  *    3. Composite encoders     enc_color_word(cmd, r, g, b)
  *    2. Per-field encoders     enc_gp0_color_r(r), enc_gp0_color_g(g), ...
- *    1. Bitfield layout consts gp0_color_red_shift = 0, gp0_color_red_mask = 0xFF
+ *    1. Bitfield layout consts gp0_color_red_shift = 0, gp0_color_red_width = 8
  *    0. Opcode IDs             gp0_cmd_poly_f3 = 0x20
  *
  *  Vendor mnemonics (gte_mtc2, gte_mfc2, etc.) are NOT in this header.
@@ -74,7 +74,7 @@ enum {
  *  ============================================================================
  *  8-bit GP0 opcodes (the upper byte of a primitive's first word). These are the BYTE only.
  *  NO macro body past this point uses a raw shift or raw mask.
- *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
+ *  Mirrors the OPCODE_SHIFT / RS_SHIFT convention from mips.h.
  * ============================================================================ */
 enum {
 	gp0_cmd_Nop             = 0x00,
@@ -116,21 +116,20 @@ enum {
 	gp0_cmd_SetDrawOffset        = 0xE5,
 	gp0_cmd_SetMaskBit           = 0xE6,
 
-	/* bitfield shifts / widths / masks ----
+	/* bitfield shifts / widths ----
 		* Generic GP0/GP1 command byte (upper 8 bits of every word sent to either port). */
 	gp0_cmd_shift = 24,
 	gp0_cmd_width =  8,
-	gp0_cmd_mask  = 0xFF,
 
 	/* Color word layout (lives in Poly_F3.color, Poly_G4.c0..c3, etc.):
 		*   bits 31..24 = command byte
 		*   bits 23..16 = BLUE
 		*   bits 15..08 = GREEN
 		*   bits 07..00 = RED     (PSX GPU is BGR, NOT RGB) */
-	gp0_color_cmd_shift   = 24, gp0_color_cmd_width   = 8, gp0_color_cmd_mask   = 0xFF,
+	gp0_color_cmd_shift   = 24, gp0_color_cmd_width   = 8,
-	gp0_color_blue_shift  = 16, gp0_color_blue_width  = 8, gp0_color_blue_mask  = 0xFF,
+	gp0_color_blue_shift  = 16, gp0_color_blue_width  = 8,
-	gp0_color_green_shift =  8, gp0_color_green_width = 8, gp0_color_green_mask = 0xFF,
+	gp0_color_green_shift =  8, gp0_color_green_width = 8,
-	gp0_color_red_shift   =  0, gp0_color_red_width   = 8, gp0_color_red_mask   = 0xFF,
+	gp0_color_red_shift   =  0, gp0_color_red_width   = 8,
 };
 
 /* ============================================================================
@@ -143,12 +142,12 @@ enum {
  * ============================================================================ */
 
 /* ---- Layer 1.5: per-field encoders ---- */
-#define enc_gp0_cmd(cmd)       (((cmd) & gp0_cmd_mask)         << gp0_cmd_shift)
+#define enc_gp0_cmd(cmd)       ((cmd) << gp0_cmd_shift)
 
-#define enc_gp0_color_cmd(cmd) (((cmd) & gp0_color_cmd_mask)   << gp0_color_cmd_shift)
+#define enc_gp0_color_cmd(cmd) ((cmd) << gp0_color_cmd_shift)
-#define enc_gp0_color_r(r)     (((r)   & gp0_color_red_mask)   << gp0_color_red_shift)
+#define enc_gp0_color_r(r)     ((r)   << gp0_color_red_shift)
-#define enc_gp0_color_g(g)     (((g)   & gp0_color_green_mask) << gp0_color_green_shift)
+#define enc_gp0_color_g(g)     ((g)   << gp0_color_green_shift)
-#define enc_gp0_color_b(b)     (((b)   & gp0_color_blue_mask)  << gp0_color_blue_shift)
+#define enc_gp0_color_b(b)     ((b)   << gp0_color_blue_shift)
 
 /* ---- Layer 2: composite encoders ---- */
 #define enc_color_word(cmd, r, g, b) (enc_gp0_color_cmd(cmd) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
@@ -211,38 +210,38 @@ enum {
 	gp1_disp_Color24    = 0x1,
 	gp1_disp_VInterlace = 0x1,
 
-	/* ---- Layer 1: GP1 display-mode + range + draw-area shifts/masks ---- */
+	/* ---- Layer 1: GP1 display-mode + range + draw-area shifts/widths ---- */
-	gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2, gp1_disp_hres_mask      = 0x3,
+	gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2,
-	gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1, gp1_disp_vres_mask      = 0x1,
+	gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1,
-	gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1, gp1_disp_color_mask     = 0x1,
+	gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1,
-	gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1, gp1_disp_interlace_mask = 0x1,
+	gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1,
 
 	/* GP1 horizontal display range: bits 0..11 = X2, bits 12..23 = X1 */
-	gp1_hrange_x1_shift = 12, gp1_hrange_x1_width = 12, gp1_hrange_x1_mask = 0xFFF,
+	gp1_hrange_x1_shift = 12, gp1_hrange_x1_width = 12,
-	gp1_hrange_x2_shift =  0, gp1_hrange_x2_width = 12, gp1_hrange_x2_mask = 0xFFF,
+	gp1_hrange_x2_shift =  0, gp1_hrange_x2_width = 12,
 
 	/* GP1 vertical display range: bits 0..9 = Y2, bits 10..19 = Y1 */
-	gp1_vrange_y1_shift = 10, gp1_vrange_y1_width = 10, gp1_vrange_y1_mask = 0x3FF,
+	gp1_vrange_y1_shift = 10, gp1_vrange_y1_width = 10,
-	gp1_vrange_y2_shift =  0, gp1_vrange_y2_width = 10, gp1_vrange_y2_mask = 0x3FF,
+	gp1_vrange_y2_shift =  0, gp1_vrange_y2_width = 10,
 
 	/* GP1 draw area (top-left or bottom-right): bits 0..9 = X, bits 10..19 = Y
-		* (10-bit signed — caller pre-signs and masks with the named mask) */
+		* (10-bit signed — caller pre-signs) */
-	gp1_draw_x_shift =  0, gp1_draw_x_width = 10, gp1_draw_x_mask = 0x3FF,
+	gp1_draw_x_shift =  0, gp1_draw_x_width = 10,
-	gp1_draw_y_shift = 10, gp1_draw_y_width = 10, gp1_draw_y_mask = 0x3FF,
+	gp1_draw_y_shift = 10, gp1_draw_y_width = 10,
 };
 
 /* ---- Layer 1.5: GP1 per-field encoders ---- */
-#define enc_gp1_disp_hres(h)      (((h) & gp1_disp_hres_mask)      << gp1_disp_hres_shift)
+#define enc_gp1_disp_hres(h)      ((h) << gp1_disp_hres_shift)
-#define enc_gp1_disp_vres(v)      (((v) & gp1_disp_vres_mask)      << gp1_disp_vres_shift)
+#define enc_gp1_disp_vres(v)      ((v) << gp1_disp_vres_shift)
-#define enc_gp1_disp_color(c)     (((c) & gp1_disp_color_mask)     << gp1_disp_color_shift)
+#define enc_gp1_disp_color(c)     ((c) << 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_shift)
 
-#define enc_gp1_hrange_x1(x1) (((x1) & gp1_hrange_x1_mask) << gp1_hrange_x1_shift)
+#define enc_gp1_hrange_x1(x1) ((x1) << gp1_hrange_x1_shift)
-#define enc_gp1_hrange_x2(x2) (((x2) & gp1_hrange_x2_mask) << gp1_hrange_x2_shift)
+#define enc_gp1_hrange_x2(x2) ((x2) << gp1_hrange_x2_shift)
-#define enc_gp1_vrange_y1(y1) (((y1) & gp1_vrange_y1_mask) << gp1_vrange_y1_shift)
+#define enc_gp1_vrange_y1(y1) ((y1) << gp1_vrange_y1_shift)
-#define enc_gp1_vrange_y2(y2) (((y2) & gp1_vrange_y2_mask) << gp1_vrange_y2_shift)
+#define enc_gp1_vrange_y2(y2) ((y2) << gp1_vrange_y2_shift)
-#define enc_gp1_draw_x(x)     (((x)  & gp1_draw_x_mask)    << gp1_draw_x_shift)
+#define enc_gp1_draw_x(x)     ((x)  << gp1_draw_x_shift)
-#define enc_gp1_draw_y(y)     (((y)  & gp1_draw_y_mask)    << gp1_draw_y_shift)
+#define enc_gp1_draw_y(y)     ((y)  << gp1_draw_y_shift)
 
 /* ---- Layer 2: GP1 composite encoders ---- */
 #define enc_gp1_disp_mode_word(h, v, c, i) (enc_gp0_cmd(gp1_cmd_DisplayMode)            | enc_gp1_disp_hres(h)  | enc_gp1_disp_vres(v) | enc_gp1_disp_color(c) | enc_gp1_disp_interlace(i))
@@ -555,14 +554,14 @@ typedef Struct_(Poly_GT4) {
  *    bits 12..31  = reserved (zero)
  * ============================================================================ */
 enum {
-	/* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
+	/* ---- Layer 1: TPage bitfield shifts / widths ---- */
-	gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4, gp0_tpage_x_mask            = 0xF,
+	gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4,
-	gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1, gp0_tpage_y_mask            = 0x1,
+	gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1,
-	gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2, gp0_tpage_semi_trans_mask   = 0x3,
+	gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2,
-	gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2, gp0_tpage_color_depth_mask  = 0x3,
+	gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2,
-	gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1, gp0_tpage_dither_mask       = 0x1,
+	gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1,
-	gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1, gp0_tpage_draw_to_disp_mask = 0x1,
+	gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1,
-	gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1, gp0_tpage_tex_disable_mask  = 0x1,
+	gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1,
 
 	/* TPage color-depth payload values (NOT bit positions — these go in
 	 * the 2-bit field at gp0_tpage_color_depth_shift). */
@@ -581,13 +580,13 @@ enum {
 };
 
 /* ---- Layer 1.5: TPage per-field encoders. Mirrors enc_gte_sf/mx/v in gte.h. ---- */
-#define enc_gp0_tpage_x(x)            (((x) & gp0_tpage_x_mask)            << gp0_tpage_x_shift)
+#define enc_gp0_tpage_x(x)            ((x) << gp0_tpage_x_shift)
-#define enc_gp0_tpage_y(y)            (((y) & gp0_tpage_y_mask)            << gp0_tpage_y_shift)
+#define enc_gp0_tpage_y(y)            ((y) << gp0_tpage_y_shift)
-#define enc_gp0_tpage_semi_trans(s)   (((s) & gp0_tpage_semi_trans_mask)   << gp0_tpage_semi_trans_shift)
+#define enc_gp0_tpage_semi_trans(s)   ((s) << gp0_tpage_semi_trans_shift)
-#define enc_gp0_tpage_color_depth(c)  (((c) & gp0_tpage_color_depth_mask)  << gp0_tpage_color_depth_shift)
+#define enc_gp0_tpage_color_depth(c)  ((c) << gp0_tpage_color_depth_shift)
-#define enc_gp0_tpage_dither(d)       (((d) & gp0_tpage_dither_mask)       << gp0_tpage_dither_shift)
+#define enc_gp0_tpage_dither(d)       ((d) << gp0_tpage_dither_shift)
-#define enc_gp0_tpage_draw_to_disp(d) (((d) & gp0_tpage_draw_to_disp_mask) << gp0_tpage_draw_to_disp_shift)
+#define enc_gp0_tpage_draw_to_disp(d) ((d) << gp0_tpage_draw_to_disp_shift)
-#define enc_gp0_tpage_tex_disable(t)  (((t) & gp0_tpage_tex_disable_mask)  << gp0_tpage_tex_disable_shift)
+#define enc_gp0_tpage_tex_disable(t)  ((t) << gp0_tpage_tex_disable_shift)
 
 /* ---- Layer 2: TPage composite encoder. Mirrors enc_gte_cmdw in gte.h ---- */
 #define enc_gp0_tpage_word(x, y, semi_trans, color_depth, dither, draw_to_disp, tex_disable) \
@@ -617,17 +616,17 @@ typedef Struct_(TexturePage) { U4 raw; };
  *    bits 24..31  = command byte — 0x20 (4bpp load) or 0x25 (8bpp load)
  * ============================================================================ */
 enum {
-	/* ---- Layer 1: CLUT bitfield shifts / widths / masks ---- */
+	/* ---- Layer 1: CLUT bitfield shifts / widths ---- */
-	gp0_clut_y_shift = 0, gp0_clut_y_width = 6, gp0_clut_y_mask = 0x3F,
+	gp0_clut_y_shift = 0, gp0_clut_y_width = 6,
-	gp0_clut_x_shift = 6, gp0_clut_x_width = 9, gp0_clut_x_mask = 0x1FF,
+	gp0_clut_x_shift = 6, gp0_clut_x_width = 9,
 	/* CLUT-load cmd-byte variants — the upper byte of the GP0 word. */
 	gp0_clut_cmd_Load4bpp = 0x20,
 	gp0_clut_cmd_Load8bpp = 0x25,
 };
 
 /* ---- Layer 1.5: CLUT per-field encoders ---- */
-#define enc_gp0_clut_x(x) (((x) & gp0_clut_x_mask) << gp0_clut_x_shift)
+#define enc_gp0_clut_x(x) ((x) << gp0_clut_x_shift)
-#define enc_gp0_clut_y(y) (((y) & gp0_clut_y_mask) << gp0_clut_y_shift)
+#define enc_gp0_clut_y(y) ((y) << gp0_clut_y_shift)
 
 /* ---- Layer 2: CLUT composite encoder ---- */
 #define enc_gp0_clut_word(cmd, x, y) (enc_gp0_cmd(cmd) | enc_gp0_clut_x(x) | enc_gp0_clut_y(y))

code/duffle/gte.atom.c

@@ -0,0 +1,385 @@
+#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(AtomBuilder_R ab, U4 r_face_cusor, U4 r_i0, U4 r_i1, U4 r_i2)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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(AtomBuilder_R ab, U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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(AtomBuilder_R ab, U4 r_vert_base, U4 r_v0, U4 r_v1, U4 r_v2) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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(AtomBuilder_R ab, U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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(AtomBuilder_R ab, U4 r_primitive_cursor) atom_dbg_skip MipsAtomComp_Proc_(ab, { gte_sw(C2_SXY2, r_primitive_cursor, O_(Poly_G4,p3)) })
+
+/* ─── STAGE 1 of normalize: SQR + mfc2 MAC1/2/3 ───
+ * Emits squared magnitude per component (in MAC1/2/3) into caller-provided scratch regs.
+ * Stage 2 of normalize consumes these directly.
+ * Words: 8.  Clobbers: IR1/2/3, MAC1/2/3.  Uses gte_cmdw_sqr (sf=0, lm=1). */
+FI_ Slice_MipsCode ac_gte_sqr_v3(AtomBuilder_R ab, U4 r_sx, U4 r_sy, U4 r_sz, U4 r_sq_x, U4 r_sq_y, U4 r_sq_z) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	mac_gte_sqr_v3s4(r_sx, r_sy, r_sz, nop),
+	gte_mv_from_data_r(r_sq_x, C2_MAC1),
+	gte_mv_from_data_r(r_sq_y, C2_MAC2),
+	gte_mv_from_data_r(r_sq_z, C2_MAC3),
+})
+
+/* ─── SQR FIRE — mtc2 3 GPRs into IR1/IR2/IR3, then fire SQR. ───
+ * The SQR command always squares IR1/IR2/IR3 — those C2 registers are fixed.
+ * The GPRs holding the source vector are caller-determined.
+ * Words: 5 (3 mtc2 + 1 nop hazard + 1 cmd). */
+FI_ Slice_MipsCode ac_gte_sqr_v3s4(AtomBuilder_R ab, Reg r_sx, Reg r_sy, Reg r_sz, MipsCode nop_slot)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(r_sx, C2_IR1),
+	gte_mv_to_data_r(r_sy, C2_IR2),
+	gte_mv_to_data_r(r_sz, C2_IR3),
+	nop_slot, gte_cmdw_sqr,
+})
+
+/* ─── STAGE 4 of normalize: mtc2 IR0..3 + GPF + mfc2 MAC + srav finalize ───
+ * Reusable standalone — given an IR0 = 1/|v| estimate (typically from a sqrtbl lookup) and a shift count 
+ * (typically (31 - LZCR)/2), multiplies IR0*IR[i] via GPF and shifts right to produce the normalized output.
+ * Used standalone for "scale vector by scalar".
+ * Words: 11.  Clobbers: IR0..3, MAC1..3.  Uses gte_cmdw_gpf (sf=0, lm=0). */
+FI_ Slice_MipsCode ac_gte_gpf_scale(AtomBuilder_R ab, 
+	U4 r_sx, U4 r_sy, U4 r_sz, 
+	U4 r_recip_est, U4 r_shift, 
+	U4 r_dx, U4 r_dy, U4 r_dz)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(r_recip_est, C2_IR0),
+	gte_mv_to_data_r(r_sx,        C2_IR1),
+	gte_mv_to_data_r(r_sy,        C2_IR2),
+	gte_mv_to_data_r(r_sz,        C2_IR3),
+	nop2, /* retire IR0..IR3 → GPF input pre-fill (matches libgte 0x80016134..0x80016138) */
+	gte_cmdw_gpf,
+	gte_mv_from_data_r(r_dx, C2_MAC1),
+	gte_mv_from_data_r(r_dy, C2_MAC2),
+	gte_mv_from_data_r(r_dz, C2_MAC3),
+	shift_aright_var(r_dx, r_dx, r_shift),
+	shift_aright_var(r_dy, r_dy, r_shift),
+	shift_aright_var(r_dz, r_dz, r_shift),
+})
+
+/* ─── TRANS MATRIX (libgte TransMatrix port) ───
+ * Atom component — auto-generates mac_trans_matrix Mac composer macro.
+ * m->t = v (struct copy; libgte's TransMatrix at 0x8001a540 is just 3 store_words, no GTE, no add).
+ * Uses 1 GPR (r_t1 = off value) per axis; per-axis load-delay-slot pattern.
+ * Words: 9.  Clobbers: r_t1. */
+FI_ Slice_MipsCode ac_trans_mt3s3s4(AtomBuilder_R ab
+	,	U4 r_mtx, U4 r_off
+	,	U4 r_t0, U4 r_t1, U4 r_t2
+) MipsAtomComp_Proc_(ab, {
+	load_word(r_t0, r_off, O_(V3_S4,x)),
+	load_word(r_t1, r_off, O_(V3_S4,y)),
+	load_word(r_t2, r_off, O_(V3_S4,z)),
+	store_word(r_t0, r_mtx, O_(MT3_S2S4,t[0])),
+	store_word(r_t1, r_mtx, O_(MT3_S2S4,t[1])),
+	store_word(r_t2, r_mtx, O_(MT3_S2S4,t[2])),
+})
+
+/* ─── LZCR ROUND EVEN + HALF-SHIFT ───
+ * Takes the raw LZCR leading-zero/ones count (from mfc2 C2_LZCR, range 1..32
+ * per PSX-SPX cop2r31) and the |v|² sum (in r_mag_sq from the MAC1+MAC2+MAC3
+ * add). Produces:
+ *   r_shift        ← LZCR rounded down to even (clear bit 0)
+ *   r_mag_sq_copy  ← |v|² sum (moved out of r_mag_sq before it's overwritten)
+ *   r_mag_sq       ← (31 - even_LZCR) / 2 = the final srav/GPF shift amount
+ *
+ * Rounding to even ensures (31 - LZCR) is always odd, so the >> 1 division
+ * is consistent — no 0.5 loss. The caller branches on LZCR < 24 to decide
+ * left-shift vs right-shift of r_mag_sq_copy, then saves the shift count.
+ *
+ * Note: C2_LZCR (cop2r31) is a fixed read-only C2 data register — the caller
+ * must read it via mfc2 from C2_LZCR; there is no register choice at the
+ * hardware level. Only the GPR that holds the result is caller-determined. */
+FI_ Slice_MipsCode ac_lzcr_round_even_half_shift(AtomBuilder_R ab,
+	U4 r_shift,
+	U4 r_mag_sq,
+	U4 r_mag_sq_copy
+)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	and_i(r_shift, r_shift, gte_lzcr_even_mask),
+	or_u(r_mag_sq_copy, r_mag_sq, 0),
+	li_s(r_mag_sq, 31),
+	sub_s(r_mag_sq, r_mag_sq, r_shift),
+	shift_aright(r_mag_sq, r_mag_sq, 1),
+})
+
+FI_ Slice_MipsCode ac_shift_aright_var_v3(AtomBuilder_R ab
+	, Reg rd_v0, Reg rd_v1, Reg rd_v2
+	, Reg rs_v0, Reg rs_v1, Reg rs_v2
+	, Reg r_shift)
+MipsAtomComp_Proc_(ab, {
+	shift_aright_var(rd_v0, rs_v0, r_shift),
+	shift_aright_var(rd_v1, rs_v1, r_shift),
+	shift_aright_var(rd_v2, rs_v2, r_shift),
+})
+
+FI_ Slice_MipsCode ac_shift_aright_var_v3_self(AtomBuilder_R ab
+	, Reg rds_v0, Reg rds_v1, Reg rds_v2
+	, Reg r_shift)
+MipsAtomComp_Proc_(ab, {
+	shift_aright_var(rds_v0, rds_v0, r_shift),
+	shift_aright_var(rds_v1, rds_v1, r_shift),
+	shift_aright_var(rds_v2, rds_v2, r_shift),
+})
+
+FI_ Slice_MipsCode ac_gte_general_purpose_interopolation(AtomBuilder_R ab
+	, Reg to_ir0,  Reg to_ir1,  Reg to_ir2, Reg to_ir3
+	, Reg fr_mac1, Reg fr_mac2, Reg fr_mac3
+	, MipsCode nop_slot1, MipsCode nop_slot2)
+MipsAtomComp_Proc_(ab, {
+	gte_mv_to_data_r(to_ir0, C2_IR0),
+	gte_mv_to_data_r(to_ir1, C2_IR1), /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */
+	gte_mv_to_data_r(to_ir2, C2_IR2),
+	gte_mv_to_data_r(to_ir3, C2_IR3), /* IR3 = src.z (reloaded) */
+	LdSlot_ nop_slot1,
+	LdSlot_ nop_slot2,
+	gte_cmdw_gpf,
+	gte_mv_from_data_r(fr_mac1, C2_MAC1),
+	gte_mv_from_data_r(fr_mac2, C2_MAC2),
+	gte_mv_from_data_r(fr_mac3, C2_MAC3),
+})
+
+FI_ Slice_MipsCode gte_mv_from_data_r_mac123(AtomBuilder_R ab
+	, Reg fr_mac1, Reg fr_mac2, Reg fr_mac3
+)
+MipsAtomComp_Proc_(ab, {
+	gte_mv_from_data_r(fr_mac1, C2_MAC1),
+	gte_mv_from_data_r(fr_mac2, C2_MAC2),
+	gte_mv_from_data_r(fr_mac3, C2_MAC3),
+})
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Atom Procs
+
+/* ─── Local copy of PSYQ's sqrtbl (1/sqrt lookup table for VectorNormal). ───
+ * Source: PSYQ 4.7 libgte sqrtbl at 0x800185B4 in hello_camera.elf.
+ *   objdump -s --start-address=0x800185B4 --stop-address=0x800185F4 hello_camera.elf
+ *   → 192 entries × 16-bit signed, in 1.12 fixed-point (max value 0x1000 = 1.0).
+ *
+ * Data is identical to the libgte original (byte-for-byte verified).
+ * 
+ * ─── Per-entry semantics (decoded from libgte msc02 VectorNormal) ───
+ * Each entry is `1/sqrt(x)` in 1.12 fixed point (value / 4096).
+ * The 192 entries span 4 octaves of the input magnitude, with 48 entries per octave:
+ *   Octave 0 (entries  0- 47): mantissa in [0x8000,  0x10000)  output ~[1.000, 0.707]
+ *   Octave 1 (entries 48- 95): mantissa in [0x10000, 0x20000)  output ~[0.707, 0.500]
+ *   Octave 2 (entries 96-143): mantissa in [0x20000, 0x40000)  output ~[0.500, 0.354]
+ *   Octave 3 (entries144-191): mantissa in [0x40000, 0x80000)  output ~[0.354, 0.251]
+ * Within each octave, 8 sub-entries interpolate over the 8 fractional bits of the mantissa
+ * (the byte `(0x80 | (i mod 8))` for the lower-byte of the aligned value).
+ * Sampling the first value of each octave:
+ *   [0]   0x1000 = 1.0000                 ; 1 / sqrt(1.0000)
+ *   [48]  0x0e4f = 0.8940                 ; 1 / sqrt(1.2500)
+ *   [96]  0x0d10 = 0.8164                 ; 1 / sqrt(1.5000)
+ *   [144] 0x0c0a = 0.7520                 ; 1 / sqrt(1.7500)
+ * And representative sub-entries within octave 0 (mantissa in [0x8000, 0x8100)):
+ *   [0]   0x1000 = 1.0000                 ; 1 / sqrt(0x8000)
+ *   [1]   0x0fe0 = 0.9922                 ; 1 / sqrt(0x8100)
+ *   [2]   0x0fc1 = 0.9846                 ; 1 / sqrt(0x8200)
+ *   [3]   0x0fa3 = 0.9773                 ; 1 / sqrt(0x8300)
+ *   [4]   0x0f85 = 0.9700                 ; 1 / sqrt(0x8400)
+ *   [5]   0x0f68 = 0.9629                 ; 1 / sqrt(0x8500)
+ *   [6]   0x0f4c = 0.9561                 ; 1 / sqrt(0x8600)
+ *   [7]   0x0f30 = 0.9492                 ; 1 / sqrt(0x8700)
+ *
+ * The algorithm's `addi -64 / sll 1 / lh` selects the entry at `(aligned - 64) * 2` for the case where `aligned` has its top bit at bit 24.
+ * After the sllv/srav pair, `aligned` always lands in `[0x80, 0x100)` 
+ * (with top bit at bit 24 → after `sub $aligned - 64`, the index sits in `[0x40, 0x80) * 2 = [0x80, 0x100)` bytes = entries [64, 128) within the sqrtbl).
+ * The earlier 64 entries (octave 0) are reached when the magnitude after shifting puts the top bit below bit 24 (the `sllv` branch),
+ * and the load upper_halves of the table bracket the input range.
+ * The later 64 entries (octaves 2-3) are the `srav` branch when the magnitude's top bit is well above bit 24.
+ *
+ * 192-entry table is reproduced verbatim from libgte (verified against libpsn00b/psxgte/vector.s:100-123 — 24 rows × 8 halfwords, last entry 0x0804). */
+internal S2 const gte_normalize_sqr_tbl[192] align_(2) = {
+	0x1000, 0x0fe0, 0x0fc1, 0x0fa3, 0x0f85, 0x0f68, 0x0f4c, 0x0f30,
+	0x0f15, 0x0efb, 0x0ee1, 0x0ec7, 0x0eae, 0x0e96, 0x0e7e, 0x0e66,
+	0x0e4f, 0x0e38, 0x0e22, 0x0e0c, 0x0df7, 0x0de2, 0x0dcd, 0x0db9,
+	0x0da5, 0x0d91, 0x0d7e, 0x0d6b, 0x0d58, 0x0d45, 0x0d33, 0x0d21,
+	0x0d10, 0x0cff, 0x0cee, 0x0cdd, 0x0ccc, 0x0cbc, 0x0cac, 0x0c9c,
+	0x0c8d, 0x0c7d, 0x0c6e, 0x0c5f, 0x0c51, 0x0c42, 0x0c34, 0x0c26,
+	0x0c18, 0x0c0a, 0x0bfd, 0x0bef, 0x0be2, 0x0bd5, 0x0bc8, 0x0bbb,
+	0x0baf, 0x0ba2, 0x0b96, 0x0b8a, 0x0b7e, 0x0b72, 0x0b67, 0x0b5b,
+	0x0b50, 0x0b45, 0x0b39, 0x0b2e, 0x0b24, 0x0b19, 0x0b0e, 0x0b04,
+	0x0af9, 0x0aef, 0x0ae5, 0x0adb, 0x0ad1, 0x0ac7, 0x0abd, 0x0ab4,
+	0x0aaa, 0x0aa1, 0x0a97, 0x0a8e, 0x0a85, 0x0a7c, 0x0a73, 0x0a6a,
+	0x0a61, 0x0a59, 0x0a50, 0x0a47, 0x0a3f, 0x0a37, 0x0a2e, 0x0a26,
+	0x0a1e, 0x0a16, 0x0a0e, 0x0a06, 0x09fe, 0x09f6, 0x09ef, 0x09e7,
+	0x09e0, 0x09d8, 0x09d1, 0x09c9, 0x09c2, 0x09bb, 0x09b4, 0x09ad,
+	0x09a5, 0x099e, 0x0998, 0x0991, 0x098a, 0x0983, 0x097c, 0x0976,
+	0x096f, 0x0969, 0x0962, 0x095c, 0x0955, 0x094f, 0x0949, 0x0943,
+	0x093c, 0x0936, 0x0930, 0x092a, 0x0924, 0x091e, 0x0918, 0x0912,
+	0x090d, 0x0907, 0x0901, 0x08fb, 0x08f6, 0x08f0, 0x08eb, 0x08e5,
+	0x08e0, 0x08da, 0x08d5, 0x08cf, 0x08ca, 0x08c5, 0x08bf, 0x08ba,
+	0x08b5, 0x08b0, 0x08ab, 0x08a6, 0x08a1, 0x089c, 0x0897, 0x0892,
+	0x088d, 0x0888, 0x0883, 0x087e, 0x087a, 0x0875, 0x0870, 0x086b,
+	0x0867, 0x0862, 0x085e, 0x0859, 0x0855, 0x0850, 0x084c, 0x0847,
+	0x0843, 0x083e, 0x083a, 0x0836, 0x0831, 0x082d, 0x0829, 0x0824,
+	0x0820, 0x081c, 0x0818, 0x0814, 0x0810, 0x080c, 0x0808, 0x0804,
+};
+
+/* ─── Full normalize (all 4 stages inline) ───
+ * Generic 4-stage GTE normalize (SQR → sum+LZCR → align+sqrtbl → GPF+srav).
+ *
+ * Parameterized by caller-provided scratch base + src/dst offsets.
+ * The caller passes r_src_offset and r_dst_offset as compile-time constants
+ * (typically derived from O_ macros in the caller's struct schema, e.g., `O_(CallerBundleScratch, fwd)`).
+ *
+ * This design lets any caller (with a scratch base + struct schema) use `normalize_v3s4_proc` 
+ * without putting magic offsets in the C-side bundle helper — the offsets come from O_ macros at the call site.
+ *
+ * Body uses 9 GPRs (r_src_ptr..r_branch_tmp):
+ *   r_src_ptr, r_dst_ptr : src/dst pointers (computed from r_scratch + caller offsets)
+ *   r_tmp                : src.x PRESERVED across stages 1-2 (NOT clobbered by mfc2 MAC2) → fed to IR1 in stage 4
+ *   r_mac1_scratch       : MAC1 result scratch (also holds aligned |v|² in stage 3)
+ *   r_mac2_scratch       : MAC2 result scratch → result.x after stage 4 sra
+ *   r_recip_est          : src.y PRESERVED across stages 1-2 → fed to IR2 in stage 4 → result.y
+ *   r_norm               : |v|² sum (stage 2) → half-shift (stage 3) → 1/|v| (stage 4 IR0)
+ *   r_shift              : shift count SAVED in stage 3 → consumed by stage 4 srav
+ *   r_branch_tmp         : src.z PRESERVED across stages 1-2 → fed to IR3 in stage 4 → result.z (also sqrtbl base addr)
+ *
+ * Atom_labels are srav_path / aligned_done 
+ * (NOT namespaced — they're internal to this proc; 
+ * the metaprogram's per-atom-name enum emission handles any collision across different atoms/files that share the same labels).
+ *
+ * Pool cost: 11 GPRs (well within the 9-10 caller-trash GPR budget when r_scratch is a wave-context carrier).
+ *
+ * Direct port of PSYQ libgte msc02.rel.text VectorNormal disassembly (0x800160a0..0x8001615c).
+ * Words: ~59 (matches libgte 0x800160a0..0x8001615c at +/- 0-2 words for BD-slot reshuffling).
+ * Sqrtbl: hardcoded to 0x800185B4 (libgte msc02.rel.data). Note: swapped to local.
+ * Pipeline: clobbers IR0..3, MAC1..3, LZCS, LZCR.
+ */
+/* MipsAtom_Proc_ wrapper: declares the static MipsCode[] body, then calls atombuilder_unroll(ab, ...) to copy the encoded instructions into the caller's MipsAtomBuilder arena.  */
+internal MipsAtom* normalize_v3s4_proc(AtomArena_R aa,	U4 r_scratch /* GPR code: scratch base carrier (e.g., R_T4 = R_ResolveScratch) */
+	,	U4 src_offset, U4 dst_offset            /* GPR codes: PARAMETERIZED offsets (caller passes O_ macros) */
+	,	Reg r_src_ptr, Reg r_dst_ptr, Reg r_tmp /* GPR codes: 3 scratch regs (src/dst computed + tmp) */
+	,	Reg r_mac1_scratch, Reg r_mac2_scratch  /* GPR codes: 2 more: MAC1/MAC2 scratch */
+	,	Reg r_recip_est                         /* GPR code: |v|² sum + shift-input + sqrtbl[index] */
+	,	Reg r_norm, Reg r_shift                 /* GPR codes: normalize working reg + final srav amount */
+	,	Reg r_branch_tmp                        /* GPR code: scratch (shift count, branch target, lookup addr) */
+)
+MipsAtom_Proc_(aa, {
+	add_si(r_src_ptr, r_scratch, src_offset),     /* r_src_ptr = &src */
+	// add_si(r_dst_ptr, r_scratch, r_dst_offset),     /* r_dst_ptr = &dst */
+
+	/* Load src.x/y/z from r_src_ptr (caller-determined address) into r_tmp/r_recip_est/r_branch_tmp.
+	 * r_tmp holds src.x throughout stages 1-2 — r_mac2_scratch is clobbered to MAC2 in stage 1.5 (line below). */
+	mac_load_v3s4(r_tmp, r_recip_est, r_branch_tmp, r_src_ptr, 0),
+
+	/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */
+	LdSlot_ mac_gte_sqr_v3s4(r_tmp, r_recip_est, r_branch_tmp, LdSlot_ nop),
+
+	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. */
+	mac_gte_mv_from_data_r_mac123(r_mac1_scratch, r_mac2_scratch, r_norm), LdSlot_ nop,
+	add_u_self(        r_norm, r_mac1_scratch),
+	add_u_self(        r_norm, r_mac2_scratch),
+	gte_mv_to_data_r(  r_norm,  C2_LZCS), LdSlot_ nop2,
+	gte_mv_from_data_r(r_shift, C2_LZCR), LdSlot_ nop,
+
+	/* Stage 3: round LZCR to even, compute half-shift, align |v|² to bit 24.
+	 * r_norm holds |v|² sum; r_shift holds the LZCR count from mfc2.
+	 * After the component: r_shift = even(LZCR), r_norm = half-shift, r_mac1_scratch = |v|². */
+	mac_lzcr_round_even_half_shift(r_shift, r_norm, r_mac1_scratch),
+	/* r_branch_tmp = LZCR - 24 (overwrites r_branch_tmp; src.z no longer needed after SQR) */
+	add_si(        r_branch_tmp, r_shift, -24),
+	branch_lt_zero(r_branch_tmp, atom_offset(aligned_done, srav_path)), BdSlot_ nop, /* bltz → srav_path (LZCR <  24 path) */
+		jump_rel(atom_offset(srav_path, aligned_done)),                                /* b → aligned_done (LZCR >= 24 path) */
+		BdSlot_ shift_lleft_var(r_mac1_scratch, r_mac1_scratch, r_branch_tmp),         /* src=sum (r_mac1_scratch), dst=same */
+	atom_label(srav_path)
+		li_s( r_branch_tmp, 24),
+		sub_s(r_branch_tmp, r_branch_tmp, r_shift),
+		shift_aright_var(r_mac1_scratch, r_mac1_scratch, r_branch_tmp), /* src=sum (r_mac1_scratch), dst=same */
+atom_label(aligned_done)
+	// Save the shift count to r_shift before the next 5 instructions overwrite r_norm (the sqrtbl lookup loads 1/|v| into r_norm, which becomes IR0 in stage 4).
+	or_u(r_shift, r_norm, 0), /* r_shift ← shift count (preserved through stage 4) */
+	/* r_mac1_scratch holds |v|² aligned (top bit at bit 7). */
+	add_si(     r_mac1_scratch, r_mac1_scratch, -64),
+	shift_lleft(r_mac1_scratch, r_mac1_scratch, 1),
+	mac_load_word_imm(r_branch_tmp, & gte_normalize_sqr_tbl), add_u_self(r_branch_tmp, r_mac1_scratch),
+	load_half(r_norm, r_branch_tmp, 0), /* r_norm = sqrtbl[aligned-64] = 1/|v| (IR0 in stage 4) */
+
+	/* r_branch_tmp held the sqrtbl base+index, NOT src.z. Reload src.z from scratch now that r_branch_tmp is free. */
+	LdSlot_ load_word(r_branch_tmp, r_src_ptr, O_(V3_S4,z)), /* r_branch_tmp = src.z (for IR3 in stage 4) */
+
+	/* Stage 4: GPF + srav finalize (r_shift = shift count, r_norm = 1/|v|). */
+	LdSlot_ mac_gte_general_purpose_interopolation(
+		r_norm, 
+		r_tmp,        /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */
+		r_recip_est, 
+		r_branch_tmp, /* IR3 = src.z (reloaded) */
+		r_mac2_scratch, r_recip_est, r_branch_tmp,
+		LdSlot_ add_si(r_dst_ptr, r_scratch, dst_offset), // pre-laoding destination to register here.
+		LdSlot_ nop
+	),
+	/* sra by r_shift = (31-LZCR)/2 (saved before sqrtbl lookup) */
+	mac_shift_aright_var_v3_self(r_mac2_scratch, r_recip_est, r_branch_tmp, r_shift),
+	/* Store result.x/y/z to r_dst_ptr (caller-determined dst address). */
+	mac_store_v3s4(r_mac2_scratch, r_recip_est, r_branch_tmp, r_dst_ptr, 0),
+
+	mac_yield()
+})
+
+#pragma endregion Atom Procs
+
+#pragma region Baked Atoms
+
+typedef Struct_(Binds_SetGteMT3S2S4) {
+	MT3_S2S4* transform;
+};
+internal MipsAtom_(set_gte_mt3s2s4) atom_info(
+		atom_bind(Binds_SetGteMT3S2S4)
+	,	atom_reads(R_TapePtr)
+){
+	/* Pop matrix address from tape into R_T3 ($11) */
+	load_word(R_T3, R_TapePtr, O_(Binds_SetGteMT3S2S4,transform)),
+	add_ui_self(    R_TapePtr, S_(Binds_SetGteMT3S2S4)),
+	/* 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

@@ -161,6 +161,8 @@ enum {
 	gte_cmd_nclip     = 0x06, /* Normal Clipping (Backface culling) */
 	gte_cmd_op        = 0x0C, /* Outer Product */
 	gte_cmd_mvmva     = 0x12, /* Matrix Vector Multiply & Add (Custom math) */
+	gte_cmd_sqr       = 0x28, /* Square vector — MAC[i] = IR[i]²; IR[i] ← MAC[i] saturated */
+	gte_cmd_gpf       = 0x3D, /* General-purpose Interpolation — MAC[i] = IR0 * IR[i]  */
 
 /* --- GTE Command Bit-Field Layout ---
  * A GTE command word (sent to COP2 with RS=1) is laid out as:
@@ -171,19 +173,47 @@ enum {
  *   +------------+--+-----+------+------+------+------+---+--------+----------+
  *                                    \_____ GTE_PAYLOAD _____/       \__ GTE_CMD __/
  *
- * Shifts/masks below are the *bit positions* and *bit widths* of each
+ * Shifts/masks below are the *bit positions* and *bit widths* of each configurable field, used by the ENC_GTE_CMD encoder. 
- * configurable field, used by the ENC_GTE_CMD encoder. 
  * Mirrors the OPCODE_SHIFT / RS_SHIFT convention used in mips.h.
  */
 
-	gte_shift_sf  = 19,  gte_width_sf  = 1,  gte_mask_sf  = 0x1,
+	gte_shift_sf  = 19,  gte_width_sf  = 1,
-	gte_shift_mx  = 17,  gte_width_mx  = 2,  gte_mask_mx  = 0x3,
+	gte_shift_mx  = 17,  gte_width_mx  = 2,
-	gte_shift_v   = 15,  gte_width_v   = 2,  gte_mask_v   = 0x3,
+	gte_shift_v   = 15,  gte_width_v   = 2,
-	gte_shift_cv  = 13,  gte_width_cv  = 2,  gte_mask_cv  = 0x3,
+	gte_shift_cv  = 13,  gte_width_cv  = 2,
-	gte_shift_lm  = 10,  gte_width_lm  = 1,  gte_mask_lm  = 0x1,
+	gte_shift_lm  = 10,  gte_width_lm  = 1,
-	gte_shift_cmd =  0,  gte_width_cmd = 6,  gte_mask_cmd = 0x3F,
+	gte_shift_cmd =  0,  gte_width_cmd = 6,
+
+	/* Fake command number (bits 24-20) — IGNORED by the GTE hardware per PSX-SPX `geometrytransformationenginegte.md` line 48.
+	 * libgte's compiler emits non-zero values in this field as a disassembly signature. */
+	gte_shift_fake_cmd = 20,
+	gte_width_fake_cmd = 5,
 };
 
+/* --- GTE Control Register Aliases (Pitfall 1) ---
+ * Three pairs of aliases map to the SAME C2 control-register slot on real silicon:
+ *   C2[24] = gte_cr_RBK (background R)         | gte_cr_OFX (screen offset X)
+ *   C2[25] = gte_cr_GBK (background G)         | gte_cr_OFY (screen offset Y)
+ *   C2[26] = gte_cr_BBK (background B)         | gte_cr_H   (projection plane distance H)
+ * Cross-alias writes inside one atom body, or across the wave-context boundary,
+ * silently clobber each other. The metaprogram's check_gte_cr_alias_writes
+ * (CHECK_RULES row) warns about each pair per source. See
+ * docs/gte_reference.md §"Control-register alias table" for the silicon
+ * rationale and the libgte outer-product convention.
+ */
+
+/* --- RT-matrix packed-slot convention (Pitfall 4) ---
+ * The silicon packs two 16-bit RT elements per 32-bit C2 slot:
+ *   C2[2] = (RT22 << 16) | RT13       (gte_cr_RT13 writes the low  half, gte_cr_RT22 writes the high half)
+ *   C2[4] = (RT33 << 16) | RT22       (gte_cr_RT22 writes the low  half — clobbers prior RT22 value if RT13 was also written)
+ * OP and MVMVA read D1/D2/D3 from these packed slots. The libgte outer-product
+ * convention (see ac_apply_matrix_lv at gte.atom.c:108-122) writes C2[2] then
+ * C2[4] in sequence; the SECOND write's low half is RT22, not RT13. An agent
+ * who writes gte_cr_RT13 then gte_cr_RT22 to the SAME source GPR clobbers the
+ * RT13 value. See docs/gte_reference.md §"RT-matrix packed-slot convention"
+ * for the canonical write pattern.
+ */
+
 /* --- 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).
@@ -243,10 +273,10 @@ enum { _C2_OPS_ = 0
  *   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
+ * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit numbers as 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).
+ * and so the encoding is next to its only consumer (this header).
  *
  * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2) live in gte_vendor_sym.h. */
 enum { _C2_TX_SUBS_ = 0
@@ -309,23 +339,24 @@ enum { _C2_TX_SUBS_ = 0
 
 /* GTE Command Format
  * Opcode is always MIPS_OP_COP2, RS is always 1 (CO).
- * The lower 25 bits are the GTE-specific command payload.
+ * Lower 25 bits are GTE-specific command payload.
  *
- * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs` pattern in mips.h:
+ * The `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.
+ * `ENC_GTE_CMD` is an all-in-one convenience for emitting a full command word.
  * 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. */
-#define enc_gte_sf(sf)   (((sf)  & gte_mask_sf ) << gte_shift_sf )
+#define enc_gte_sf(sf)          ((sf)  << gte_shift_sf )
-#define enc_gte_mx(mx)   (((mx)  & gte_mask_mx ) << gte_shift_mx )
+#define enc_gte_mx(mx)          ((mx)  << gte_shift_mx )
-#define enc_gte_v(v)     (((v)   & gte_mask_v  ) << gte_shift_v  )
+#define enc_gte_v(v)            ((v)   << gte_shift_v  )
-#define enc_gte_cv(cv)   (((cv)  & gte_mask_cv ) << gte_shift_cv )
+#define enc_gte_cv(cv)          ((cv)  << gte_shift_cv )
-#define enc_gte_lm(lm)   (((lm)  & gte_mask_lm ) << gte_shift_lm )
+#define enc_gte_lm(lm)          ((lm)  << gte_shift_lm )
-#define enc_gte_cmd(cmd) (((cmd) & gte_mask_cmd) << gte_shift_cmd)
+#define enc_gte_cmd(cmd)        ((cmd) << gte_shift_cmd )
+#define enc_gte_fake_cmd(x)     ((x)   << gte_shift_fake_cmd)
 
 /* Composite: all six GTE fields + the COP2/CO base. */
 #define enc_gte_cmdw(sf, mx, v, cv, lm, cmd) ( \
@@ -363,11 +394,11 @@ enum { _C2_TX_SUBS_ = 0
  *  (the perspective divide happens regardless of `sf`).
  *
  *  If we emit a strictly-spec-compliant word (`sf=0`, reserved bits clear),
- *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops — 
+ *  PCSX-Redux's GTE checks those bits more strictly than the silicon does and RTPT silently no-ops.
- *  the floor's screen coordinates come out as raw projection-of-rotation (Z never divided),
+ *  The floor's screen coordinates come out as raw projection-of-rotation (Z never divided), 
  *  `nclip` ends up wrong, and the triangle is culled.
  *
- *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to match the working bit pattern everyone has shipped for 25 years.
+ *  So for RTPS and RTPT we OR-in the `0x28` "PsyQ compat" pattern to match the working bit pattern.
  *  NCLIP / OP / MVMVA stay spec-clean — their reserved bits really are zero in the original PsyQ source.
  * --------------------------------------------------------------------------
  */
@@ -378,11 +409,92 @@ enum { _C2_TX_SUBS_ = 0
 #define gte_cmdw_nclip  (gte_cmd_base | enc_gte_cmd(gte_cmd_nclip))
 #define gte_cmdw_op     (gte_cmd_base | enc_gte_cmd(gte_cmd_op   ))
 #define gte_cmdw_outer_product gte_cmdw_op      /* "outer product" -- NOCASH/Sdk terminology */
-#define gte_cmdw_wedge        gte_cmdw_op      /* "wedge product"  -- geometric-algebra terminology */
+#define gte_cmdw_wedge         gte_cmdw_op      /* "wedge product"  -- geometric-algebra terminology.
+	* RGA(Lengyel): the GTE OP is a 3D signed-16-bit D x IR cross, not a generic RGA exterior product.
+	* The wedge alias is the 3D complement interpretation of the same 3 scalars (MAC1..MAC3). */
 #define gte_cmdw_mvmva  (gte_cmd_base | enc_gte_cmd(gte_cmd_mvmva))
 
+/* MVMVA with sf=0 (no shift, full-integer), cv=3 (no translation), v=3 (IR vector input).
+ * Reads input from IR1/2/3 (loaded via mtc2 rt, C2_IRx). MAC1/2/3 = RT row · IR (full product, no >>12).
+ * Per PSX-SPX: SAR (sf*12) with sf=0 = SAR 0 = no shift. */
+#define gte_cmdw_mvmva_sf0_ir  (gte_cmd_base | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA with sf=1 (>>12 shift, 4.12 fixed-point), cv=3 (no translation), v=3 (IR): for ApplyMatrixLV.
+ * Reads input from IR1/2/3 (loaded via mtc2 rt, C2_IRx). MAC1/2/3 = (RT row · IR) >> 12.
+ * Per PSX-SPX: SAR (sf*12) with sf=1 = SAR 12 = arithmetic right-shift by 12.
+ * This matches the libgte C-side ApplyMatrixLV output (R*pos >> 12). */
+#define gte_cmdw_mvmva_ir    (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=0, mx=3 (Light matrix), v=3 (IR), cv=3 (no TR).
+ * For pass1 of the C11 two-pass decomposition. Reads L matrix.
+ * Since L matrix is typically zero, pass1 contributes 0 to the combine. */
+#define gte_cmdw_mvmva_sf0_mx3_v3_cv3  (gte_cmd_base | enc_gte_sf(0) | enc_gte_cv(3) | enc_gte_v(3) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=1 (>>12), mx=3 (Light matrix), v=2 (V0), cv=0 (with TR).
+ * Matches the C11 ApplyMatrixLV pass 2 command word (0x49E012) exactly.
+ * The combine is (pass1 << 3) + pass2. */
+#define gte_cmdw_mvmva_pass2_c11  (gte_cmd_base | enc_gte_sf(1) | enc_gte_v(2) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=0, mx=3, v=2, cv=0. Matches the C11 pass 1 command. */
+#define gte_cmdw_mvmva_pass1_c11  (gte_cmd_base | enc_gte_v(2) | enc_gte_mx(3) | enc_gte_cmd(gte_cmd_mvmva))
+#define gte_cmdw_mvmva_no_tr gte_cmdw_mvmva_ir
+
+/* MVMVA pass 2 — C11 ApplyMatrixLV command.
+ * Decoded: op_cop2 | CO | fake_cmd=4 | sf=1 (>>12) | mx=0 (RT matrix) | v=3 (IR) | cv=3 (no translation) | lm=0 | cmd=MVMVA.
+ * Reads (RT row · IR) >> 12 into MAC1/2/3. Per-field composition (no opaque literal)
+ * keeps the bit layout visible at the call site + matches the libgte C-side byte-exact. */
+#define gte_cmdw_mvmva_c11_pass2  (gte_cmd_base | enc_gte_fake_cmd(4) | enc_gte_sf(1) | enc_gte_v(3) | enc_gte_mx(0) | enc_gte_cv(3) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* MVMVA: sf=1 (>>12), mx=0 (RT matrix), v=0 (V0), cv=3 (no TR). */
+#define gte_cmdw_mvmva_sf1_mx0_v0_cv3  (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_v(0) | enc_gte_mx(0) | enc_gte_cmd(gte_cmd_mvmva))
+
+/* RTPS with sf=1 (12-bit shift, no translation): matches the output of libgte's
+ * ApplyMatrixLV when the GTE pipeline expects R*pos >> 12. The shift produces
+ * values like (-270, 710, 1713) which match the C11 reference path. */
+#define gte_cmdw_rtps_sf1   (gte_cmd_base | enc_gte_sf(1) | enc_gte_cv(3) | enc_gte_cmd(gte_cmd_rtps))
+
+/* SQR / GPF cosmetic-bits compat helpers.
+ * Each command's `_compat` macro ORs in the `fake_cmd` field value libgte happens to emit.
+ * The hardware ignores these bits (per PSX-SPX line 48). */
+#define gte_cmdw_sqr_fake_sig  enc_gte_fake_cmd(0x0A)
+#define gte_cmdw_gpf_fake_sig  enc_gte_fake_cmd(0x19)
+
+/* SQR — Square Vector.
+ * PSX-SPX `geometrytransformationenginegte.md` §"SQR":
+ *   [MAC1,MAC2,MAC3] = [IR1*IR1, IR2*IR2, IR3*IR3] SHR (sf*12)
+ *   [IR1,IR2,IR3]    = [MAC1,MAC2,MAC3] (saturated to 0x7FFF when lm=1)
+ * Sourced verbatim from libgte msc02 VectorNormal disassembly at 0x800160b0:
+ *   0x4AA00428 = gte_cmd_base | gte_cmdw_sqr_compat | enc_gte_lm(1) | enc_gte_cmd(0x28)
+ *   bit 19     sf=0
+ *   bit 10     lm=1
+ *   bits 5-0   cmd=0x28=SQR
+ *   bits 24-20 = 0x0A (libgte "nonsense SDK command number" signature) */
+#define gte_cmdw_sqr   (gte_cmd_base | enc_gte_cmd(gte_cmd_sqr) | enc_gte_lm(1) | gte_cmdw_sqr_fake_sig)
+
+/* GPF — General-purpose Interpolation.
+ * PSX-SPX `geometrytransformationenginegte.md` §"GPF":
+ *   [MAC1,MAC2,MAC3] = (([IR1,IR2,IR3] * IR0) + [MAC1,MAC2,MAC3]) SAR (sf * 12)
+ *   [IR1,IR2,IR3]    = [MAC1,MAC2,MAC3]
+ * Sourced verbatim from libgte msc02 VectorNormal disassembly at 0x8001613c:
+ *   0x4B90003D = gte_cmd_base | gte_cmdw_gpf_compat | enc_gte_cmd(0x3D)
+ *   bit 19     sf = 0
+ *   bit 10     lm = 0
+ *   bits 5-0   cmd = 0x3D = GPF
+ *   bits 24-20 = 0x19 (libgte "nonsense SDK command number" signature) */
+#define gte_cmdw_gpf   (gte_cmd_base | enc_gte_cmd(gte_cmd_gpf) | gte_cmdw_gpf_fake_sig)
+
+/* Mask to round LZCR (leading-zero/ones count, range 1..32 per PSX-SPX cop2r31)
+ * down to even. The normalize_v3s4 half-shift logic computes (31 - LZCR) >> 1;
+ * clearing bit 0 ensures the subtraction result is always odd,
+ * so the >> 1 division is consistent (no 0.5 loss). */
+enum {
+	gte_lzcr_even_mask = 0xFFFE, /* all bits except bit 0 */
+};
+
 #define gte_cmdw_rotate_translate_perspective_single gte_cmdw_rtps
 #define gte_cmdw_rotate_translate_perspective_triple gte_cmdw_rtpt
+/* RGA(Lengyel): RTPS/RTPT consume the matrix expansion of a rigid transformation (rotation matrix + translation vector) loaded into the RT/TR control registers.
+ * For unitized points the same result equals the motor antiproduct; the GTE executes the LA form, not a symbolic antiproduct. */
 
 /* PsyQ compatibility bits for AVSZ3 (Bits 20, 22, 24 must be set) */
 #define gte_cmdw_psyq_avsz3_compat (0x15 << 20)
@@ -433,7 +545,6 @@ enum {
 #define gte_lw_v2_z(base)   enc_gte_lw(gte_in_v2_z,  (base), GTE_Z_Offset)
 
 /* gte_load_vN(r_ptr, base) — placeholder-punned lwc2 loaders
- *
  * Emits `.word` constants encoding `lwc2 $N, off(<base>)` for the chosen GTE vector register, where `<base>` is the GPR number you pass in
  * (typically one of R_T4..R_T9 for the standard "3-pointer" pattern).
  *

code/duffle/lottes_tape.h

@@ -1,36 +1,70 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
+
 #	include "dsl.h"
 #	include "gcc_asm.h"
 # include "mips.h"
 # include "gte.h"
 # include "memory.h"
-#	include "atom_dsl.h"
+#	include "dsl.atom.h"
-#	include "gen/duffle.macs.h"
-#	include "gen/duffle.offsets.h"
 #endif
 
-typedef U4 const MipsCode; // Underlying type to mips asm words.
+#pragma region Tape Drive
-typedef Slice_(MipsCode);
+/* -----------------------------------------------------------------------------------------------------------
-
+ *  TAPE DRIVE ABI
-typedef U4 const MipsAtom; // Underlying type to an array of mips asm words that must terminate with an ac_yield.
+ * -----------------------------------------------------------------------------------------------------------
-#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+ * Note(Ed): One of the main purposes of this codebase is to help me learn this, 
-
+ * as such the information below may not* be entirely realized or finalized conceptually.
-// Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
+ * -----------------------------------------------------------------------------------------------------------
-//   MipsAtomComp_(ac_X) { body }
+ * This ABI and its associated legos were directly inspired by researching the work of 
-// expands to:
+ * Timothy Lottes and Onat Türkçüoğlu; along with many others. It's the simplest bootstrap of a 
-//   MipsCode ac_X[] align_(4) = { body };
+ * directly executed chain of assemby arrays (Atoms) that terminate with a yield sequence to the next atom.
-#define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+ * These eventually lead to a terminal atom for the tape which is defined below as "tape_exit".
-
+ * 
-// Used for components with value-args (e.g., ac_format_f3_color).
+ * It behaves as one of the simplest runtime harnesses ontop of a host-enviornment's execution engine 
-//   FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+ * to author and compose programs with. From here various conventions can be further applied. 
-// expands to:
+ * To make things easier  to understand it may be better to focus on what this ABI does not have. 
-//   FI_ Slice_MipsCode ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+ * It does not have have any branching within the tape but relative branches within atoms or between atoms.
-#define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
+ * Branching nearly is always downstream. Automatic stack usage is non-existent.
-
+ * Push/Pop, FIFO, or Arena/Bump data structures are used by atoms explicitly.
-// Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
+ * In it's current form with the C11 macro DSL, the user also has fullfill manual register allocation per atom.
-
+ * 
-/* Register aliases */
+ * 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 cannot ignore excessive argument shuffle across workload or
+ * waves and thier phases. Crossing ABI boundaries to other runtimes that do has obviouss penalties.
+ * 
+ * Learning data-oriented code becomes a natural progression. Your not fighting a stack-based procedural 
+ * paradigm that wants to argument shuffle. There is no ambiguity due to the lack of constraints, for example,
+ * on how the user may "call" a procedure in traditional random dispatch runtimes. The user does have to
+ * hammer down "rules" or patterns for massaging the compiler to dissolve those call frames; just 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, or maybe something more familar like an immediate mode library 
+ * for various systems such as GUIs. 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 has just started functioning. I'm abusing C11 & a lua metaprogram to help establish 
+ * a hybrid toolchain to ideate on a traditional text-based authoring UX for this paradigm.
+ * If pcsx-redux provides viable hot-reload and persistent data storage beyond save-states
+ * (just copying ram to filesystem), I can author a color forth to mess around with.
+ * With either an editor in-emulator or on the actual machine itself. Assembly is tedius,
+ * but I think this codebase most likely has a pretty ergonomic flavor worst case...
+ * */
+/* Register Allocation Info */
 enum {
 	R_AtomJmp  = R_T8 atom_reg,  /* debug-visible; tape yield handshake scratch */
 	R_TapePtr  = R_T9 atom_reg,  /* The Instruction Stream Pointer */
@@ -59,28 +93,73 @@ enum {
 	R_TScratch6  = R_T6,
 	R_TScratch7  = R_T7,
 	R_TScratch8 =  R_T8,
-	R_TScratch10 = R_V0,
+	R_TScratch10 = R_V0, // Tend to be used with gte DMAs
-	R_TScratch11 = R_V1,
+	R_TScratch11 = R_V1, // Tend to be used with gte DMAs
 	// Note(Ed): We can technically clobber these, but don't unless we hit a bottleneck.
-	// R_TScratch12 = R_A0,
+	// A 0-2
-	// R_TScratch13 = R_A1,
+	// S 0-7
-	// R_TScratch14 = R_A3,
-	// TODO(Ed): Review S0-S7, they are technically avaialble, we just have to snapshot them at the ABI boundary.
-	// TODO(Ed): This is technically a waste of cycles for most work? so maybe only do this for expensive atoms on-demand or atom phases.
-// TODO(Ed): Sort out the other available registers... (Not sure how much is left avail)
 };
 
-#pragma region Tape Drive
+typedef U2 Reg; // Register parameter used with atom or atom component procedures
-/* ---------------------------------------------------------------------------
+
- *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
+typedef U4 const MipsCode; // Underlying type to mips asm words.
- * ---------------------------------------------------------------------------*/
+typedef Slice_(MipsCode);
-typedef Slice_(MipsAtom); typedef Slice_MipsAtom Tape;
+
+typedef U4 const MipsAtom; 
+typedef Slice_(MipsAtom); 
+// Sometimes a user will define a bundle of atoms that represent a procedure of work as:
+//   MipsAtom* <identifier>[...];
+// Unfortuantely if using slice_from_array it will make the slice's pointer: MipsAtom** so this enforce its defined as MipsAtom*
+// TODO(Ed): Alternatively we can make the MipsAtom an opaque pointer to the atom... so that the blow returns 'MipsAtom'.
+#define atombundle_from_array(array) (Slice_MipsAtom){.ptr=array[0],.len=Array_len(array)}
+
+// Underlying type to an ptr to an array of mips asm words that must terminate with an ac_yield.
+#define MipsAtom_(sym) MipsCode sym [] align_(4) =
+
+// Used for atoms with value-args
+//   internal MipsAtom* X_proc(AtomArena_R aa, args) MipsAtom_Proc_(X, aa, { body })
+// expands to:
+//   internal MipsAtom* X_proc(AtomArena_R aa, args) { MipsCode atom_comp_code[] align_(4) = { body }; return atomarena_push(aa, slice_from_array(MipsCode, atom_comp_code)); }
+// The atom name is derived by the Lua metaprogram from the preceding
+// `MipsAtom* X_proc(...)` declaration (backward walk from the macro site,
+// strips the `_proc` suffix).
+#define MipsAtom_Proc_(aa, ...) { MipsCode atom_comp_code[] align_(4) = __VA_ARGS__; return atomarena_push(aa, slice_from_array(MipsCode, atom_comp_code)); }
+
+// Used for components with no args (e.g., ac_load_tri_indices) or identifier-args (hardcoded register names).
+//   MipsAtomComp_(ac_X) { body }
+// expands to:
+//   MipsCode ac_X[] align_(4) = { body };
+#define MipsAtomComp_(sym) MipsCode sym [] align_(4) =
+
+// Used for components with value-args (mandatory `ab` (atom-builder) arg).
+//   FI_ void ac_X(MipsAtomBuilder_R ab, args) MipsAtomComp_Proc_(ab, { body })
+// expands to:
+//   FI_ void ac_X(MipsAtomBuilder_R ab, args) {
+//       MipsCode atom_comp_code[] align_(4) = { body };
+//       atombuilder_push(ab, slice_from_array(MipsCode, atom_comp_code));
+//   }
+// The body must NOT include mac_yield() (the parent atom yields).
+// The component name is derived by the Lua metaprogram from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration (backward walk from the macro site).
+// Inline-only callers (the generated `mac_<name>` aliases) skip the `ab` arg via metaprogram filtering; escape callers (ac_<name> invoked as a function) pass a long-lived builder.
+#define MipsAtomComp_Proc_(ab, ...) { MipsCode atom_comp_code[] align_(4) = __VA_ARGS__; atombuilder_push(ab, slice_from_array(MipsCode, atom_comp_code)); }
+
+/* Line-table anchor: gcc only adds a file to the .debug_line file table when the contains line-numbered content.
+	Files containing only atoms and atom components.
+   Place `ATOM_FILE_LINE_MARKER();` once at file scope in any `.atom.c` that defines atoms.
+	 Macro expands to a file-scope `internal U4 const` declaration keeps the file in the line table.
+	 The constant is in `.rodata` so the linker may eliminate it.
+	 Two-level concat + `__LINE__` suffix makes the identifier unique per call site
+	 (identifier embeds the source line, so duplicates across `#include`d files don't collide). */
+#define ATOM_FILE_DEBUGGER_LINE_MARKER(file_name) internal U4 const tmpl(atom_file_debugger_line_marker,file_name) = 0
+
+typedef Slice_MipsAtom Tape;
 
 /* The 'Exit' Atom */
-atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(R_RA), nop };
 
-//TODO(Ed): Do we backup R_S0-7 here? Have it in a heavier tape run as a opt-in? Same with V0-1 and A0-3?
+// TODO(Ed): When we have a substantial workload/throughput, profile each of these to see impact at ABI boundaries.
-/* Generalized Tape Engine Runner */
+
+/* Tape Runner (Default) */
 FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
 		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
@@ -91,29 +170,50 @@ FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape
 	asm_rpins, r_use(tape_ptr)
 	asm_clobber: 
 		rlit(R_AT),
-		rlit(R_V0), rlit(R_V1),
+		rlit(R_V0), rlit(R_V1), // We clobber these for GTE ACs (that don't expose register selection, might expose them in the future...)
 		rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
 		rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8),
 		clb_mem_drain 
 ); }
 
+/* Tape Runner (Static and Arg Clobbers) */
+FI_ void tape_run_a02_s07(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
+	asm_words(
+		  load_word(  R_AtomJmp, R_TapePtr, 0) /* Bootstrap the first jump */
+		, add_ui_self(R_TapePtr, S_(MipsAtom)) /* Advance tape */
+		, call_reg(   R_AtomJmp)               /* jalr $t9 */
+		, nop                                  /* Branch delay slot */
+	)
+	asm_rpins, r_use(tape_ptr)
+	asm_clobber: 
+		rlit(R_AT),
+		rlit(R_V0), rlit(R_V1), rlit(R_A0), rlit(R_A1), rlit(R_A2),
+		rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+		rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8),
+		rlit(R_S0), rlit(R_S1), rlit(R_S2), rlit(R_S3), rlit(R_S4),
+		rlit(R_S5), rlit(R_S6), rlit(R_S7),
+		clb_mem_drain 
+); }
+
+// Procedural authoring of tapes:
 typedef Relative_(FArena) Struct_(TapeBuilder) { U4 ptr; U4 capacity; U4 used; };
 FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start; tb->used = 0; }
 FI_ TapeBuilder tb_make_old(             FArena* arena) { return (TapeBuilder){ arena->start, 0 }; }
-FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }; }
+FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ u4_(mem.ptr), mem.len, 0 }; }  /* capacity in elements (matches used units) */
 
-FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
+FI_ void tb_emit(TapeBuilder* tb, MipsAtom* 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_ void tb_emit_bundle(TapeBuilder_R tb, Slice_MipsAtom atoms) { mem_copy(u4_(tb->ptr), u4_(atoms.ptr), S_slice(atoms)); tb->used += atoms.len; }
+
 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
@@ -142,192 +242,146 @@ atom_dbg_skip MipsAtomComp_(ac_yield_tail) {
 	add_ui_self(R_TapePtr, S_(MipsCode)),
 	jump_reg(  R_AtomJmp), nop,
 };
-
-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
-};
-
-/* 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)),
-};
-
-/* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-atom_dbg_skip MipsAtomComp_(ac_gte_load_tri_verts) {
-	shift_lleft(R_AT, R_T0, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-	shift_lleft(R_AT, R_T1, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-	shift_lleft(R_AT, R_T2, v3s2_byteoff), add_u_self(R_AT, R_VertBase), load_word(R_V0, R_AT, O_(V3_S2,x)), load_word(R_V1, R_AT, O_(V3_S2,z)), gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
-};
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_F3 (5 words). For Poly_G4, use ac_insert_ot_tag_g4. */
-MipsAtomComp_(ac_insert_ot_tag_f3) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (5 - 1) << 24 = 4 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 11; Correctly inserts a primitive into the Ordering Table linked list.
- * Hardcoded for Poly_G4 (9 words). For Poly_F3, use ac_insert_ot_tag_f3. */
-MipsAtomComp_(ac_insert_ot_tag_g4) {
-	shift_lleft( R_T1, R_T1, S_(U4)/2),                        // T1 = otz * S_(U4) (otz arg is implicit R_T1)
-	add_u_self(  R_T1, R_OtBase),                              // T1 = & OrderingTable[OTZ]
-	load_word(   R_AT, R_T1,         O_(PolyTag,code)),        // AT = old_ot_head
-	load_upper_i(R_V0, (S_(Poly_G4)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits), // V0 = (9 - 1) << 24 = 8 << 24
-	mask_upper(  R_AT, R_AT,         S_(PolyTag_len_bits)),    // Strip upper 8 bits (length from prev cell) → keep only low 24
-	or_u(        R_AT, R_AT, R_V0),                            // Merge length
-	store_word(  R_AT, R_PrimCursor, O_(PolyTag,code)),        // prim->tag = packed(prim_length, old_addr)
-	shift_lleft( R_AT, R_PrimCursor, S_(PolyTag_len_bits)),    // AT = (prim_length << 24) | old_addr
-	shift_lright(R_AT, R_AT,         S_(PolyTag_len_bits)),
-	store_word(  R_AT, R_T1,         O_(PolyTag,code)),        // OrderingTable[OTZ] = PrimCursor
-};
-
-/* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
- * byte offset. Internal helper used by the *_format_*_color macros. */
-FI_ Slice_MipsCode ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
-atom_dbg_skip MipsAtomComp_Proc_(ac_pack_color_word, {
-	load_upper_i(R_AT, (cmd) << 8  | (b)),
-	or_i_self(   R_AT, ((g)  << 8) | (r)),
-	store_word(  R_AT, R_PrimCursor, (off)),
-})
-
-/* Words: 3; Emits the F3 command+color word (cmd byte | BLUE | GREEN | RED)
- * Args: _r, _g, _b are 8-bit RGB byte values (not raw 16-bit fields). */
-FI_ Slice_MipsCode ac_format_f3_color(U1 r, U1 g, U1 b)
-atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
- * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_f3) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_F3,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_F3,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_F3,p2)),
-};
-
-/* Words: 12; Emits the four (code|color) words of a Poly_G4.
- * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-FI_ Slice_MipsCode ac_format_g4_color(
-	U1 r0, U1 g0, U1 b0,
-  U1 r1, U1 g1, U1 b1,
-  U1 r2, U1 g2, U1 b2,
-  U1 r3, U1 g3, U1 b3)
-MipsAtomComp_Proc_(ac_format_g4_color, {
-	mac_pack_color_word(O_(Poly_G4,c0), gp0_cmd_poly_g4, r0,g0,b0),
-	mac_pack_color_word(O_(Poly_G4,c1), 0,               r1,g1,b1),
-	mac_pack_color_word(O_(Poly_G4,c2), 0,               r2,g2,b2),
-	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3),
-})
-
-/* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
- * G4 triangle portion to p0/p1/p2.
- * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2 get overwritten with v3
- * (RTPS writes only to SXY2, but to keep the three registers aligned with v0/v1/v2 you must store before RTPS). */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p012) {
-	gte_sw(C2_SXY0, R_PrimCursor, O_(Poly_G4,p0)),
-	gte_sw(C2_SXY1, R_PrimCursor, O_(Poly_G4,p1)),
-	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p2)),
-};
-
-/* Words: 1; Stores the V3 screen coord to the G4's p3 slot.
- * PIPELINE: post-RTPS (SXY2 holds v3.screen because RTPS writes its single-vertex result to SXY2;
- * SXY0 still holds v0.screen from the earlier RTPT.
- */
-atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
-
 #pragma endregion Macro Atom Components
 
-#pragma region Mips Atom Builder
+#pragma region 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; };
 
 // FArena Related
-typedef Relative_(FArena) Struct_(MipsAtomBuilder) { U4 start; U4 capacity; U4 used; };
+typedef Relative_(FArena) Struct_(AtomBuilder) { U4 start; U4 capacity; U4 used; };
-// Whatever the builder is writting to should most likely coresspond
-// to something that can fit within instruction cache?
 
-FI_ void atombuilder_unroll(MipsAtomBuilder_R ab, Slice_MipsCode_R code) {
+// Usual way to resolve an atom after the bulder is done.
-	assert(ab->capacity - ab->used - code->len);
+#define atom_from_atombuilder(ab) C_(MipsAtom*, (ab).start)
-	mem_copy(ab->start, u4_(code->ptr), code->len);
+
-	mem_bump(ab->start, ab->capacity, & ab->used, code->len);
+FI_ void atombuilder_push(AtomBuilder_R ab, Slice_MipsCode code) {
+	assert(ab->capacity - ab->used - code.len);
+	U4 dest = ab->start + ab->used * S_(MipsCode); U4 size = S_slice(code);
+	mem_copy(dest, u4_(code.ptr), size); ab->used += size;
 }
-#define atombuilder_unroll_mac(ab, mac) atombuilder_unroll(ab, slice_arg_from_array(Slice_MipsCode, mac))
+#define atombuilder_push_mac(ab, mac) atombuilder_push(ab, slice_arg_from_array(Slice_MipsCode, mac))
 
-// When done authoring, utilize this to cap-off the atom
+// When done authoring, utilize this to cap-off the atom (if not utilizing a MipsAtom_Proc).
-FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
+FI_ void atombuilder_end(AtomBuilder_R ab) { atombuilder_push(ab, slice_from_array(MipsCode, ac_yield)); }
-	mem_copy(ab->start, u4_(ac_yield), S_(ac_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
-}
-
-#define mipsatom_from_builder(ab) (Slice_MipsCode){ab.start, ab.used}
 
+FI_ void tb_emit_atombuilder(TapeBuilder_R tb, AtomBuilder_R ab) { tb_emit(tb, atom_from_atombuilder(ab[0])); }
 #pragma endregion Mips Atom Builder
 
+#pragma region Atom Arena
+// Just a dedicated FArena that is meant to mem_copy and return atom definitions made with MipsAtom_Proc_
+
+typedef Relative_(FArena) Struct_(AtomArena) { U4 start; U4 capacity; U4 used; };
+
+#define atomarena_unused_start(ab) ((ab).start + (ab).used)
+FI_ void atomarena_init(AtomArena_R arena, Slice mem) {  assert(arena != nullptr);
+	arena->start    = u4_(mem.ptr);
+	arena->capacity = mem.len;
+	arena->used     = 0;
+}
+FI_ AtomArena atomarena_make(Slice mem) { AtomArena a; atomarena_init(& a, mem); return a; }
+FI_ MipsAtom* atomarena_push(AtomArena_R aa, Slice_MipsCode code) {
+	assert(aa->capacity - aa->used - code.len);
+	U4 dest = atomarena_unused_start(aa[0]); U4 size = S_slice(code);
+	mem_copy(dest, u4_(code.ptr), size); aa->used += size;
+	return C_(MipsAtom*, dest);
+}
+FI_ void atomarena_reset(AtomArena_R aa) { aa->used = 0; }
+#pragma region Atom Arena
+
+#pragma region RegFile (Register File Allocator)
+// A specialized allocator utilized to help the user track which registers are bound to values
+// that must be preserved for the arena's bounds.
+// TODO(Ed): Technically we can do this at comp-time with the metaprogram, but we may have namespace conflicts.
+// Unless we follow a convention for #define <Scope_Prefix> or something per register allocation boundary.
+
+/* ABI + tape reserves that are never handed out by alloc. */
+U4 const regfile_abi_mask =
+	(1u << R_0)  | (1u << R_AT)  |
+	(1u << R_K0) | (1u << R_K1)  |
+	(1u << R_GP) | (1u << R_SP)  |
+	(1u << R_FP) | (1u << R_RA)  |
+	(1u << R_T8) | (1u << R_T9); /* AtomJmp + TapePtr */
+
+typedef Struct_(RegFile) {
+	A2_U2 GPR;
+	A2_U2 GTE;
+};
+#define regfile(pin_mask) {.GPR={u4_lo(pin_mask), u4_hi(pin_mask)} }
+FI_ void regfile_init(RegFile_R rf) {
+	/* pack the 32-bit ABI mask into the two U2s */
+	rf->GPR[0] = u4_lo(regfile_abi_mask);
+	rf->GPR[1] = u4_hi(regfile_abi_mask);
+	rf->GTE[0] = rf->GTE[1] = 0;
+}
+FI_ RegFile regfile_make(void) { RegFile rf; regfile_init(& rf); return rf; }
+
+typedef Struct_(RegFile_RInfo) {
+	U2_R section;
+	U2   mask;
+	B2   occupied;
+};
+FI_ RegFile_RInfo regfile_rinfo(A2_U2 file, Reg r_id) {
+	U2   s_id     = r_id >> 4;
+	U2_R section  = & file[s_id];
+	U2   mask     = u2_(1u << (r_id & 15));
+	B2   occupied = (section[0] & mask) != 0;
+	return (RegFile_RInfo){section, mask, occupied};
+}
+FI_ Reg regfile__alloc_helper(A2_U2 file, Reg r_id) {
+	Reg result = 0; RegFile_RInfo info = regfile_rinfo(file, r_id);
+	if (info.occupied == false) {
+		info.section[0] |= info.mask;
+		result = r_id;
+	}
+	return result; 
+}
+I_ Reg regfile_alloc(RegFile_R rf) {
+	U2 allocated = 0;
+	for index_iter(Reg, r_id, R_T0, <=, R_T7) {
+		allocated = regfile__alloc_helper(rf->GPR, r_id); Jmp_nZero_(allocated,resolved);
+	}
+	allocated = regfile__alloc_helper(rf->GPR, R_V0); Jmp_nZero_(allocated,resolved);
+	allocated = regfile__alloc_helper(rf->GPR, R_V1);
+	assert(allocated != 0);
+resolved: return allocated;
+}
+FI_ Reg regfile_pin(RegFile_R rf, Reg r_id) {
+	RegFile_RInfo info = regfile_rinfo(rf->GPR, r_id);
+	assert(info.occupied == false);
+	info.section[0] |= info.mask;
+	return r_id;
+}
+FI_ void regfile_pin_mask(RegFile_R rf, U4 mask) {
+	B4 occupied = u4_r(rf->GPR)[0] & mask;
+	assert(occupied == false);
+	u4_r(rf->GPR)[0] |= mask;
+}
+FI_ void regfile_free_mask(RegFile_R rf, U4 mask) {
+	if (regfile_abi_mask & mask) return;
+	u4_r(rf->GPR)[0] &= ~mask;
+}
+FI_ void regfile_free_reg(RegFile_R rf, Reg r_id) {
+	/* never free the ABI set */
+	if (regfile_abi_mask & (1u << r_id)) return;
+	RegFile_RInfo info = regfile_rinfo(rf->GPR, r_id);
+	info.section[0] &= ~info.mask;
+}
+FI_ void regfile_reset(RegFile_R rf) {
+	rf->GPR[0] = u4_lo(regfile_abi_mask);
+	rf->GPR[1] = u4_hi(regfile_abi_mask);
+}
+FI_ void regfile_reset_mask(RegFile_R rf, U4 mask) {
+	rf->GPR[0] = u4_lo(mask);
+	rf->GPR[1] = u4_hi(mask);
+}
+#pragma endregion RegFileArena (Register File Allocator)
+
+#pragma region Mips Atom Procs
+
+#pragma endregion Mips Atom Procs
+
 #pragma region Baked Mips Atoms
 // These atoms are resolved at compile time and are (usually) statically linked readonly data.
 
-enum {
-	bios_flushcache = 0x44,
-	bios_table_addr = 0xA0,
-};
-
-/* Flushes the Instruction Cache (PSX A-function 0x44 via BIOS stub at 0xA0).
- * Sequence (per MIPS ABI; arguments in arg registers, RA pushed to stack):
- *   1. sp -= 8;  sw $ra, 4($sp)        ; save RA
- *   2. $a0 = bios_flushcache (arg0)
- *   3. $t0 = bios_table_addr           ; t0 = &BIOS A-function table
- *   4. jalr $t0, $ra                   ; call BIOS(flushcache)
- *      nop                             ; branch delay slot
- *   5. lw $ra, 4($sp);  jr $ra         ; restore & return
- *   6. sp += 8
- */
-internal MipsAtom_(mips_flush_icache) {
-	add_ui(rstack_ptr, rstack_ptr, -MipsStackAlignment), // sp -= 8
-	store_word(rret_addr, rstack_ptr, S_(U4)),           // sw  $ra,   4($sp) 
-	add_ui(rret_0, rdiscard, bios_flushcache),           // addiu $a0, $0, 0x44 
-	add_ui(rtmp_0, rdiscard, bios_table_addr),           // addiu $t0, $0, 0xA0 
-	jump_link(rtmp_0, rret_addr), nop,                   // jalr  $t0, $ra, BD slot
-	load_word(rret_addr, rstack_ptr, S_(U4)),            // lw   $ra, 4($sp) 
-	jump_reg(rret_addr),                                 // jr   $ra 
-	add_ui(rstack_ptr, rstack_ptr, MipsStackAlignment),  // sp += 8 (BD) 
-	mac_yield(),
-};
-
-typedef Struct_(Binds_SetGteWorld) {
-	M3_S2* transform;
-};
-internal MipsAtom_(set_gte_world) atom_info(
-		atom_bind(Binds_SetGteWorld)
-	,	atom_reads(R_TapePtr)
-){
-	/* Pop matrix address from tape into R_T3 ($11) */
-	load_word(R_T3, R_TapePtr, O_(Binds_SetGteWorld,transform)),
-	add_ui_self(    R_TapePtr, S_(Binds_SetGteWorld)),
-	/* Load 3x3 Rotation + 3x1 Translation from R_T3 into GTE CONTROL Regs (ctc2) */
-	load_word(R_T0, R_T3, 0),  load_word(R_T1, R_T3,  4),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT11), gte_mv_to_ctrl_r(R_T1, gte_cr_RT12),
-	load_word(R_T0, R_T3, 8),  load_word(R_T1, R_T3, 12), load_word(R_T2, R_T3, 16),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_RT13), gte_mv_to_ctrl_r(R_T1, gte_cr_RT21), gte_mv_to_ctrl_r(R_T2, gte_cr_RT22),
-	load_word(R_T0, R_T3, 20), load_word(R_T1, R_T3, 24), load_word(R_T2, R_T3, 28),
-	gte_mv_to_ctrl_r(R_T0, gte_cr_TRX),  gte_mv_to_ctrl_r(R_T1, gte_cr_TRY),  gte_mv_to_ctrl_r(R_T2, gte_cr_TRZ),
-	mac_yield()
-};
-
 #pragma endregion Baked Mips Atoms

code/duffle/math.atom.c

@@ -0,0 +1,55 @@
+#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_imm
+
+FI_ Slice_MipsCode ac_load_v2s2(AtomBuilder_R ab, U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_half( rs_x, r_base, offset + O_(V3_S2,x)),
+	load_half( rs_y, r_base, offset + O_(V3_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_v2s2(AtomBuilder_R ab, U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_half(rt_x, base, offset + O_(V2_S2,x)),
+	store_half(rt_y, base, offset + O_(V2_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_load_v3s4(AtomBuilder_R ab, U4 rs_x, U4 rs_y, U4 rs_z, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_word( rs_x, r_base, offset + O_(V3_S4,x)),
+	load_word( rs_y, r_base, offset + O_(V3_S4,y)),
+	load_word( rs_z, r_base, offset + O_(V3_S4,z)),
+})
+// TODO(Ed): we could generate these mappings properly..
+#define ac_load_p3s4 ac_load_v3s4
+#define mac_load_p3s4 mac_load_v3s4
+
+FI_ Slice_MipsCode ac_store_v3s4(AtomBuilder_R ab, U4 rt_x, U4 rt_y, U4 rt_z, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	store_word(rt_x, base, offset + O_(V3_S4,x)),
+	store_word(rt_y, base, offset + O_(V3_S4,y)),
+	store_word(rt_z, base, offset + O_(V3_S4,z)),
+})
+// TODO(Ed): we could generate these mappings properly..
+#define ac_store_p3s4 ac_store_v3s4
+#define mac_store_p3s4 mac_store_v3s4
+
+FI_ Slice_MipsCode ac_sub_v3s4(AtomBuilder_R ab, U4 rds_x, U4 rds_y, U4 rds_z, U4 rt_x,  U4 rt_y,  U4 rt_z) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	sub_s(rds_x, rds_x, rt_x),
+	sub_s(rds_y, rds_y, rt_y),
+	sub_s(rds_z, rds_z, rt_z),
+})
+
+FI_ Slice_MipsCode ac_store_rects2(AtomBuilder_R ab, U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	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

@@ -7,11 +7,24 @@
 #define max(A, B)       (((A) > (B)) ? (A) : (B))
 #define clamp_bot(X, B) max(X, B)
 
+/* Convention
+<Type> ## <Width> _ <Component Type> ## <Component Width>
+For types with compound data (Ex: Rotation Matrix & Translation):
+<TypeA> ## <TypeB> ## <Width> _ <ComponentTypeA> ## <ComponentWidthA> ## <ComponentTypeB> ## <ComponentWidthB>
+
+A: Array
+V: Vector
+R: Range
+M: Matrix
+T: Translation
+*/
+
 enum { 
 	v3s2_byteoff = 3, // log2(8), used with shift_left_logical op for index via byte offset.
 };
 
 typedef Array_(U1, 2);
+typedef Array_(U2, 2);
 typedef Array_(U4, 2);
 typedef Array_(S2, 2);
 typedef Array_(S2, 3);
@@ -26,23 +39,43 @@ 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; };
+typedef Struct_(V3_S2)      { S2 x; S2 y; S2 z; S2 pad; }; // PSY-Q: SVECTOR
-typedef Struct_(V3_S4)      { S4 x; S4 y; S4 z; S4 pad; };
+typedef Struct_(V3_S4)      { S4 x; S4 y; S4 z; S4 pad; }; // PSY-Q: VECTOR. RGA(Lengyel): Euclidean vector or direction. A zero-weight RGA point is stored as a V3_S4 with the implicit weight dropped.
 typedef Struct_(V4_S2)      { S2 x; S2 y; S2 z; S2 w; };
 typedef Struct_(V4_S4)      { S4 x; S4 y; S4 z; S4 w; };
 
-typedef Struct_(R2_S2)      { V2_S2 p0; V2_S2 p1; };
+// typedef Struct_(P3_S4)      { S4 x; S4 y; S4 z; S4 w1; }; // RGA(Lengyel): Affine point with implicit weight one. Storage alias of V3_S4. Use P3_S4 when the value is a point.
-typedef Struct_(R2_S4)      { V2_S4 p0; V2_S4 p1; };
+typedef V3_S4 P3_S4;
 
-typedef Struct_(Rect_S2) { S2 x; S2 y; S2 width; S2 height; };
+typedef Struct_(R1_U2) { U2 p0; U2 p1; };
-typedef Struct_(Rect_S4) { S4 x; S4 y; S4 width; S4 height; };
+typedef Struct_(R1_S2) { S2 p0; S2 p1; };
 
-typedef Struct_(M3_S2) { A3x3_S2 m; A3_S4 t; };
+typedef Struct_(R2_S2)      { V2_S2 p0; V2_S2 p1; }; // Range-2 Signed 2-Byte (16-bit)
+typedef Struct_(R2_S4)      { V2_S4 p0; V2_S4 p1; }; // Range-2 Signed 4-Byte (32-bit)
 
+typedef Struct_(Rect_S2)    { S2 x; S2 y; S2 width; S2 height; };
+typedef Struct_(Rect_S4)    { S4 x; S4 y; S4 width; S4 height; };
+
+typedef Struct_(MT3_S2S4)   { A3x3_S2 m; A3_S4 t; }; // PSY-Q: MATRIX. RGA(Lengyel): Matrix expansion of a rigid transformation. GTE utilizes this representation; corresponding motor not constructed here.
+
+/* RGA(Lengyel) reserved names (deferred):
+ *   P4_S4  - future flat point with explicit weight (Lengyel/TML FlatPoint3D analog).
+ *   B3_S4  - future 3D bivector (callers store a Complement(Wedge(...)) as a V3_S4).
+ *   Mo8_S4 - future motor. Not introduced until a course operation actually needs composition, interpolation, or inversion. */
+
+typedef Array_(V2_U1, 2);
 typedef Array_(V2_S2, 2);
 typedef Array_(V2_S2, 3);
 typedef Array_(V2_S2, 4);
 
+#define r1u2(p0,p1) (R1_U2){p0,p1}
+
+enum {
+	fp_one = (1 << 12),
+};
+
+#define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)
+
 #define v2s2(x,y)     (V2_S2){x,y}
 #define v3s2(x,y,z)   (V3_S2){x,y,z,0}
 #define v3s4(x,y,z)   (V3_S4){x,y,z,0}
@@ -61,10 +94,28 @@ FI_ void add_a3s4_fp(A3_S4_R out_a, A3_S4 b) {
 	(out_a[0])[2] += b[2] >> 1;
 }
 
-FI_ void add_v3s4(V3_S4_R out_a, V3_S4 b) { 
+FI_ void sub_a3s4(A3_S4_R out_a, A3_S4 b) {
-	add_a3s4(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+	(out_a[0])[0] -= b[0];
+	(out_a[0])[1] -= b[1];
+	(out_a[0])[2] -= b[2];
 }
 
-FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) { 
+FI_ void sub_a3s4_fp(A3_S4_R out_a, A3_S4 b) {
-	add_a3s4_fp(pcast(A3_S4_R, out_a), pcast(A3_S4, b)); 
+	(out_a[0])[0] -= b[0] >> 1;
+	(out_a[0])[1] -= b[1] >> 1;
+	(out_a[0])[2] -= b[2] >> 1;
 }
+
+FI_ void mul_a3s4(A3_S4_R out_a, A3_S4 b) {
+	(out_a[0])[0] *= b[0];
+	(out_a[0])[1] *= b[1];
+	(out_a[0])[2] *= b[2];
+}
+
+FI_ void add_v3s4   (V3_S4_R out_a, V3_S4 b) {  add_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+FI_ void add_v3s4_fp(V3_S4_R out_a, V3_S4 b) {  add_a3s4_fp(C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+
+FI_ void sub_v3s4   (V3_S4_R out_a, V3_S4 b) {  sub_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+FI_ void sub_v3s4_fp(V3_S4_R out_a, V3_S4 b) {  sub_a3s4_fp(C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }
+
+FI_ void mul_v3s4   (V3_S4_R out_a, V3_S4 b) {  mul_a3s4   (C_ptr(A3_S4_R, out_a), C_ptr(A3_S4, b));  }

code/duffle/memory.h

@@ -18,7 +18,7 @@ I_ U4 align_pow2(U4 x, U4 b) {
 
 #define align_struct(type_width) ((U4)(((type_width) + 3) & ~3))
 
-FI_ void mem_bump(U4 start, U4 cap, U4*R_ used, U4 amount) {
+FI_ void mem_bump(U4 cap, U4*R_ used, U4 amount) {
 	assert(amount <= (cap - used[0])); 
 	used[0] += amount; 
 }
@@ -58,13 +58,13 @@ typedef Struct_(Str8)        { UTF8* ptr; U4 len; };
 typedef Struct_(Slice_Str8)  { Str8* ptr; U4 len; };
 #define slit(string_literal) (Str8){ (UTF8*) string_literal, S_(string_literal) - 1 }
 
-typedef Struct_(Slice) { U4 ptr, len; }; // Untyped Slice
+typedef Struct_(Slice) { B1* ptr; U4 len; }; // Untyped Slice (byte-addressable; .len in elements)
-FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){ptr, len}; }
+FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){(B1*)ptr, len}; }
 
 #define Slice_(type)       Struct_(tmpl(Slice,type)) { type* ptr; U4 len; }
 typedef Slice_(B1);
 #define slice_assert(s)    do { assert((s).ptr != 0); assert((s).len > 0); } while(0)
-#define slice_end(slice)   ((slice).ptr + (slice).len)
+#define slice_end(slice)   ((slice).ptr + S_slice(slice) / S_(B1))  /* byte-ptr arithmetic; .len is in elements per slice convention */
 #define S_slice(s)         ((s).len * S_((s).ptr[0]))
 
 #define slice_ut(ptr,len)  slice_ut_(u4_(ptr),     u4_(len))
@@ -72,23 +72,30 @@ typedef Slice_(B1);
 #define slice_to_ut(s)     slice_ut_(u4_((s).ptr), S_slice(s))
 
 #define slice_iter(container, iter)     (T_((container).ptr) iter = (container).ptr; iter != slice_end(container); ++ iter)
-#define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = array_decl(type,__VA_ARGS__), .len = array_len( array_decl(type,__VA_ARGS__)) }
+#define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = Array_decl(type,__VA_ARGS__), .len = Array_len( Array_decl(type,__VA_ARGS__)) }
-#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = S_(array) }
+#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = Array_len(array) }
 
-FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(s.ptr, s.len); }
+FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(u4_(s.ptr), s.len); }
 #define  slice_zero(s)        slice_zero_(slice_to_ut(s))
 
 FI_ void slice_copy_(Slice dest, Slice src) {
-	assert(dest.len >= src.len);
+	assert(S_slice(dest) >= S_slice(src));
 	slice_assert(dest);
 	slice_assert(src);
-	mem_copy(dest.ptr, src.ptr, src.len);
+	mem_copy(u4_(dest.ptr), u4_(src.ptr), S_slice(src));
 }
 #define slice_copy(dest, src) do {  \
 	static_assert(T_same(dest, src)); \
 	slice_copy_(slice_to_ut(dest), slice_to_ut(src)); \
 } while(0)
 
+FI_ Slice slice_bump(U4_R used, U4 start, U4 len, U4 amount) {
+	assert(len - used[0] - amount);
+	U4 ptr = start + used[0]; used[0] += amount;
+	return slice_ut(ptr, amount);
+}
+
+typedef Slice_(U1);
 typedef Slice_(U4);
 
 #pragma endregion Slice
@@ -98,18 +105,19 @@ typedef Slice_(U4);
 typedef Opt_(farena)    { U4 alignment, type_width; };
 typedef Struct_(FArena) { U4 start, capacity, used; };
 FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
-	arena->start    = mem.ptr;
+	arena->start    = u4_(mem.ptr);
 	arena->capacity = mem.len;
 	arena->used     = 0;
 }
 FI_ FArena farena_make(Slice mem) { FArena a; farena_init(& a, mem); return a; }
-I_  Slice  farena_push(FArena_R arena, U4 amount, Opt_farena o) {
+FI_ Slice farena_bump(FArena_R a, U4 amount) { return slice_bump(& a->used, a->start, a->capacity, amount); }
+I_  Slice farena_push(FArena_R arena, U4 amount, Opt_farena o) {
 	if (amount == 0) { return (Slice){}; }
 	U4 desired   = amount * (o.type_width == 0 ? 1 : o.type_width);
 	U4 to_commit = align_pow2(desired, o.alignment ?  o.alignment : MEM_ALIGNMENT_DEFAULT);
 	U4 ptr       = arena->start + arena->used;
-	mem_bump(arena->start, arena->capacity, & arena->used, to_commit);
+	mem_bump(arena->capacity, & arena->used, to_commit);
-	return (Slice){ ptr, to_commit };
+	return (Slice){ (B1*)ptr, to_commit };
 }
 FI_ void farena_reset (FArena_R arena) { arena->used = 0; }
 FI_ void farena_rewind(FArena_R arena, U4 save_point) {
@@ -117,6 +125,7 @@ FI_ void farena_rewind(FArena_R arena, U4 save_point) {
 	arena->used -= save_point - arena->start;
 }
 FI_ U4 farena_save(FArena arena) { return arena.used; }
+FI_ U4 farena_unused_start(FArena arena) { return arena.start + arena.used; }
 #define farena_push_(arena, amount, ...)                                          farena_push((arena), (amount), opt_(farena, __VA_ARGS__))
 #define farena_push_type(arena, type, ...)                              C_(type*, farena_push((arena), 1,        opt_(farena, .type_width=S_(type), __VA_ARGS__)).ptr)
 #define farena_push_array(arena, type, amount, ...) (tmpl(Slice,type)){ C_(type*, farena_push((arena), (amount), opt_(farena, .type_width=S_(type), __VA_ARGS__)).ptr), (amount) }

code/duffle/mips.atom.c

@@ -0,0 +1,45 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#		include "gen/macs.h"
+#		include "gen/offsets.h"
+#	include "bios.h"
+#	include "mips.h"
+#	include "lottes_tape.h"
+#endif
+
+ATOM_FILE_DEBUGGER_LINE_MARKER(mips_atom_c);
+
+#pragma region MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_load_word_imm(AtomBuilder_R ab, Reg dst, U4 imm)
+atom_dbg_skip MipsAtomComp_Proc_(ab, {
+	load_upper_i(dst, u4_hi(imm)),
+	or_i_self(   dst, u4_lo(imm)),
+})
+
+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Baked Atoms
+
+/* 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(R_SP, R_SP, -MipsStackAlignment), // sp -= 8
+	store_word(R_RA, R_SP, S_(U4)),           // sw  $ra,   4($sp) 
+	add_ui(R_V0, R_0, bios_flushcache),           // addiu $a0, $0, 0x44 
+	add_ui(R_T0, R_0, bios_table_addr),           // addiu $t0, $0, 0xA0 
+	jump_link(R_T0, R_RA), nop,                   // jalr  $t0, $ra, BD slot
+	load_word(R_RA, R_SP, S_(U4)),            // lw   $ra, 4($sp) 
+	jump_reg(R_RA),                                 // jr   $ra 
+	add_ui(R_SP, R_SP, MipsStackAlignment),  // sp += 8 (BD) 
+	mac_yield(),
+};
+
+#pragma endregion Baked Atoms

code/duffle/mips.h

@@ -136,31 +136,31 @@ enum {
 
 /* Semantic Aliases for MIPS Registers (O32 ABI) */
 
-	, rdiscard      = R_0  /* Hardwired to 0 */
+	// , rdiscard      = R_0  /* Hardwired to 0 */
-	, rasm_tmp      = R_AT /* Assembler temporary (destroyed by some assembler pseudoinstructions!) */
+	// , rasm_tmp      = R_AT /* Assembler temporary (destroyed by some assembler pseudoinstructions!) */
-	, rret_0        = R_V0 /* Function return value */
+	// , rret_0        = R_V0 /* Function return value */
-	, rret_1        = R_V1 /* Second return value (e.g., 64-bit) */
+	// , rret_1        = R_V1 /* Second return value (e.g., 64-bit) */
-	, rarg_0        = R_A0 /* First function argument */
+	// , rarg_0        = R_A0 /* First function argument */
-	, rarg_1        = R_A1 /* Second function argument */
+	// , rarg_1        = R_A1 /* Second function argument */
-	, rarg_2        = R_A2 /* Third function argument */
+	// , rarg_2        = R_A2 /* Third function argument */
-	, rarg_3        = R_A3 /* Fourth function argument */
+	// , rarg_3        = R_A3 /* Fourth function argument */
-	, rtmp_0        = R_T0 /* Temporary (Caller saved) */
+	// , rtmp_0        = R_T0 /* Temporary (Caller saved) */
-	, rtmp_1        = R_T1 /* Temporary (Caller saved) */
+	// , rtmp_1        = R_T1 /* Temporary (Caller saved) */
-	, rtmp_2        = R_T2 /* Temporary (Caller saved) */
+	// , rtmp_2        = R_T2 /* Temporary (Caller saved) */
-	, rtmp_3        = R_T3 /* Temporary (Caller saved) */
+	// , rtmp_3        = R_T3 /* Temporary (Caller saved) */
-	, rtmp_4        = R_T4 /* Temporary (Caller saved) — common GTE base pointer */
+	// , rtmp_4        = R_T4 /* Temporary (Caller saved) — common GTE base pointer */
-	, rtmp_9        = R_T9 /* Temporary (Caller saved) — common GTE base pointer */
+	// , rtmp_9        = R_T9 /* Temporary (Caller saved) — common GTE base pointer */
-	, rstatic_0     = R_S0 /* Static (Callee saved, preserved across calls) */
+	// , rstatic_0     = R_S0 /* Static (Callee saved, preserved across calls) */
-	, rstatic_1     = R_S1
+	// , rstatic_1     = R_S1
-	, rstatic_2     = R_S2
+	// , rstatic_2     = R_S2
-	, rstatic_3     = R_S3
+	// , rstatic_3     = R_S3
-	, rstatic_4     = R_S4
+	// , rstatic_4     = R_S4
-	, rstatic_5     = R_S5
+	// , rstatic_5     = R_S5
-	, rstatic_6     = R_S6
+	// , rstatic_6     = R_S6
-	, rstatic_7     = R_S7
+	// , rstatic_7     = R_S7
-	, rsaved_0      = R_S0 /* Alias for rstatic_0 (alternate vocabulary) */
+	// , rsaved_0      = R_S0 /* Alias for rstatic_0 (alternate vocabulary) */
-	, rstack_ptr    = R_SP /* Stack Pointer */
+	// , rstack_ptr    = R_SP /* Stack Pointer */
-	, rret_addr     = R_RA /* Return Address (populated by JAL) */
+	// , rret_addr     = R_RA /* Return Address (populated by JAL) */
 
 /* --- MIPS CPU Opcodes (Bits 31-26) --- */
 
@@ -259,22 +259,22 @@ enum { _BitOffsets = 0
 	, SHAMT_SHIFT  = 6  /* Shift Amount */
 	, FC_SHIFT     = 0
 
-/* Bit Masks to prevent overflow into adjacent fields */
+/* IMM_MASK is the 16-bit two's-complement truncation for the immediate field.
+ * It is NOT a range guard — it is load-bearing for negative branch offsets
+ * (the metaprogram emits raw signed offsets; the mask truncates them to the
+ * 16-bit representation the hardware expects). The static analysis
+ * `immediate_field_width` check validates ranges at build time. */
 
-	, OPCODE_MASK  = 0x3F
-	, REG_MASK     = 0x1F
-	, SHAMT_MASK   = 0x1F /* Shift Amount */
-	, FC_MASK      = 0x3F
 	, IMM_MASK     = 0xFFFF
 };
 
-#define enc_op(op)       (((op)    & OPCODE_MASK) << OPCODE_SHIFT)
+#define enc_op(op)       ((op)    << OPCODE_SHIFT)
-#define enc_rs(rs)       (((rs)    & REG_MASK)    << RS_SHIFT)
+#define enc_rs(rs)       ((rs)    << RS_SHIFT)
-#define enc_rt(rt)       (((rt)    & REG_MASK)    << RT_SHIFT)
+#define enc_rt(rt)       ((rt)    << RT_SHIFT)
-#define enc_rd(rd)       (((rd)    & REG_MASK)    << RD_SHIFT)
+#define enc_rd(rd)       ((rd)    << RD_SHIFT)
-#define enc_shamt(shamt) (((shamt) & SHAMT_MASK)  << SHAMT_SHIFT)
+#define enc_shamt(shamt) ((shamt) << SHAMT_SHIFT)
-#define enc_fc(fc)       (((fc)    & FC_MASK)     << FC_SHIFT)
+#define enc_fc(fc)       ((fc)    << FC_SHIFT)
-#define enc_imm(imm)     (((imm)   & IMM_MASK))
+#define enc_imm(imm)     ((imm) & IMM_MASK)
 
 /* MIPS R-Type Instruction Format (Register-to-Register) */
 #define enc_r(op, rs, rt, rd, shamt, fc) (enc_op(op) | enc_rs(rs) | enc_rt(rt) | enc_rd(rd) | enc_shamt(shamt) | enc_fc(fc))
@@ -318,7 +318,10 @@ enum { _BitOffsets = 0
 #define load_half(rt, base, off)   enc_i(op_lh,    (base), (rt), (off))
 #define load_byte_u(rt, base, off) enc_i(op_lbu,   (base), (rt), (off))
 #define load_half_u(rt, base, off) enc_i(op_lhu,   (base), (rt), (off))
+#define LdSlot_
+
 #define store_word(rt, base, off)  enc_i(op_sw,    (base), (rt), (off))
+
 #define add_ui(rt, rs, imm)        enc_i(op_addiu, (rs),   (rt), (imm))
 #define and_i(rt, rs, imm)         enc_i(op_andi,  (rs),   (rt), (imm))
 // #define and_si                     and_i
@@ -348,6 +351,12 @@ enum { _BitOffsets = 0
 #define shift_lright(rd, rt, shamt) enc_r(op_special, R_0,  (rt), (rd), (shamt), fc_srl)
 #define shift_aright(rd, rt, shamt) enc_r(op_special, R_0,  (rt), (rd), (shamt), fc_sra)
 
+/* Shift Variable — register-shift forms.
+ * shift_lleft_var(rd, rt, rs)  → sllv rd, rt, rs   (shamt in low 5 bits of rs)
+ * shift_aright_var(rd, rt, rs) → srav rd, rt, rs */
+#define shift_lleft_var(rd, rt, rs)  enc_r(op_special, (rs), (rt), (rd), 0, fc_sllv)
+#define shift_aright_var(rd, rt, rs) enc_r(op_special, (rs), (rt), (rd), 0, fc_srav)
+
 #define shift_lleft_self(rd_rt, shamt) enc_r(op_special, R_0, (rd_rt), (rd_rt), (shamt), fc_sll)
 
 #define mask_upper(rd, rt, shamt) shift_lleft(rd, rt, shamt), shift_lright(rd, rt, shamt)
@@ -366,20 +375,21 @@ enum { _BitOffsets = 0
  * 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.
+ * TODO(Ed): Review this.. technically we can resolve aboslute jumps on baked atoms? (Even proedurally generated ones...)
  */
 #define jump(off)                  enc_i(op_j,     R_0, R_0, (off))
 
+// Annotate an instruction as filling a branch-delay slot.
+#define BdSlot_
+
 /* 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`.
+ * 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.
@@ -397,13 +407,7 @@ enum { _BitOffsets = 0
  * sub_s / sub_u  → sub / subu
  * mult_s / mult_u → mult / multu   (writes HI/LO; result in LO)
  * div_s / div_u   → div / divu     (LO = quot, HI = rem)
- *
+ */
- * NOTE: dsl.h defines `add_s`/`sub_s`/`mut_s`/`gt_s`/etc. as _Generic-based signed integer-arithmetic helpers for U1/U2/U4.
- * Those live in a different conceptual layer (generic arithmetic on DSL types) and would collide with the instruction encoders here.
- * The `#undef` below lets the gas-style names below win; if a file needs both, the dsl.h versions can be reached via their long forms
- * (e.g. `def_signed_op`-style or the underlying `add_s1/s2/s4`). */
-#undef add_s
-#undef sub_s
 #define add_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_add)
 #define add_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_addu)
 #define sub_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_sub)
@@ -413,6 +417,7 @@ enum { _BitOffsets = 0
 #define div_s(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_div)
 #define div_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_divu)
 
+// TODO(Ed): Change convention of 'self' to ds for (destination is source)?
 #define add_u_self(rd_rs, rt)      add_u(rd_rs, rd_rs, rt)
 
 /* --- Arithmetic I-type (immediate) --- */
@@ -455,9 +460,13 @@ enum { _BitOffsets = 0
 #define shift_amount(rd, rt, n)     shift_lleft(rd, rt, n)
 
 /* nop — sll $0, $0, 0 */
-#define nop  shift_lleft(rdiscard, rdiscard, 0)
+#define nop  shift_lleft(R_0, R_0, 0)
 #define nop2 nop, nop
 
+// li_s — load signed 16-bit immediate into GPR (addiu rt, $0, imm — sign-extends).
+#define li_s(rt, imm)       add_ui((rt), R_0, (imm))
+// #define load_imm_s(rt, imm) add_ui((rt), R_0, (imm))
+
 #define load_imm_1w(rt, imm)    add_ui((rt),  R_0, (imm))
 #define load_imm_1w_s0(rt, imm) add_si((rt)), R_0, (imm))
 

code/duffle/pad.atom.c

@@ -0,0 +1,197 @@
+#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 MACs (Mips Atom Components)
+
+FI_ Slice_MipsCode ac_pad_set_centered_axes(AtomBuilder_R ab, Reg state, Reg scratch) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+		load_upper_i(scratch, (PadAxis_Centered >> 16) & 0xFFFF),
+    or_i_self(   scratch,  PadAxis_Centered        & 0xFFFF),
+		// mac_load_word_imm(scratch, PadAxis_Centered),
+    store_word(       scratch, state, O_(PadState,axes)),
+})
+
+FI_ Slice_MipsCode ac_pad_set_id_byte(AtomBuilder_R ab, Reg state, Reg r_id, U1 id_value) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    add_ui(    r_id, R_0, id_value),
+    store_byte(r_id, state, O_(PadState,id)),
+})
+
+FI_ Slice_MipsCode ac_pad_set_status(AtomBuilder_R ab, U4 r_tmp, U1 r_state, U4 pad_status) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    add_ui(    r_tmp, R_0, pad_status),
+    store_word(r_tmp, r_state, O_(PadState,status)),
+})
+
+/* Invert r_buttons (active-low → active-high) and store to PadState.buttons.
+ * r_buttons must already be loaded (the caller is responsible for filling the load-delay slot of
+ * the preceding load_half_u with an instruction that doesn't read r_buttons). */
+FI_ Slice_MipsCode ac_pad_store_inverted_buttons(AtomBuilder_R ab, U1 r_buttons, U1 r_pad_state) atom_dbg_skip MipsAtomComp_Proc_(ab, {
+    nor_u(       r_buttons, r_buttons, R_0),
+    store_half(  r_buttons, r_pad_state, O_(PadState,buttons)),
+})
+
+#pragma endregion MACs (Mips Atom Components)
+
+#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 atom_type(PadState*),
+	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)
+, atom_writes(R_PadRaw, R_PadState, R_RawStatus, R_RawId)
+) {
+	/* === 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, O_(PadBiosRaw,status)),
+	load_byte_u(R_RawId,     R_PadRaw, O_(PadBiosRaw,id)),
+
+atom_label(snap_root) /* === Case 1: Disconnected (status == 0xFF). */
+	add_ui(R_T4, R_0, PadRawStatus_Timeout), 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. */
+	mac_pad_set_status(R_T4,  R_PadState, PadStatus_Disconnected),
+	store_half(        R_0,   R_PadState, O_(PadState,buttons)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	mac_pad_set_id_byte(R_PadState, R_RawId, PadRawStatus_Timeout),
+	jump_rel(atom_offset(disconnected, snap_end)),
+		/* BD-slot: load next atom's entry point (replaces the nop).
+		 * 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 (status=0, id=0 — pre-IRQ-empty buffer). */
+	mac_pad_set_status(R_T4,  R_PadState, PadStatus_Pending),
+	store_half(        R_0,   R_PadState, O_(PadState,buttons)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	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, PadRawId_Digital), 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_T5 holds the 0x80808080 axes constant (loaded into the load-delay slot of the buttons-load).
+	 * R_T5 is then "dead" — only consumed at the analog_pad range check downstream. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_Digital),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw, buttons)), /* R_T4 = raw_buttons; */
+	mac_load_word_imm(R_T5, PadAxis_Centered), /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	// load_upper_i(R_T5, PadAxis_Centered_Hi), or_i_self(R_T5, PadAxis_Centered_Lo), 
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),            /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	store_word(R_T5,    R_PadState, O_(PadState, axes)),         /* single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y) */
+	mac_pad_set_id_byte(R_PadState, R_T4, PadRawId_Digital),
+
+	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, PadRawId_AnalogStick), 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
+	* R_T5 holds left_xy (loaded into the load-delay slot of the buttons-load via the left-axis load_half_u).
+	* R_T4 holds right_xy (loaded into the load-delay slot of the left-load).
+	* R_T5 is then "dead" — reused for the id-byte value load in mac_pad_write_id_byte.
+	* The buttons invert+store happens BEFORE R_T4 is overwritten by the right_xy load. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_AnalogStick),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw,buttons)),  /* R_T4 = raw_buttons; delay slot at the next instruction */
+	load_half_u(       R_T5, R_PadRaw, O_(PadBiosRaw,left)),     /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),            /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	load_half_u( R_T4,  R_PadRaw,   O_(PadBiosRaw,right)),       /* fills R_T5's load-delay slot (doesn't read R_T5); overwrites R_T4 (was buttons) with right_xy */
+	store_half(  R_T5,  R_PadState, O_(PadState,  left)),
+	store_half(  R_T4,  R_PadState, O_(PadState,  right)),
+	mac_pad_set_id_byte(R_PadState, R_T5, PadRawId_AnalogStick),
+	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, PadRawId_AnalogPadMask),
+	add_ui(   R_T5, R_0,     PadRawId_AnalogPadValue),
+	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).
+	* The id byte is raw id from the BIOS buffer (R_RawId already holds raw[1]).
+	* Buttons invert + store happens before R_T4 is overwritten by the right_xy load. */
+	mac_pad_set_status(R_T4, R_PadState, PadStatus_AnalogPad),
+	load_half_u(       R_T4, R_PadRaw, O_(PadBiosRaw,buttons)), /* R_T4 = raw_buttons; delay slot at the next instruction */
+	load_half_u(       R_T5, R_PadRaw, O_(PadBiosRaw,left)),    /* fills the buttons-load's delay slot (doesn't read R_T4) */
+	mac_pad_store_inverted_buttons(R_T4, R_PadState),           /* R_T4 settled: nor + sh writes ~raw_buttons to state.buttons */
+	load_half_u(R_T4,    R_PadRaw,   O_(PadBiosRaw,right)),     /* fills R_T5's load-delay slot (doesn't read R_T5); overwrites R_T4 with right_xy */
+	store_half( R_T5,    R_PadState, O_(PadState,  left)),
+	store_half( R_T4,    R_PadState, O_(PadState,  right)),
+	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)),
+	mac_pad_set_centered_axes(R_PadState, R_T4),
+	mac_pad_set_id_byte(R_PadState, R_RawId, PadUnknownId_Sentinel),
+	/* 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.c

@@ -0,0 +1,78 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "dsl.h"
+#	include "gcc_asm.h"
+#	include "mips.h"
+#	include "bios.h"
+#	include "pad.h"
+#endif
+
+/* Uses ONE 8-byte frame allocated via the compiler's standard prologue.
+ * 4 wasted-arg words for B(12h) InitPAD2 are at [SP+0..15] but are not explicitly allocated.
+ * 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 + 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_mem_drain; 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(    R_A2, R_A1, R_0),                 /* $a2 = $a1 = raw1 */
+			add_ui(  R_A1, R_0, bios_pad_buffer_size), /* $a1 = 0x22 */
+			add_ui(  R_A3, R_0, bios_pad_buffer_size), /* $a3 = 0x22 */
+			add_ui(  R_T1, R_0, bios_init_pad_2),      /* $t1 = 0x12 */
+			add_ui(  R_T2, R_0, bios_btable_addr),     /* $t2 = 0xB0 */
+			call_reg(R_T2),                            /* 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(  R_T1, R_0, bios_start_pad_2), /* $t1 = 0x13 */
+			add_ui(  R_T2, R_0, bios_btable_addr), /* $t2 = 0xB0 (re-load) */
+			call_reg(R_T2),                             /* 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
+	);
+}

code/duffle/pad.h

@@ -1,28 +1,30 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
 #	include "dsl.h"
+#	include "math.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. */
+ * The decoder atom computes buttons = (~raw_buttons) & 0xFFFF; 
-enum {
+ * active-low-to-active-high inversion is applied bit-by-bit. */
-	Bit_(Pad_Select,  0),
+typedef Enum_(U2, PadBtns) {
-	Bit_(Pad_L3,      1),
+	Bit_(Pad_Select,    0),
-	Bit_(Pad_R3,      2),
+	Bit_(Pad_L3,        1),
-	Bit_(Pad_Start,   3),
+	Bit_(Pad_R3,        2),
-	Bit_(Pad_Up,      4),
+	Bit_(Pad_Start,     3),
-	Bit_(Pad_Right,   5),
+	Bit_(Pad_Up,        4),
-	Bit_(Pad_Down,    6),
+	Bit_(Pad_Right,     5),
-	Bit_(Pad_Left,    7),
+	Bit_(Pad_Down,      6),
-	Bit_(Pad_L2,      8),
+	Bit_(Pad_Left,      7),
-	Bit_(Pad_R2,      9),
+	Bit_(Pad_L2,        8),
-	Bit_(Pad_L1,     10),
+	Bit_(Pad_R2,        9),
-	Bit_(Pad_R1,     11),
+	Bit_(Pad_L1,       10),
+	Bit_(Pad_R1,       11),
 	Bit_(Pad_Triangle, 12),
-	Bit_(Pad_Circle,  13),
+	Bit_(Pad_Circle,   13),
-	Bit_(Pad_Cross,   14),
+	Bit_(Pad_Cross,    14),
-	Bit_(Pad_Square,  15),
+	Bit_(Pad_Square,   15),
 };
 
 enum {
@@ -32,18 +34,22 @@ enum {
 	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,
 };
+// BIOS pad buffer layout (docs/psx-spx/docs/kernelbios.md (InitPAD2 returns 0x22 = 34 bytes per port)).
+// Bytes 0..7 are the named snapshot region; bytes 8..33 are reserved (the BIOS writes the buffer raw; we only read bytes 0..7 via O_(PadBiosRaw, ...)).
 typedef Struct_(PadBiosRaw) {
-	U1 bytes[PAD_BIOS_RAW_SIZE];
+	U1    status;  /* offset 0   (PadRawStatus_Ok / PadRawStatus_Timeout) */
+	U1    id;      /* offset 1   (PadRawId_Digital / PadRawId_AnalogStick / 0x7x AnalogPad) */
+	U2    buttons; /* offset 2-3 (active-low 16-bit button map) */
+	V2_U1 right;   /* offset 4-5 (right stick x, y) */
+	V2_U1 left;    /* offset 6-7 (left  stick x, y) */
+	U1    reserved[PAD_BIOS_RAW_SIZE - 8]; /* offset 8..33 */
 };
 
 typedef Enum_(U4, PadStatus) {
@@ -56,18 +62,54 @@ typedef Enum_(U4, PadStatus) {
 	PadStatus_Invalid,
 };
 
-/* PadState — per-port normalized runtime state.
+/* Distinct from the game-facing PadStatus enum: PadRawStatus_Ok and PadRawStatus_Timeout are raw BIOS values;
- * Field order is chosen so that the 4 axes (left_x, left_y, right_x, right_y)
+ * PadStatus_* are game-facing post-decode states. PadUnknownId_Sentinel is written by the decoder 
- * 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.
+ * when the controller id does not match any known controller type.
- * The struct size stays 12 bytes (unchanged from the prior order,
+ * PadAxisCentered_Word: Four-byte 0x80 pattern used to clear / center 
- * which left the C compiler to insert 1 byte of trailing pad to reach the 4-byte struct alignment). */
+ * four byte axes at PadState.left_x through PadState.right_y. */
-typedef Struct_(PadState) {
+typedef Enum_(U1, PadRawStatus) {
-	PadStatus status;  /* offset 0,  size 4 (U4) */
+	PadRawStatus_Ok      = 0x00,
-	U2        buttons; /* offset 4,  size 2 */
+	PadRawStatus_Timeout = 0xFF,
-	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 */
 };
+typedef Enum_(U1, PadRawId) {
+	PadRawId_Digital           = 0x41,
+	PadRawId_AnalogStick       = 0x53,
+	PadRawId_AnalogPadMask     = 0xF0,
+	PadRawId_AnalogPadValue    = 0x70,
+};
+typedef Enum_(U1, PadUnknownId) {
+	PadUnknownId_Sentinel      = 0xFF,
+};
+typedef Enum_(U4, PadAxisCentered) {
+	PadAxis_Centered_Hi = 0x8080,
+	PadAxis_Centered_Lo = 0x8080,
+	PadAxis_Centered    = 0x80808080U,
+};
+typedef Enum_(U1, PadDeadZone) {
+	PadDeadZone_LowBound  = 0x70,  /* left_x <  LowBound  → active;  delta = 0x80 - left_x > 0 (rightward pull) */
+	PadDeadZone_Center    = 0x80,  /* analog rest position; left_x == Center → delta = 0 (no rotation) */
+	PadDeadZone_HighBound = 0x90,  /* left_x >  HighBound → active;  delta = 0x80 - left_x < 0 (leftward pull) */
+};
+
+
+typedef Struct_(PadAxes) {
+	V2_U1  left;  /* offset 8-9 */
+	V2_U1  right; /* offset 10-11 */
+};
+// 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.
+typedef Struct_(PadState) {
+	PadStatus status;   /* offset 0,    (U4) */
+	PadBtns   buttons;  /* offset 4,    */
+	U1        id;       /* offset 6,    */
+	byte_pad(1);        /* offset 7,  explicit pad to align the axes block */
+	union {
+		A2_V2_U1 axes;    /* offset 8-11 store_target (4-byte aligned)*/
+		struct {
+			V2_U1  left;    /* offset 8-9 */
+			V2_U1  right;   /* offset 10-11 */
+		};
+	};
+};
+
+internal void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1);

code/duffle/psyq.atom.c

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

code/duffle/psyq.h

@@ -1,8 +1,8 @@
 #ifdef INTELLISENSE_DIRECTIVES
 #	pragma once
-#	include "duffle/dsl.h"
+#	include "dsl.h"
-#	include "duffle/math.h"
+#	include "math.h"
-#	include "duffle/gp.h"
+#	include "gp.h"
 #endif
 
 typedef Struct_(DrawEnv_Packed) { U4 tag; U4 code[15]; };
@@ -64,9 +64,9 @@ typedef Struct_(Tile) {
 	Linear Algebra
 */
 
-M3_S2* m3s2_rotation   (V3_S2* vec, M3_S2* mat) asm("RotMatrix");
+MT3_S2S4* mt3s2s4_rotation   (V3_S2*    vec, MT3_S2S4* mat) asm("RotMatrix");
-M3_S2* m3s2_translation(M3_S2* mat, V3_S4* vec) asm("TransMatrix");
+MT3_S2S4* mt3s2s4_translation(MT3_S2S4* mat, V3_S4*    vec) asm("TransMatrix");
-M3_S2* m3s2_scale      (M3_S2* mat, V3_S4* vec) asm("ScaleMatrix");
+MT3_S2S4* mt3s2s4_scale      (MT3_S2S4* mat, V3_S4*    vec) asm("ScaleMatrix");
 
 // Rotation, Translation, Perspective
 
@@ -99,5 +99,23 @@ FI_ S4 rtp_avg_nclip_a4_v3s2(
 	);
 }
 
-void gte_matrix_set_rotation   (M3_S2* mat) asm("SetRotMatrix");
+void gte_matrix_set_rotation   (MT3_S2S4* mat) asm("SetRotMatrix");
-void gte_matrix_set_translation(M3_S2* mat) asm("SetTransMatrix");
+void gte_matrix_set_translation(MT3_S2S4* mat) asm("SetTransMatrix");
+
+// Einheit, Metrication to unit vector. "Normalization", not Orthogonal "Normal, Normalis". Directionalization.
+// RGA(Lengyel): Normalize the bulk of a zero-weight direction. This is not finite-point unitization (which forces w=1).
+S4 normalize_v3s4(V3_S4* v0, V3_S4* v1) asm("VectorNormal");
+
+// RGA(Lengyel): Apply the matrix expansion of a rigid transformation.
+// Motor antiproduct is equivalent for unitized points; LA form is what GTE consumes.
+V3_S4* mul_m3s2_v3s4(MT3_S2S4* m, V3_S4* v, V3_S4* result) asm("ApplyMatrixLV");
+
+// RGA(Lengyel): Store the full translation column. The motor translator would store half this displacement in m.xyz.
+MT3_S2S4* trans_m3s2(MT3_S2S4* m, V3_S4* off) asm("TransMatrix");
+
+MT3_S2S4* gte_comp_coord_m3s2(MT3_S2S4* m0, MT3_S2S4* m1, MT3_S2S4* result) asm("CompMatrixLV");
+
+// RGA(Lengyel): Complement(Wedge(a,b)), i.e. the Euclidean 3D complement of the exterior product, stored as a V3_S4.
+// The underlying GTE OP is a specialized signed-16-bit D x IR command; the wedge interpretation is a 3D dual of the same 3 scalars.
+void cross_v3s4(V3_S4* v0, V3_S4* v1, V3_S4* result) asm("OuterProduct12");
+

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.
 //
@@ -15,6 +15,8 @@
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
 
 WORD_COUNT(nop,                              1)
+WORD_COUNT(atom_label,                        0)
+WORD_COUNT(atom_offset,                       0)
 WORD_COUNT(load_upper_i,                     1)
 WORD_COUNT(jump_reg,                         1)
 WORD_COUNT(jump_link,                        1)
@@ -54,6 +56,15 @@ WORD_COUNT(gte_sw,                           1)
 WORD_COUNT(gte_cmdw_rtpt,                    1)
 WORD_COUNT(gte_cmdw_nclip,                   1)
 WORD_COUNT(gte_avg_sort_z3,                  1)
+WORD_COUNT(gte_cmdw_sqr,                     1)
+WORD_COUNT(gte_cmdw_gpf,                     1)
+WORD_COUNT(shift_lleft_var,                  1)
+WORD_COUNT(shift_aright_var,                 1)
+WORD_COUNT(li_s,                             1)
+WORD_COUNT(and_i,                            1)
+WORD_COUNT(add_si,                           1)
+WORD_COUNT(branch_lt_zero,                   1)
+WORD_COUNT(sub_s,                            1)
 WORD_COUNT(sub_u,                            1)
 WORD_COUNT(nop2,                             2)
 

code/hello_camera/gen/auto_reg.h

@@ -0,0 +1,13 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua (passes/auto_reg.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
+// Per-phase register allocations resolved by the lua pass.
+// R_<Sym>_Code = <chosen GPR's _Code constant> for every marker in this directory.
+
+#define R_GpTmp_Code R_V0_Code
+

code/hello_camera/gen/macs.h

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

code/hello_camera/gen/offsets.h

@@ -0,0 +1,68 @@
+// 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_input_cube_rotation (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: pad_input_cam (40 words) ---
+
+#define _atom_offset_left_x_exit_left_x               3
+#define _atom_offset_right_x_exit_right_x             3
+#define _atom_offset_up_y_exit_up_y                   3
+#define _atom_offset_down_y_exit_down_y               3
+#define _atom_offset_cross_z_exit_cross_z             3
+#define _atom_offset_circle_z_exit_circle_z           3
+
+enum {
+    atom_offset_left_x_exit_left_x = _atom_offset_left_x_exit_left_x,
+    atom_offset_right_x_exit_right_x = _atom_offset_right_x_exit_right_x,
+    atom_offset_up_y_exit_up_y = _atom_offset_up_y_exit_up_y,
+    atom_offset_down_y_exit_down_y = _atom_offset_down_y_exit_down_y,
+    atom_offset_cross_z_exit_cross_z = _atom_offset_cross_z_exit_cross_z,
+    atom_offset_circle_z_exit_circle_z = _atom_offset_circle_z_exit_circle_z,
+};
+
+// --- 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 +1,961 @@
+#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 "gen/auto_reg.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(AtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	// 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),
+})
+
+I_ Slice_MipsCode ac_put_draw_env(AtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ab, {
+	/*
+	* 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 Atom Procs
+// Modular Atoms
+
+enum {
+	// TODO(Ed): We can resolve scratch at anytime its fixed to a specific address.
+	R_ResolveScratch = R_T4 atom_reg atom_type(U4*),
+#define R_ResolveScratch_Code R_T4_Code
+};
+typedef Struct_(Binds_ResolveLookAt) {
+	MT3_S2S4* look_at;
+	P3_S4*    eye;
+	P3_S4*    target;
+	V3_S4*    up_in;
+};
+
+/* ─── ResolveLookAtScratch — offset schema for the resolve_look_at bundle's */
+typedef Struct_(ResolveLookAtScratch) {
+	V3_S4 fwd;       /* offset  +0  (16 bytes — 4 S4 fields incl. internal pad) */
+	V3_S4 uz;        /* offset +16 (16 bytes) */
+	V3_S4 right;     /* offset +32 (16 bytes) */
+	V3_S4 ux;        /* offset +48 (16 bytes) */
+	V3_S4 up;        /* offset +64 (16 bytes) */
+	V3_S4 uy;        /* offset +80 (16 bytes) */
+	P3_S4 eye;       /* offset +96 (16 bytes; storage alias of V3_S4) */
+	P3_S4 target;    /* offset +112 (16 bytes; storage alias of V3_S4) */
+	V3_S4 up_in;     /* offset +128 (16 bytes) */
+};
+
+/* ─── resolve_look_at bundle chain atoms ──────────────────────────── */
+
+typedef Struct_(Binds_ResolveLookAtSub) {
+	P3_S4* target;  /* U4 (C-side P3_S4* — read by atom 0 directly; NOT a scratchpad address) */
+	P3_S4* eye;     /* U4 (C-side P3_S4* — read by atom 0 directly; staged into scratchpad by atom 0) */
+	V3_S4* up_in;   /* U4 (C-side V3_S4* — read by atom 0 directly; staged into scratchpad by atom 0) */
+	ResolveLookAtScratch* scratchpad;
+};
+
+/* Atom 0 in the bundle: input_and_sub. Stages C-side inputs into the scratchpad and computes fwd = target - eye.
+ * Staging work:
+ *   * Stage eye.x/y/z   → scratch (for atom 6's translation column)
+ *   * Stage up_in.x/y/z → scratch (for atom 2's outer-product operand)
+ *   * Compute fwd = target - eye, store fwd.x/y/z → scratch+0/+4/+8 (for atom 1)
+ * GPR codes (assigned by resolve_look_at_init):
+ *   r_target_ptr : R_T0
+ *   r_eye_ptr    : R_T1
+ *   r_up_in_ptr  : R_T2
+ *   r_scratch    : R_T4 (R_ResolveScratch; wave-context carrier)
+ *   r_tmp0       : R_T3 (stage eye/up_in + load eye.y)
+ *   r_tmp1       : R_T5 (stage eye/up_in + load eye.z)
+ *   r_tmp2       : R_T6 (stage eye/up_in + load target.x)
+ *   r_tmp3       : R_T7 (stage eye/up_in + load target.y)
+ *   R_AT         : hardcoded (load eye.y / eye.z / target.z)
+ *   R_V0         : hardcoded (load eye.z / target.z)
+ * Pool cost: 8 GPRs + R_T4 (carrier) + R_AT + R_V0 (hardcoded) = 11 GPRs.
+ */
+internal MipsAtom* resolve_look_at__input_and_sub_proc(AtomArena_R aa,	
+			// TODO(Ed): We can resolve scratch at anytime its fixed to a specific address.
+			U4 r_scratch
+	,	U4 r_target_ptr,U4 r_eye_ptr, U4 r_up_in_ptr
+	,	U4 r_tmp0,      U4 r_tmp1,    U4 r_tmp2, U4 r_tmp3
+) MipsAtom_Proc_(aa, {
+	load_word(r_target_ptr, R_TapePtr, O_(Binds_ResolveLookAtSub,target)),
+	load_word(r_eye_ptr,    R_TapePtr, O_(Binds_ResolveLookAtSub,eye)),
+	load_word(r_up_in_ptr,  R_TapePtr, O_(Binds_ResolveLookAtSub,up_in)),
+	load_word(r_scratch,    R_TapePtr, O_(Binds_ResolveLookAtSub,scratchpad)),
+	add_ui_self(            R_TapePtr, S_(Binds_ResolveLookAtSub)),
+
+	// Stage eye.x/y/z into the scratchpad (atom 6 reads these for the translation column).
+	mac_load_p3s4( r_tmp0, r_tmp1, r_tmp2, r_eye_ptr, 0),
+	mac_store_p3s4(r_tmp0, r_tmp1, r_tmp2, r_scratch, O_(ResolveLookAtScratch,eye)),
+
+	/* Stage up_in.x/y/z into the scratchpad. */
+	mac_load_p3s4( r_tmp0, r_tmp1, r_tmp2, r_up_in_ptr, 0),
+	mac_store_p3s4(r_tmp0, r_tmp1, r_tmp2, r_scratch, O_(ResolveLookAtScratch,up_in)),
+
+	/* Compute fwd = target - eye. */
+	mac_load_p3s4(r_tmp0, r_tmp1, r_tmp2, r_target_ptr, 0),
+	mac_load_p3s4(r_tmp3, R_AT, R_V0, r_eye_ptr, 0),
+	mac_sub_v3s4(
+		r_tmp0, r_tmp1, r_tmp2,
+		r_tmp3, R_AT,   R_V0),
+	mac_store_v3s4(r_tmp0, r_tmp1, r_tmp2, r_scratch, O_(ResolveLookAtScratch,fwd)),
+
+	mac_yield()
+})
+
+/* Atom 2: cross uz × up_in → right. */
+internal MipsAtom* resolve_look_at__cross_uz_up_in_to_right_proc(AtomArena_R aa, U4 r_scratch
+	,	U4 r_a, U4 r_b, U4 r_c /* load a.x/y/z; result out.x/y/z */
+	,	U4 r_d                 /* load b.x */
+	,	U4 r_f, U4 r_g, U4 r_h /* r_f = &right (out ptr), r_g = &uz, r_h = &up_in */
+) MipsAtom_Proc_(aa, {
+	/* FIX: build packed RT22+RT33 with proper sign extension. */
+	add_si(r_g, r_scratch, O_(ResolveLookAtScratch,uz)),    /* r_g = &uz */
+	add_si(r_h, r_scratch, O_(ResolveLookAtScratch,up_in)), /* r_h = &up_in */
+	add_si(r_f, r_scratch, O_(ResolveLookAtScratch,right)), /* r_f = &right (out) */
+	nop,
+
+	/* Load a (uz).x/y/z into r_a/r_b/r_c. */
+	load_word(r_a, r_g, O_(V3_S4,x)),
+	load_word(r_b, r_g, O_(V3_S4,y)),
+	load_word(r_c, r_g, O_(V3_S4,z)),
+	nop,
+
+	/* Load b (up_in).x/y/z into r_d + R_AT/R_V0 (R_AT/R_V0 are hardcoded scratch). */
+	load_word(r_d,  r_h, O_(V3_S4,x)),
+	load_word(R_AT, r_h, O_(V3_S4,y)),
+	load_word(R_V0, r_h, O_(V3_S4,z)),
+	nop,
+
+	/* Save the two RT control-register slots OP will clobber. We reuse
+	 * r_g/r_h (scratch pointers, no longer needed) as the save targets. */
+	gte_mv_from_ctrl_r(r_g, gte_cr_RT11),    /* r_g = C2 r0 (RT11|RT12) */
+	gte_mv_from_ctrl_r(r_h, gte_cr_RT22),    /* r_h = C2 r4 (RT22|RT33) */
+
+	/* Load uz.x/uz.y/uz.z into COP2 control registers.
+	 * OP reads D1 = RT11 from $0.low, D2 = RT22 from $2.high, D3 = RT33 from $4.high.
+	 * RT22 is in BOTH $2.high AND $4.low (shared bit position). OP reads from $2.high.
+	 * So set RT22 via ctc2 r_b, $2 (sets $2.high = a.y.high = RT22, $2.low = a.y.low = RT13).
+	 * Then set RT33 via ctc2 r_c, $4 (sets $4.high = a.z.high = RT33, $4.low = a.z.low).
+	 * The $2 and $4 writes don't clobber each other (separate registers).
+	 * The 2nd ctc2 DOES clobber $4.low (becomes a.z.low, NOT a.y.high), but since OP
+	 * reads RT22 from $2.high (which the 2nd ctc2 doesn't touch), D2 is still a.y.high.
+	 * This is libpsyx's OuterProduct12 convention EXACTLY. */
+	gte_mv_to_ctrl_r(r_b, gte_cr_RT13),    /* $2 = r_b = a.y. RT13=a.y.low, RT22=a.y.high. */
+	gte_mv_to_ctrl_r(r_c, gte_cr_RT22),    /* $4 = r_c = a.z. RT22=a.z.low, RT33=a.z.high. */
+
+	/* Load uz into the RT diagonal. */
+	gte_mv_to_ctrl_r(r_a, gte_cr_RT11),   /* D1 = RT11 = uz.x (low 16 of $0, sign-extended by OP). */
+	nop2,                                 /* CTC2 retirement (CPU→COP2 2-slot delay) */
+
+	/* Load up_in into IR (the second operand for OP). */
+	gte_mv_to_data_r(r_d,  C2_IR1),       /* IR1 = up_in.x */
+	gte_mv_to_data_r(R_AT, C2_IR2),       /* IR2 = up_in.y */
+	gte_mv_to_data_r(R_V0, C2_IR3),       /* IR3 = up_in.z */
+	nop2,                                 /* MTC2 retirement (CPU→COP2 2-slot delay) */
+
+	gte_cmdw_outer_product, /* OP: MAC1/2/3 = uz × up_in
+		*   MAC1 = IR3*D2 - IR2*D3 = up_in.z*uz.y.high - up_in.y*uz.z.high
+		*   MAC2 = IR1*D3 - IR3*D1 = up_in.x*uz.z.high - up_in.z*uz.x
+		*   MAC3 = IR2*D1 - IR1*D2 = up_in.y*uz.x - up_in.x*uz.y.high
+		* For up_in = (0, -fp_one, 0):
+		*   MAC1 = 0 - (-fp_one)*uz.z.high = fp_one*uz.z.high
+		*   MAC2 = 0 - 0 = 0
+		*   MAC3 = (-fp_one)*uz.x - 0 = -fp_one*uz.x */
+
+	/* Restore the RT slots we clobbered. */
+	gte_mv_to_ctrl_r(r_g, gte_cr_RT11),   /* restore C2 r0 (RT11|RT12) */
+	gte_mv_to_ctrl_r(r_h, gte_cr_RT22),   /* restore C2 r4 (RT22|RT33) */
+
+	/* mfc2 MAC1/2/3 → r_a/r_b/r_c (out.x/y/z). */
+	gte_mv_from_data_r(r_a, C2_MAC1),
+	gte_mv_from_data_r(r_b, C2_MAC2),
+	gte_mv_from_data_r(r_c, C2_MAC3),
+	nop,  /* MFC2 retirement */
+
+	/* Right-shift MAC by 12 to convert from GTE's S12.20 fixed-point scale back to libpsyx OuterProduct12 convention (S12.0, fp_one=4096=1<<12).
+	 * Without this, MAC values (~16M for unit-vector cross products) overflow the GTE's 16-bit IR registers when atom 3 normalizes via mtc2. */
+	shift_aright(r_a, r_a, 12),
+	shift_aright(r_b, r_b, 12),
+	shift_aright(r_c, r_c, 12),
+
+	/* Store out.x/y/z to r_f (out ptr = scratch+32). */
+	store_word(r_a, r_f, O_(V3_S4,x)),
+	store_word(r_b, r_f, O_(V3_S4,y)),
+	store_word(r_c, r_f, O_(V3_S4,z)),
+
+	mac_yield()
+})
+
+/* Atom 4: cross uz × ux → up. */
+internal MipsAtom* resolve_look_at__cross_uz_ux_to_up_proc(AtomArena_R aa,	U4 r_scratch
+	,	U4 r_a, U4 r_b, U4 r_c /* load a.x/y/z; result out.x/y/z */
+	,	U4 r_d                 /* load b.x */
+	,	U4 r_f, U4 r_g, U4 r_h /* r_f = &up (out ptr), r_g = &uz, r_h = &ux */
+) MipsAtom_Proc_(aa, {
+	/* Compute the three scratch pointers from r_scratch. */
+	add_si(r_g, r_scratch, O_(ResolveLookAtScratch,uz)), /* r_g = &uz */
+	add_si(r_h, r_scratch, O_(ResolveLookAtScratch,ux)), /* r_h = &ux */
+	add_si(r_f, r_scratch, O_(ResolveLookAtScratch,up)), /* r_f = &up (out) */
+	nop,
+
+	/* Load a (uz).x/y/z into r_a/r_b/r_c. */
+	load_word(r_a, r_g, O_(V3_S4,x)),
+	load_word(r_b, r_g, O_(V3_S4,y)),
+	load_word(r_c, r_g, O_(V3_S4,z)),
+	nop,
+
+	/* Load b (ux).x/y/z into r_d + R_AT/R_V0. */
+	load_word(r_d,  r_h, O_(V3_S4,x)),
+	load_word(R_AT, r_h, O_(V3_S4,y)),
+	load_word(R_V0, r_h, O_(V3_S4,z)),
+	nop,
+
+	/* OP reads D1/D2/D3 from RT11/RT22/RT33 ($0/$2/$4), not V0/V1/V2.
+	 * Mirror atom 1: cfc2 RT save, ctc2 RT diagonal from uz, mtc2 IR from ux,
+	 * ctc2 RT restore. */
+
+	/* Save the two RT control-register slots OP will clobber (reusing
+	 * r_g/r_h — they're no longer needed as scratch pointers). */
+	gte_mv_from_ctrl_r(r_g, gte_cr_RT11),    /* r_g = C2 $0 (RT11|RT12) */
+	gte_mv_from_ctrl_r(r_h, gte_cr_RT22),    /* r_h = C2 $4 (RT22|RT33) */
+
+	/* Load uz into the RT diagonal — same packing as atom 1.
+	 * OP reads D1 = RT11 from $0.low, D2 = RT22 from $2.high, D3 = RT33 from $4.high.
+	 * RT22 is shared between $2.high and $4.low — the ctc2 sequence to $2 then $4
+	 * sets RT22 to uz.y.high (via $2), then to uz.z.low (via $4). OP reads
+	 * RT22 from $2.high which the second ctc2 doesn't touch, so D2 stays uz.y.high.
+	 * (This is libpsyx OuterProduct12 convention EXACTLY.) */
+	gte_mv_to_ctrl_r(r_b, gte_cr_RT13),    /* $2 = uz.y. RT13=uz.y.low, RT22=uz.y.high. */
+	gte_mv_to_ctrl_r(r_c, gte_cr_RT22),    /* $4 = uz.z. RT22=uz.z.low, RT33=uz.z.high. */
+	gte_mv_to_ctrl_r(r_a, gte_cr_RT11),   /* $0 = uz.x. RT11=uz.x. */
+	nop2,                                 /* CTC2 retirement (CPU→COP2 2-slot delay) */
+
+	/* Load ux into the IR registers (the second operand for OP). */
+	gte_mv_to_data_r(r_d,  C2_IR1),       /* IR1 = ux.x */
+	gte_mv_to_data_r(R_AT, C2_IR2),       /* IR2 = ux.y */
+	gte_mv_to_data_r(R_V0, C2_IR3),       /* IR3 = ux.z */
+	nop2,                                 /* MTC2 retirement (CPU→COP2 2-slot delay) */
+
+	gte_cmdw_outer_product,
+
+	/* Restore the RT slots we clobbered. */
+	gte_mv_to_ctrl_r(r_g, gte_cr_RT11),   /* restore C2 $0 (RT11|RT12) */
+	gte_mv_to_ctrl_r(r_h, gte_cr_RT22),   /* restore C2 $4 (RT22|RT33) */
+
+	gte_mv_from_data_r(r_a, C2_MAC1),
+	gte_mv_from_data_r(r_b, C2_MAC2),
+	gte_mv_from_data_r(r_c, C2_MAC3),
+	nop,
+	/* Right-shift MAC by 12 to convert from GTE's S12.20 scale back to libpsyx
+	 * OuterProduct12 convention (S12.0, fp_one=4096). See atom 1 for rationale. */
+	shift_aright(r_a, r_a, 12),
+	shift_aright(r_b, r_b, 12),
+	shift_aright(r_c, r_c, 12),
+	store_word(r_a, r_f, O_(V3_S4,x)),
+	store_word(r_b, r_f, O_(V3_S4,y)),
+	store_word(r_c, r_f, O_(V3_S4,z)),
+
+	mac_yield()
+})
+
+typedef Struct_(Binds_ResolveLookAtPopAndTrans) {
+	U4 look_at;  /* U4 (MT3_S2S4* — destination matrix address) */
+};
+/* Atom 6 in the bundle: write look_at->m[][] from ux/uy/uz, then compute the translation column t[] = R * (-eye).
+ *
+ * GPR codes (assigned by resolve_look_at_init):
+ *   r_look_at   : MT3_S2S4* (popped from tape; output matrix destination)
+ *   r_pux       : pointer to ux   (offset O_(ResolveLookAtScratch,ux))
+ *   r_puy       : pointer to uy   (offset O_(ResolveLookAtScratch,uy))
+ *   r_puz       : pointer to uz   (offset O_(ResolveLookAtScratch,uz))
+ *   r_peye      : pointer to eye  (offset O_(ResolveLookAtScratch,eye))
+ *   r_tmp0/1/2  : atom-local scratch (load + MVMVA + store temps)
+ *
+ * 4 pointer regs (r_pux/r_puy/r_puz/r_peye) are DEDICATED — they hold the scratch addresses for the entire body.
+ * They are computed in-body via `add_si(r_px, r_scratch, O_(ResolveLookAtScratch, field))` so no tape-data pointer is needed.
+ *
+ * Struct layout (per duffle/math.h):
+ *   MT3_S2S4 { A3x3_S2 m; A3_S4 t; }  →  m[][] is S2 packed (9 × 2 = 18 bytes at offset 0) 
+ *                                        t[0/1/2] is S4     (3 × 4 = 12 bytes at offset 18)
+ *
+ * Translation column: GTE MVMVA with the world rotation matrix pre-set
+ * (helper emits set_gte_world before the bundle, per the bundle design).
+ * MVMVA computes R * pos (with cv=0/mx=0/sf=0/v=0); MAC1/2/3 = R * (-eye).
+ * Pool cost: r_look_at (1) + r_scratch (R_T4 carrier) + 4 ptr regs + 3 tmp regs = 9 GPRs.
+ */
+internal MipsAtom* resolve_look_at__populate_proc(AtomArena_R aa
+	,	U4 r_look_at
+	,	U4 r_scratch
+	,	U4 r_pux, U4 r_puy, U4 r_puz
+	,	U4 r_tmp0, U4 r_tmp1, U4 r_tmp2
+) MipsAtom_Proc_(aa, {
+	/* Pop look_at* (the matrix output) — advance R_TapePtr by 4 bytes. */
+	load_word(r_look_at, R_TapePtr, O_(Binds_ResolveLookAtPopAndTrans,look_at)),
+	add_ui_self(         R_TapePtr, S_(Binds_ResolveLookAtPopAndTrans)),
+
+	/* Compute the 3 scratch pointers in their dedicated GPRs (eye isn't needed by 6a — 6b reads it). */
+	add_si(r_pux, r_scratch, O_(ResolveLookAtScratch,ux)),  /* r_pux = &ux */
+	add_si(r_puy, r_scratch, O_(ResolveLookAtScratch,uy)),  /* r_puy = &uy */
+	add_si(r_puz, r_scratch, O_(ResolveLookAtScratch,uz)),  /* r_puz = &uz */
+	nop,
+
+	/* ── m[0] = (S2)ux ── */
+	load_word(r_tmp0, r_pux, O_(V3_S4,x)),
+	load_word(r_tmp1, r_pux, O_(V3_S4,y)),
+	load_word(r_tmp2, r_pux, O_(V3_S4,z)),
+	nop,
+	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[0][0])),
+	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[0][1])),
+	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[0][2])),
+
+	/* ── m[1] = (S2)uy ── */
+	load_word(r_tmp0, r_puy, O_(V3_S4,x)),
+	load_word(r_tmp1, r_puy, O_(V3_S4,y)),
+	load_word(r_tmp2, r_puy, O_(V3_S4,z)),
+	nop,
+	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[1][0])),
+	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[1][1])),
+	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[1][2])),
+
+	/* ── m[2] = (S2)uz ── */
+	load_word(r_tmp0, r_puz, O_(V3_S4,x)),
+	load_word(r_tmp1, r_puz, O_(V3_S4,y)),
+	load_word(r_tmp2, r_puz, O_(V3_S4,z)),
+	nop,
+	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[2][0])),
+	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[2][1])),
+	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[2][2])),
+
+	/* Zero t[0..2] — atom 6c writes the final values here. */
+	store_word(R_0, r_look_at, O_(MT3_S2S4,t[0])),
+	store_word(R_0, r_look_at, O_(MT3_S2S4,t[1])),
+	store_word(R_0, r_look_at, O_(MT3_S2S4,t[2])),
+
+	mac_yield()
+})
+
+/* Atom 6b in the bundle: matrix-vector product off = R * (-eye) >> 12.
+ * Uses RTPS with V0 loaded from scratch via lwc2. The RT matrix is
+ * pre-loaded by atom 6a.5 (resolve_look_at__load_rt).
+ * Stores off to scratch+96 (overwriting the packed pos).
+ *
+ * GPR codes (assigned by resolve_look_at_init):
+ *   r_scratch : R_ResolveScratch (R_T4) — scratch base
+ *   r_peye    : pointer to eye (slot +96, reused as off destination)
+ *   r_tmp0/1/2: -eye + GTE transfer scratch
+ *
+ * Pool cost: r_scratch (carrier) + 1 ptr reg + 3 tmp regs = 5 GPRs.
+ */
+internal MipsAtom* resolve_look_at__matrix_vector_proc(AtomArena_R aa
+	,	U4 r_scratch
+	,	U4 r_peye
+	,	U4 r_look_at
+	,	U4 r_tmp0, U4 r_tmp1, U4 r_tmp2
+) MipsAtom_Proc_(aa, {
+	/* === EXACT C11 ApplyMatrixLV replication ===
+	 * The C11 does:
+	 *   1. ctc2 RT matrix (5 ctc2s to C2[0..4])
+	 *   2. lw v.x/y/z from memory
+	 *   3. S15 decomposition (negu + sra 15 + negu + andi 0x7FFF + negu)
+	 *   4. mtc2 HIGH bits to IR1/2/3, nop, MVMVA pass1 (sf=0, mx=0, v=3, cv=3)
+	 *   5. mfc2 MACs
+	 *   6. mtc2 LOW bits to IR1/2/3, nop, MVMVA pass2 (sf=1, mx=0, v=3, cv=3)
+	 *   7. mfc2 MACs
+	 *   8. Combine: (pass1 << 3) + pass2
+	 *
+	 * For S16-fitting pos (|pos| < 32768), pos >> 15 = 0, so pass1 = 0.
+	 * The combine simplifies: result = 0 + pass2 = pass2.
+	 * So we skip the S15 decomposition and just do pass 2 directly.
+	 * We still use v=3 (IR input) and mx=0 (RT matrix) like the C11. */
+
+	/* Pop look_at* from tape. */
+	load_word(r_look_at, R_TapePtr, O_(Binds_ResolveLookAtPopAndTrans,look_at)),
+	add_ui_self(         R_TapePtr, S_(Binds_ResolveLookAtPopAndTrans)),
+
+	/* r_peye = &eye (slot +96, reused as off destination). */
+	add_si(r_peye, r_scratch, O_(ResolveLookAtScratch,eye)),
+	nop,
+
+	/* === Load RT matrix from look_at into C2[0..4] via ctc2 ===
+	 * Exact s  ame sequence as set_gte_mt3s2s4 / C11's ApplyMatrixLV. */
+	load_word(  r_tmp0, r_look_at,  0), nop, gte_mv_to_ctrl_r(r_tmp0, gte_cr_RT11),
+	load_word(  r_tmp0, r_look_at,  4), nop, gte_mv_to_ctrl_r(r_tmp0, gte_cr_RT12),
+	load_word(  r_tmp0, r_look_at,  8), nop, gte_mv_to_ctrl_r(r_tmp0, gte_cr_RT13),
+	load_word(  r_tmp0, r_look_at, 12), nop, gte_mv_to_ctrl_r(r_tmp0, gte_cr_RT21),
+	load_half_u(r_tmp0, r_look_at, 16), nop, gte_mv_to_ctrl_r(r_tmp0, gte_cr_RT22),
+	nop2,  /* CTC2 retirement (2 slots × 5 ctc2s) */
+
+	/* Load pos = -eye after the matrix load releases r_tmp0. */
+	load_word(r_tmp0, r_peye, O_(P3_S4,x)),
+	load_word(r_tmp1, r_peye, O_(P3_S4,y)),
+	load_word(r_tmp2, r_peye, O_(P3_S4,z)),
+	nop,
+	sub_u(r_tmp0, R_0, r_tmp0),  /* pos.x = -eye.x */
+	sub_u(r_tmp1, R_0, r_tmp1),
+	sub_u(r_tmp2, R_0, r_tmp2),
+
+	/* === mtc2 pos (as S16) to IR1/2/3 ===
+	 * The GTE takes low 16 bits. pos fits in S16. For negative pos, the
+	 * 32-bit sign-extended value's low 16 bits = correct S16. */
+	/* Mask pos to 16 bits to be safe. For S16-fitting pos, pos & 0xFFFF
+	 * gives the correct S16 value (sign bit preserved). */
+	/* r_tmp0/1/2 already have pos values. */
+	gte_mv_to_data_r(r_tmp0, C2_IR1),
+	gte_mv_to_data_r(r_tmp1, C2_IR2),
+	gte_mv_to_data_r(r_tmp2, C2_IR3),
+	nop2,  /* MTC2 retirement (2 slots) */
+
+	/* === MVMVA pass 2 — C11 ApplyMatrixLV command ===
+	 * sf=1, mx=0 (RT), v=3 (IR), cv=3. Reads RT × IR >> 12. */
+	gte_cmdw_mvmva_c11_pass2,
+	nop,  /* GTE interlock */
+
+	/* === mfc2 MAC1/2/3 → r_tmp0/1/2 === */
+	gte_mv_from_data_r(r_tmp0, C2_MAC1),
+	gte_mv_from_data_r(r_tmp1, C2_MAC2),
+	gte_mv_from_data_r(r_tmp2, C2_MAC3),
+	nop,
+
+	/* === Store off → scratch+96 (overwriting pos) === */
+	store_word(r_tmp0, r_peye, O_(V3_S4,x)),
+	store_word(r_tmp1, r_peye, O_(V3_S4,y)),
+	store_word(r_tmp2, r_peye, O_(V3_S4,z)),
+
+	mac_yield()
+})
+
+/* Atom 6c in the bundle: copy scratch+96 (off, written by atom 6b) → look_at->t[].
+ * Uses mac_trans_matrix component (m->t = v, libgte TransMatrix semantics = struct copy).
+ *
+ * GPR codes (assigned by resolve_look_at_init):
+ *   r_look_at : MT3_S2S4* (popped from tape; output matrix destination)
+ *   r_scratch : R_ResolveScratch (R_T4) — scratch base
+ *   r_off_ptr : pointer to off (= &scratch.eye, reused slot)
+ *   r_tmp0    : transfer reg for mac_trans_matrix
+ *
+ * Pool cost: r_look_at (1) + r_scratch (carrier) + r_off_ptr + 1 clobber = 4 GPRs.
+ */
+I_ MipsAtom* resolve_look_at__trans_matrix_proc(AtomArena_R aa
+	,	U4 r_look_at,	U4 r_scratch,	U4 r_off_ptr
+	,	U4 r_tmp0, U4 r_tmp1, U4 r_tmp2
+) MipsAtom_Proc_(aa, {
+	/* Pop look_at* from tape. */
+	// load_word(r_Vlook_at, R_TapePtr, O_(Binds_ResolveLookAtPopAndTrans,look_at)),
+	// add_ui_self(         R_TapePtr, S_(Binds_ResolveLookAtPopAndTrans)),
+
+	/* r_off_ptr = &off (= &scratch.eye since atom 6b overwrote eye with off). */
+	add_si(r_off_ptr, r_scratch, O_(ResolveLookAtScratch,eye)),
+	nop,
+
+	/* Copy off → look_at.t[] (mac_trans_matrix: m->t = v). */
+	mac_trans_mt3s3s4(r_look_at, r_off_ptr, r_tmp0, r_tmp1, r_tmp2),
+
+	mac_yield()
+})
+
+#pragma endregion Atom Procs
+
+#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) + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,display_area) + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)       + O_(DoubleBuffer,display[0])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace)   + O_(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) + O_(DoubleBuffer,display[1])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)     + O_(DoubleBuffer,display[1])),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace) + O_(DoubleBuffer,display[1])),
+
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area)         + O_(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]) + O_(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) + O_(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)     + O_(DoubleBuffer,draw[0])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + O_(DoubleBuffer,draw[0])),
+
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,drawing_offset[0].x)  + O_(DoubleBuffer,draw[1])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + O_(DoubleBuffer,draw[1])),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + O_(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) + O_(DoubleBuffer,draw[0])),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + O_(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)          + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + O_(DoubleBuffer,draw[0])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + O_(DoubleBuffer,draw[1])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + O_(DoubleBuffer,draw[1])),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + O_(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) + O_(DoubleBuffer,draw[0])),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + O_(DoubleBuffer,draw[1])),
+
+	mac_yield(),
+};
+
+/* gp_screen_init's GPR setup. Tests the mixed user-pinning + auto-reg pattern:
+ *   - R_IO_BaseAddr = R_T4 (user-pinned via atom_reg; pre-existing)
+ *   - R_GP1_Offset  = R_T2 (user-pinned via atom_reg; NEW -- for GPIO_PORT1_OFFSET)
+ *   - R_ScreenX     = R_T5 (user-pinned via atom_reg; used as a transfer and GTE setup reg)
+ *   - R_GpTmp       = auto-allocated by the lua pass and used for several GPU transfers;
+ *                    the C preprocessor resolves it to the chosen free pool GPR.
+ *
+ * For gp_screen_init, the auto-reg pool exclusions are:
+ *   user_pinned (from the corpus register_alias_registry) : R_T0..R_T7 (all 8 user-pinned across hello_camera.atom.c)
+ *   body-parsed physical registers                         : aliases resolve through the registry;
+ *                                                           the body uses R_ScreenX, not raw R_T5
+ *   source_pool after both subtractions                   : {R_V0, R_V1} only
+ *   R_GpTmp gets R_V0 (the first-fit choice). Its repeated GPU-transfer use proves that the
+ *   auto-reg allocation is active while the R_ScreenX references prove the pinned alias is used.
+ *   R_TapePtr (R_T9), R_AtomJmp (R_T8), R_AT are excluded from the POOL by construction in
+ *   passes/auto_reg.lua -- see the "obvious exclusions" comment block at the top of that file.
+ */
+enum {
+	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+	R_GP1_Offset  = R_T2 atom_reg, /* Caller-pinned: GPIO_PORT1_OFFSET = 0x10 */
+	atom_auto_reg(gp_screen_init, R_GpTmp),  /* Auto-allocated scratch; resolved to a free pool GPR by the lua pass. C-preprocessor expands to R_GpTmp = R_GpTmp_Code with an atom_auto_reg trailing comment. */
+#define R_IO_BaseAddr_Code R_T4_Code
+#define R_GP1_Offset_Code  R_T2_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_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(01h) ClearFIFO; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(02h) AckIRQ; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(03h) Display ON; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_GpTmp,   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. Uses auto-allocated R_GpTmp. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_GpTmp,   R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(05h) StartDisplayArea (X=0, Y=0); uses auto-allocated R_GpTmp. */
+
+	/* GP1: DisplayMode + Display Ranges. */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+
+	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
+	load_upper_i(R_ScreenX, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFX_Code),
+	load_upper_i(R_ScreenX, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFY_Code),
+
+	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
+	add_ui(R_ScreenX, R_0, ScreenZ), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_H_Code),
+
+	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
+	mac_gcmd_push(gp1_word_DisplayOn(), R_GpTmp, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* Uses auto-allocated R_GpTmp. */
+	mac_yield(),
+};
+
+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_input_cube_rotation) 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, 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, 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, PadDeadZone_HighBound), 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, PadDeadZone_Center), /* 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, PadDeadZone_HighBound),
+
+	/* 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, PadDeadZone_Center), /* 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_Cam         = R_T4 atom_reg,
+	R_CamPadState = R_T5 atom_reg,
+};
+typedef Struct_(Binds_PadInputCam) {
+	PadState* state;
+	Camera*   cam;
+};
+internal MipsAtom_(pad_input_cam) atom_info(atom_bind(Binds_PadInputCam)
+, atom_reads( R_Cam, R_CamPadState, R_TapePtr)
+, atom_writes(R_Cam)
+) {
+	/* Bind pop: state → R_CamPadState (R_T5), cam → R_Cam (R_T4), advance R_TapePtr by 8. */
+	load_word(R_CamPadState, R_TapePtr, O_(Binds_PadInputCam,state)),
+	load_word(R_Cam,         R_TapePtr, O_(Binds_PadInputCam,cam)),
+	add_ui_self(             R_TapePtr, S_(Binds_PadInputCam)),
+
+	/* Load pad[0].buttons into R_T0; nop fills the load-delay slot. */
+	load_word(R_T0, R_CamPadState, O_(PadState,buttons)),
+	load_word(R_T1, R_Cam, O_(Camera,pos.x)), // BD-Slot.
+
+	// D-pad Left → cam.pos.x -= 50.  and_i fulfills BD-slot for load on R_Cam.
+	and_i(R_T3, R_T0, Pad_Left), branch_le_zero(R_T3, atom_offset(left_x, exit_left_x)), mac_yield_load(),
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.x)), 
+	atom_label(exit_left_x)
+
+	/* D-pad Right → cam.pos.x += 50. Reuses R_T1 from Left. */
+	and_i(R_T3, R_T0, Pad_Right), branch_le_zero(R_T3, atom_offset(right_x, exit_right_x)), nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.x)),
+	atom_label(exit_right_x)
+
+	/* D-pad Up → cam.pos.y -= 50. Load pos.y BEFORE the andi. */
+	load_word(R_T1, R_Cam, O_(Camera,pos.y)),
+	and_i(R_T3, R_T0, Pad_Up), branch_le_zero(R_T3, atom_offset(up_y, exit_up_y)), nop,
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.y)),
+	atom_label(exit_up_y)
+
+	/* D-pad Down → cam.pos.y += 50. Reuses R_T1 from Up. */
+	and_i(R_T3, R_T0, Pad_Down), branch_le_zero(R_T3, atom_offset(down_y, exit_down_y)), nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.y)),
+	atom_label(exit_down_y)
+
+	/* D-pad Cross → cam.pos.z -= 50. Load pos.z BEFORE the andi. */
+	load_word(R_T1, R_Cam, O_(Camera,pos.z)),
+	and_i(R_T3, R_T0, Pad_Cross), branch_le_zero(R_T3, atom_offset(cross_z, exit_cross_z)), nop,
+		add_si(R_T1, R_T1, -50), store_word(R_T1, R_Cam, O_(Camera,pos.z)),
+	atom_label(exit_cross_z)
+
+	/* D-pad Circle → cam.pos.z += 50. Reuses R_T1 from Cross. */
+	and_i(R_T3, R_T0, Pad_Circle), branch_le_zero(R_T3, atom_offset(circle_z, exit_circle_z)), nop,
+		add_si(R_T1, R_T1,  50), store_word(R_T1, R_Cam, O_(Camera,pos.z)),
+	atom_label(exit_circle_z)
+
+	mac_yield_tail(),
+};
+
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* 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. */
+		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(R_OtBase, R_PrimCursor, S_(Poly_G4)),
+			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(R_OtBase, R_PrimCursor, S_(Poly_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_camera/hello_camera.c

@@ -0,0 +1,564 @@
+#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/bios.h"
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/pad.c"
+#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 "gen/auto_reg.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_Loc = 0x1F800000,
+};
+#define C_scratch(type) C_(type, Scratchpad_Loc)
+
+enum {
+	Scratchpad_Len           = 1024,
+	MemTape_Len              = 512,
+	ResolveLookAtArena_Words = 1024,
+	ResolveLookAtArena_Size  = ResolveLookAtArena_Words * S_(MipsCode),
+};
+typedef Struct_(SMemory) {
+	PrimitiveArena          primitives;
+	A2_OrderingTable_Buffer ordering_tbl;
+	DoubleBuffer            screen_buf;
+	S4                      active_buf_id;
+
+	U4 MemTape[MemTape_Len];
+
+	MT3_S2S4 tform_world;
+	MT3_S2S4 tform_view;
+
+	Camera cam;
+
+	Ent_Cube  cube;
+	Ent_Floor floor;
+
+	PadBiosRaw pad_raw[2];
+	PadState   pad[2];
+
+	// TODO(Ed): We don't need this we can just cast at any point an address to a desired view of scratchpad, we have the address.
+	U4_V scratchpad; // d-cache
+
+	U1        resolve_look_at_mem[ResolveLookAtArena_Size];
+	MipsAtom* resolve_look_at_atom_addrs[10];
+};
+global SMemory smem;
+extern SMemory smem;
+
+#define pad0_btn_(btn) btn & smem.pad[0].buttons
+#define pad1_btn_(btn) btn & smem.pad[1].buttons
+
+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)))
+
+I_ void resolve_look_at_c11(MT3_S2S4* look_at, P3_S4* eye, P3_S4* target, V3_S4* up_in) {
+// RGA(Lengyel): Build matrix expansion of a rigid transformation. Corresponding motor is not constructed; we write the LA form for GTE.
+// Preconditions: eye != target, up_in not collinear with (target - eye).
+	V3_S4 right, up, forward;
+	V3_S4 ux, uy, uz;
+	V3_S4 pos, off;
+
+	forward = target[0]; sub_v3s4(& forward, eye[0]); // RGA(Lengyel): Affine point - point = zero-weight direction.
+	normalize_v3s4(& forward, & uz);                  // RGA(Lengyel): Normalize the direction bulk. Not finite-point unitization.
+
+	cross_v3s4(& uz,   up_in, & right); normalize_v3s4(& right, & ux); // RGA(Lengyel): Complement(Wedge(forward, up_in)) -> right axis.
+	cross_v3s4(& uz, & ux,    & up);    normalize_v3s4(& up,    & uy); // RGA(Lengyel): Complement(Wedge(forward, right)) -> up axis.
+
+	// RGA(Lengyel): matrix expansion of the world-to-camera rotation (basis rows).
+	look_at->m[0][0] = ux.x; look_at->m[0][1] = ux.y; look_at->m[0][2] = ux.z;
+	look_at->m[1][0] = uy.x; look_at->m[1][1] = uy.y; look_at->m[1][2] = uy.z;
+	look_at->m[2][0] = uz.x; look_at->m[2][1] = uz.y; look_at->m[2][2] = uz.z;
+
+	pos = eye[0]; mul_v3s4(& pos, v3s4(-1,-1,-1)); // RGA(Lengyel): -eye in world coordinates (spatial bulk only; implicit weight is dropped).
+
+	// RGA(Lengyel): R * (-eye) is the full matrix translation column. 
+	// Motor translator would store half this displacement in m.xyz; GTE consumes full column.
+	mul_m3s2_v3s4(look_at, & pos, & off);
+	trans_m3s2(   look_at,        & off);
+}
+FI_ void camera_look_at_c11(Camera* c, P3_S4* target, V3_S4* up_in) { resolve_look_at_c11(& c->look_at, & c->pos, target, up_in); }
+
+/* Pre-build all 7 chain atoms of the resolve_look_at bundle into the static arena.
+ * 4 unique procs in hello_camera.atom.c (chain atoms 0, 2, 4, 6); atoms 1, 3, 5
+ * share the GENERIC normalize_v3s4_proc from gte.atom.c
+ * 	0: resolve_look_at__input_and_sub_proc
+ * 	1: normalize_v3s4_proc                  (fwd → uz; offsets 0, 16)
+ * 	2: resolve_look_at__cross_uz_up_in_to_right_proc
+ * 	3: normalize_v3s4_proc                  (right → ux; offsets 32, 48)
+ * 	4: resolve_look_at__cross_uz_ux_to_up_proc
+ * 	5: normalize_v3s4_proc                  (up → uy; offsets 64, 80)
+ * 	6: resolve_look_at__populate_and_translate_proc
+ */
+internal void resolve_look_at_init(void) {
+	/* Wrap the static arena in a MipsAtomBuilder. */
+	AtomArena   ab = atomarena_make(slice_ut_arr(smem.resolve_look_at_mem));
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.resolve_look_at_atom_addrs));
+
+	U4 pin_mask = regfile_abi_mask | (1 << R_ResolveScratch);
+	RegFile rf  = regfile(pin_mask);
+
+	// defer(regfile_reset_mask(& rf, pin_mask)) {
+	// 	U4 r_target_ptr = regfile_alloc(& rf); 
+	// 	U4 r_eye_ptr    = regfile_alloc(& rf);
+	// 	U4 r_up_in_ptr  = regfile_alloc(& rf);
+	// 	U4 r_tmp0       = regfile_alloc(& rf);
+	// 	U4 r_tmp1       = regfile_alloc(& rf);
+	// 	U4 r_tmp2       = regfile_alloc(& rf);
+	// 	U4 r_tmp3       = regfile_alloc(& rf);
+	// 	tb_emit_(resolve_look_at__input_and_sub_proc(& ab,
+	// 		R_ResolveScratch,
+	// 		r_target_ptr, r_eye_ptr, r_up_in_ptr,
+	// 		r_tmp0, r_tmp1, r_tmp2, r_tmp3));
+	// }
+	U4 r_target_ptr = regfile_alloc(& rf); 
+	U4 r_eye_ptr    = regfile_alloc(& rf);
+	U4 r_up_in_ptr  = regfile_alloc(& rf);
+	U4 r_tmp0       = regfile_alloc(& rf);
+	U4 r_tmp1       = regfile_alloc(& rf);
+	U4 r_tmp2       = regfile_alloc(& rf);
+	U4 r_tmp3       = regfile_alloc(& rf);
+	smem.resolve_look_at_atom_addrs[0] = resolve_look_at__input_and_sub_proc(& ab,
+		R_ResolveScratch,
+		r_target_ptr, r_eye_ptr, r_up_in_ptr,
+		r_tmp0, r_tmp1, r_tmp2, r_tmp3);
+
+	/* === ATOM 1: normalize fwd→uz === */
+	U4 r_src_offset  = O_(ResolveLookAtScratch, fwd);
+	U4 r_dst_offset  = O_(ResolveLookAtScratch, uz);
+	U4 r_src_ptr     = R_T0;
+	U4 r_dst_ptr     = R_T1;
+	U4 r_tmp         = R_T2;
+	U4 r_mac1        = R_T3;
+	U4 r_mac2        = R_T5;
+	U4 r_recip       = R_T6;
+	U4 r_norm        = R_T7;
+	U4 r_shift       = R_V0;
+	U4 r_branch      = R_V1;
+	// tb_emit_(
+	smem.resolve_look_at_atom_addrs[1] = normalize_v3s4_proc(& ab,
+		R_ResolveScratch,
+		r_src_offset, r_dst_offset,
+		r_src_ptr, r_dst_ptr, r_tmp,
+		r_mac1, r_mac2, r_recip, r_norm,
+		r_shift, r_branch);
+	// );
+
+	/* === ATOM 2: cross uz×up_in→right === */
+	U4 r_a_2 = R_T0;
+	U4 r_b_2 = R_T1;
+	U4 r_c_2 = R_T2;
+	U4 r_d_2 = R_T3;
+	U4 r_f_2 = R_T5;  /* out ptr (HARDCODED in body: scratch+32) */
+	U4 r_g_2 = R_T6;  /* a ptr = scratch+16 */
+	U4 r_h_2 = R_T7;  /* b ptr = scratch+128 */
+	smem.resolve_look_at_atom_addrs[2] = resolve_look_at__cross_uz_up_in_to_right_proc(& ab,
+		R_ResolveScratch,
+		r_a_2, r_b_2, r_c_2, r_d_2, r_f_2, r_g_2, r_h_2);
+
+	/* === ATOM 3: normalize right→ux === */
+	U4 r_src_offset_3  = O_(ResolveLookAtScratch, right);
+	U4 r_dst_offset_3  = O_(ResolveLookAtScratch, ux);
+	U4 r_src_ptr_3     = R_T0;
+	U4 r_dst_ptr_3     = R_T1;
+	U4 r_tmp_3         = R_T2;
+	U4 r_mac1_3        = R_T3;
+	U4 r_mac2_3        = R_T5;
+	U4 r_recip_3       = R_T6;
+	U4 r_norm_3        = R_T7;
+	U4 r_shift_3       = R_V0;
+	U4 r_branch_3      = R_V1;
+	smem.resolve_look_at_atom_addrs[3] = normalize_v3s4_proc(& ab,
+		R_ResolveScratch,
+		r_src_offset_3, r_dst_offset_3,
+		r_src_ptr_3, r_dst_ptr_3, r_tmp_3,
+		r_mac1_3, r_mac2_3, r_recip_3, r_norm_3,
+		r_shift_3, r_branch_3);
+
+	/* === ATOM 4: cross uz×ux→up === */
+	U4 r_a_4 = R_T0;
+	U4 r_b_4 = R_T1;
+	U4 r_c_4 = R_T2;
+	U4 r_d_4 = R_T3;
+	U4 r_f_4 = R_T5;  /* out ptr (HARDCODED: scratch+64) */
+	U4 r_g_4 = R_T6;  /* a ptr = scratch+16 */
+	U4 r_h_4 = R_T7;  /* b ptr = scratch+48 */
+	smem.resolve_look_at_atom_addrs[4] = resolve_look_at__cross_uz_ux_to_up_proc(& ab,
+		R_ResolveScratch,
+		r_a_4, r_b_4, r_c_4, r_d_4, r_f_4, r_g_4, r_h_4);
+
+	/* === ATOM 5: normalize up→uy === */
+	U4 r_src_offset_5  = O_(ResolveLookAtScratch, up);
+	U4 r_dst_offset_5  = O_(ResolveLookAtScratch, uy);
+	U4 r_src_ptr_5     = R_T0;
+	U4 r_dst_ptr_5     = R_T1;
+	U4 r_tmp_5         = R_T2;
+	U4 r_mac1_5        = R_T3;
+	U4 r_mac2_5        = R_T5;
+	U4 r_recip_5       = R_T6;
+	U4 r_norm_5        = R_T7;
+	U4 r_shift_5       = R_V0;
+	U4 r_branch_5      = R_V1;
+	smem.resolve_look_at_atom_addrs[5] = normalize_v3s4_proc(& ab,
+		R_ResolveScratch,
+		r_src_offset_5, r_dst_offset_5,
+		r_src_ptr_5, r_dst_ptr_5, r_tmp_5,
+		r_mac1_5, r_mac2_5, r_recip_5, r_norm_5,
+		r_shift_5, r_branch_5);
+
+	/* === ATOM 6a: populate (m[][] from ux/uy/uz, t[]=0) === */
+	U4 r_look_at_6a = R_T0;  /* tape pop → look_at* */
+	U4 r_scratch_6a = R_ResolveScratch;
+	U4 r_pux_6a      = R_T1;
+	U4 r_puy_6a      = R_T3;
+	U4 r_puz_6a      = R_T5;
+	U4 r_tmp0_6a     = R_T2;
+	U4 r_tmp1_6a     = R_T6;
+	U4 r_tmp2_6a     = R_V0;
+	smem.resolve_look_at_atom_addrs[6] = resolve_look_at__populate_proc(& ab,
+		r_look_at_6a, r_scratch_6a,
+		r_pux_6a, r_puy_6a, r_puz_6a,
+		r_tmp0_6a, r_tmp1_6a, r_tmp2_6a);
+
+	/* === ATOM 6a.5: set_gte_mt3s2s4 (BAKED — ctc2 RT matrix) ===
+	 * This is a BAKED atom from gte.atom.c. Its body hardcodes R_T3 as
+	 * the matrix pointer (popped from tape). It does NOT need GPR
+	 * assignment from us — it has its own internal GPR usage.
+	 * We just take its address. */
+	smem.resolve_look_at_atom_addrs[7] = (MipsAtom*) & set_gte_mt3s2s4;
+
+	/* === ATOM 6b: matrix_vector (RT * (-eye) >> 12) ===
+	 * Uses mac_apply_matrix_lv component macro which internally uses
+	 * r_t0 for the RT matrix load + V0 load, then r_t0/r_t1/r_t2
+	 * for the mfc2/store. We pass our GPRs. */
+	U4 r_scratch_6b = R_ResolveScratch;
+	U4 r_peye_6b    = R_T1;  /* scratch+96 (packed V0 dst, then off dst) */
+	U4 r_look_at_6b = R_T0;  /* tape pop → look_at* */
+	U4 r_tmp0_6b    = R_T2;
+	U4 r_tmp1_6b    = R_T3;
+	U4 r_tmp2_6b    = R_T5;
+	smem.resolve_look_at_atom_addrs[8] = resolve_look_at__matrix_vector_proc(& ab,
+		r_scratch_6b, r_peye_6b, r_look_at_6b,
+		r_tmp0_6b, r_tmp1_6b, r_tmp2_6b);
+
+	/* === ATOM 6c: trans_matrix (off → look_at->t[]) === */
+	U4 r_look_at_6c = R_T0;  /* tape pop → look_at* */
+	U4 r_scratch_6c = R_ResolveScratch;
+	U4 r_off_ptr_6c = R_T1;  /* &scratch.eye (= off dst) */
+	U4 r_tmp0_6c    = R_T2;
+	smem.resolve_look_at_atom_addrs[9] = resolve_look_at__trans_matrix_proc(& ab,
+		r_look_at_6c, r_scratch_6c, r_off_ptr_6c, r_tmp0_6c, R_T3, R_T4);
+
+	/* Sanity check: arena didn't overflow. */
+	assert(ab.used <= ResolveLookAtArena_Size);
+}
+
+/* Emit the resolve_look_at bundle into the tape. Called once per frame from update().
+ * The 7 chain atoms are pre-built at init time (resolve_look_at_init) and referenced by address via smem.resolve_look_at_atom_addrs[].
+ * Per-frame work: 7 tb_emit (atom pointer emissions) + 5 tb_data (C-side pointers for atom 0 + look_at for atom 6).
+ *
+ * Binds_ contract (the field-name labels are for human readability):
+ *   Atom 0  input_and_sub              target(4) eye(4) up_in(4) scratch_base(4) = 4 words
+ *   Atoms 1-5                           (no tape data — atom uses r_scratch + offset internally)
+ *   Atom 6  populate_and_translate     look_at(4)                                = 1 word
+ *                                       ----
+ *                                       5 tb_data words total per frame.
+ */
+I_ void resolve_look_at(
+		TapeBuilder_R tb
+	,	MT3_S2S4* look_at
+	,	P3_S4*    eye
+	,	P3_S4*    target
+	,	V3_S4*    up_in
+){
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[0]); {
+		tb_data(tb, u4_(target));
+		tb_data(tb, u4_(eye));
+		tb_data(tb, u4_(up_in));
+		tb_data(tb, u4_(smem.scratchpad));
+	}
+
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[1]); { }
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[2]); { }
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[3]); { }
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[4]); { }
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[5]); { }
+
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[6]); {
+		tb_data(tb, u4_(look_at));
+	}
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[7]); {
+		tb_data(tb, u4_(look_at));
+	}
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[8]); {
+		tb_data(tb, u4_(look_at));
+	}
+	tb_emit(tb, smem.resolve_look_at_atom_addrs[9]); {
+		// tb_data(tb, u4_(look_at));
+	}
+}
+
+GCC_OPTIMIZATION_DISABLE
+void update(PrimitiveArena* pa, U4* ordering_buf) 
+{
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
+	// Pad Input
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			// Grab latest state from bios.
+			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]);
+
+			tb_emit_(pad_input_cam);
+				tb_data_(state, & smem.pad[0]);
+				tb_data_(cam, & smem.cam);
+
+			// tb_emit_(pad_input_cube_rotation);
+			// 	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; //????
+
+	B4 use_c11_path = false;
+	if (use_c11_path) {
+		camera_look_at_c11(& smem.cam, & smem.cube.pos, & v3s4(0, -fp_one, 0));
+	}
+	if (use_c11_path == false)
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			resolve_look_at(& tb, & smem.cam.look_at, & smem.cam.pos, & smem.cube.pos, & v3s4(0, -fp_one, 0));
+		}
+	}
+
+	// Draw cube
+	if (1)
+	{
+		mt3s2s4_rotation   (& smem.cube.rot,    & smem.tform_world);
+		mt3s2s4_translation(& smem.tform_world, & smem.cube.pos);
+		mt3s2s4_scale      (& smem.tform_world, & smem.cube.scale);
+
+		// Combine world and look_at matrix.
+		gte_comp_coord_m3s2(& smem.cam.look_at, & smem.tform_world, & smem.tform_view);
+		gte_matrix_set_rotation   (& smem.tform_view);
+		gte_matrix_set_translation(& smem.tform_view);
+
+		// 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_a02_s07(tb_slice(tb));// Fire off the tape (bigger-clobber variant).
+
+		// smem.cube.rot.y += 30;
+	}
+	// Draw floor
+	if (1)
+	{
+		mt3s2s4_rotation   (& smem.floor.rot,   & smem.tform_world);
+		mt3s2s4_translation(& smem.tform_world, & smem.floor.pos);
+		mt3s2s4_scale      (& smem.tform_world, & smem.floor.scale);
+
+		// Combine world and look_at matrix.
+		gte_comp_coord_m3s2(& smem.cam.look_at, & smem.tform_world, & smem.tform_view);
+
+		gte_matrix_set_rotation   (& smem.tform_view);
+		gte_matrix_set_translation(& smem.tform_view);
+
+		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_mt3s2s4);
+			// 	tb_data(& tb, u4_(& smem.tform_view));
+
+			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_a02_s07(tb_slice(tb));// Fire off the tape (bigger-clobber variant).
+
+		// 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};
+	// TODO(Ed): remove this field we don't need it in smem.
+	smem.scratchpad = C_(U4_V, Scratchpad_Loc);
+	// smem.primitives.used = 0;
+	// smem.active_buf_id   = 0;
+	smem.cam.pos = v3s4(500, -1000, -1500);
+	/*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]);
+
+		/* Pre-build the resolve_look_at bundle atoms into the static arena. */
+		resolve_look_at_init();
+
+		/* 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),
+};
+
+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; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	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; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};
+
+typedef Struct_(Camera) {
+	P3_S4    pos; // RGA(Lengyel): affine point with implicit weight one. Storage alias of V3_S4.
+	V3_S2    rot;
+	MT3_S2S4 look_at;
+};

code/hello_gte/hello_gte.c

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

code/hello_gte/hello_gte.tape.c

@@ -1,10 +1,10 @@
 #ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/macs.h"
-#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/gen/offsets.h"
 #	include "duffle/atom_dsl.h"
 #	include "duffle/lottes_tape.h"
 #	include "duffle/word_count.metadata.h"
-#	include "gen/hello_gte.offsets.h"
+#	include "gen/offsets.h"
 #	include "hello_gte.h"
 #endif
 

code/hello_joypad/gen/macs.h

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

code/hello_joypad/gen/offsets.h

@@ -1,8 +1,11 @@
 // Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
-// Source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.tape.c
+// Directory: C:\projects\Pikuma\ps1\code\hello_joypad\
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.c
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.h
+//   source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.atom.c
 #pragma once
 
-#pragma region hello_joypad.tape
+#pragma region hello_joypad
 
 
 // --- atom: cube_g4_face (76 words) ---
@@ -25,18 +28,18 @@ enum {
     atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
 };
 
-// --- atom: pad_bios_snapshot (77 words) ---
+// --- atom: pad_bios_snapshot (78 words) ---
 
 #define _atom_offset_snap_root_skip_disconnected      8
-#define _atom_offset_disconnected_snap_end            60
+#define _atom_offset_disconnected_snap_end            61
 #define _atom_offset_case_2_id_dispatch               8
-#define _atom_offset_pending_snap_end                 50
+#define _atom_offset_pending_snap_end                 51
 #define _atom_offset_id_dispatch_try_analog_stick     11
-#define _atom_offset_id_dispatch_snap_end             37
+#define _atom_offset_id_dispatch_snap_end             38
 #define _atom_offset_try_analog_stick_try_analog_pad  12
-#define _atom_offset_analog_stick_snap_end            23
+#define _atom_offset_analog_stick_snap_end            24
 #define _atom_offset_try_analog_pad_try_unsupported   11
-#define _atom_offset_analog_pad_snap_end              9
+#define _atom_offset_analog_pad_snap_end              10
 
 enum {
     atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
@@ -51,14 +54,14 @@ enum {
     atom_offset_analog_pad_snap_end = _atom_offset_analog_pad_snap_end,
 };
 
-// --- atom: pad_apply_input (59 words) ---
+// --- 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        23
+#define _atom_offset_dead_zone_skip_exit_stick        24
-#define _atom_offset_end_low_exit_stick               11
+#define _atom_offset_end_low_exit_stick               12
 
 enum {
     atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
@@ -69,5 +72,5 @@ enum {
     atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
 };
 
-#pragma endregion hello_joypad.tape
+#pragma endregion hello_joypad
 

code/hello_joypad/hello_joypad.atom.c

@@ -1,53 +1,31 @@
 #ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
+#	pragma once
-#	include "duffle/gen/duffle.offsets.h"
+#		include "duffle/gen/macs.h"
-#	include "duffle/atom_dsl.h"
+#		include "duffle/gen/offsets.h"
+#	include "duffle/dsl.atom.h"
 #	include "duffle/lottes_tape.h"
 #	include "duffle/mips.h"
 #	include "duffle/gte.h"
 #	include "duffle/gp.h"
 #	include "duffle/pad.h"
 #	include "duffle/word_count.metadata.h"
-#	include "psyq.h"
+#	include "duffle/psyq.h"
-#	include "gen/hello_joypad.offsets.h"
+#	include "duffle/math.atom.c"
-#	include "gen/hello_joypad.macs.h"
+#	include "duffle/mips.atom.c"
+#	include "duffle/gte.atom.c"
+#	include "duffle/gp.atom.c"
+#	include "duffle/psyq.atom.c"
+#		include "gen/offsets.h"
+#		include "gen/macs.h"
 #	include "hello_joypad.h"
 #endif
 
+ATOM_FILE_DEBUGGER_LINE_MARKER(hello_joypad_atom_c);
+
 #pragma region MACs (Mips Atom components)
 
-FI_ Slice_MipsCode ac_load_v2s2(U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_load_v2s2, {
+FI_ Slice_MipsCode ac_put_disp_env(MipsAtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
-	load_half( rs_x, r_base, O_(V3_S2,x)),
+MipsAtomComp_Proc_(ab, {
-	load_half( rs_y, r_base, O_(V3_S2,y)),
-})
-
-FI_ Slice_MipsCode ac_store_v2s2(U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_v2s2, {
-	store_half(rt_x, base, offset + O_(V2_S2,x)),
-	store_half(rt_y, base, offset + O_(V2_S2,y)),
-})
-
-FI_ Slice_MipsCode ac_store_rects2(U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rects2, {
-	store_half(rt_x,      base, offset + O_(Rect_S2,x)),
-	store_half(rt_y,      base, offset + O_(Rect_S2,y)),
-	store_half(rt_width,  base, offset + O_(Rect_S2,width)),
-	store_half(rt_height, base, offset + O_(Rect_S2,height)),
-})
-
-FI_ Slice_MipsCode ac_store_rgb8(U1 rr, U1 rg, U1 rb, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rgb8, {
-	store_byte(rr, base, offset + O_(DrawEnv,initial_bg_color.r)),
-	store_byte(rg, base, offset + O_(DrawEnv,initial_bg_color.g)),
-	store_byte(rb, base, offset + O_(DrawEnv,initial_bg_color.b)),	
-})
-
-FI_ Slice_MipsCode ac_gcmd_push(U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
-MipsAtomComp_Proc_(ac_gcmd_push, {
-	load_upper_i(reg_transfer, cmd >> 16),
-	or_i_self(   reg_transfer, cmd & 0xFFFF),
-	store_word(  reg_transfer, reg_base, port),
-})
-
-FI_ Slice_MipsCode ac_put_disp_env(U4 reg_transfer, U4 reg_base, U2 port)
-MipsAtomComp_Proc_(ac_put_disp_env, {
 	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
 	// 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),
@@ -57,8 +35,8 @@ MipsAtomComp_Proc_(ac_put_disp_env, {
 	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)
+FI_ Slice_MipsCode ac_put_draw_env(MipsAtomBuilder_R ab, U4 reg_transfer, U4 reg_base, U2 port)
-MipsAtomComp_Proc_(ac_put_draw_env, {
+MipsAtomComp_Proc_(ab, {
 	/*
 	* ORIGIN: each code word corresponds to the EXACT value libpsyx's PutDrawEnv function would compute for the same DrawEnv settings.
 	* References:
@@ -138,7 +116,7 @@ internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
 	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_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)),
 
@@ -151,7 +129,7 @@ internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
 	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), 
+	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)),
 
@@ -166,7 +144,7 @@ internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
 	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), 
+	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)),
 
@@ -202,6 +180,17 @@ internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads
 	mac_yield(),
 };
 
+enum {
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
+	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
+	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
+#define R_PrimCursor_Code  R_T7_Code
+#define R_FaceCursor_Code  R_T4_Code
+#define R_VertBase_Code    R_T5_Code
+#define R_OtBase_Code      R_T6_Code
+};
+
 typedef Struct_(Binds_CubeTri) {
 	U4     PrimCursor;
 	V4_S2* FaceCursor;
@@ -232,23 +221,23 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
 	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
 	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
 
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase, R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
 	gte_cmdw_nclip,
 
 	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
 	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
 		/* 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
+		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded —
-		 * because the OT entry that points to this prim is created later, only on the body path. */
+		 * harmless because the OT entry that points to this prim is created later, only on the body path. */
 		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
 		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
 		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
 		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
 
-		mac_gte_store_g4_p012(),
+		mac_gte_store_g4_p012(R_PrimCursor),
 		gte_cmdw_rotate_translate_perspective_single,
-		mac_gte_store_g4_p3(),
+		mac_gte_store_g4_p3(R_PrimCursor),
 
 		gte_cmdw_avg_sort_z4,
 		gte_mv_from_data_r(R_T1, C2_OTZ),
@@ -256,8 +245,8 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
 		set_lt_u(    R_AT, R_T1, R_AT),
 
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
-			mac_insert_ot_tag_g4(),
+			mac_insert_ot_tag_g4(R_OtBase, R_PrimCursor),
-			mac_format_g4_color(
+			mac_format_g4_color(R_PrimCursor,
 				/* c0 magenta */ 0xFF, 0x00, 0xFF,
 				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
 				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
@@ -297,16 +286,16 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
 	, atom_writes(R_PrimCursor, R_FaceCursor)
 ) {
-	mac_load_tri_indices(  R_T0, R_T1, R_T2),
+	mac_load_tri_indices(R_FaceCursor, R_T0, R_T1, R_T2),
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_VertBase,   R_T0, R_T1, R_T2),
 	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
 	gte_cmdw_nclip,
 
 	/* 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. 
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot.
 		/* Format Primitive */
-		mac_gte_store_f3(),
+		mac_gte_store_f3(R_PrimCursor),
 
 		/* Calculate Depth */
 		gte_avg_sort_z3,
@@ -315,8 +304,8 @@ MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
 		add_ui(      R_AT, R_0,  OrderingTbl_Len),
 		set_lt_u(    R_AT, R_T1, R_AT),
 		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop,
-			mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_format_f3_color(R_PrimCursor, 0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-			mac_insert_ot_tag_f3(),                 /* Insert into Ordering Table Linked List */
+			mac_insert_ot_tag_f3(R_OtBase, R_PrimCursor),         /* Insert into Ordering Table Linked List */
 			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
 				// Note(Ed): No bounds checking, should be checked before atom runs.
 		// end: branch(bounds_chk)
@@ -504,6 +493,9 @@ atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the
 	store_byte(  R_T4, R_PadState, O_(PadState,id)),
 	/* Fall through to snap_end. */
 
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
 atom_label(snap_end)
 	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the case-exit branch. */
 	mac_yield_tail(),
@@ -635,6 +627,9 @@ atom_label(dead_high_active)
 	add_u(       R_T0, R_T0, R_T4),
 	store_half(  R_T0, R_FloorRot, O_(V3_S2,y)),
 
+atom_label(no_jump_fallthrough)
+	mac_yield_load(),
+
 atom_label(exit_stick)
 	/* NOT mac_yield() — R_AtomJmp was already loaded in the BD-slot of the dead-zone/exit branch. */
 	mac_yield_tail(),

code/hello_joypad/hello_joypad.c

@@ -1,9 +1,17 @@
+#pragma region Vendors
 #include <stdio.h>
 #include <stdlib.h>
 #include <assert.h>
 // #include "libgpu.h"
 // #include "libetc.h"
 // #include "libgte.h"
+#pragma endregion Vendors
+
+#pragma region Duffle Headers
+#	include "duffle/gen/macs.h"
+#	include "duffle/gen/offsets.h"
+
+#include "duffle/word_count.metadata.h"
 
 #include "duffle/dsl.h"
 #include "duffle/memory.h"
@@ -15,96 +23,30 @@
 #include "duffle/gte.h"
 #include "duffle/pad.h"
 
-#	include "duffle/gen/duffle.macs.h"
+#include "duffle/dsl.atom.h"
-#	include "duffle/gen/duffle.offsets.h"
-#include "duffle/atom_dsl.h"
 #include "duffle/lottes_tape.h"
-#include "duffle/word_count.metadata.h"
 
-#include "psyq.h"
+#include "duffle/psyq.h"
+#pragma endregion Duffle Headers
+
+#pragma region Duffle TUs
+#include "duffle/math.atom.c"
+#include "duffle/mips.atom.c"
+#include "duffle/gte.atom.c"
+#include "duffle/gp.atom.c"
+#include "duffle/psyq.atom.c"
+#pragma endregion Duffle TUs
+
+#pragma region Joypade Headers
+#	include "gen/macs.h"
+#	include "gen/offsets.h"
 
-#	include "gen/hello_joypad.macs.h"
-#	include "gen/hello_joypad.offsets.h"
 #include "hello_joypad.h"
+#pragma region Joypad Headers
 
-#include "psyq.c"
+#pragma region Hello Joypad TUs
-#include "hello_joypad.tape.c"
+#include "hello_joypad.atom.c"
-
+#pragma endregion Hello Joypad TUs
-
-typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
-typedef Array_(OrderingTable_Buffer, 2);
-
-typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
-typedef Array_(PrimitiveBuffer, 2);
-typedef Struct_(PrimitiveArena) {
-	A2_PrimitiveBuffer buf;
-	U4                 used;
-};
-
-#define Cube_num_verts 8
-typedef Array_(V3_S2, Cube_num_verts);
-#define Cube_num_faces 6
-typedef Array_(V4_S2, Cube_num_faces);
-I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
-	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
-		{ -128, -128, -128 },
-		{  128, -128, -128 },
-		{  128, -128,  128 },
-		{ -128, -128,  128 },
-		{ -128,  128, -128 },
-		{  128,  128, -128 },
-		{  128,  128,  128 },
-		{ -128,  128,  128 }
-	};
-	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
-		{ 3, 2, 0, 1 },
-		{ 0, 1, 4, 5 },
-		{ 4, 5, 7, 6 },
-		{ 1, 2, 5, 6 },
-		{ 2, 3, 6, 7 },
-		{ 3, 0, 7, 4 },
-	};
-	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
-	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
-	return;
-}
-typedef Struct_(Ent_Cube) {
-	V3_S4 accel;
-	V3_S4 vel;
-	V3_S4 pos;
-	V3_S4 scale;
-	V3_S2 rot;
-	A8_V3_S2 verts;
-	A6_V4_S2 faces;
-};
-
-#define Floor_num_verts 4
-typedef Array_(V3_S2, Floor_num_verts);
-#define Floor_num_faces 2
-typedef Array_(V3_S2, Floor_num_faces);
-I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
-	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
-		{ -900, 0, -900 },
-		{ -900, 0,  900 },
-		{  900, 0, -900 },
-		{  900, 0,  900 },
-	};
-	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
-		{ 0, 1, 2 },
-		{ 1, 3, 2 },
-	};
-	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
-	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
-};
-typedef Struct_(Ent_Floor) {
-	V3_S4 accel;
-	V3_S4 pos;
-	V3_S4 scale;
-	V3_S2 rot;
-	A4_V3_S2 verts;
-	A2_V3_S2 faces;
-};
-
 
 enum { 
 	Scratchpad_Len = 1024, 
@@ -212,50 +154,6 @@ NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
 	);
 }
 
-void gp_screen_init_c11(DoubleBuffer* screen_buf, S4* active_buf_id)
-{
-	reset_graph(0);
-
-	// Set the current initial buffer
-	active_buf_id[0] = 0;
-
-	// Just setting env data, not interacting with console hw.
-	// First buffer area
-	displayenv_init(& r_(screen_buf->display)[0], 0, 0,           ScreenRes_X, ScreenRes_Y);
-	drawenv_init   (& r_(screen_buf->draw   )[0], 0, ScreenRes_Y, ScreenRes_X, ScreenRes_Y);
-	// Second buffer area
-	displayenv_init(& r_(screen_buf->display)[1], 0, ScreenRes_Y, ScreenRes_X, ScreenRes_Y);
-	drawenv_init   (& r_(screen_buf->draw   )[1], 0, 0,           ScreenRes_X, ScreenRes_Y);
-	// Set the back/drawing buffer
-	screen_buf->draw[0].enable_auto_clear = true;
-	screen_buf->draw[1].enable_auto_clear = true;
-	// Set the background clear color
-	screen_buf->draw[0].initial_bg_color = rgb8( .r = 7, .g = 7,  .b = 7 );
-	screen_buf->draw[1].initial_bg_color = rgb8( .r = 7, .g = 7,  .b = 7 );
-	// screen_buf->draw[1].initial_bg_color = rgb8( .r = 47, .g = 13, .b = 0 );
-	displayenv_put(& r_(screen_buf->display)[ active_buf_id[0] ]);
-	drawenv_put   (& r_(screen_buf->draw   )[ active_buf_id[0] ]);
-
-	// Initialize and setup the GTE geometry offsets
-	geom_init();
-	geom_set_offset(ScreenRes_CenterX, ScreenRes_CenterY);
-	geom_set_screen(ScreenZ);
-
-	set_display_enabled(1); // gp_DisplayEnabled
-}
-
-void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_buf, PrimitiveArena* pa) {
-	draw_sync(0);
-	vsync(0);
-	displayenv_put(& r_(screen_buf->display)[active_buf_id[0] ]);
-	drawenv_put   (& r_(screen_buf->draw)   [active_buf_id[0] ]);
-	{
-		draw_orderingtbl(ordering_buf + OrderingTbl_Len - 1);
-		pa->used = 0;
-	}
-	active_buf_id[0] = ! active_buf_id[0]; // Swap current buffer
-}
-
 GCC_OPTIMIZATION_DISABLE
 void update(PrimitiveArena* pa, U4* ordering_buf) 
 {
@@ -484,6 +382,19 @@ 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};
@@ -527,3 +438,4 @@ int main(void)
 	};
 	return 0;
 }
+GCC_OPTIMIZATION_ENABLE

code/hello_joypad/hello_joypad.h

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

code/hello_joypad/psyq.c

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

code/hello_joypad/sio.tape.c

@@ -24,8 +24,8 @@
  * 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)
+FI_ Slice_MipsCode ac_pad_sio_write_pad_state(MipsAtomBuilder_R ab, U4 status_val, U4 state_ptr_reg, U4 scratch_reg)
-MipsAtomComp_Proc_(ac_pad_sio_write_pad_state, {
+MipsAtomComp_Proc_(ac_pad_sio_write_pad_state, ab, {
 	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

scripts/build_psyq.ps1

@@ -180,12 +180,9 @@ function link-modules { param([string[]]$link_modules, [string]  $elf, [string[]
 	$link_args += ($f_link_pass_through_prefix + $f_link_mapfile + $map)
 
 	$link_args += ($f_link_pass_through_prefix + $f_link_start_group)
-	# raw_sio_pad_poll_20260802 — Task 5.1c surgical library-list trim.
+	# 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math, mcrd, mcx, press, sio, snd, spu, tap)
-	# The 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math,
+	# had LOAD lines in the map but ZERO .o files pulled in — they were unused.
-	# mcrd, mcx, press, sio, snd, spu, tap) had LOAD lines in the map but
+	# 5 kept libraries (api, c, etc, gpu, gte) are required by the C-side calls in hello_joypad.c (reset_graph, draw_sync, vsync, etc.).
-	# ZERO .o files pulled in — they were unused. The 5 kept libraries
-	# (api, c, etc, gpu, gte) are required by the C-side calls in
-	# hello_joypad.c (reset_graph, draw_sync, vsync, etc.).
 	$libraries = @(
 		"api",
 		"c",
@@ -227,9 +224,7 @@ function ps1-meta { param(
 		[string[]]$passes = @('--pre-link'),
 		[string[]]$extra_args = @()
 	)
-	# `--unity-root` and `--source` are
+	# `--unity-root` and `--source` are mutually exclusive. Exactly one of `$unity_root` / `$sources` must be supplied; the other must be absent.
-	# mutually exclusive. Exactly one of `$unity_root` / `$sources` must
-	# be supplied; the other must be absent.
 	if ($null -ne $unity_root -and $unity_root -ne '') 
 	{
 		if ($null -ne $sources -and $sources.Count -gt 0) {
@@ -265,6 +260,73 @@ function ps1-meta { param(
 	}
 }
 
+function inject-dwarf { param(
+	[string]$elf,
+	[string]$path_gen
+)
+	$base_name               = [System.IO.Path]::GetFileNameWithoutExtension($elf)
+	$path_dwarf_line_bin     = join-path $path_gen  "$base_name.dwarf_line.bin"
+	$path_dwarf_aranges_bin  = join-path $path_gen  "$base_name.dwarf_aranges.bin"
+	$path_dwarf_rnglists_bin = join-path $path_gen  "$base_name.dwarf_rnglists.bin"
+	$path_dwarf_info_bin     = join-path $path_gen  "$base_name.dwarf_info.bin"
+	$path_dwarf_abbrev_bin   = join-path $path_gen  "$base_name.dwarf_abbrev.bin"
+	$path_dwarf_str_bin      = join-path $path_gen  "$base_name.dwarf_str.bin"
+	$path_dwarf_loc_bin      = join-path $path_gen  "$base_name.dwarf_loc.bin"
+	$path_dwarf_loclists_bin = join-path $path_gen  "$base_name.dwarf_loclists.bin"
+	$path_inject_elf         = join-path $path_build "$base_name.dwarf-injected.elf"
+
+	if (-not (Test-Path $path_dwarf_line_bin))     { return }
+	if (-not (Test-Path $path_dwarf_aranges_bin))  { return }
+	if (-not (Test-Path $path_dwarf_rnglists_bin)) { return }
+
+	Write-Host "[build] DWARF-injecting $elf -> $path_inject_elf"
+	Copy-Item -LiteralPath $elf -Destination $path_inject_elf -Force
+
+	# Objcopy call 1: 3x --update-section for the PC-mapping tables (line, aranges, rnglists).
+	$objcopy_args_dwarf_pc = @(
+		"--update-section=.debug_line=$path_dwarf_line_bin",
+		"--update-section=.debug_aranges=$path_dwarf_aranges_bin",
+		"--update-section=.debug_rnglists=$path_dwarf_rnglists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_pc $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-pc splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Objcopy call 2: 3x --update-section + 2x --add-section for the debug-data tables (info, abbrev, str, loc, loclists).
+	$objcopy_args_dwarf_info = @(
+		"--update-section=.debug_info=$path_dwarf_info_bin",
+		"--update-section=.debug_abbrev=$path_dwarf_abbrev_bin",
+		"--update-section=.debug_str=$path_dwarf_str_bin",
+		"--add-section=.debug_loc=$path_dwarf_loc_bin",
+		"--add-section=.debug_loclists=$path_dwarf_loclists_bin"
+	)
+	& $Objcopy @objcopy_args_dwarf_info $path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] objcopy dwarf-info splice failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+		return
+	}
+
+	# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
+	# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
+	# The original ELF and PS-EXE remain byte/flag unchanged.
+	& $Objcopy `
+		--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
+		--set-section-flags ".data=alloc,load,data,code,contents" `
+		$path_inject_elf 2>&1 | Out-Null
+	if ($LASTEXITCODE -ne 0) {
+		Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $path_inject_elf"
+		Remove-Item -LiteralPath $path_inject_elf -ErrorAction SilentlyContinue
+	}
+	else {
+		Write-Host "[build] DWARF-injected ELF: $path_inject_elf"
+	}
+}
+# inject-dwarf
+
 function build-hello_psyqo {
 	$includes += @()
 
@@ -391,61 +453,7 @@ function build-hello_gte {
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
 	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
 
-	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
+	inject-dwarf $elf $path_build_gen
-	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
-	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
-	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
-	{
-		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
-		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		$f_args = @(
-			"--update-section=.debug_line=$dwarfLineBin",
-			"--update-section=.debug_aranges=$dwarfArangesBin",
-			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
-		)
-		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		$dwarfInfoBin     = join-path $path_build_gen 'hello_gte.dwarf_info.bin'
-		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_gte.dwarf_abbrev.bin'
-		$dwarfStrBin      = join-path $path_build_gen 'hello_gte.dwarf_str.bin'
-		$dwarfLocBin      = join-path $path_build_gen 'hello_gte.dwarf_loc.bin'
-		$dwarfLoclistsBin = join-path $path_build_gen 'hello_gte.dwarf_loclists.bin'
-		$g_args = @(
-			"--update-section=.debug_info=$dwarfInfoBin",
-			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
-			"--update-section=.debug_str=$dwarfStrBin",
-			"--add-section=.debug_loc=$dwarfLocBin",
-			"--add-section=.debug_loclists=$dwarfLoclistsBin"
-		)
-		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
-		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
-		# The original ELF and PS-EXE remain byte/flag unchanged.
-		& $Objcopy `
-			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
-			--set-section-flags ".data=alloc,load,data,code,contents" `
-			$injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-		} 
-		else {
-			Write-Host "[build] DWARF-injected ELF: $injectElf"
-		}
-	}
 }
 # build-hello_gte
 
@@ -496,63 +504,60 @@ function build-hello_joypad {
 	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
 	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
 
-	$dwarfLineBin     = join-path $path_build_gen 'hello_joypad.dwarf_line.bin'
+	inject-dwarf $elf $path_build_gen
-	$dwarfArangesBin  = join-path $path_build_gen 'hello_joypad.dwarf_aranges.bin'
-	$dwarfRnglistsBin = join-path $path_build_gen 'hello_joypad.dwarf_rnglists.bin'
-	$injectElf        = join-path $path_build 'hello_joypad.dwarf-injected.elf'
-	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
-	{
-		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
-		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
-		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
-		$f_args = @(
-			"--update-section=.debug_line=$dwarfLineBin",
-			"--update-section=.debug_aranges=$dwarfArangesBin",
-			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
-		)
-		& $Objcopy @f_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy F' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		$dwarfInfoBin     = join-path $path_build_gen 'hello_joypad.dwarf_info.bin'
-		$dwarfAbbrevBin   = join-path $path_build_gen 'hello_joypad.dwarf_abbrev.bin'
-		$dwarfStrBin      = join-path $path_build_gen 'hello_joypad.dwarf_str.bin'
-		$dwarfLocBin      = join-path $path_build_gen 'hello_joypad.dwarf_loc.bin'
-		$dwarfLoclistsBin = join-path $path_build_gen 'hello_joypad.dwarf_loclists.bin'
-		$g_args = @(
-			"--update-section=.debug_info=$dwarfInfoBin",
-			"--update-section=.debug_abbrev=$dwarfAbbrevBin",
-			"--update-section=.debug_str=$dwarfStrBin",
-			"--add-section=.debug_loc=$dwarfLocBin",
-			"--add-section=.debug_loclists=$dwarfLoclistsBin"
-		)
-		& $Objcopy @g_args $injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] objcopy G' splice failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-			return;
-		}
-
-		# Baked atoms execute from RAM but are emitted as C data arrays, so their ELF sections lack SHF_EXECINSTR.
-		# GDB discards line rows for non-code sections. Mark only the debug-copy sections executable.
-		# The original ELF and PS-EXE remain byte/flag unchanged.
-		& $Objcopy `
-			--set-section-flags ".rodata=alloc,load,readonly,code,contents" `
-			--set-section-flags ".data=alloc,load,data,code,contents" `
-			$injectElf 2>&1 | Out-Null
-		if ($LASTEXITCODE -ne 0) {
-			Write-Warning "[build] atom-section flag update failed (exit $LASTEXITCODE); removing $injectElf"
-			Remove-Item -LiteralPath $injectElf -ErrorAction SilentlyContinue
-		} 
-		else {
-			Write-Host "[build] DWARF-injected ELF: $injectElf"
-		}
-	}
 }
-build-hello_joypad
+# 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_define + 'BUILD_DEBUG')
+	$compile_args += $f_optimize_none
+	# $compile_args += $f_optimize_intrinsics
+	# $compile_args += $f_optimize_size
+	# $compile_args += $f_optimize_debug
+	$compile_args += ($f_include + $path_code)
+	compile-unit $src_c $module_c $includes $compile_args
+
+	$elf = join-path $path_build 'hello_camera.elf'
+	$exe = join-path $path_build 'hello_camera.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
+
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
+
+	inject-dwarf $elf $path_build_gen
+}
+build-hello_camera
 
 # NO idea if this works yet...
 function Send-ToEmulator { param( [string]$exePath )

scripts/duffle_paths.lua

@@ -47,15 +47,14 @@ local function find_repo_root()
 	return root
 end
 
---- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and `package.cpath` (for `lpeg.dll`).
---- `package.cpath` (for `lpeg.dll`).
 ---
 --- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$PATH` mods. 
 --- It just sets `package.path` and `package.cpath` (the standard Lua way to register module search dirs). 
 --- lpeg is built by `update_deps.ps1` to `toolchain/lpeg/`, 
 --- which we wire into `package.cpath` here (so `require("lpeg")` from `duffle.lua` resolves without any global state).
 function M.setup()
-	local repo_root = find_repo_root()
+	local  repo_root = find_repo_root()
 	if not repo_root then
 		-- Unreachable in practice: find_repo_root() derives the repo root from this script's
 		-- own source path via debug.getinfo(1, "S").source (no subprocess, no git CLI, <1ms).

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 = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -60,19 +60,19 @@ M.DW_AT = {
 }
 
 M.DW_FORM = {
-	addr        = 0x01,
+	addr           = 0x01,
-	data1       = 0x0B,
+	data1          = 0x0B,
-	data2       = 0x05,
+	data2          = 0x05,
-	data4       = 0x06,
+	data4          = 0x06,
-	string      = 0x08,
+	string         = 0x08,
-	strp        = 0x0E,
+	strp           = 0x0E,
-	exprloc     = 0x18,
+	exprloc        = 0x18,
-	ref4        = 0x13,
+	ref4           = 0x13,
-	udata       = 0x0F,
+	udata          = 0x0F,
-	ref_sig8    = 0x20,
+	ref_sig8       = 0x20,
 	implicit_const = 0x21,
-	flag_present = 0x19,
+	flag_present   = 0x19,
-	sec_offset  = 0x17,
+	sec_offset     = 0x17,
 }
 
 M.DW_ATE = {
@@ -104,34 +104,16 @@ M.MIPS_BYTES_PER_WORD = 0x04
 -- ----------------------------------------------------------------------------
 -- ELF32 (System V ABI gABI v1.2)
 -- ----------------------------------------------------------------------------
---- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors,
+--- **Wire-offset contract:** format offsets, fixed-width reader offsets, LEB/parser cursors, and section-relative values are zero-based wire offsets. 
---- and section-relative values are zero-based wire offsets. Only Lua string APIs receive
+--- Only Lua string APIs receive a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
---- a `+ 1` conversion at their boundary (`byte`, `sub`, and `find`).
+--- ELF/DWARF field offsets are expressed in hex so they map directly to the zero-based byte positions in the binary file.
 ---
---- ELF/DWARF field offsets are expressed in hex so they map directly to the
+--- The ELF32 header / section / sym layout tables are within scripts/elf32.lua.
---- zero-based byte positions in the binary file.
+--- The metaprogram re-exports the DWARF32 initial-length terminator.
 
-
+--- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
---- spec: System V ABI gABI v1.2 §"ELF Header" (Table 1) + §"Section Header Table"
+M.dw_dwarf32_terminator = E.dw_dwarf32_terminator
-M.ELF32 = {
+-- TODO(Ed): Remove re-export.
-	magic_offset          = 0x00,         -- 4-byte magic "\127ELF" at file offset 0x00
-	magic                 = "\127ELF",
-	class_offset          = 0x04,         -- 1-byte; 1 = ELF32, 2 = ELF64
-	class_elf32           = 1,
-	endian_offset         = 0x05,         -- 1-byte; 1 = little-endian, 2 = big-endian
-	endian_little         = 1,
-	header_bytes          = 0x34,         -- spec: gABI v1.2 §"ELF Header" — ELF32 header is 52 bytes total
-	e_shoff_offset        = 0x20,         -- 4-byte LE; section-header table file offset
-	e_shentsize_offset    = 0x2E,         -- 2-byte LE; section-header entry size in bytes
-	e_shnum_offset        = 0x30,         -- 2-byte LE; number of section headers
-	e_shstrndx_offset     = 0x32,         -- 2-byte LE; index of section-name string table
-	sh_size_bytes         = 0x28,         -- spec: gABI v1.2 §"Section Header Table" — each entry is 40 bytes
-	sh_name_offset        = 0x00,         -- 4-byte LE; offset into .shstrtab
-	sh_type_offset        = 0x04,         -- 4-byte LE; section type (SHT_*)
-	sh_offset_offset      = 0x10,         -- 4-byte LE; section's file offset
-	sh_size_offset        = 0x14,         -- 4-byte LE; section's size in bytes
-	dw_dwarf32_terminator = 0xFFFFFFFF,   -- spec: DWARF4 spec §7.4 — 32-bit DWARF initial-length terminator
-}
 
 -- ----------------------------------------------------------------------------
 -- DWARF4 .debug_aranges (per DWARF5 spec §7.4 — Address Range Table)
@@ -192,8 +174,7 @@ M.DWARF_LINE_OPS = {
 	DW_LNE_end_sequence = 1,    -- spec: §6.2.5.3
 	DW_LNE_set_address  = 2,    -- spec: §6.2.5.3
 	-- Standard opcode header (§6.2.5.1)
-	-- opcode_base + line_range are 1-byte header fields; hex so they map
+	-- opcode_base + line_range are 1-byte header fields; hex so they map directly to the line-program header byte sequence.
-	-- directly to the line-program header byte sequence.
 	-- line_base stays signed decimal (=-5) since 0xFB obscures the spec semantics.
 	opcode_base              = 0x0D,
 	line_base                = -5,
@@ -249,41 +230,35 @@ M.DWARF5_DEBUG_LINE = {
 --- Read a 4-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- Equivalent to `string.unpack("<I4", buf, off + 1)` but avoids the table-return shape + works under LuaJIT 2.1
 --- (which has partial `string.unpack` coverage).
----
 --- **Convention:** `off` is a zero-based wire offset; `+ 1` is applied only at the `string.byte` boundary.
 ---
---- **Byte weights** are written as `0x100`, `0x10000`, `0x1000000` (i.e. 2^8, 2^16, 2^24) so the LE byte positions are visually explicit:
+--- Thin forwarder: the canonical implementation lives in scripts/elf32.lua.
---- byte 0 contributes its value directly; byte 1 is shifted left by 8
+--- The "second caller lifts" pattern keeps the metaprogram side fluent
---- (= 0x100); byte 2 by 16 (= 0x10000); byte 3 by 24 (= 0x1000000).
+--- (`M.read_u32_le(buf, off)`) while the body is deduped.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u32_le(buf, off)
-	local byte_off = off + 1
+	return E.read_u32_le(buf, off)
-	return buf:byte(byte_off)
-		+ buf:byte(byte_off + 0x01) * 0x00000100
-		+ buf:byte(byte_off + 0x02) * 0x00010000
-		+ buf:byte(byte_off + 0x03) * 0x01000000
 end
 
 --- Read a 2-byte little-endian unsigned integer from `buf` at zero-based wire offset `off`.
 --- (`off` is zero-based; `+ 1` is applied only at the `string.byte` boundary.)
+--- Thin forwarder — see `M.read_u32_le` for the rationale.
 --- @param buf string
 --- @param off integer  -- zero-based wire offset
 --- @return integer
 function M.read_u16_le(buf, off)
-	local byte_off = off + 1
+	return E.read_u16_le(buf, off)
-	return buf:byte(byte_off) + buf:byte(byte_off + 0x01) * 0x00000100
 end
 
 -- Pure-Lua 5.3 LEB128 readers (no `bit` library). `2^shift` arithmetic matches the existing parser.
 -- Offsets are 0-based; returns (value, next_pos).
--- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua
+-- Promoted from `local function` to M.* exports so passes/dwarf_injection.lua can import them as file-scope locals per the 2nd-caller lift precedent
--- can import them as file-scope locals per the 2nd-caller lift precedent
 -- (the uleb128 + sleb128 encoders were promoted the same way).
 function M.read_uleb128_at(buf, pos)
 	local value, shift = 0, 0
-	local len = #buf
+	local len          = #buf
 	while pos < len do
 		local b = buf:byte(pos + 1)
 		value = value + (b % 0x80) * (2 ^ shift)
@@ -427,13 +402,13 @@ local function read_form_value(buf, str_buf, pos, form)
 		-- The constant is declared in the abbrev; no value bytes in the DIE.
 		return nil, pos
 	elseif form == M.DW_FORM.ref_sig8 then
-		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): an 8-byte value identifying a type
+		-- DW_FORM_ref_sig8 (DWARF5 §7.4.2): An 8-byte value identifying a type by signature.
-		-- by signature. The low 4 bytes (LE) are the type signature (content hash);
+		-- The low  4 bytes (LE) are the type signature (content hash);
-		-- the high 4 bytes (LE) are a CU-relative offset into the matching type unit.
+		-- The high 4 bytes (LE) are a CU-relative offset into the matching type unit.
-		-- Consumers use the low 4 to look up the type unit (see M.find_type_unit_by_signature)
+		-- 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.
+		--          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;
+		-- Return the low  4 as the primary value to preserve the (value, next_pos) shape;
-		-- the high 4 is exposed via M.read_ref_sig8 (which returns both halves).
+		--        the high 4 is exposed via M.read_ref_sig8 (which returns both halves).
 		local _, _, next_pos = M.read_ref_sig8(buf, pos)
 		return M.read_u32_le(buf, pos), next_pos
 	else
@@ -455,26 +430,26 @@ function M.read_ref_sig8(buf, pos)
 	return M.read_u32_le(buf, pos), M.read_u32_le(buf, pos + 4), pos + 8
 end
 
--- DWARF5 §7.5.6 (Type Entries).
+--- DWARF5 §7.5.6 (Type Entries).
--- Walk all units in `info` and return the 0-based offset of the first unit
+--- Walk all units in `info` and return the 0-based offset of the first unit whose `DW_AT_type_signature`
--- whose `DW_AT_type_signature` (8-byte value at the end of the unit header) equals `target_sig`.
+--- (8-byte value at the end of the unit header) equals `target_sig`.
--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
+--- The signature is interpreted as two 32-bit halves (low/high) per the read_ref_sig8 contract;
--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
+--- we match both halves (i.e. the 8-byte value as a whole). Returns nil if no matching unit exists.
---
+---
--- Unit header layout (from pos 0):
+--- Unit header layout (from pos 0):
---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
+---   unit_length(4) + version(2) + unit_type(1) + address_size(1) + debug_abbrev_offset(4)
---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
+---   followed by type_unit_specific fields: type_signature(8) + type_offset(4)
--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
+--- The type_signature is at byte offset 8 of the body (right after debug_abbrev_offset).
--- @param info string  -- the .debug_info section bytes
+--- @param info          string   -- the .debug_info section bytes
--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
+--- @param target_sig_lo integer  -- low 4 bytes (LE) of the desired signature
--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
+--- @param target_sig_hi integer  -- high 4 bytes (LE) of the desired signature
--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
+--- @return integer|nil, integer|nil  -- unit offset, type_offset within the unit
 function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
-	local pos = 0
+	local pos         = 0
 	local section_len = #info
 	while pos + 4 < section_len do
 		local unit_length = M.read_u32_le(info, pos)
-		if unit_length == 0xFFFFFFFF then
+		if    unit_length == 0xFFFFFFFF then
 			return nil, nil  -- DWARF64 not supported
 		end
 		-- unit_length is the body size, NOT including the 4-byte unit_length field itself.
@@ -503,9 +478,9 @@ function M.find_type_unit_by_signature(info, target_sig_lo, target_sig_hi)
 			--   byte 8-15: type_signature (8)
 			--   byte 16-19: type_offset (4)
 			local unit_type = info:byte(body_start + 2 + 1)  -- 0-based +2 = unit_type in 1-indexed
-			if unit_type == 0x02 then  -- DW_UT_type
+			if    unit_type == 0x02 then  -- DW_UT_type
 				local sig_lo, sig_hi, _ = M.read_ref_sig8(info, body_start + 8)  -- 0-based +8 = type_signature in 1-indexed
-				if sig_lo == target_sig_lo and sig_hi == target_sig_hi then
+				if    sig_lo == target_sig_lo and sig_hi == target_sig_hi then
 					local type_offset = M.read_u32_le(info, body_start + 16)  -- 0-based +16 = type_offset in 1-indexed
 					return pos, type_offset
 				end
@@ -552,7 +527,7 @@ end
 --- (we walk all `e_shnum` headers regardless of how many names are requested, to find the .shstrtab first).
 --- For frequent callers, pass the union of all needed sections in one call.
 --  Can add `.debug_info` + `.debug_loc` + `.debug_str_offsets` to the list without writing a 2nd ELF walker.
---- @param elf_path Path
+--- @param elf_path      Path
 --- @param section_names string[]  -- list of section names to read
 --- @return table<string, string>
 function M.read_elf_sections(elf_path, section_names)
@@ -571,75 +546,64 @@ function M.read_elf_sections(elf_path, section_names)
 		return result
 	end
 
-	local f = io.open(elf_path, "rb")
+	local  f = io.open(elf_path, "rb")
 	if not f then
 		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] io.open failed: %s\n", elf_path))
 		return result
 	end
 
-	-- Read the ELF32 header.
+	local file_size
-	local header = f:read(M.ELF32.header_bytes)
+	do
-	if not header or #header < M.ELF32.header_bytes then
+		f:seek("end", 0)
-		io.stderr:write("[elf_dwarf.read_elf_sections] ELF too small for ELF32 header\n")
+		file_size = f:seek("cur", 0)
+	end
+	local adapter = {
+		read_u8_at = function(offset)
+			f:seek("set", offset)
+			local b = f:read(1)
+			if not b then return nil end
+			return b:byte()
+		end,
+		read_u16_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			if not b1 or not b2 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+		end,
+		read_u32_at = function(offset)
+			f:seek("set", offset)
+			local b1 = f:read(1)
+			local b2 = f:read(1)
+			local b3 = f:read(1)
+			local b4 = f:read(1)
+			if not b1 or not b2 or not b3 or not b4 then return nil end
+			return b1:byte() + b2:byte() * 0x100
+				+ b3:byte() * 0x10000 + b4:byte() * 0x1000000
+		end,
+		read_size = function() return file_size end,
+	}
+
+	-- Delegate the header parse + section walk to E.*.
+	local hdr, hdr_err = E.parse_elf32_headers(adapter)
+	if not hdr then
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] header parse failed: %s\n", tostring(hdr_err)))
 		f:close()
 		return result
 	end
 
-	-- Sanity-check magic + class + endianness.
+	local sections, walk_err = E.walk_sections(adapter, hdr)
-	if header:sub(M.ELF32.magic_offset + 1, M.ELF32.magic_offset + 0x04) ~= M.ELF32.magic then
+	if not sections then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not an ELF file\n")
+		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] section walk failed: %s\n", tostring(walk_err)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.class_offset + 1) ~= M.ELF32.class_elf32 then
-		io.stderr:write(string.format("[elf_dwarf.read_elf_sections] not ELF32 (class=%d)\n", header:byte(M.ELF32.class_offset + 1)))
-		f:close()
-		return result
-	end
-	if header:byte(M.ELF32.endian_offset + 1) ~= M.ELF32.endian_little then
-		io.stderr:write("[elf_dwarf.read_elf_sections] not little-endian; unsupported\n")
 		f:close()
 		return result
 	end
 
-	-- Parse section-header table location + dimensions from the header.
+	-- Resolve the requested sections.
-	local e_shoff     = M.read_u32_le(header, M.ELF32.e_shoff_offset)
+	for _, s in ipairs(sections) do
-	local e_shentsize = M.read_u16_le(header, M.ELF32.e_shentsize_offset)
+		if wanted[s.name] then
-	local e_shnum     = M.read_u16_le(header, M.ELF32.e_shnum_offset)
+			local bytes = E.read_section_bytes(adapter, s)
-	local e_shstrndx  = M.read_u16_le(header, M.ELF32.e_shstrndx_offset)
+			if bytes then result[s.name] = bytes end
-
-	-- Read the section-header string table (.shstrtab) so we can resolve section names from their `sh_name` offsets.
-	f:seek("set", e_shoff + e_shstrndx * e_shentsize)
-	local strtab_hdr = f:read(e_shentsize)
-	if not strtab_hdr or #strtab_hdr < e_shentsize then
-		io.stderr:write("[elf_dwarf.read_elf_sections] could not read .shstrtab header\n")
-		f:close()
-		return result
-	end
-	local strtab_offset = M.read_u32_le(strtab_hdr, M.ELF32.sh_offset_offset)
-	local strtab_size   = M.read_u32_le(strtab_hdr, M.ELF32.sh_size_offset)
-	f:seek("set", strtab_offset)
-	local strtab = f:read(strtab_size) or ""
-
-	-- Walk all section headers; collect (offset, size) for the wanted names.
-	local function read_section_bytes(sh_offset, sh_size)
-		f:seek("set", sh_offset)
-		return f:read(sh_size) or ""
-	end
-
-	for sh_idx = 0, e_shnum - 1 do
-		f:seek("set", e_shoff + sh_idx * e_shentsize)
-		local sh = f:read(e_shentsize)
-		if not sh or #sh < e_shentsize then break end
-		local sh_name   = M.read_u32_le(sh, M.ELF32.sh_name_offset)
-		local sh_offset = M.read_u32_le(sh, M.ELF32.sh_offset_offset)
-		local sh_size   = M.read_u32_le(sh, M.ELF32.sh_size_offset)
-
-		-- Extract the name (null-terminated C string in strtab).
-		local name_end = strtab:find("\0", sh_name + 1, true) or (sh_name + 1)
-		local name     = strtab:sub(sh_name + 1, name_end - 1)
-		if wanted[name] then
-			result[name] = read_section_bytes(sh_offset, sh_size)
 		end
 	end
 
@@ -656,49 +620,87 @@ end
 --- - We filter on STB_GLOBAL (high nibble of st_info = 1) to match `nm`'s default (external symbols only). STB_WEAK excluded.
 --- - The `code_` prefix is stripped (MipsAtom_ macros emit bare atom names, no `code_` prefix).
 --- - `st_size > 0` filter excludes undefined/imported symbols.
+---
 --- @param elf_path Path
 --- @return table<string, {integer, integer}>
 function M.read_nm(elf_path)
 	local addrs = {}
 
-	-- Read .symtab + .strtab via the existing ELF walker (no subprocess).
+	-- Existence check first; an empty or missing ELF returns an empty map.
-	local sections = M.read_elf_sections(elf_path, {".symtab", ".strtab"})
+	if lfs.attributes(elf_path, "mode") ~= "file" then
-	local symtab   = sections[".symtab"]
-	local strtab   = sections[".strtab"]
-	if not symtab or not strtab or #symtab == 0 or #strtab == 0 then
-		-- No symbol table (e.g. stripped ELF). Return empty.
 		return addrs
 	end
 
-	-- Iterate the 16-byte ELF32 symtab entries.
+	local f = io.open(elf_path, "rb")
-	-- Each entry (zero-based): st_name at 0, st_value at 4, st_size at 8, st_info at 12, st_other at 13, st_shndx at 14.
+	if not f then
-	local SYM_ENTRY_BYTES = 0x10
+		return addrs
-	local SYM_ST_NAME     = 0x00
+	end
-	local SYM_ST_VALUE    = 0x04
+
-	local SYM_ST_SIZE     = 0x08
+	-- Build the file adapter for E.*.
-	local SYM_ST_INFO     = 0x0C
+	local file_size
-	local n_syms = #symtab / SYM_ENTRY_BYTES
+	do
-	for i = 0, n_syms - 1 do
+		f:seek("end", 0)
-		local entry_off  = i * SYM_ENTRY_BYTES
+		file_size = f:seek("cur", 0)
-		local st_info    = symtab:byte(entry_off + SYM_ST_INFO + 1)
+	end
-		-- High nibble = binding (STB_LOCAL=0, STB_GLOBAL=1, STB_WEAK=2).
+	local adapter = {
-		-- Use math.floor(/16) instead of bit.rshift for LuaJIT 2.1 compat
+		read_u8_at = function(offset)
-		-- (LuaJIT's `>>` is 5.3+, but math.floor(x/16) works on all versions).
+			f:seek("set", offset)
-		local binding    = math.floor(st_info / 16)
+			local b = f:read(1)
-		if binding == 0 or binding == 1 then            -- STB_LOCAL or STB_GLOBAL
+			if not b then return nil end
-			local st_size   = M.read_u32_le(symtab, entry_off + SYM_ST_SIZE)
+			return b:byte()
-			if st_size > 0 then
+		end,
-				local st_name_off = M.read_u32_le(symtab, entry_off + SYM_ST_NAME)
+		read_u16_at = function(offset)
-				-- Extract the name from .strtab (null-terminated C string).
+			f:seek("set", offset)
-				local name_end = strtab:find("\0", st_name_off + 1, true) or (st_name_off + 1)
+			local b1 = f:read(1)
-				local name     = strtab:sub(st_name_off + 1, name_end - 1)
+			local b2 = f:read(1)
-				-- Filter: keep all symbol-table symbols (atoms emit their name as the bare `<name>` — MipsAtom_ macros strip the `code_` prefix).
+			if not b1 or not b2 then return nil end
-				-- The atoms_source_map pass already filters out non-atom symbols via the source-map.txt cross-ref.
+			return b1:byte() + b2:byte() * 0x100
-				if name and #name > 0 then
+		end,
-					local st_value = M.read_u32_le(symtab, entry_off + SYM_ST_VALUE)
+		read_u32_at = function(offset)
-					addrs[name] = { st_value, st_size }
+			f:seek("set", offset)
-				end
+			local b1 = f:read(1)
-			end
+			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
 
@@ -779,8 +781,8 @@ function M.sleb128(n)
 		local b = n % (LEB_DATA_MASK + 1) -- extract low 7 bits
 		n = (n - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
 		-- Termination: remaining value bits fit in the sign bit of the last byte.
-		if     n == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if n == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
-		if     n == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if n == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
 		if more then b = b + LEB_CONT_BIT end
 		bytes[#bytes + 1] = string.char(b)
 	end
@@ -811,14 +813,14 @@ end
 --- @param n integer  -- any integer (negative allowed)
 --- @return integer
 function M.sleb128_size(n)
-	local more = true
+	local more  = true
 	local bytes = 0
-	local v = n
+	local v     = n
 	while more do
 		local b = v % (LEB_DATA_MASK + 1) -- extract low 7 bits
 		v = (v - b) / (LEB_DATA_MASK + 1) -- arithmetic shift right by 7
-		if     v == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
+		if v == 0  and b <  SLEB_SIGN_BIT then more = false end  -- positive terminator
-		if     v == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
+		if v == -1 and b >= SLEB_SIGN_BIT then more = false end  -- negative terminator
 		if more then b = b + LEB_CONT_BIT end
 		bytes = bytes + 1
 	end
@@ -836,12 +838,11 @@ end
 ---   * The `.debug_line` section may contain MULTIPLE line-program units
 ---     File indices are 1-based, **per unit**; we concatenate all units and the index ranges from 1..N₁ in unit 1, N₁+1..N₁+N₂ in unit 2, etc.
 ---     Per-unit indices (the way gcc emits them, and the way `DW_LNS_set_file` references them in the line program)
----     are returned via the `basename_to_index` map only when the unit boundary happens to align with the metaprogram's per-atom
+---     are returned via the `basename_to_index` map only when the unit boundary happens to align with the metaprogram's per-atom `inv.call_file`
----     `inv.call_file` (true today for hello_joypad — the C unit is the LAST unit, and atom-side file indices fit 1-based).
 ---   * Per spec, the `.debug_line_str` section (DWARF5 §7.5.6) holds the strings referenced by `DW_FORM_line_strp`.
 ---     The legacy DWARF3 format embeds strings directly with null terminators. This helper handles BOTH.
----   * File entries may have multiple forms (gcc -gdwarf-5 with `DW_LNCT_directory_index`
+---   * File entries may have multiple forms (gcc -gdwarf-5 with `DW_LNCT_directory_index` emits 2 forms: path + dir_index). 
----     emits 2 forms: path + dir_index). The helper supports:
+---     The helper supports:
 ---     - DW_FORM_line_strp (DWARF5; offset into .debug_line_str)
 ---     - DW_FORM_string (DWARF4-compat; inline null-terminated in .debug_line)
 ---     - DW_FORM_udata (ULEB128)
@@ -851,51 +852,47 @@ end
 ---
 --- Behavior on failure: writes to stderr and returns nil.
 --- Helpers consumed by `passes/dwarf_injection.lua::init_file_index_lookup(elf_path)` calls this once at pass start to populate the module-level `basename_to_index` map;
---- downstream `resolve_provenance_file_index(path)` consumers
+--- downstream `resolve_provenance_file_index(path)` consumers consult the map directly.
---- (which replaced the former hardcoded `ATOM_SOURCE_FILE_INDEX` + `PROVENANCE_BASENAME_TO_FILE_INDEX` table per `conductor/tracks/dwarf_file_index_lookup_20260731/`)
---- consult the map directly.
 ---
---- @param elf_path string  -- absolute path to the post-link ELF (typically the gcc-emitted `.elf` BEFORE dwarf_injector's splice;
+--- @param elf_path string  -- absolute path to the post-link ELF (typically the gcc-emitted `.elf` BEFORE dwarf_injector's splice; both shapes work since the splice preserves `.debug_line`)
----                            both shapes work since the splice preserves `.debug_line`)
 --- @return table|nil, table|nil, table|nil
----        basename_to_index: { [basename] = 1-based-per-unit-file-index, ... }
+---   basename_to_index: { [basename] = 1-based-per-unit-file-index, ... }
----        basenames:         { [1-based-per-unit-file-index] = basename, ... }
+---   basenames:         { [1-based-per-unit-file-index] = basename, ... }
----        paths:             { [1-based-per-unit-file-index] = full path (mixed slashes), ... }
+---   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 line     = sections[".debug_line"]
-	local lstr      = sections[".debug_line_str"] or ""
+	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 basenames         = {}
 	local basename_to_index = {}
-	local paths            = {}
+	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)
+	---   * the ULEB128 number  (DW_FORM_udata)
-	---   * nil + skip-bytes (DW_FORM_data16; we don't surface the MD5)
+	---   * 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 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)
+			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, 
+			-- We do NOT stderr-write because the crt0.s DWARF5 line unit (gcc-as emitted) uses DW_FORM_addr (0x01) for what is effectively a path entry, which is non-standard.
-			-- which is non-standard.
 			-- The C-unit's DWARF3 paths are read via the parallel DWARF3 path and never see this error.
 			-- Callers should consult `basename_to_index` for the paths they care about and ignore this unit if it produced none.
 			return nil, p
@@ -916,9 +913,9 @@ function M.read_line_unit_file_table(elf_path)
 		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
+			if    nul > body_end then break end
 			local len = nul - up - 1
-			if len == 0 then up = nul break end
+			if    len == 0 then up = nul break end
 			dirs[#dirs + 1] = buf:sub(up + 1, nul - 1)
 			up = nul
 		end
@@ -926,14 +923,14 @@ function M.read_line_unit_file_table(elf_path)
 		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
+			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 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 _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 _size, up_next3 = M.read_uleb128_at(buf, up); up = up_next3
 			local idx = #unit_basenames + 1
-			local bs   = path:match("[^/\\]+$") or path
+			local bs  = path:match("[^/\\]+$") or path
 			unit_paths[idx]     = path
 			unit_basenames[idx] = bs
 			dirs[1] = dirs[1] or ""  -- safety: gcc emits "" sentinel dir at 0
@@ -1006,7 +1003,7 @@ function M.read_line_unit_file_table(elf_path)
 	local section_end = #line
 	while p + 4 <= section_end do
 		local unit_length = M.read_u32_le(line, p)
-		if unit_length == 0xFFFFFFFF then
+		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

scripts/passes/annotation.lua

@@ -22,11 +22,11 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -45,28 +45,28 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 --- @field warnings table[]
 
 --- @class AtomAnnotation
---- @field line    integer       -- source line of the atom_info call
+--- @field line    integer       -- Source line of the atom_info call
---- @field macro   string        -- the macro name (always "atom_info" in the new shape)
+--- @field macro   string        -- Macro name (always "atom_info" in the new shape)
---- @field name    string        -- the atom name
+--- @field name    string        -- Atom name
---- @field kind    string        -- always "info"
+--- @field kind    string        -- Always "info"
 --- @field binds   string|nil    -- Binds_X name if any
 --- @field reads   string[]      -- R_* names (read targets)
 --- @field writes  string[]      -- R_* names (write targets)
---- @field errors  string[]|nil  -- parse-time errors from scan_source (atom_info body malformed)
+--- @field errors  string[]|nil  -- Parse-time errors from scan_source (atom_info body malformed)
 
---- @class DebugSkipMarker  -- sub-shape of scan_source.lua's @class DebugSkipMarker
+--- @class DebugSkipMarker  -- Sub-shape of scan_source.lua's @class DebugSkipMarker
---- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
+--- @field marker_kind    string                 -- Exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
 --- @field marker_line    integer
---- @field args           string|nil             -- trimmed text inside the parens (nil when has_parens is false)
+--- @field args           string|nil             -- Trimmed text inside the parens (nil when has_parens is false)
 --- @field has_parens     boolean
 --- @field is_bare        boolean                -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
 --- @field pending        boolean                -- true while awaiting the following declaration
---- @field superseded_by_marker_line integer|nil -- set on a marker that was bumped out of the pending slot
+--- @field superseded_by_marker_line integer|nil -- Set on a marker that was bumped out of the pending slot
 --- @field target_kind    string|nil             -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
 
 --- @class Finding
---- @field line integer -- source line (or 0 for pass-level)
+--- @field line integer -- Source line (or 0 for pass-level)
---- @field msg  string  -- finding message
+--- @field msg  string  -- Finding message
 
 --- @class Findings
 --- @field errors   Finding[]
@@ -74,14 +74,14 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 --- @field info     Finding[]
 
 --- @class PipeCtx
---- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
+--- @field atom_index     table<string, AtomAnnotation> -- Name -> AtomAnnotation (only kind=="atom")
---- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
+--- @field binds_index    table<string, BindsStruct>    -- Name -> BindsStruct
---- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
+--- @field annot_counts   table<string, integer>        -- Name -> annotation count (for unique_annotation check)
---- @field types          table<string, RegTypeDefault> -- from scan_source
+--- @field types          table<string, RegTypeDefault> -- From scan_source
---- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- From scan_source
---- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field seen_defaults  table<string, integer>        -- Duplicate atom_dbg_reg_default detection
 --- @field seen_field     table<string, integer>        -- Binds_* -> count of fields (set/checked by check_binds_no_duplicate_fields)
---- @field _scan          SourceScan                    -- full scan payload (typed-view sub-calls live here) 
+--- @field _scan          SourceScan                    -- Full scan payload (typed-view sub-calls live here) 
 
 --- @class AnnotatedResult
 --- @field atoms    AtomEntry[]
@@ -95,11 +95,10 @@ local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-check functions (the CHECK_RULES table's payload)
 -- ════════════════════════════════════════════════════════════════════════════
---
 --- The dispatcher in `validate()` routes each result by convention: existence checks write errors[] and shape checks write warnings[].
 --- `macro_word_drift` writes errors[] for missing or mismatched metadata and info[] for a match.
 
---- Check: every annotated atom must have a matching MipsAtom_(name) declaration.
+--- Check: Every annotated atom must have a matching MipsAtom_(name) declaration.
 --- @param a        AtomAnnotation
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
@@ -112,8 +111,8 @@ local function check_atom_decl_exists(a, pipe_ctx, findings)
 	end
 end
 
---- Check: every atom may have AT MOST ONE annotation.
+--- Check:     Every atom may have AT MOST ONE annotation.
---- Post-loop: needs full-corpus `annot_counts` from pipe_ctx.
+--- Post-loop: Needs full-corpus `annot_counts` from pipe_ctx.
 --- @param pipe_ctx PipeCtx
 --- @param findings Findings
 local function check_unique_annotation(pipe_ctx, findings)
@@ -146,7 +145,7 @@ end
 --- Check: TAPE_WORDS(mac_X, N) ↔ WORD_COUNT(mac_X, N) drift.
 --- Three outcomes: missing (error), mismatch (error), match (info).
 --- @param m  MacroEntry
---- @param wc table<string, integer> -- the shared word-count table (from ctx.shared.word_counts)
+--- @param wc table<string, integer> -- Shared word-count table (from ctx.shared.word_counts)
 --- @param findings Findings
 local function check_macro_word_drift(m, wc, findings)
 	local  declared = wc[m.name]
@@ -304,7 +303,7 @@ local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
 	end
 end
 
--- Check: debug-skip markers must satisfy shape + placement constraints.
+-- Check: Debug-skip markers must satisfy shape + placement constraints.
 --- Walks the priority list once; each marker produces at most one error, so one source defect yields one finding.
 --- Priority order (first defect wins):
 ---   1. marker_kind ~= "atom_dbg_skip" -> legacy/renamed spelling (use `atom_dbg_skip`)
@@ -315,12 +314,11 @@ end
 ---   6. unsupported target_kind     -> marker precedes an unrelated declaration
 --- Valid markers stamp `debug_skip` on whole-atom, bare-component, and proc-component declaration records in scan_source.lua.
 --- @param marker DebugSkipMarker
---- @param _pipe_ctx PipeCtx     -- unused today; kept for plex-shape consistency with per_annot
+--- @param _pipe_ctx PipeCtx     -- Unused; kept for consistency with per_annot // TODO(Ed): Remove?
 --- @param findings Findings
 local function check_skip_marker(marker, _pipe_ctx, findings)
 	local kind = marker.marker_kind
 	local line = marker.marker_line
-
 	-- Left `scan.debug_skip_markers` with production records for `atom_dbg_skip` only; other identifiers take the walker's unrelated branch.
 
 	if marker.has_parens then
@@ -371,8 +369,6 @@ local function check_skip_marker(marker, _pipe_ctx, findings)
 end
 
 --- Warn when a source references an unregistered alias.
----
---- R_TapePtr, R_AtomJmp, R_PrimCursor, R_FaceCursor, R_VertBase, and R_OtBase opt in through `#define atom_reg` in lottes_tape.h.
 --- When a source uses an unregistered R_X, this check emits one pass-level info entry for that source and directs C-ABI register names to explicit alias registration.
 --- @param _src     SourceFile
 --- @param pipe_ctx PipeCtx
@@ -426,8 +422,7 @@ local CHECK_RULES = {
 -- ════════════════════════════════════════════════════════════════════════════
 -- Validation
 -- ════════════════════════════════════════════════════════════════════════════
---
+-- Pure check: Read from src.scan, run validations, emit findings. The scan was done once upstream.
--- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
 
 --- Builds one pass-wide pipe_ctx from the merged `corpus.*` registries and source-ordered `corpus.atom_infos`; per-source declarations and bodies remain in `src.scan`.
 --- The module ownership contract above requires callers to construct `ctx.shared.corpus` through `build_ctx`; the error message below enforces that gate.
@@ -473,7 +468,7 @@ end
 --- Validate one source against its pre-scanned SourceScan payload + the corpus-wide pipe_ctx.
 --- @param ctx             PassCtx
 --- @param src             SourceFile
---- @param corpus_pipe_ctx PipeCtx|nil  -- built once per pass from corpus registries; nil builds the same projection here.
+--- @param corpus_pipe_ctx PipeCtx|nil  -- Built once per pass from corpus registries; nil builds the same projection here.
 --- @return AnnotatedResult
 local function validate(ctx, src, corpus_pipe_ctx)
 	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
@@ -503,14 +498,8 @@ local function validate(ctx, src, corpus_pipe_ctx)
 	end
 
 	-- Build a per-source pipe_ctx: shared lookups come from `corpus_pipe_ctx`, while declarations, bodies, types, views, defaults, and occurrences come from `src.scan`.
-	local seen_defaults = {}
+	local seen_defaults   = {}; for reg, _ in pairs (scan.types      or {}) do seen_defaults[reg] = (seen_defaults[reg] or 0) + 1 end
-	for reg, _ in pairs(scan.types or {}) do
+	local atom_infos_list = {}; for _, ai  in ipairs(scan.atom_infos or {}) do atom_infos_list[#atom_infos_list + 1] = ai         end
-		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
-	end
-	local atom_infos_list = {}
-	for _, ai in ipairs(scan.atom_infos or {}) do
-		atom_infos_list[#atom_infos_list + 1] = ai
-	end
 
 	local pipe_ctx = {
 		atom_index               = {},

scripts/passes/atoms_source_map.lua

@@ -1,8 +1,8 @@
 --- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
 ---
 --- Writer: this pass, given `atom.paths` (the per-atom mutable surface owned by `emission_model`). Readers:
---- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and the gdb-runtime
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and 
---- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
+--- the gdb-runtime wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
 ---
 --- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: 
 ---  one `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
@@ -28,7 +28,6 @@
 ---   ...
 ---   ENDATOM
 ---   ```
----
 --- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -39,8 +38,8 @@
 -- (works both standalone + when require'd). `duffle_paths.lua` sets package.path then returns `require("duffle")` 
 -- at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle    = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local elf_dwarf = require("elf_dwarf")
+local elf_dwarf      = require("elf_dwarf")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -69,8 +68,8 @@ local FORMAT_VERSION = 1
 --- @param atom table
 --- @return table[], integer
 local function canonical_word_entries(atom)
-	local paths = atom.paths or {}
+	local paths      = atom.paths or {}
-	local events = paths.word_events or {}
+	local events     = paths.word_events or {}
 	local word_items = {}
 	for _, item in ipairs(paths.items or {}) do
 		if item.kind == "word" then word_items[#word_items + 1] = item end
@@ -95,41 +94,38 @@ end
 --- Render one atom's provenance stanza. Format 1 line shapes:
 ---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>" BODY <line>` (component invocation)
 ---   `WORD N  CALL <src-path>:<src-line>  RAW` (raw `.word` outside any mac_* component)
---- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was declared in `corpus.word_counts` 
---- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
+--- (populated by word_count_eval + components passes).
---- @param src table
+--- @param src  table
 --- @param atom table
---- @param wc table -- identity alias of corpus.word_counts
+--- @param wc   table -- identity alias of corpus.word_counts
 --- @return string[], integer
 local function emit_provenance_stanza(src, atom, wc)
-	local lines    = {}
+	local lines          = {}
-	local rel_path = src.path:gsub("\\\\", "/")
+	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 inv         = entry.invocation
 		local macro_count = inv and wc["mac_" .. inv.component_name]
 		if inv and macro_count ~= nil then
-			lines[#lines + 1] = string.format(
+			lines[#lines + 1] = string.format('WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d'
-				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				, entry.pos, rel_path, entry.line, inv.component_name
-				entry.pos, rel_path, entry.line, inv.component_name,
+				, inv.def_path or "", inv.def_line or 0, entry.body_line)
-				inv.def_path or "", inv.def_line or 0, entry.body_line)
 		else
-			lines[#lines + 1] = string.format(
+			lines[#lines + 1] = string.format("WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
-				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
 		end
 	end
 
-	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
+	lines[1]          = lines[1]:gsub(" 0$", " " .. tostring(total))
 	lines[#lines + 1] = "ENDATOM"
 	return lines, total
 end
 
 --- Render the full provenance file content for one source.
 --- @param src table
---- @param wc table
+--- @param wc  table
 --- @return string
 local function render_provenance(src, wc)
 	local lines = {}
@@ -162,8 +158,8 @@ end
 --- @param wc     table
 --- @return string[], integer
 local function emit_atom_stanza(src, atom)
-	local lines    = {}
+	local lines          = {}
-	local rel_path = src.path:gsub("\\\\", "/")
+	local rel_path       = src.path:gsub("\\\\", "/")
 	local entries, total = canonical_word_entries(atom)
 
 	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
@@ -179,10 +175,10 @@ end
 --- Render the full source map file content for one source (one .atoms.sourcemap.txt per source). Mirrors offsets.lua's
 --- `project_atoms` shape: scan.atoms + scan.raw_atoms, no kind filter.
 --- @param src table
---- @param wc table
+--- @param wc  table
 --- @return string
 local function render_source_map(src)
-	local lines = {}
+	local lines       = {}
 	lines[#lines + 1] = "# FORMAT_VERSION " .. FORMAT_VERSION
 	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
 
@@ -216,16 +212,16 @@ end
 --- @param ctx PassCtx
 --- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
 local function build_atom_table(ctx)
-	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local addrs   = elf_dwarf.read_nm(ctx.flags.elf_path)
-	local corpus = ctx.shared and ctx.shared.corpus
+	local corpus  = ctx.shared and ctx.shared.corpus
 	local matched = {}
 
 	for _, src in ipairs(corpus.source_order or {}) do
 		local file_base = src.path:match("([^/\\\\]+)$") or src.path
 		local function append(atom)
 			if not atom.paths then return end
-			local name = atom.raw_name or atom.name
+			local  name = atom.raw_name or atom.name
-			local info = addrs[name]
+			local  info = addrs[name]
 			if not info then return end
 			local entries, total = canonical_word_entries(atom)
 			matched[#matched + 1] = {
@@ -248,9 +244,10 @@ local function build_atom_table(ctx)
 	return matched
 end
 
---- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`, the convenience
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`,
---- vars set in `emit_gdb_runtime` provide printf args, and each command is a static sequence of `printf` / `tbreak` /
+--- the convenience vars set in `emit_gdb_runtime` provide printf args, and 
---- `if ... end` blocks. The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
+--- each command is a static sequence of `printf` / `tbreak` / `if ... end` blocks.
+--- The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
 ---
 --- Why hardcoded per-atom: gdb's `$` substitution doesn't concat inside var names — `$__atom_name_$__i` in a `while`
 --- loop resolves to one literal identifier, not `name_i`. Compile-time emission is the only path.
@@ -395,8 +392,7 @@ local function append_gdb_commands(lines, matched)
 end
 
 --- Emit the gdb-runtime file (post-link). Pure gdb scripting — addresses come from `mipsel-none-elf-nm -S`, get embedded
---- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb source-time.
---- source-time.
 --- @param ctx PassCtx
 local function emit_gdb_runtime(ctx)
 	if not (ctx.flags and ctx.flags.gdb_runtime) then return end

scripts/passes/auto_reg.lua

@@ -0,0 +1,355 @@
+--- passes/auto_reg.lua — Per-phase automatic GPR allocator + gen/auto_reg.h emitter.
+---
+--- Reads the per-source + corpus-level `atom_auto_regs` + `phase_auto_regs` registries populated by `passes/scan_source.lua`.
+--- Runs a deterministic first-fit allocator in the `R_T0..R_T7 + R_V0..R_V1` pool (10 physical GPRs).
+--- Emits one `#define R_<Sym>_Code R_Tn_Code` per marker into per-directory `gen/auto_reg.h`.
+---
+--- User-pinned GPRs : The corpus's `register_alias_registry` is consulted to exclude GPRs the user has pinned via 
+--- `atom_reg` + `_Code` defs (e.g. carriers like `R_ResolveScratch = R_T4 atom_reg`). 
+--- These GPRs are unavailable to EVERY atom's source pool.
+--- Carriers are preserved across atoms by context discipline and must never be reallocated.
+--- Per-atom body parsing also catches alias references (R_<Alias>) and hardcoded R_Tn references,
+--- so the user can write either `R_T4` or `R_ResolveScratch` in an atom body and the pass will
+--- exclude R_T4 from that atom's pool.
+---
+--- Conflict detection: If the user hardcodes `R_Tn` in an atom body that shares a phase with an auto-reg that picked `R_Tn`,
+--- emit `phase_register_clash` as an info finding (no build stop).
+--- Should be unreachable after the user-pinning + body-parsing fix above; kept as a defensive safety net.
+---
+--- Pool exhaustion: If a phase declares more `R_<Sym>` mappings than the 10-register pool can hold,
+--- emit `phase_register_pool_exhausted` as a build-stopping error.
+
+--- @class AutoRegResult
+--- @field outputs  table[]  -- {kind=, path=} entries
+--- @field errors   table[]  -- {line=, msg=}  entries (build-stops)
+--- @field warnings table[]  -- {line=, msg=}  entries (build-continues)
+
+local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+--- ════════════════════════════════════════════════════════════════════════════
+--- THE GPR ALLOCATION POOL — what is allocatable, and (more importantly) WHY
+--- ════════════════════════════════════════════════════════════════════════════
+---
+--- The auto-reg pass picks physical GPRs for `atom_auto_reg(...)` / `phase_auto_reg(...)` markers.
+--- It allocates from a FIXED 10-register pool.
+--- This comment block makes the inclusion AND exclusion criteria obvious so a reader doesn't have 
+--- to grep lottes_tape.h + mips.h to understand the design.
+---
+--- ── WHAT'S IN THE POOL (10 GPRs, all caller-trash per the O32 ABI) ────────
+---     R_T0..R_T7  (GPR codes 8..15), R_V0..R_V1 (GPR codes 2..3)  
+---      The workhorse of every atom body. The uesr should be aware of atom allocation across atoms they chain. 
+---      If they have a collision it means either they didn't saturate the register file optimally for a phase,
+---      or the may have made the workload to large for the run.
+---
+--- ── WHAT'S NOT IN THE POOL — and WHY (the "obvious exclusions") ────────────
+---   R_T9  (GPR code 25)  — R_TapePtr, the tape instruction stream pointer.
+---     Owned by the tape runtime (in tape_run / tape_run_a02_s07).
+---     `rgcc(R_TapePtr)` register-variable ties the C compiler's view to $t9 across the whole tape_run.
+---     The auto-reg pass MUST NOT clobber this; doing so would desync the C-side tape pointer from the
+---     hardware pointer and crash on the next tape_run.
+---
+---   R_T8  (GPR code 24)  — R_AtomJmp, the atom-jump register used by the 4-word yield handshake.
+---     Every `mac_yield()` / `mac_yield_tail` does `load_word R_AtomJmp, R_TapePtr, 0` then
+---     `jump_reg R_AtomJmp`. The auto-reg pass MUST NOT clobber this either, or the atom dispatcher breaks.
+---     Owned by the tape runtime, same family as R_TapePtr.
+---
+---   R_AT  (GPR code 1)   — Assembler temporary. Reserved by the MIPS O32 ABI for pseudoinstruction expansion
+---     (lottes_tape.h:86, mips.h:93). The ISA's psuedo instructions use it as a scratch temporary.
+---
+---   R_A0..A3 (codes 4..7)   — Function arguments. Used in tape_run_a02_s07, see below.
+---   R_S0..S7 (codes 16..23) — Callee-saved. Preserved across C-ABI calls by convention.
+---     The `tape_run_a02_s07` variant clobbers them deliberately, but the default `tape_run` does NOT.
+---     Kept out of POOL to preserve the conservative default.
+---     Add them in a separate "big clobber" pool if/when needed.
+---
+---   R_K0/K1       (codes 26..27) — Kernel / interrupt handler reserves. Never touched by user code; OS-internal.
+---   R_GP/SP/FP/RA (codes 28..31) — Stack frame + return-address. Owned by the C compiler; never allocatable.
+---   R_0           (code 0)       — Hardwired zero. Cannot be written.
+---
+local POOL = {
+    "R_T0", "R_T1", "R_T2", "R_T3",
+    "R_T4", "R_T5", "R_T6", "R_T7",
+    "R_V0", "R_V1",
+}
+
+-- Map from integer MIPS GPR code (the `code` field on AliasEntry) to the physical GPR ident in POOL.
+-- The standard MIPS O32 ABI register numbering matches mips.h's R_*_Code #defines (mips.h).
+-- Only the POOL entries matter for auto_reg — non-pool aliases
+-- (R_AT=1, R_A0..A3=4..7, R_T8=24, R_T9=25, R_K0/K1=26..27, R_GP/SP/FP/RA=28..31) 
+-- are deliberately omitted — see the comment block above for the WHY of each exclusion.
+local INT_CODE_TO_POOL_GPR = {
+    [2]  = "R_V0", [3]  = "R_V1",
+    [8]  = "R_T0", [9]  = "R_T1", [10] = "R_T2", [11] = "R_T3",
+    [12] = "R_T4", [13] = "R_T5", [14] = "R_T6", [15] = "R_T7",
+}
+
+-- Stable sort for deterministic allocation order.
+local function stable_sort_keys(tbl)
+    local keys = {}
+    for k in pairs(tbl) do keys[#keys + 1] = k end
+    table.sort(keys)
+    return keys
+end
+
+-- Allocate one phase's auto-reg mappings.
+-- Returns (allocated_map, errors). On pool exhaustion, errors is populated and the function halts.
+local function allocate_phase(phase_label, decls)
+    -- Deep-copy POOL into a fresh sequence table. The original `table.unpack and table.unpack(POOL) or { unpack(POOL) }`
+    -- idiom wraps the unpacked values in a single inner table under LuaJIT 5.1 (`table.unpack` is nil; the `or` returns one value),
+    -- which corrupts the pool into `{ {R_T0, R_T1, ...} }` — making `table.remove(pool, 1)` return the inner table on iteration.
+    local pool   = {}
+    for i = 1, #POOL do pool[i] = POOL[i] end
+    local result = {}
+    local errors = {}
+    for _, sym in ipairs(stable_sort_keys(decls)) do
+        local next_gpr = table.remove(pool, 1)
+        if not next_gpr then
+            errors[#errors + 1] = {
+                line = 0,
+                msg  = string.format("phase_register_pool_exhausted: "
+									.. "phase '%s' requested symbol '%s' but the pool has no remaining registers "
+									.. "(max 10 per phase: R_T0..R_T7 + R_V0..R_V1). Split the phase or use hardcoded GPRs."
+                  , phase_label, sym),
+            }
+            return result, errors
+        end
+        result[sym] = next_gpr
+    end
+    return result, errors
+end
+
+-- Build two projections from corpus.register_alias_registry:
+--   user_pinned   -- { [physical_gpr_ident] = true } -- GPRs unavailable to auto_reg globally (wave-context carriers, file-scope pinned aliases)
+--   alias_to_gpr  -- { [alias_ident]        = physical_gpr_ident } -- for body parsing
+-- Both projections are derived from the same set of entries: every AliasEntry in register_alias_registry has `has_atom_reg = true` 
+-- (only those entries are added to the registry; see passes/scan_source.lua parse_enum_entry).
+-- Each entry's `code` is the integer MIPS GPR number (0..31); INT_CODE_TO_POOL_GPR translates it back to the physical GPR ident.
+-- Aliases whose `code` points to a non-POOL GPR (e.g. R_S0, R_T8, R_K1) are ignored — 
+-- they don't affect the auto_reg pool, and they're already excluded from POOL above.
+local function build_user_pins(corpus)
+    local user_pinned  = {}
+    local alias_to_gpr = {}
+    if not corpus.register_alias_registry then return user_pinned, alias_to_gpr end
+    for alias_name, alias_entry in pairs(corpus.register_alias_registry) do
+        if alias_entry.has_atom_reg and alias_entry.code then
+            local gpr = INT_CODE_TO_POOL_GPR[alias_entry.code]
+            if gpr then
+                user_pinned[gpr]         = true
+                alias_to_gpr[alias_name] = gpr
+            end
+        end
+    end
+    return user_pinned, alias_to_gpr
+end
+
+-- Find every physical GPR referenced in the atom body, via EITHER:
+--   (a) A hardcoded physical GPR ident (R_T\d+|R_V\d+|R_A\d+|R_S\d+) — the existing regex;
+--   (b) An alias ident (R_<Alias>) resolved via alias_to_gpr back to its physical GPR ident.
+-- Returns { [physical_gpr_ident] = count }. Clash-detection and source-pool-exclusion logic
+-- only needs the presence of each GPR (boolean test), but keeping count preserves the
+-- original find_hardcoded_rn shape so callers can switch without churn.
+-- The alias pattern is sorted lexicographically to keep the regex deterministic.
+local function find_used_gprs(body_text, alias_to_gpr)
+    local found = {}
+    -- (a) Hardcoded physical GPRs (R_T0..R_T7, R_V0..R_V1, R_A0..R_A3, R_S0..R_S7).
+    for gpr in body_text:gmatch("(R_T%d+|R_V%d+|R_A%d+|R_S%d+)") do
+        found[gpr] = (found[gpr] or 0) + 1
+    end
+    -- (b) Alias references (R_<Alias>) resolved to physical GPRs via the registry.
+    --      Sorted by name so the regex is byte-stable across runs.
+    if alias_to_gpr and next(alias_to_gpr) then
+        local aliases = {}
+        for alias_name in pairs(alias_to_gpr) do
+            aliases[#aliases + 1] = alias_name
+        end
+        table.sort(aliases)
+        local pattern = "(" .. table.concat(aliases, "|") .. ")"
+        for alias_name in body_text:gmatch(pattern) do
+            local gpr = alias_to_gpr[alias_name]
+            if gpr and not found[gpr] then
+                found[gpr] = 1
+            end
+        end
+    end
+    return found
+end
+
+-- Emit one gen/auto_reg.h header per directory.
+local function emit_auto_reg_h(out_dir, dir, sources, mappings)
+    if not mappings or next(mappings) == nil then return end
+    local out_path = out_dir .. "/" .. "auto_reg.h"
+    duffle.ensure_dir(out_dir)
+    local lines = {
+        "#ifdef INTELLISENSE_DIRECTIVES",
+        "#pragma once",
+        "#endif",
+        "// Auto-generated by ps1_meta.lua (passes/auto_reg.lua) — DO NOT EDIT",
+        "// Directory: " .. dir:gsub("/", "\\"),
+    }
+    for _, src in ipairs(sources) do
+        lines[#lines + 1] = "//   source: " .. src.path
+    end
+    lines[#lines + 1] = "// Per-phase register allocations resolved by the lua pass."
+    lines[#lines + 1] = "// R_<Sym>_Code = <chosen GPR's _Code constant> for every marker in this directory."
+    lines[#lines + 1] = ""
+    for _, sym in ipairs(stable_sort_keys(mappings)) do
+        local gpr         = mappings[sym]
+        local gpr_code    = gpr .. "_Code"
+        lines[#lines + 1] = "#define " .. sym .. "_Code " .. gpr_code
+    end
+    lines[#lines + 1] = ""
+    duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
+    print("  -> " .. out_path)
+    return out_path
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Pass entry
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+--- @param ctx PassCtx
+--- @return AutoRegResult
+function M.run(ctx)
+    local outputs  = {}
+    local errors   = {}
+    local warnings = {}
+
+    local corpus = ctx.shared and ctx.shared.corpus
+    if type(corpus) ~= "table" then
+        error("auto_reg.run requires ctx.shared.corpus", 0)
+    end
+
+    -- 0. Build the user-pinned GPR exclusion set + alias-to-GPR resolution map.
+    --  Wave-context carriers (e.g. `R_ResolveScratch = R_T4 atom_reg` in hello_camera.atom.c)
+    --  MUST NOT be allocated to any auto-reg marker — they're preserved across atoms by the wave-context discipline.
+		--  The corpus's register_alias_registry is the source of truth for these opt-in pins.
+    --  Body references to those aliases (via alias_to_gpr) are also excluded on a per-atom basis in step 2 below.
+    local user_pinned, alias_to_gpr = build_user_pins(corpus)
+
+    -- 1. Allocate phase pools first (phase declarations take precedence over per-atom declarations).
+    local phase_allocations = {}
+    for phase_label, decls in pairs(corpus.phase_auto_regs or {}) do
+        local mapping, errs = allocate_phase(phase_label, decls)
+        for sym, gpr in pairs(mapping) do
+            phase_allocations[phase_label]      = phase_allocations[phase_label] or {}
+            phase_allocations[phase_label][sym] = gpr
+        end
+        for _, e in ipairs(errs) do
+            errors[#errors + 1] = e
+        end
+    end
+
+    -- 2. Allocate per-atom auto-regs. If the atom scope matches a phase, reuse the phase pool.
+    -- Otherwise, allocate a private pool for the atom.
+    -- The phase membership is in `corpus.atom_phases[phase_label].atoms` (an array of atom names declared via `atom_phase(<phase>)`
+		-- in the atom's `atom_info` line). Build a reverse map `atom_name -> phase_label` so the lookup is O(1) per atom scope.
+    local atom_name_to_phase = {}
+    for phase_label, entry in pairs(corpus.atom_phases or {}) do
+        for _, atom_name in ipairs(entry.atoms or {}) do
+            atom_name_to_phase[atom_name] = phase_label
+        end
+    end
+
+    local atom_allocations = {}
+    for atom_scope, decls in pairs(corpus.atom_auto_regs or {}) do
+        local phase_label = atom_name_to_phase[atom_scope]
+        -- Build the atom's source pool: start with the full POOL, subtract:
+        --   (a) every GPR already committed (phase allocations + prior atom allocations)
+        --   (b) every USER-PINNED GPR (wave-context carriers + file-scope pinned aliases)
+        --   (c) every GPR referenced in the atom's body — either hardcoded R_X or alias R_Xxx
+        --       (the latter resolved via alias_to_gpr; this catches cases where the user wrote R_ResolveScratch instead of R_T4 directly)
+        -- Atoms whose scope matches a phase share the global pool with the phase allocations;
+        -- the original `source_pool = phase_allocations[phase_label]` form used the phase
+        -- allocation MAP as a pool, but that map has no array part, so `table.remove(source_pool, 1)`
+        -- returned nil and every atom-with-phase marker errored with `phase_register_pool_exhausted`.
+        local used = {}
+        for _, m in pairs(phase_allocations) do for _, gpr in pairs(m) do used[gpr] = true end end
+        for _, m in pairs(atom_allocations)  do for _, gpr in pairs(m) do used[gpr] = true end end
+        -- (c) Body references — scan the atom body for hardcoded + alias-resolved GPRs.
+        --     Folded into `used` so the source_pool exclusion is a single check.
+        local atom = corpus.atoms_by_name and corpus.atoms_by_name[atom_scope]
+        if atom and atom.body then
+            local body_used = find_used_gprs(atom.body, alias_to_gpr)
+            for gpr in pairs(body_used) do used[gpr] = true end
+        end
+        local source_pool = {}
+        for _, gpr in ipairs(POOL) do
+            -- Exclude (a) prior commitments, (b) USER-PINNED GPRs (wave-context carriers
+            -- declared via atom_reg + _Code defs, preserved across atoms globally).
+            if not used[gpr] and not user_pinned[gpr] then
+                source_pool[#source_pool + 1] = gpr
+            end
+        end
+        local result = {}
+        for _, sym in ipairs(stable_sort_keys(decls)) do
+            local next_gpr = table.remove(source_pool, 1)
+            if not next_gpr then
+                errors[#errors + 1] = {
+                    line = 0,
+                    msg  = string.format("phase_register_pool_exhausted: atom '%s' requested symbol '%s' "
+											.. "but no free registers remain in its scope pool."
+                    	, atom_scope, sym),
+                }
+            else
+                result[sym] = next_gpr
+            end
+        end
+        atom_allocations[atom_scope] = result
+    end
+
+    -- 3. Conflict-with-hardcoded detection (defensive — should be unreachable now).
+    -- The source_pool exclusion in step 2 (b) + (c) already accounts for both user-pinned GPRs
+    -- and body-referenced GPRs (hardcoded R_Tn OR alias R_<Alias>).
+		-- An auto-reg allocation that matched an existing body reference would be impossible by construction.
+		-- This warning is kept as a defensive safety net for cases the body scanner might miss
+		-- (e.g. macros that expand to register references the scanner cannot resolve).
+    -- For each resolved (scope, sym) -> R_Tn mapping, scan the atom body source for used GPRs.
+    for atom_scope, decls in pairs(atom_allocations) do
+        local atom = corpus.atoms_by_name and corpus.atoms_by_name[atom_scope]
+        if atom and atom.body then
+            local used_in_body = find_used_gprs(atom.body, alias_to_gpr)
+            for sym, allocated_gpr in pairs(decls) do
+                if used_in_body[allocated_gpr] and used_in_body[allocated_gpr] > 0 then
+                    warnings[#warnings + 1] = {
+                        line  = atom.line or 0,
+                        msg   = string.format("phase_register_clash: atom '%s' has hardcoded '%s' in its body AND an auto-reg marker '%s' "
+													.. "that was allocated to '%s' (same phase). Resolve by removing the hardcoded reference or renaming the auto-reg."
+                        	, atom_scope, allocated_gpr, sym, allocated_gpr),
+                    }
+                end
+            end
+        end
+    end
+
+    -- 4. Emit per-directory gen/auto_reg.h.
+    -- For each source directory that has atom_auto_regs or phase_auto_regs entries, emit one header.
+    local sources_by_dir = corpus.sources_by_dir or {}
+    for dir, sources in pairs(sources_by_dir) do
+        local per_dir_mappings = {}
+        for _, src in ipairs(sources) do
+            -- Collect every (sym -> gpr) entry that originated from a source in this directory.
+            -- `src.scan.atom_auto_regs` is keyed by ATOM SCOPE NAME; `pairs(t)` iterates KEYS so `scope_name` here is the scope ident (e.g. "cube_g4_face").
+						-- The previous `for _, scan_atom_auto` form silently assigned the VALUE (a `{sym = sym}` table) to the variable,
+						-- which made `atom_allocations[scan_atom_auto]` a table-indexed lookup that never resolved.
+            for scope_name in pairs(src.scan and src.scan.atom_auto_regs or {}) do
+                for sym, gpr in pairs(atom_allocations[scope_name] or {}) do
+                    per_dir_mappings[sym] = gpr
+                end
+            end
+            for scope_name in pairs(src.scan and src.scan.phase_auto_regs or {}) do
+                for sym, gpr in pairs(phase_allocations[scope_name] or {}) do
+                    per_dir_mappings[sym] = gpr
+                end
+            end
+        end
+        local out_dir  = dir .. "/gen"
+        local out_path = emit_auto_reg_h(out_dir, dir, sources, per_dir_mappings)
+        if out_path then outputs[#outputs + 1] = { auto_reg_h = out_path } end
+    end
+    return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/components.lua

@@ -3,13 +3,15 @@
 --- Ownership: `corpus.word_counts`, `corpus.components`, and `corpus.component_body_index`.
 --- 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,
+--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations (kind="comp_bare" / "comp_proc"),
 --- 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.
+--- `MipsAtom_Proc_(X, ab, { body })` declarations (kind="atom_proc") are ATOMS, not components, and are deliberately excluded —
+--- atoms get emitted via `tb_emit(tb, code_<name>)` linker symbols, not inlined as `mac_*` macros.
 ---
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
+--- 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.
---- Lua 5.3 compatible.
+--- 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
@@ -41,43 +43,44 @@ local MAC_PREFIX_LEN  = 4
 local BYTE_NEWLINE    = 10
 local BYTE_SLASH      = 47
 
--- Source dir basename used as the output `.macs.h` filename.
+-- Output gen subdirectory + filename (per-directory aggregation; the directory name is the namespace).
 local GEN_SUBDIR      = "gen"
+local MACS_FILENAME   = "macs.h"
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
+--- @field sources            SourceFile[]           -- All source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table                  -- Cross-pass shared state
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
+--- @field project_root       string                 -- Project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
+--- @field upstream           table<string, table>   -- Per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @field verbose            boolean                -- log diagnostic info
+--- @field verbose            boolean                -- Log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
---- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field errors   table[]  -- {line=, msg=}  entries; build-stops
---- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+--- @field warnings table[]  -- {line=, msg=}  entries; build-succeeds
 
 --- @class Component
---- @field name       string     -- atom name (without `ac_` prefix)
+--- @field name       string     -- Atom name (without `ac_` prefix)
---- @field body       string     -- brace-delimited body (without the braces)
+--- @field body       string     -- Brace-delimited body (without the braces)
---- @field args       string|nil -- function-args string (function form only)
+--- @field args       string|nil -- Function-args string (function form only)
---- @field line       integer    -- source line of the declaration
+--- @field line       integer    -- Source line of the declaration
---- @field comment    string|nil -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @field comment    string|nil -- Scanner-owned `declaration_comment`; the components pass reads it from the scanner record
---- @field kind       string     -- "comp_bare" | "comp_proc"
+--- @field kind       string     -- "comp_bare" | "comp_proc"  (atom_proc is NOT a component — see `project_components`)
---- @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)
@@ -93,48 +96,21 @@ local M = {}
 -- so this file reads it forward rather than re-walking the source.
 -- ════════════════════════════════════════════════════════════════════════════
 
---- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation of the given name.
+--- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation.
 --- 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
+--- After the `sym` arg was dropped from MipsAtomComp_Proc_, the component name
----   `FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
+--- and the args both come from the preceding `FI_ Slice_MipsCode ac_X(args)`
---- We find the LAST occurrence of `"ac_X("` before `before_pos` and extract the args from inside the parens.
+--- declaration. The shared `duffle.find_function_decl_for` helper does the
---- We then verify the preceding context ends with `Slice_MipsCode`
+--- backward walk; this function returns just the args.
---- (the function-decl keyword with possible qualifiers between).
 ---
 --- @param source     string
---- @param name       string
+--- @param name       string  (retained for signature stability; unused — the walk derives the name)
 --- @param before_pos integer
 --- @return string|nil
 local function find_function_args_for(source, name, before_pos)
-	-- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
+	local _, args_inner = duffle.find_function_decl_for(source, before_pos, #MIPS_ATOM)
-	local name_open  = name .. "("
+	return args_inner
-	local last_idx   = nil
-	local scan_pos   = 1
-	while true do
-		-- Pass `before_pos + 1` so string.find only returns positions < before_pos + 1
-		-- (string.find's 4th arg `plain` is true; we use the 3rd arg `init` for the upper bound).
-		local found = source:find(name_open, scan_pos, true)
-		if not found or found >= before_pos then break end
-		last_idx = found
-		scan_pos = found + #name_open
-	end
-	if not last_idx then return nil end
-
-	-- Verify the preceding context ends with "MipsAtom" (with possible qualifiers between).
-	local before   = source:sub(1, last_idx - 1)
-	local trimmed  = duffle.trim(before)
-	if trimmed:sub(-#MIPS_ATOM) ~= MIPS_ATOM then
-		-- Preceding context is not a function declaration.
-		return nil
-	end
-
-	local open_paren = last_idx + #name  -- position of "("
-	-- scan: MipsAtom ac_X(
-	local inner      = duffle.read_parens(source, open_paren)
-	-- scan: MipsAtom ac_X(<args>)
-	if not inner then return nil end
-	return inner
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -200,7 +176,16 @@ end
 local function project_components(source, scan)
 	local out = {}
 	for _, a in ipairs(scan.atoms) do
+		-- Only `MipsAtomComp_(ac_X)` (kind="comp_bare") and `MipsAtomComp_Proc_(ac_X, ...)` (kind="comp_proc")
+		-- are COMPONENTS — they get inlined via `mac_<name>` aliases inside atom bodies.
+		-- `MipsAtom_Proc_` (kind="atom_proc") is an ATOM (ends with `mac_yield()`); it gets emitted via
+		-- `tb_emit(tb, code_<name>)` (linker symbol), NOT inlined as a macro. Including `atom_proc` here
+		-- would incorrectly emit `mac_<name>` aliases for atoms, polluting `gen/macs.h`.
+		-- See `docs/duffle_dsl_primer.md` §"mac_* aliases" for the contract.
 		if a.kind == "comp_bare" or a.kind == "comp_proc" then
+			-- Function-args lookup is meaningful for `MipsAtomComp_Proc_` components
+			-- (the macro sits inside `FI_ Slice_MipsCode ac_X(...)`); the alias expansion
+			-- discards the `ab` (atom-builder) arg the same way both forms do.
 			local args = find_function_args_for(source, a.raw_name, a.ident_pos)
 			-- Comment ownership: scan_source.lua stamps `declaration_comment` on the record by walking backward past any associated bare marker.
 			-- The pass reads `declaration_comment` directly.
@@ -231,7 +216,6 @@ 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)
@@ -300,7 +284,9 @@ local function word_count_rec(name, comp_by_name, wc, cache)
 			local trimmed = t.tok
 			if trimmed ~= "" then
 				local lookup = strip_mac_prefix(duffle.read_ident(trimmed, 1))
-				if lookup and comp_by_name[lookup] then
+				if lookup == "atom_label" or lookup == "atom_offset" then
+					-- Pure metaprogram anchors; emit zero words.
+				elseif lookup and comp_by_name[lookup] then
 					-- It's a `mac_X(...)` call. Recurse.
 					n = n + word_count_rec(lookup, comp_by_name, wc, cache)
 				elseif lookup and wc and wc[lookup] then
@@ -350,7 +336,6 @@ end
 
 --- (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>
@@ -392,9 +377,7 @@ local function cycle_cost_rec(name, comp_by_name, latency, cache)
 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
+--- 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.
---- 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>
@@ -410,9 +393,9 @@ local function gp0_contrib_rec(name, comp_by_name, cache)
 		local tokens = cc.body_tokens
 		for _, t in ipairs(tokens) do
 			local trimmed = t.tok
-			if trimmed ~= "" then
+			if    trimmed ~= "" then
 				local ident = duffle.read_ident(trimmed, 1)
-				if ident and ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
+				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)
@@ -460,8 +443,8 @@ end
 --- @param s string
 --- @return string[]
 local function split_comment_lines(s)
-	local out  = {}
+	local out   = {}
-	local pos  = 1
+	local pos   = 1
 	local s_len = #s
 	while pos <= s_len do
 		local nl = s:find("\n", pos, true)
@@ -476,12 +459,25 @@ local function split_comment_lines(s)
 end
 
 --- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
+--- For `MipsAtomComp_Proc_` components, the leading `ab` (atom-builder) arg is dropped:
+--- the generated `mac_<name>` macros are inline-expansion aliases for baked atoms; their bodies don't reference `ab`
+--- (the builder is only consumed by the procedural `atombuilder_unroll` line that `MipsAtomComp_Proc_` appends after the body).
+--- Inline callers therefore don't need to thread a builder context.
 --- @param args_str string|nil
 --- @return string
 local function signature_from_args(args_str)
 	local arg_names = extract_arg_names(args_str)
 	if    arg_names and #arg_names > 0 then
-		return table.concat(arg_names, ", ")
+		-- Drop the leading `ab` (atom-builder) first arg if present.
+		-- Convention: `MipsAtomComp_Proc_` components always declare `ab` as the first function-arg
+		-- (type `MipsAtomBuilder_R`), mirroring the macro signature in `lottes_tape.h`.
+		if arg_names[1] == "ab" then
+			table.remove(arg_names, 1)
+		end
+		if #arg_names > 0 then
+			return table.concat(arg_names, ", ")
+		end
+		return "..."  -- `ab` was the only arg; fall through to variadic
 	end
 	return "..."
 end
@@ -523,7 +519,7 @@ local function build_component_lines(c, counts)
 
 	-- Marker comment: emitted once for every skipped component.
 	-- The marker is scanner-owned (declared by `atom_dbg_skip` immediately before the declaration in the source);
-	-- the components pass projects `c.debug_skip` and emits the marker as a generated comment.
+	-- This pass projects `c.debug_skip` and emits the marker as a generated comment.
 	if c.debug_skip then
 		lines[#lines + 1] = "/* atom_dbg_skip */"
 	end
@@ -536,9 +532,9 @@ local function build_component_lines(c, counts)
 
 	local tokens = duffle.split_top_level_commas(c.body)
 	for i = 1, #tokens do tokens[i] = duffle.trim(tokens[i]) end
-	local sig    = signature_from_args(c.args)
+	local sig = signature_from_args(c.args)
 	-- Direct lookup against the per-source precomputed `counts` table (built once by count_all_components).
-	local n      = counts[c.name]
+	local n   = counts[c.name]
 
 	if n > 0 then
 		emit_macro_body(lines, c, sig, tokens)
@@ -557,9 +553,15 @@ 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       -- 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.
@@ -567,7 +569,7 @@ 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.
@@ -580,30 +582,30 @@ local function header_boilerplate(src)
 	}
 end
 
---- Compute the output path for one source's `.macs.h` file.
+--- Compute the per-directory output path for `.macs.h`.
---- The pre-rework convention uses the *directory* basename (not the source file basename)
+---  e.g. any source in `code/duffle/` produces `code/duffle/gen/macs.h` regardless of source filename.
----  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+--- The directory name is the namespace; the filename does not repeat it.
---- This matches what the C codebase #includes.
+--- @param dir string  -- Absolute source directory
---- @param src SourceFile
+--- @return string  -- Output directory
---- @return string  -- the output directory
+--- @return string  -- Full output path
---- @return string  -- the full output path
+local function compute_macs_h_path(dir)
-local function compute_macs_h_path(src)
+	local  out_dir  = dir .. "/" .. GEN_SUBDIR
-	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
+	local  out_path = out_dir .. "/" .. MACS_FILENAME
-	local  out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
 	return out_dir, out_path
 end
 
---- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
+--- Emit a per-directory `.macs.h` header with the aggregated `mac_X` macros + `WORD_COUNT` entries.
 --- Writes in BINARY mode so LF line endings are preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
 --- @param ctx        PassCtx
---- @param src        SourceFile
+--- @param dir        string                  -- Absolute source directory
---- @param components Component[]
+--- @param sources    SourceFile[]            -- Sources contributing to this directory (for the header comment)
---- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param components Component[]             -- Aggregated components from all sources in this directory
---- @return string|nil  -- path to the written file (nil if no components)
+--- @param counts     table<string, integer>  -- Precomputed word counts (from count_all_components)
-local function emit_component_macros_h(ctx, src, components, counts)
+--- @return string|nil -- Path to the written file (nil if no components)
+local function emit_component_macros_h(ctx, dir, sources, components, counts)
 	if #components == 0 then return nil end
-	local out_dir, out_path = compute_macs_h_path(src)
+	local out_dir, out_path = compute_macs_h_path(dir)
-	local lines             = header_boilerplate(src)
+	local lines             = header_boilerplate(dir, sources)
 
 	for _, c in ipairs(components) do
 		for _, l in ipairs(build_component_lines(c, counts)) do
@@ -638,11 +640,11 @@ local function update_canonical_word_counts(corpus, components, counts)
 end
 
 --- @class ComponentDef
---- @field name       string  -- bare name (without ac_/mac_ prefix)
+--- @field name       string  -- Bare name (without ac_/mac_ prefix)
---- @field line       integer -- definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
+--- @field line       integer -- Definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
---- @field path       string  -- absolute source path of the definition
+--- @field path       string  -- Absolute source path of the definition
---- @field kind       string  -- "comp_bare" | "comp_proc"
+--- @field kind       string  -- "comp_bare" | "comp_proc"  (atom_proc is NOT a component)
---- @field debug_skip boolean -- mirror of the scanner-owned `a.debug_skip`; consumers read this directly
+--- @field debug_skip boolean -- Mirror of the scanner-owned `a.debug_skip`; consumers read this directly
 
 --- (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").
@@ -675,7 +677,7 @@ local function update_canonical_components(corpus, src, components, metadata)
 			-- Identical-shape declarations (same path + line) reuse the first-wins entry without a collision record.
 			local existing = corpus.components[c.name]
 			if existing.path ~= rel_path or existing.line ~= c.line then
-				local kind     = c.kind or "comp_bare"
+				local kind       = c.kind or "comp_bare"
 				local first_kind = existing.kind or "comp_bare"
 				corpus.collisions[#corpus.collisions + 1] = {
 					kind              = "component",
@@ -740,24 +742,39 @@ function M.run(ctx)
 	--   * `corpus.component_body_index[name]` — body / line_of / source index
 	-- The pass writes to the corpus only; consumers read from the corpus directly.
 
-	for _, src in ipairs(corpus.source_order) do
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/macs.h`.
-		-- project_components reads from src.scan + does backward lookups on src.text
+	-- The directory itself is the namespace. `corpus.sources_by_dir` preserves source-order within each bucket (matches `corpus.source_order`).
-		local components = project_components(src.text, src.scan)
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
-		if #components > 0 then
+	for dir, sources in pairs(sources_by_dir) do
-			-- Compute all component word counts once per source.
+		-- Aggregate components from every source in this directory.
-			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
+		-- `project_components` returns nil for sources with no `MipsAtomComp_` declarations; we skip those.
-			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
+		local aggregated_components = {}
-			local counts = count_all_components(components, corpus.word_counts)
+		local metadata_per_source   = {}
-			-- Derive cycle_cost + gp0_contrib from the original `MipsAtomComp_` body tokens
+		for _, src in ipairs(sources) do
-			-- (NOT from the generated `mac_*` variants — those are written to disk above).
+			local per_source = project_components(src.text, src.scan) or {}
-			local metadata = compute_components_metadata(components, duffle.INSTRUCTION_LATENCY)
+			for _, c in ipairs(per_source) do
-			local macs_path = emit_component_macros_h(ctx, src, components, counts)
+				aggregated_components[#aggregated_components + 1] = c
+			end
+			if #per_source > 0 then
+				metadata_per_source[src] = compute_components_metadata(per_source, duffle.INSTRUCTION_LATENCY)
+			end
+		end
+		if #aggregated_components > 0 then
+			-- Compute word counts across the aggregated set. `corpus.word_counts` carries the
+			-- same-source + prior-directory entries so the recursive lookup sees both.
+			local counts    = count_all_components(aggregated_components, corpus.word_counts)
+			local macs_path = emit_component_macros_h(ctx, dir, sources, aggregated_components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				-- Populate the projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
+				-- Populate the projections AFTER disk emission (byte-identical `.macs.h` contract).
-				update_canonical_word_counts(corpus, components, counts)
+				update_canonical_word_counts(corpus, aggregated_components, counts)
-				update_canonical_components(corpus, src, components, metadata)
+				for _, src in ipairs(sources) do
-				update_canonical_component_body_index(corpus, src, components, src.scan)
+					local per_source = project_components(src.text, src.scan) or {}
+					if #per_source > 0 then
+						update_canonical_components(corpus, src, per_source, metadata_per_source[src])
+						update_canonical_component_body_index(corpus, src, per_source, src.scan)
+					end
+				end
 			end
 		end
 	end

scripts/passes/dwarf_injection.lua

@@ -73,8 +73,8 @@ local DW_RLE_start_length  = DWARF5_RNGLISTS.start_length
 
 -- File-index lookup for the existing main line unit (Unit 2).
 -- Populated at pass start by `init_file_index_lookup(elf_path)` from the runtime ELF (see `elf_dwarf.read_line_unit_file_table`).
-local _file_index_by_basename = nil  -- [basename] = 1-based line-table file index
+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 _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
@@ -107,8 +107,8 @@ local ABBREV_TYPED_VIEW_POINTER = 0x6E   -- 110: DW_TAG_pointer_type no children
 -- DWARF5 §7.7.3 loclist opcodes.
 local DW_LLE_end_of_list    = 0x00
 local DW_LLE_start_length   = 0x08
-local DW_OP_reg0            = 0x50   -- base reg op; regN = 0x50 + N
+local DW_OP_reg0            = 0x50 -- base reg op; regN = 0x50 + N
-local DW_OP_breg0           = 0x70   -- base breg op; bregN = 0x70 + N (SLEB offset)
+local DW_OP_breg0           = 0x70 -- base breg op; bregN = 0x70 + N (SLEB offset)
 local DW_OP_piece           = 0x93
 local MIPS_LOAD_DELAY_BYTES = 0x08 -- 1 load word + 1 BD-slot word
 
@@ -184,15 +184,17 @@ end
 --- Resolve an absolute provenance path to the line-unit file index used by the emitting line program.
 --- Normalizes mixed `/` and `\` separators to a basename and looks it up against the runtime-computed file table populated by `init_file_index_lookup`.
 ---
---- Fails loudly on an unknown provenance basename: adding a new component source file will produce a clear error message naming the missing basename and listing the .debug_line file table contents, 
+--- Returns 0 (the DWARF `set_file(0)` "no file change" sentinel) when the basename is not in the file table.
---- so the user can either confirm the gcc include order, the unity-root, or the `.debug_line` file table contents.
+--- This is a normal occurrence: the compiler only adds a file to the `.debug_line` file table when the file has line-numbered content (i.e., code).
---- Silent fallback would mask the new-file case by misattributing component rows to an arbitrary source file.
+--- Files containing only static-array data (e.g. `MipsAtomComp_` declarations in `gp.atom.c`, `psyq.atom.c`, `pad.atom.c` — the OT-tag inserts, etc.) produce no line numbers,
+--- so gcc omits them from the file table.
+--- The DWARF emitter then keeps the previous line-program file state instead of pointing at a file that has no entries to walk.
+--- A stderr warning is emitted per-miss so the user can audit which files the compiler dropped.
 --- @param path string  -- absolute provenance path (mixed slashes accepted)
---- @return integer  -- 1-based line-unit file index
+--- @return integer  -- 1-based line-unit file index, or 0 on miss (DWARF no-change sentinel)
 local function resolve_provenance_file_index(path)
 	if _file_index_by_basename == nil then
-		error("[dwarf_injection] resolve_provenance_file_index called before init_file_index_lookup. "
+		error("[dwarf_injection] resolve_provenance_file_index called before init_file_index_lookup. Is M.run being entered correctly (with --elf)?")
-			.. "Is M.run being entered correctly (with --elf)?")
 	end
 	if path == nil or path == "" then
 		error("[dwarf_injection] resolve_provenance_file_index: empty path")
@@ -201,19 +203,17 @@ local function resolve_provenance_file_index(path)
 	local normalized = path:gsub("\\", "/")
 	-- Take the last path component (the basename).
 	local basename = normalized:match("([^/]+)$") or normalized
-	local idx = _file_index_by_basename[basename]
+	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
-	-- Build an error message listing the known basenames for fast diagnostics.
+	-- File is in the corpus but gcc omitted it from the .debug_line file table (data-only content).
-	local known = {}
+	-- Return 0 = DWARF `set_file(0)` no-change sentinel so the line program keeps its prior file state.
-	for k in pairs(_file_index_by_basename) do known[#known + 1] = k end
+	io.stderr:write(string.format("[dwarf_injection] line-table miss: '%s' (basename '%s') not in .debug_line file table; "
-	table.sort(known)
+		.. "falling back to set_file(0)\n", path, basename))
-	error(string.format("[dwarf_injection] resolve_provenance_file_index: unknown provenance basename '%s' (from '%s'). "
+	return 0
-		.. "Known basenames in the .debug_line file table (%d): %s"
-		, basename, path, #known, table.concat(known, ", ")))
 end
 
 local DW_FORM_addr           = 0x01
@@ -228,7 +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.)
@@ -448,9 +447,9 @@ end
 --- Statement-state rules:
 ---   * A marked whole atom emits one opaque is_stmt=false range row and no nested component rows; its subprogram symbol/range remains available.
 ---   * Per-row policy at every other PC:
----       - Call-site row of any invocation's first word:  is_stmt = true   (unconditional; `want_call = true`).
+---       - Call-site row of any invocation's first word:     is_stmt = true   (unconditional; `want_call = true`).
 ---       - Body row of any invocation (first or subsequent): is_stmt = not inv.debug_skip  (`want_body = not inv.debug_skip`).
----       - RAW word (no containing invocation):            is_stmt = true   (unconditional).
+---       - RAW word (no containing invocation):              is_stmt = true   (unconditional).
 ---   * The previous per-word `marked_idx` ancestor walk and the GDB 12 zero-instruction-prologue duplicate row at atom entry are DELETED; the new
 ---     first-word emission IS the entry statement.
 ---   * Whole-atom suppression wins over component markers; no nested inversion.
@@ -612,10 +611,8 @@ local function build_atom_sequence(atom)
 		-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call).
 		for ai, anc in ipairs(entry_1_ancestry) do
 			assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-			assert(anc.body_lines[1] ~= nil
+			assert(anc.body_lines[1] ~= nil,              "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
-				, "dwarf_injection: body_lines[1] missing on first-word entry for inv=" .. tostring(anc.component_name))
+			assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation "            .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
-			assert(anc.call_path and anc.call_path ~= ""
-				, "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
 			emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line,    true)
 			local is_outermost = (ai == 1)
 			if not (is_outermost and anc.debug_skip) then
@@ -640,8 +637,7 @@ local function build_atom_sequence(atom)
 			-- all OTHER ancestors emit body_lines[1] with is_stmt = not debug_skip.
 			--
 			-- This re-emits the outer ancestor's call-site + body rows at the inner's first word PC
-			-- for debugger context: source-level stepping now shows the outer body line
+			-- for debugger context: source-level stepping now shows the outer body line (not the inner body line) when stepping into the inner. PROBLEM B fix.
-			-- (not the inner body line) when stepping into the inner. PROBLEM B fix.
 			-- The body_lines[1] row references body_first_line_of[anc.id] (= the body's first content line in the parent's source), 
 			-- NOT anc.body_lines[1] (= the line of the first WORD, which is wrong when the outer's body starts with a nested call: 
 			-- gdb 12.1 picks the displayed line as the LAST row at the same PC in byte-stream order,
@@ -649,12 +645,9 @@ local function build_atom_sequence(atom)
 			local ancestry = ancestry_idx[idx]
 			for ai, anc in ipairs(ancestry) do
 				assert(anc.body_lines, "missing body_lines: emitter did not run emission-model")
-				assert(anc.body_lines[1] ~= nil
+				assert(anc.body_lines[1] ~= nil, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d", anc.component_name, anc.start_pos, #(anc.body_lines or {})))
-					, string.format("missing body_lines[1] for inv=%s start_pos=%d len=%d",
+				assert(anc.call_path and anc.call_path ~= "", "dwarf_injection: inv.call_path is missing on invocation " .. tostring(anc.component_name) .. "; emitter did not run emission-model.")
-						anc.component_name, anc.start_pos, #(anc.body_lines or {})))
+				emit_row(resolve_provenance_file_index(anc.call_path), anc.call_line, true)
-				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
 					emit_row(resolve_provenance_file_index(anc.def_path),  body_first_line_of[anc.id] or anc.body_lines[1], not anc.debug_skip)
@@ -669,10 +662,8 @@ local function build_atom_sequence(atom)
 			-- Marked invocations emit non-statement body rows at every body word; unmarked invocations emit statement body rows.
 			assert(inv.body_lines, "missing body_lines: emitter did not run emission-model")
 			local words_into = idx - inv.start_pos
-			assert(inv.body_lines[words_into] ~= nil
+			assert(inv.body_lines[words_into] ~= nil, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d", words_into, inv.component_name, inv.start_pos, #(inv.body_lines or {}), idx))
-				, string.format("missing body_lines[%d] for inv=%s start_pos=%d len=%d idx=%d",
+			emit_row(resolve_provenance_file_index(inv.def_path), inv.body_lines[words_into], not inv.debug_skip)
-					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).
 			emit_row(call_file_idx, entry.line, true)
@@ -712,9 +703,9 @@ end
 ---     `{comp_name, call_file, call_line, comp_file, comp_line, start_pos, end_pos, body_lines, debug_skip}`. `body_lines[k]`
 ---     is the k-th word's source line within the component body.
 ---
---- @param corpus    table -- the corpus from `ctx.shared.corpus`
+--- @param corpus table -- From `ctx.shared.corpus`
---- @param addrs     table -- ELF symbols keyed by atom name from `elf_dwarf.read_nm`
+--- @param addrs  table -- ELF symbols keyed by atom name from `elf_dwarf.read_nm`
---- @return table[]  -- list of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
+--- @return table[]  -- List of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
 local function build_atom_table(corpus, addrs)
 	-- Cross-ref: keep only atoms present in BOTH the nm symbol table AND `corpus.atoms_by_name`. Output is sorted by ascending addr.
 	local atoms_by_name = corpus.atoms_by_name or {}
@@ -730,8 +721,8 @@ local function build_atom_table(corpus, addrs)
 		local word_events      = paths.word_events or {}
 		local invocations_proj = paths.invocations or {}
 		-- Build the dense entries list from `word_events`.
-		--   `word_events[i].i`     = the 0-based `.word` position
+		--   `word_events[i].i`  = the 0-based `.word` position
-		--   `call_line`            = the root atom's physical source line for that word (stamped by emission_model)
+		--   `call_line`         = the root atom's physical source line for that word (stamped by emission_model)
 		local entries = {}
 		for idx, ev in ipairs(word_events) do
 			entries[#entries + 1] = {
@@ -773,9 +764,8 @@ local function build_atom_table(corpus, addrs)
 
 	local out = {}
 	-- Walk every source's atom list (which preserves source order + per-source src_path).
-	-- Cross-ref with the nm symbol table; atoms absent from `addrs` are skipped (an atom
+	-- Cross-ref with the nm symbol table; atoms absent from `addrs` are skipped
-	-- declared in source but not emitted as a symbol is a metaprogram or atom-info bug, not
+	-- (an atom declared in source but not emitted as a symbol is a metaprogram or atom-info bug, not a source-correlation bug — emit_no_emit would catch it upstream).
-	-- a source-correlation bug — emit_no_emit would catch it upstream).
 	for _, src in ipairs((corpus and corpus.source_order) or {}) do
 		local src_path = src.path or ""
 		for _, atom_rec in ipairs(((src.scan or {}).atoms) or {}) do
@@ -843,12 +833,11 @@ end
 --- (no `atom_reg` opt-in) are silently skipped — the resulting rbind record will be incomplete and the atom will fail to bind a usable piece chain.
 --- This is intentional: silently falling back to a hardcoded GPR would mask the missing opt-in.
 ---
---- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level
+--- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level statements (each entry is a single `load_*` call or other statement).
---- statements (each entry is a single `load_*` call or other statement).
+--- @param body_tokens table[]  -- The atom's pre-tokenized body statements (from atom.body_tokens)
---- @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 binds_name  string   -- expected Binds_X name (skip pairs with mismatching binds)
+--- @param registries  table    -- Merged registries from collect_per_source_registries
---- @param registries  table    -- merged registries from collect_per_source_registries
+--- @return table[]             -- List of {reg = <MIPS index>, field = <field name>}
---- @return table[]             -- list of {reg = <MIPS index>, field = <field name>}
 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 {}
@@ -856,7 +845,7 @@ local function parse_body_load_pairs(body_tokens, binds_name, registries)
 	-- 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 "")
+		local tok         = duffle.trim(t.tok or "")
 		local kind, inner = tok:match(load_pattern)
 		if kind then
 			local args = duffle.split_top_level_commas(inner)
@@ -882,9 +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}}
@@ -893,9 +880,9 @@ end
 --- The piece chain uses (DW_OP_regN, DW_OP_piece, ULEB128(field_size)).
 ---
 --- Binds fields come from `scan.binds`; the per-source `scan.binds[i].fields` already carries the typed-field record after the scan-source generalization.
---- @param corpus     table   -- the corpus from `ctx.shared.corpus`
+--- @param corpus     table   -- From `ctx.shared.corpus`
---- @param atom_table table[] -- the cross-ref'd atom table from build_atom_table
+--- @param atom_table table[] -- Cross-ref'd atom table from build_atom_table
---- @param registries table   -- merged registries from collect_per_source_registries
+--- @param registries table   -- Merged registries from collect_per_source_registries
 --- @return table, table      -- (rbind_atoms, rbind_structs)
 local function parse_rbind_atoms(corpus, atom_table, registries)
 	registries = registries or {}
@@ -929,7 +916,7 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 	local body_tokens_by_atom = {}
 	for _, src in ipairs((corpus and corpus.source_order) or {}) do
 		local scan = src.scan
-		if scan then
+		if    scan then
 			for _, atom in ipairs(scan.atoms or {}) do
 				body_tokens_by_atom[atom.name] = atom.body_tokens
 			end
@@ -948,8 +935,8 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 
 	for atom_name, ai in pairs(ai_by_atom) do
 		if ai.binds then
-			local struct     = rbind_structs[ai.binds]
+			local struct    = rbind_structs[ai.binds]
-			local body_toks  = body_tokens_by_atom[atom_name]
+			local body_toks = body_tokens_by_atom[atom_name]
 			if struct and body_toks then
 				local pairs = parse_body_load_pairs(body_toks, ai.binds, registries)
 				if #pairs > 0 then
@@ -957,7 +944,7 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 						binds     = ai.binds,
 						fields    = struct.fields, -- {name, offset} from scan.binds
 						bytes     = struct.bytes,
-						regs      = pairs,         -- ordered list of {reg, field}
+						regs      = pairs,         -- Ordered list of {reg, field}
 						info_line = ai.info_line,
 					}
 					table.insert(struct.atom_names, atom_name)
@@ -984,7 +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
@@ -995,9 +983,7 @@ local function build_dwarf_line_section(existing, atom_table)
 
 	-- Build the sequences.
 	local sequences = {}
-	for _, atom in ipairs(atom_table) do
+	for _, atom in ipairs(atom_table) do sequences[#sequences + 1] = build_atom_sequence(atom) end
-		sequences[#sequences + 1] = build_atom_sequence(atom)
-	end
 	local appended = table.concat(sequences)
 
 	-- Walk DWARF32 line units and retain the final unit's bounds.
@@ -1005,8 +991,8 @@ local function build_dwarf_line_section(existing, atom_table)
 	local unit_pos, last_pos, last_length, last_end = 0, nil, nil, nil
 	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)
+		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
@@ -1050,13 +1036,13 @@ local function build_dwarf_aranges_section(existing, atom_table)
 	-- We bump the unit's length field accordingly.
 	--
 	-- Unit structure (DWARF4 §7.21):
-	--   unit_length (4)
+	--   unit_length       (4)
-	--   version (2)
+	--   version           (2)
 	--   debug_info_offset (4)  -- CU DIE offset in .debug_info
-	--   address_size (1)
+	--   address_size      (1)
-	--   segment_size (1)
+	--   segment_size      (1)
-	--   entries... (4-byte addr + 4-byte length)
+	--   entries...        (4-byte addr + 4-byte length)
-	--   terminator (8 bytes: addr=0, length=0)
+	--   terminator        (8 bytes: addr=0, length=0)
 
 	-- Walk all units and emit each one (preserving existing structure).
 	-- For the LAST unit, replace the terminator with my entries + new term.
@@ -1067,7 +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
@@ -1531,13 +1517,12 @@ end
 ---        DW_AT_location  = piece-chain (DW_FORM_exprloc)
 ---        DW_AT_type      = ref4 → structure_type DIE
 ---
---- This function does NOT emit the final 0 byte (root terminator). build_debug_info_section splices bytes ahead of the root terminator
+--- This function does NOT emit the final 0 byte (root terminator).
---- and preserves existing DIE bytes exactly.
+--- build_debug_info_section splices bytes ahead of the root terminator and preserves existing DIE bytes exactly.
 ---
 --- ref4 basis: DW_FORM_ref4 is CU-relative (offset from the first byte of the CU header).
 --- Our inserted DIEs live in the main CU, so every ref4 = (target section offset) - main_cu_offset.
 --- Per-die section offsets are tracked via the running `next_offset` cursor (= section offset of the NEXT byte to emit).
----
 --- @param main_cu_offset integer    -- 0-based section offset of the main CU's unit_length field
 --- @param main_cu_end_excl integer  -- 0-based section offset of the first byte AFTER the main CU
 --- @param atom_table table[]        -- atoms (with atom.rbind set if rbind; atom.invocations set if mac_X(...) calls)
@@ -1657,6 +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 },
@@ -1793,7 +1780,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	emit(uleb128(ABBREV_TYPED_VIEW_POINTER))  -- DW_TAG_pointer_type (abbrev 110; NOT 9; void chain target)
 	emit(elf_dwarf.write_u32_le(ref4_of(void_chain_offset)))  -- 4-byte ref4: points at the void base_type's tag byte
 	-- type_chain_offsets["void|1"] is what step (f) of the per-RR_<R_Name> chain looks up.
-	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type. The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
+	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type.
+	-- The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
 	-- Capture the pointer_type's offset (the last-thing-emitted DIE start) and overwrite the lookup.
 	-- The pointer_type was emitted as: uleb(9) (1 byte) + 4-byte ref4 = 5 bytes. Its tag byte is at void_chain_offset + 8 (the base_type's 8 bytes: 1 tag + 5 name + 1 byte_size + 1 encoding).
 	local ptr_void_offset = void_chain_offset + 8
@@ -1909,12 +1897,13 @@ 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]
 		if this_ctx and this_ctx.rbind_atom then
 			local rbind = atom_by_name_global[this_ctx.rbind_atom]
-			if rbind and rbind.rbind and rbind.rbind.fields then
+			if    rbind and rbind.rbind and rbind.rbind.fields then
 				atom_view_ctx_fields = {}
 				for _, f in ipairs(rbind.rbind.fields) do atom_view_ctx_fields[f.name] = f end
 				if rbind.rbind.regs then
@@ -1934,11 +1923,11 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		end
 		if my_phase_label then
 			local group = (registries.atom_phases or {})[my_phase_label]
-			if group and group.atoms then
+			if    group and group.atoms then
 				for _, group_atom_name in ipairs(group.atoms) do
 					if group_atom_name ~= atom.name then
 						local cand = atom_by_name_global[group_atom_name]
-						if cand and cand.rbind and cand.rbind.fields then
+						if    cand and cand.rbind and cand.rbind.fields then
 							atom_view_phase_fields = {}
 							for _, f in ipairs(cand.rbind.fields) do atom_view_phase_fields[f.name] = f end
 							if cand.rbind.regs then
@@ -1959,7 +1948,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			-- (a) per-atom callsite atom_type(R_X, <T>): most specific; user explicit override for THIS atom only.
 			function(r_name, alias_code)
 				local override = atom_view and atom_view.reg_type_overrides and atom_view.reg_type_overrides[r_name]
-				if override and override.pointer_depth and override.pointer_depth > 0 then
+				if    override and override.pointer_depth and override.pointer_depth > 0 then
 					return type_chain_offsets[override.type_name .. "|" .. override.pointer_depth]
 				end
 			end,
@@ -1967,7 +1956,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local ctx_field_name = reg_to_field_ctx and reg_to_field_ctx[alias_code]
 				local ctx_f          = ctx_field_name and atom_view_ctx_fields and atom_view_ctx_fields[ctx_field_name]
-				if ctx_f and ctx_f.pointer_depth and ctx_f.pointer_depth > 0 then
+				if    ctx_f and ctx_f.pointer_depth and ctx_f.pointer_depth > 0 then
 					return type_chain_offsets[ctx_f.type_name .. "|" .. ctx_f.pointer_depth]
 				end
 			end,
@@ -1975,7 +1964,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local field_name = reg_to_field[alias_code]
 				local f         = field_name and field_type_by_name[field_name]
-				if f and f.pointer_depth and f.pointer_depth > 0 then
+				if    f and f.pointer_depth and f.pointer_depth > 0 then
 					return type_chain_offsets[f.type_name .. "|" .. f.pointer_depth]
 				end
 			end,
@@ -1983,12 +1972,13 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 			function(r_name, alias_code)
 				local phase_field_name = reg_to_field_phase and reg_to_field_phase[alias_code]
 				local phase_f          = phase_field_name and atom_view_phase_fields and atom_view_phase_fields[phase_field_name]
-				if phase_f and phase_f.pointer_depth and phase_f.pointer_depth > 0 then
+				if    phase_f and phase_f.pointer_depth and phase_f.pointer_depth > 0 then
 					return type_chain_offsets[phase_f.type_name .. "|" .. phase_f.pointer_depth]
 				end
 			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
@@ -1997,8 +1987,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 
 		-- Iterate `by_alias` in sorted order; Lua's pairs() is non-deterministic, so sorting ensures byte-identical DWARF output across builds.
 		for _, r_name in ipairs(by_alias_order) do
-			local alias = by_alias[r_name]
+			local alias      = by_alias[r_name]
-			local rr_name = "RR_" .. strip_r_prefix(r_name)
+			local rr_name    = "RR_" .. strip_r_prefix(r_name)
 			local alias_code = alias.code
 			emit(uleb128(ABBREV_VARIABLE))
 			emit(rr_name .. "\0") -- DW_FORM_string (DW_AT_name)
@@ -2020,7 +2010,7 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 		-- Two PC ranges cover every field: [atom.addr, last_load+8) describes each field as tape memory (DW_OP_bregN + offset) piece,
 		-- and [last_load+8, atom.end) describes each field as a GPR (DW_OP_regN) piece.
 		if atom.rbind then
-			local binds_name = atom.rbind.binds
+			local binds_name      = atom.rbind.binds
 			local loclists_offset = loclists_offsets[atom.name] or 0
 			emit(uleb128(ABBREV_BIND_VAR_LOCLIST))
 			emit("bind_args\0")                  -- DW_FORM_string (DW_AT_name)
@@ -2073,7 +2063,7 @@ end
 ---
 --- Fails safely by returning existing sections unchanged if the table walker can't find the table terminator (malformed input).
 ---
---- @param existing string  -- existing .debug_abbrev bytes, byte-for-byte
+--- @param existing           string   -- existing .debug_abbrev bytes, byte-for-byte
 --- @param main_abbrev_offset integer  -- 0-based offset into `existing` of the main CU's abbrev table
 --- @return string, integer  -- (new_abbrev_bytes, offset_where_duplicate_table_starts = #existing)
 local function build_debug_abbrev_section(existing, main_abbrev_offset)
@@ -2091,7 +2081,7 @@ local function build_debug_abbrev_section(existing, main_abbrev_offset)
 end
 
 --- Build the new .debug_str: existing strings + new strings appended.
---- @param existing string  -- existing .debug_str bytes, byte-for-byte
+--- @param existing   string  -- existing .debug_str bytes, byte-for-byte
 --- @param atom_table table[]
 --- @param registries table  -- merged registries from collect_per_source_registries
 --- @return string  -- existing bytes plus the deterministic appended strings
@@ -2101,7 +2091,6 @@ 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.
@@ -2111,14 +2100,14 @@ end
 ---
 --- The crt CU (everything before main_cu_start) is preserved.
 
---- @param existing string            -- existing .debug_info section bytes
+--- @param existing          string   -- existing .debug_info section bytes
---- @param main_cu_start integer      -- 0-based offset of the main CU's unit_length field
+--- @param main_cu_start     integer  -- 0-based offset of the main CU's unit_length field
---- @param main_cu_end_excl integer   -- 0-based offset of the first byte AFTER the main CU
+--- @param main_cu_end_excl  integer  -- 0-based offset of the first byte AFTER the main CU
 --- @param new_abbrev_offset integer  -- 0-based offset into the new .debug_abbrev of the duplicate main table
---- @param atom_table table[]
+--- @param atom_table        table[]
---- @param rbind_structs table     -- {[binds_name] = {bytes, fields, atom_names}}
+--- @param rbind_structs     table    -- {[binds_name] = {bytes, fields, atom_names}}
---- @param loclists_offsets table  -- {[atom_name] = section-relative offset}
+--- @param loclists_offsets  table    -- {[atom_name] = section-relative offset}
---- @param registries table        -- merged registries from collect_per_source_registries
+--- @param registries        table    -- merged registries from collect_per_source_registries
 --- @return string  -- the rebuilt .debug_info bytes
 local function build_debug_info_section(existing, main_cu_start, main_cu_end_excl, new_abbrev_offset, atom_table, rbind_structs, loclists_offsets, registries)
 	-- 1) Build the inserted children bytes (just before the main CU's root terminator).
@@ -2176,13 +2165,13 @@ local SECTION_WRITERS = {
 -- Write a list of `{name, data}` section records to disk via SECTION_WRITERS.
 -- @param results table[]  -- list of `{name=, data=}` records to write
 -- @param ctx PassCtx
--- @param basename string   -- output file basename (e.g. "hello_gte")
+-- @param basename string  -- output file basename (e.g. "hello_gte")
--- @return table             -- list of {name_bin = path} entries to append to M.run's outputs
+-- @return table           -- list of {name_bin = path} entries to append to M.run's outputs
 local function write_sections(results, ctx, basename)
 	local outputs = {}
 	for _, r in ipairs(results) do
 		local path = SECTION_WRITERS[r.name](ctx.out_root, basename)
-		local f = io.open(path, "wb")
+		local  f   = io.open(path, "wb")
 		if not f then
 			io.stderr:write(string.format("[dwarf_injection] failed to open %s for write\n", path))
 		else
@@ -2209,7 +2198,7 @@ function M.run(ctx)
 	end
 
 	-- Guard: --elf is required.
-	local elf_path = ctx.flags and ctx.flags.elf_path
+	local  elf_path = ctx.flags and ctx.flags.elf_path
 	if not elf_path or elf_path == "" then
 		io.stderr:write("[dwarf_injection] --elf flag missing\n")
 		return { outputs = {}, errors = {}, warnings = {} }
@@ -2220,7 +2209,8 @@ 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",
@@ -2235,12 +2225,11 @@ function M.run(ctx)
 	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 corpus     = (ctx.shared and ctx.shared.corpus) or {}
 	local registries = collect_per_source_registries(corpus)
-	-- Read nm symbols (the ONLY disk-side input to the atom table) and join
+	-- Read nm symbols (the ONLY disk-side input to the atom table) and join them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
-	-- them against `corpus.atoms_by_name` + `atom.paths` for word rows + invocation ancestry.
 	-- Disk source-map/provenance text is not consulted (those are diagnostic artifacts; semantic inputs are in memory).
-	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local addrs      = elf_dwarf.read_nm(ctx.flags.elf_path)
 	local atom_table = build_atom_table(corpus, addrs)
 
 	-- Detect rbind atoms + index Binds_* struct fields from the corpus.
@@ -2264,7 +2253,7 @@ function M.run(ctx)
 		duffle.ensure_dir(ctx.out_root)
 
 		-- Step 0: layout validation. Bail out safely if the .debug_info layout doesn't match what we expect (crt CU + DWARF5 main CU + final 0 byte).
-	-- A layout mismatch means the gcc emission changed; the safest response is to leave existing sections unchanged and emit no synthetic data, so the build's debug-info step never silently produces broken DWARF.
+		-- A layout mismatch means the gcc emission changed; the safest response is to leave existing sections unchanged and emit no synthetic data, so the build's debug-info step never silently produces broken DWARF.
 		local existing_info   = existing_sections[".debug_info"]   or ""
 		local existing_abbrev = existing_sections[".debug_abbrev"] or ""
 		local main_cu_start, main_cu_end_excl, main_abbrev_offset = find_main_cu_layout(existing_info)
@@ -2299,7 +2288,7 @@ function M.run(ctx)
 		local new_info  = build_debug_info_section(existing_info, main_cu_start, main_cu_end_excl, new_abbrev_offset, atom_table, rbind_structs, loclists_offsets, registries)
 
 		-- Step 2b: rebuild .debug_str now that we know which RR_<R_Name> entries get emitted.
-	-- This aligns with build_debug_info_section's by_alias loop.
+		-- This aligns with build_debug_info_section's by_alias loop.
 		local new_str = build_debug_str_section(existing_sections[".debug_str"] or "", atom_table, registries)
 
 		-- Step 3-5: independent sections.

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.
@@ -59,10 +58,9 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	-- `root_body_line` is the physical source line of the ATOM HEADER byte containing the opening `{`; that byte is one byte BEFORE `atom_record.body_off`.
 	-- The walker assigns line 2 to the body's first content line because line 1 is the trailing `\n` after `{`. Body-text line k therefore maps to `root_body_line + (k - 1)`.
 	-- `body_off - 1` points at the opening `{`, whose line index identifies the header line. `body_off` points after `{` and would shift every word row forward by one line.
-	local root_body_line = root_line_of(atom_record.body_off - 1)
+	local root_body_line  = root_line_of(atom_record.body_off - 1) or atom_record.line or 0
-		or atom_record.line or 0
 	local component_index = corpus.component_body_index or {}
-	local word_items = {}
+	local word_items      = {}
 
 	for _, item in ipairs(projection.items) do
 		if item.kind == "word" then word_items[#word_items + 1] = item end
@@ -102,7 +100,7 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	end
 
 	-- Normalize `inv.call_line` to a physical source line.
-	--   * ROOT invocations (`parent_id == 0`) carry body-relative `call_line` values from `M.LineIndex(body_text)`; convert them once with `root_body_line`.
+	--   * ROOT  invocations (`parent_id == 0`) carry body-relative `call_line` values from `M.LineIndex(body_text)`; convert them once with `root_body_line`.
 	--   * INNER invocations (`parent_id ~= 0`) carry physical `call_line` values from the component's `line_of`; retain them unchanged.
 	for _, inv in ipairs(projection.invocations) do
 		if inv.parent_id == 0 then
@@ -114,14 +112,14 @@ local function stamp_root_provenance(projection, atom_record, src, corpus)
 	-- `atoms_source_map` and `dwarf_injection` read `inv.body_lines[k]` directly from the invocation record created here.
 	-- Component words already carry physical `item.line` values from the walker's COMPONENT line index, so `body_line_for` returns them unchanged.
 	for _, inv in ipairs(projection.invocations) do
-		local sw = inv.start_word
+		local sw  = inv.start_word
-		local ew = inv.end_word
+		local ew  = inv.end_word
 		local bls = {}
 		for i = sw, ew do
 			local it = projection.items and projection.items[i]
-			if it and it.kind == "word" then
+			if    it and it.kind == "word" then
 				local fake_event = { invocation_ids = { inv.id } }
-				bls[#bls + 1] = body_line_for(fake_event, it) or 0
+				bls[#bls + 1]    = body_line_for(fake_event, it) or 0
 			end
 		end
 		inv.body_lines = bls
@@ -190,11 +188,11 @@ function M.run(ctx)
 	if type(corpus.source_order) ~= "table" then error("emission_model: ctx.shared.corpus.source_order is required", 0) end
 
 	-- Project once, collect errors + warnings for one atom.
-	-- Kind must be one of: atom | raw_atom | comp_bare | comp_proc.
+	-- Kind must be one of: atom | atom_proc | raw_atom | comp_bare | comp_proc.
 	local function process_atom(atom, src)
 		if not (atom and atom.body) then return end
 		local kind = atom.kind
-		if kind ~= "atom" and kind ~= "raw_atom" and kind ~= "comp_bare" and kind ~= "comp_proc" then
+		if kind ~= "atom" and kind ~= "atom_proc" and kind ~= "raw_atom" and kind ~= "comp_bare" and kind ~= "comp_proc" then
 			return
 		end
 		local proj = project_atom(atom, src, corpus)
@@ -217,7 +215,7 @@ function M.run(ctx)
 	end
 
 	-- Walk `corpus.source_order`; within each source, visit atoms followed by raw_atoms.
-	-- Recognized kinds (atom | raw_atom | comp_bare | comp_proc) each receive the atom.paths projection via duffle.project_emission.
+	-- Recognized kinds (atom | atom_proc | raw_atom | comp_bare | comp_proc) each receive the atom.paths projection via duffle.project_emission.
 	-- Components are macros inlined into atom bodies; focused tests and isolated component analyses consume atom.paths directly.
 	for _, src in ipairs(corpus.source_order) do
 		local scan = src.scan or {}

scripts/passes/offsets.lua

@@ -3,12 +3,19 @@
 --- 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.
+---
+--- (Task 12.16 note: atom-namespaced enum names — e.g., `atom_offset__normalize_v3s4__srav_path__aligned_done` —
+--- were considered to prevent cross-atom label collisions, but the C-side `atom_offset(F, T)` macro in
+--- `code/duffle/dsl.atom.h` doesn't know the current atom_name at expansion time, so any namespacing
+--- on the metaprogram side breaks the C build. Reverted. The C-side would need a per-atom
+--- `CURRENT_ATOM` #define (set by `MipsAtom_`/`MipsAtom_Proc_` macros) plus an updated `atom_offset`
+--- macro that uses it. That's a coordinated refactor — deferred to a future track.)
 ---
 --- 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
@@ -16,12 +23,11 @@
 
 -- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -39,35 +45,35 @@ local OFFSET_MACRO_COL = 44
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
+--- @field scan     table   -- Pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table -- Cross-pass shared state
---- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared.corpus      table -- Corpus projection
 --- @field shared.word_counts table
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string -- Output root (e.g. "build/gen")
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
---- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field errors   table[]  -- {line=, msg=}  entries; build-stops
---- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+--- @field warnings table[]  -- {line=, msg=}  entries; build-succeeds
 
 --- @class BranchOffset
---- @field tag              string      -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field tag               string      -- Marker tag (e.g. "F" in `atom_offset(F, T)`)
---- @field target           string      -- the target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field target            string      -- Target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field branch_word      integer     -- branch word position within the atom body
+--- @field branch_word       integer     -- Branch word position within the atom body
---- @field offset           integer     -- computed per consuming instruction (see `compute_offsets`)
+--- @field offset            integer     -- Computed per consuming instruction (see `compute_offsets`)
---- @field consuming_encoder string|nil -- the instruction consuming the offset (e.g. "branch_le_zero", "jump", "call_addr")
+--- @field consuming_encoder string|nil  -- Instruction consuming the offset (e.g. "branch_le_zero", "jump", "call_addr")
 --- @field consuming_arg_pos integer|nil -- 1-based arg position within the consuming instruction's arg list
 
 --- @class AtomData
---- @field name        string         -- atom name
+--- @field name        string         -- Atom name
---- @field total_words integer        -- total word count of the atom body
+--- @field total_words integer        -- Total word count of the atom body
---- @field offsets     BranchOffset[] -- per-branch offset list
+--- @field offsets     BranchOffset[] -- Per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Canonical marker projection
@@ -99,7 +105,7 @@ local function project_markers(markers)
 	local state = { labels = {}, branches = {} }
 	for _, marker in ipairs(markers or {}) do
 		local project = MARKER_PROJECTORS[marker.kind]
-		if project then project(state, marker) end
+		if    project then project(state, marker) end
 	end
 	return state.labels, state.branches
 end
@@ -109,7 +115,6 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- 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). 
@@ -121,7 +126,7 @@ end
 ---
 --- 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 labels   table<string, integer>
 --- @param branches table[]
 --- @return BranchOffset[]
 local function compute_offsets(labels, branches)
@@ -195,44 +200,48 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
---- Generate the per-source .offsets.h header.
+--- Generate the per-directory .offsets.h header.
---- @param source_path string
+--- @param dir          string     -- the absolute source directory
---- @param atoms_data  AtomData[]
+--- @param sources      table[]    -- sources contributing to this directory (for the header comment)
+--- @param atoms_data   AtomData[]
 --- @return string
-local function generate_header(source_path, atoms_data)
+local function generate_header(dir, sources, atoms_data)
-	local basename = duffle.basename_no_ext(source_path)
+	local dir_basename = duffle.basename_no_ext(dir)
 
 	local lines = {}
 	local function add(s) lines[#lines + 1] = s end
 
 	add("// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT")
-	add("// Source: " .. source_path)
+	add("// Directory: " .. dir:gsub("/", "\\") .. "\\")
+	for _, src in ipairs(sources) do
+		add("//   source: " .. src.path:gsub("/", "\\"))
+	end
 	add("#pragma once")
 	add("")
-	add("#pragma region " .. basename)
+	add("#pragma region " .. dir_basename)
 	add("")
 	add("")
 	for _, atom in ipairs(atoms_data) do
 		emit_atom_offsets(add, atom)
 	end
-	add("#pragma endregion " .. basename)
+	add("#pragma endregion " .. dir_basename)
 	add("")
 	return table.concat(lines, "\n") .. "\n"
 end
 
 local M = {}
 
---- (internal) Process one source: render offsets from canonical atom paths.
+--- (internal) Aggregate atoms from every source in one directory, render the per-directory `offsets.h`.
 --- Returns the offsets_h path if a header was written, or nil.
---- @param ctx PassCtx
+--- @param 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
+		local  paths = atom and atom.paths
 		if not paths then return end
 		local labels, branches = project_markers(paths.markers)
 		atoms_data[#atoms_data + 1] = {
@@ -242,19 +251,22 @@ local function process_source(ctx, src)
 		}
 	end
 
-	for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+	for _, src in ipairs(sources) do
-	for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+		local scan = src.scan or {}
+		for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+		for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	end
 	if #atoms_data == 0 then return nil end
 
-	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
+	local out_path = dir .. "/gen/offsets.h"
 	duffle.ensure_dir(duffle.dirname(out_path))
-	duffle.write_file(out_path, generate_header(src.path:gsub("/", "\\"), atoms_data))
+	duffle.write_file(out_path, generate_header(dir, sources, atoms_data))
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
+--- For each canonical source-directory, emits a per-directory `gen/offsets.h`
---- containing constants for every marker recorded in atom.paths.
+--- containing constants for every marker recorded in atom.paths across every source in that directory.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -263,12 +275,17 @@ function M.run(ctx)
 	local warnings = {}
 
 	local corpus = ctx.shared and ctx.shared.corpus
-	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+	if type(corpus) ~= "table" then
-		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+		error("offsets.run requires ctx.shared.corpus", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order.", 0)
 	end
 
-	for _, src in ipairs(corpus.source_order) do
+	-- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`.
-		local out_path = process_source(ctx, src)
+	local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
+	for dir, sources in pairs(sources_by_dir) do
+		local out_path = process_directory(ctx, dir, sources)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }
 		end

scripts/passes/report.lua

@@ -1,22 +1,21 @@
---- passes/report.lua — Per-MODULE annotation report renderer +
+--- passes/report.lua — Per-MODULE annotation report renderer + project-wide summary writer.
---- project-wide summary writer.
 ---
 --- Two output files per build:
 ---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
 ---   - `build/gen/annotation_validation.txt` — the project summary.
 ---
 --- The annotation pass emits `errors.h` files per module and the canonical `corpus.sources_by_dir` projection groups sources by directory.
---- This pass iterates the canonical dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
+--- This pass iterates the dir projection directly and re-validates each source via `annotation.validate()` to get the detailed per-source results.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
--- Bootstrap: see `ps1_meta.lua` for the rationale.
+-- Bootstrap: See `ps1_meta.lua` for the rationale.
--- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
+-- Bootstrap: Load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
 -- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
+-- Bootstrap: Load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
@@ -37,7 +36,7 @@ local atoms_source_map = dofile(_bootstrap_dir .. "atoms_source_map.lua")
 
 -- Section separators used in the rendered text reports.
 -- The thin rules are hand-tuned to align with the per-section content width; do not change without also checking the section renderers below.
-local RULE_THICK = "========================================================"
+local RULE_THICK               = "========================================================"
 local SECTION_HEADER_ATOMS     = "── Atoms ────────────────────────────────────────────────"
 local SECTION_HEADER_ANNOTS    = "── Annotations ──────────────────────────────────────────"
 local SECTION_HEADER_BINDS     = "── Binds_* structs ──────────────────────────────────────"
@@ -59,83 +58,83 @@ local PASS_NAME = "report"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field sources            SourceFile[]         -- all source files in the build
+--- @field sources            SourceFile[]         -- All source files in the build
---- @field metadata_path      string               -- path to word_count.metadata.h
+--- @field metadata_path      string               -- Path to word_count.metadata.h
---- @field shared             table                -- cross-pass shared state
+--- @field shared             table                -- Cross-pass shared state
---- @field out_root           string               -- output root (e.g. "build/gen")
+--- @field out_root           string               -- Output root (e.g. "build/gen")
---- @field project_root       string               -- project root (e.g. "code/")
+--- @field project_root       string               -- Project root (e.g. "code/")
---- @field upstream           table<string, table> -- per-pass upstream outputs
+--- @field upstream           table<string, table> -- Per-pass upstream outputs
 --- @field flags              table                -- CLI flags + per-pass stash
---- @field verbose            boolean              -- if true, log diagnostic info
+--- @field verbose            boolean              -- If true, log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
---- @field errors   table[]  -- {line=, msg=} entries; build-stops
+--- @field errors   table[]  -- {line=, msg=}  entries; build-stops
---- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
+--- @field warnings table[]  -- {line=, msg=}  entries; build-succeeds
 
 -- Shapes produced by `passes/annotation.lua`'s `M.validate()`.
 
 --- @class AtomEntry
---- @field name string   -- atom name (e.g. "cube_g4_face")
+--- @field name string   -- Atom name (e.g. "cube_g4_face")
---- @field line integer  -- source line of the atom declaration
+--- @field line integer  -- Source line of the atom declaration
 
 --- @class AnnotEntry
---- @field line    integer  -- source line
+--- @field line    integer     -- Source line
---- @field macro   string   -- the macro name (e.g. "atom_reads")
+--- @field macro   string      -- Macro name (e.g. "atom_reads")
---- @field name    string   -- the atom name (if a `name(...)` was given)
+--- @field name    string      -- Atom name (if a `name(...)` was given)
---- @field kind    string   -- "atom_info" | "atom_bind" | ...
+--- @field kind    string      -- "atom_info" | "atom_bind" | ...
 --- @field binds   string|nil  -- Binds_X name if any
 --- @field reads   string[]    -- R_* names (read targets)
 --- @field writes  string[]    -- R_* names (write targets)
---- @field error   string|nil  -- error message if annotation was malformed
+--- @field error   string|nil  -- Error message if annotation was malformed
 
 --- @class BindsField
---- @field name   string   -- field name
+--- @field name   string   -- Field name
---- @field offset integer  -- byte offset within the Binds_X struct
+--- @field offset integer  -- Byte offset within the Binds_X struct
 
 --- @class BindsStruct
---- @field name   string         -- struct name (e.g. "Binds_Floor")
+--- @field name   string         -- Struct name (e.g. "Binds_Floor")
---- @field line   integer        -- source line of the typedef
+--- @field line   integer        -- Source line of the typedef
---- @field bytes  integer        -- total byte size
+--- @field bytes  integer        -- Total byte size
---- @field fields BindsField[]   -- the field list
+--- @field fields BindsField[]   -- The field list
 
 --- @class MacroEntry
---- @field name  string   -- macro name (e.g. "WORD_COUNT(my_macro, 4)")
+--- @field name  string   -- Macro name (e.g. "WORD_COUNT(my_macro, 4)")
---- @field line  integer  -- source line
+--- @field line  integer  -- Source line
---- @field words integer  -- declared word count
+--- @field words integer  -- Declared word count
 
 --- @class Finding
---- @field line integer  -- source line
+--- @field line integer  -- Source line
---- @field msg  string   -- finding message
+--- @field msg  string   -- Finding message
 
 --- @class AnnotationResult
---- @field source   string         -- set by this pass; original source path
+--- @field source   string         -- Set by this pass; original source path
---- @field atoms    AtomEntry[]    -- atom declarations in this source
+--- @field atoms    AtomEntry[]    -- Atom declarations in this source
---- @field annots   AnnotEntry[]   -- annotation entries
+--- @field annots   AnnotEntry[]   -- Annotation entries
---- @field macros   MacroEntry[]   -- macro word-count declarations
+--- @field macros   MacroEntry[]   -- Macro word-count declarations
 --- @field binds    BindsStruct[]  -- Binds_* struct declarations
---- @field errors   Finding[]      -- errors from validation
+--- @field errors   Finding[]      -- Errors from validation
---- @field warnings Finding[]      -- warnings from validation
+--- @field warnings Finding[]      -- Warnings from validation
---- @field info     table          -- info summary (not rendered here)
+--- @field info     table          -- Info summary (not rendered here)
 
 --- @class ModuleEntry
---- @field dir          string   -- absolute directory path
+--- @field dir          string   -- Absolute directory path
---- @field dir_basename string   -- basename (e.g. "duffle", "gte_hello")
+--- @field dir_basename string   -- Basename (e.g. "duffle", "gte_hello")
---- @field atoms_count  integer  -- pre-counted atoms for filtering
+--- @field atoms_count  integer  -- Pre-counted atoms for filtering
 
 --- @class ModuleReport
---- @field dir       string             -- module directory
+--- @field dir       string             -- Module directory
---- @field sources   SourceFile[]       -- sources in this module
+--- @field sources   SourceFile[]       -- Sources in this module
---- @field results   AnnotationResult[] -- per-source validate() results
+--- @field results   AnnotationResult[] -- Per-source validate() results
 
 --- @class ProjectReport
---- @field results AnnotationResult[]  -- all per-source results
+--- @field results AnnotationResult[]  -- All per-source results
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-MODULE annotation report (aggregated across all sources in a dir)
@@ -153,9 +152,16 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- Render the thin project-wide summary (`build/atom_meta_report.summary.md`).
---- @param all_results { module:string, atoms:integer, annots:integer, binds:integer,
+--- @param all_results { 
----                       macros:integer, findings:integer, errors:integer,
+--- 	module:string,
----                       warnings:integer, info:integer }[]
+--- 	atoms:integer, 
+--- 	annots:integer, 
+--- 	binds:integer,
+---		macros:integer,   
+---		findings:integer,
+---		errors:integer,
+--- 	warnings:integer,
+--- 	info:integer }[]
 --- @return string
 local function render_project_summary(all_results)
 	local lines = {
@@ -165,13 +171,10 @@ local function render_project_summary(all_results)
 		"| module | atoms | annots | binds | macros | findings | errors | warnings | info |",
 		"|--------|-------|--------|-------|--------|----------|--------|----------|------|",
 	}
-	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0,
+	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0, findings = 0, errors = 0, warnings = 0, info = 0 }
-		findings = 0, errors = 0, warnings = 0, info = 0 }
 	for _, e in ipairs(all_results) do
-		lines[#lines + 1] = string.format(
+		lines[#lines + 1] = string.format("| %s | %d | %d | %d | %d | %d | %d | %d | %d |"
-			"| %s | %d | %d | %d | %d | %d | %d | %d | %d |",
+			, e.module, e.atoms, e.annots, e.binds, e.macros, e.findings, e.errors, e.warnings, e.info)
-			e.module, e.atoms, e.annots, e.binds, e.macros,
-			e.findings, e.errors, e.warnings, e.info)
 		totals.atoms    = totals.atoms    + e.atoms
 		totals.annots   = totals.annots   + e.annots
 		totals.binds    = totals.binds    + e.binds
@@ -181,23 +184,21 @@ local function render_project_summary(all_results)
 		totals.warnings = totals.warnings + e.warnings
 		totals.info     = totals.info     + e.info
 	end
-	lines[#lines + 1] = string.format(
+	lines[#lines + 1] = string.format("| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |"
-		"| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |",
+		, totals.atoms, totals.annots, totals.binds, totals.macros, totals.findings, totals.errors, totals.warnings, totals.info)
-		totals.atoms, totals.annots, totals.binds, totals.macros,
-		totals.findings, totals.errors, totals.warnings, totals.info)
 	return table.concat(lines, "\n") .. "\n"
 end
 
 --- 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         string
 --- @param dir_sources SourceFile[]
---- @param wc table<string, integer>
+--- @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 = {
+	local lines        = {
 		"# " .. dir_basename .. " — atoms (verbose source map)",
 		"> Per-word call-site + provenance. Auto-generated.",
 		"",
@@ -249,14 +250,14 @@ end
 --- 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           string
---- @param dir_sources SourceFile[]
+--- @param dir_sources   SourceFile[]
 --- @param annot_results AnnotationResult[]
---- @param sa_results table                 -- corpus.static_analysis_results[dir_basename]
+--- @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 = {
+	local lines        = {
 		"# " .. dir_basename .. " — atom meta report",
 		"> Auto-generated by ps1_meta.lua (passes/report.lua). Do not edit.",
 		"",
@@ -326,8 +327,8 @@ local function render_module_meta_report(dir, dir_sources, annot_results, sa_res
 				local binds  = a.binds  or "—"
 				local reads  = (#a.reads  > 0 and table.concat(a.reads,  ",")) or "—"
 				local writes = (#a.writes > 0 and table.concat(a.writes, ",")) or "—"
-				add(string.format("| %s | %d | %s | %s | %s | %s |",
+				add(string.format("| %s | %d | %s | %s | %s | %s |"
-					src_name, a.line, a.name, binds, reads, writes))
+					, src_name, a.line, a.name, binds, reads, writes))
 			end
 		end
 	end
@@ -418,7 +419,7 @@ local function render_module_meta_report(dir, dir_sources, annot_results, sa_res
 	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   = ""
+		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, ", ") .. "]"
@@ -426,7 +427,7 @@ local function render_module_meta_report(dir, dir_sources, annot_results, sa_res
 		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))
+			p.branches   or 0, p.paths      or 0, notes))
 	end
 	add("")
 

scripts/passes/scan_source.lua

@@ -2,8 +2,8 @@
 ---
 --- 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)
+---   MipsAtom_Proc_         (kind = "atom_proc", body inside last {})
 ---   MipsAtomComp_          (kind = "comp_bare")
 ---   MipsAtomComp_Proc_     (kind = "comp_proc", body inside last {})
 ---   atom_dbg_skip          — bare whole-atom/component debug-step marker; following declaration disambiguates
@@ -14,8 +14,6 @@
 --- 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).
@@ -23,7 +21,7 @@
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when required).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- Forward declarations for helpers used by earlier parsers (parse_enum_body_fields needs parse_enum_int_literal;
 -- parse_typedef_binds needs duffle.find_byte).
@@ -37,7 +35,7 @@ local parse_enum_int_literal
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceScan
---- @field atoms             AtomEntry[]     -- MipsAtom_ + MipsAtomComp_ + MipsAtomComp_Proc_
+--- @field atoms             AtomEntry[]     -- MipsAtom_ + MipsAtom_Proc_ + MipsAtomComp_ + MipsAtomComp_Proc_
 --- @field raw_atoms         AtomEntry[]     -- MipsCode code_<name> { body } (offsets pass only)
 --- @field binds             BindsEntry[]    -- typedef Struct_(Binds_X) { fields } (fields pre-parsed)
 --- @field atom_infos        AtomInfoEntry[] -- MipsAtom_(name) atom_info(...) (sub-calls pre-parsed)
@@ -50,15 +48,15 @@ local parse_enum_int_literal
 --- @field line_of           fun(pos: integer): integer       -- shared LineIndex closure
 
 --- @class DebugSkipMarker
---- @field marker_kind    string      -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive; any other ident reaches the unrelated fallback and is never associated with a declaration.
+--- @field marker_kind    string      -- Exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive; any other ident reaches the unrelated fallback and is never associated with a declaration.
---- @field marker_line    integer     -- line of the marker ident start
+--- @field marker_line    integer     -- Line of the marker ident start
---- @field marker_pos     integer     -- byte position of the marker ident start (the comment walker anchors here)
+--- @field marker_pos     integer     -- Byte position of the marker ident start (the comment walker anchors here)
 --- @field is_bare        boolean     -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
 --- @field has_parens     boolean     -- true iff a `(...)` follows the marker ident (diagnostic-only)
---- @field args           string|nil  -- trimmed args inside the `(...)` (nil when has_parens is false)
+--- @field args           string|nil  -- Trimmed args inside the `(...)` (nil when has_parens is false)
 --- @field pending        boolean     -- true while awaiting the following declaration
 --- @field superseded_by_marker_line integer|nil -- set when a newer marker bumped this one out of the pending slot
---- @field target_kind    string|nil  -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed (nil if no declaration ever followed)
+--- @field target_kind    string|nil  -- "atom" | "atom_proc" | "comp_bare" | "comp_proc" | "unrelated" once observed (nil if no declaration ever followed)
 --- @field proc_prelude   boolean|nil -- true after the marker crossed an `FI_` prelude and awaits `MipsAtomComp_Proc_`
 
 --- @class RegTypeDefault
@@ -71,15 +69,15 @@ 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 +86,11 @@ local parse_enum_int_literal
 --- @field info_line        integer                -- line of the atom_info call
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
---- @field scan     table   -- pre-scanned SourceScan payload (set by this pass)
+--- @field scan     table   -- Pre-scanned SourceScan payload (set by this pass)
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -111,15 +109,15 @@ local parse_enum_int_literal
 
 --- @class AtomEntry
 --- @field line              integer
---- @field name              string     -- atom name (for components: without ac_ prefix)
+--- @field name              string       -- Atom name (for components: without ac_ prefix)
---- @field body              string     -- brace-delimited body (without the braces)
+--- @field body              string       -- Brace-delimited body (without the braces)
---- @field body_off          integer    -- char offset of body[1] in source
+--- @field body_off          integer      -- Char offset of body[1] in source
---- @field kind              string     -- "atom" | "comp_bare" | "comp_proc" | "raw_atom"
+--- @field kind              string       -- "atom" | "atom_proc" | "comp_bare" | "comp_proc" | "raw_atom"
---- @field raw_name          string     -- un-stripped name (for components: with ac_ prefix)
+--- @field raw_name          string       -- Un-stripped name (for components: with ac_ prefix)
---- @field ident_pos         integer    -- position of the MipsAtom_/MipsAtomComp_ ident start
+--- @field ident_pos         integer      -- Position of the MipsAtom_/MipsAtomComp_ ident start
---- @field after_paren       integer    -- position past the closing paren
+--- @field after_paren       integer      -- Position past the closing paren
---- @field debug_skip        boolean    -- true when an `atom_dbg_skip` bare marker immediately precedes this declaration (sole-owner stamp; see push_debug_skip_marker)
+--- @field 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 +137,20 @@ local QUALIFIER_KEYWORDS = {
 local AC_PREFIX     = "ac_"
 local AC_PREFIX_LEN = 3
 
--- Strip the "ac_" prefix from a component name.
+-- The function-decl keyword that precedes a MipsAtomComp_Proc_ call.
--- Returns the input unchanged if it doesn't start with the prefix.
+-- Used by the backward walk in duffle.find_function_decl_for.
--- @param raw_name string
+local SLICE_MIPS_CODE    = "Slice_MipsCode"
--- @return string
+local SLICE_MIPS_CODE_LEN = #SLICE_MIPS_CODE
+
+-- The return type that precedes a MipsAtom_Proc_ function declaration.
+-- Used by the backward walk in duffle.find_atom_proc_decl_for.
+local MIPS_ATOM_PTR    = "MipsAtom*"
+local MIPS_ATOM_PTR_LEN = #MIPS_ATOM_PTR
+
+--- 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)
@@ -156,7 +164,7 @@ end
 local function push_debug_skip_marker(out, marker)
 	local markers = out.debug_skip_markers
 	local prior   = markers[#markers]
-	if prior and prior.pending then
+	if    prior and prior.pending then
 		prior.pending = false
 		prior.superseded_by_marker_line = marker.marker_line
 	end
@@ -178,25 +186,25 @@ end
 -- Returns (body, after_brace, body_off) on success, or (nil, fallback_pos) on no brace.
 -- `fallback_pos` defaults to `after_paren + 1` (the common "advance by 1" case).
 local function find_body_braces(source, after_paren, fallback)
-	local brace = duffle.scan_to_char(source, "{", after_paren)
+	local  brace = duffle.scan_to_char(source, "{", after_paren)
 	if not brace then return nil, fallback or (after_paren + 1) end
 	local body, after_brace = duffle.read_braces(source, brace)
 	return body, after_brace, brace + 1
 end
 
--- Walk backward from `start_pos` capturing contiguous `/* */` block(s) and
+--- Walk backward from `start_pos` capturing contiguous `/* */` block(s) and
--- `//` line(s) that immediately precede it. The caller (preceding_declaration_comment)
+--- `//` line(s) that immediately precede it. The caller (preceding_declaration_comment)
--- supplies `start_pos` so the walker does not need to detect marker shape or prelude layout.
+--- supplies `start_pos` so the walker does not need to detect marker shape or prelude layout.
--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
+--- The scanner already knows the marker_pos + decl ident_pos and threads that knowledge forward.
---
+---
--- The walker captures:
+--- The walker captures:
---   - Block comment close `*/` followed by walking back to `/*`.
+---   - Block comment close `*/` followed by walking back to `/*`.
---   - `//` line comments (the line containing the current non-ws position starts with `//`).
+---   - `//` line comments (the line containing the current non-ws position starts with `//`).
--- It stops at the first non-ws char that does not begin a comment block or line.
+--- It stops at the first non-ws char that does not begin a comment block or line.
--- Empty string if no comment is adjacent.
+--- Empty string if no comment is adjacent.
--- @param source string
+--- @param source    string
--- @param start_pos integer -- exclusive upper bound for the captured block
+--- @param start_pos integer -- exclusive upper bound for the captured block
--- @return string
+--- @return string
 local function preceding_comment_walk_backward(source, start_pos)
 	local pieces  = {}
 	local scan_pos = start_pos
@@ -249,13 +257,13 @@ local function preceding_comment_walk_backward(source, start_pos)
 	return table.concat(pieces, "\n")
 end
 
--- Resolve the start position for the declaration-comment walk.
+--- Resolve the start position for the declaration-comment walk.
--- When a debug-skip marker is pending, the walker must start from the position immediately before the marker ident
+--- When a debug-skip marker is pending, the walker must start from the position immediately before the marker ident
--- (so it walks backward past the marker text and any `FI_ MipsAtom ac_X(args)` proc-prelude layout — neither of which is visible if we start from the declaration ident_pos).
+--- (so it walks backward past the marker text and any `FI_ MipsAtom ac_X(args)` proc-prelude layout — neither of which is visible if we start from the declaration ident_pos).
--- When no marker is pending, the walker starts from the declaration ident_pos directly.
+--- When no marker is pending, the walker starts from the declaration ident_pos directly.
--- @param pending_marker DebugSkipMarker|nil
+--- @param pending_marker DebugSkipMarker|nil
--- @param ident_pos integer -- declaration ident position
+--- @param ident_pos      integer -- declaration ident position
--- @return integer
+--- @return integer
 local function comment_walk_start(pending_marker, ident_pos)
 	if pending_marker then
 		return pending_marker.marker_pos - 1
@@ -263,16 +271,16 @@ local function comment_walk_start(pending_marker, ident_pos)
 	return ident_pos - 1
 end
 
--- Attach the pending marker to the next declaration.
+--- Attach the pending marker to the next declaration.
--- The declaration form disambiguates whole atoms from components; the resolved `debug_skip` is stamped directly on the declaration record
+--- The declaration form disambiguates whole atoms from components; the resolved `debug_skip` is stamped directly on the declaration record
--- (sole-owner discipline; see push_debug_skip_marker).
+--- (sole-owner discipline; see push_debug_skip_marker).
---
+---
--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
+--- A marker is POSITIVE (stamps `debug_skip = true` on the declaration) iff:
---   marker_kind == "atom_dbg_skip" AND is_bare == true
+---   marker_kind == "atom_dbg_skip" AND is_bare == true
--- Any other spelling or shape (parenthesized form, legacy name) is recorded as a raw marker for annotation validation but never stamps `debug_skip`.
+--- Any other spelling or shape (parenthesized form, legacy name) is recorded as a raw marker for annotation validation but never stamps `debug_skip`.
--- @param out SourceScan
+--- @param out         SourceScan
--- @param target_kind string|nil -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
+--- @param target_kind string|nil -- "atom" | "atom_proc" | "comp_bare" | "comp_proc" | "unrelated" once observed
--- @return boolean|nil -- true iff the marker is the positive bare form
+--- @return boolean|nil -- true iff the marker is the positive bare form
 local function attach_debug_skip_marker(out, target_kind)
 	local markers = out.debug_skip_markers
 	local marker  = markers[#markers]
@@ -390,13 +398,13 @@ local function walk_body_fields(body, build_field)
 	while body_pos <= body_len do
 		body_pos = duffle.skip_ws_and_cmt(body, body_pos)
 		if body_pos > body_len then break end
-		local first, first_end = duffle.read_ident(body, body_pos)
+		local  first, first_end = duffle.read_ident(body, body_pos)
 		if not first then
 			body_pos = body_pos + 1
 		else
-			local after_first = duffle.skip_ws_and_cmt(body, first_end)
+			local after_first     = duffle.skip_ws_and_cmt(body, first_end)
 			local result, new_pos = build_field(first, first_end, after_first)
-			if result then fields[#fields + 1] = result end
+			if    result then fields[#fields + 1] = result end
 			body_pos = new_pos or first_end
 			-- Skip a single trailing `,` or `;`.
 			if body_pos <= body_len and (body:sub(body_pos, body_pos) == "," or body:sub(body_pos, body_pos) == ";") then
@@ -442,7 +450,7 @@ local function parse_enum_body_fields(body)
 		local value
 		local new_pos
 		if body:sub(after_name, after_name) == "=" then
-			local val_pos = duffle.skip_ws_and_cmt(body, after_name + 1)
+			local val_pos    = duffle.skip_ws_and_cmt(body, after_name + 1)
 			local v, end_pos = parse_enum_int_literal(body, val_pos)
 			if v ~= nil then
 				value   = v
@@ -461,16 +469,16 @@ end
 -- Returns a positive integer byte_size when the chain bottoms out at a builtin, or nil if the chain is broken, exceeds TYPE_CHAIN_MAX_DEPTH, or contains a cycle.
 local function resolve_typedef_byte_size(type_name, type_name_registry, visited, depth)
 	if depth > TYPE_CHAIN_MAX_DEPTH then return nil end
-	if visited[type_name] then return nil end
+	if visited[type_name]           then return nil end
 	visited[type_name] = true
 
 	-- Check the builtin primitive map FIRST.
 	-- This handles undeclared builtin idents (e.g. `__UINT32_TYPE__` appears as underlying_type in `typedef __UINT32_TYPE__ TSet_(V4_S2);` 
 	-- even though the fixture never declares `__UINT32_TYPE__` itself).
 	local builtin = BUILTIN_BYTE_SIZES[type_name]
-	if builtin ~= nil then return builtin end
+	if    builtin ~= nil then return builtin end
 
-	local entry = type_name_registry[type_name]
+	local  entry = type_name_registry[type_name]
 	if not entry then return nil end
 
 	-- Confident: this entry was already resolved by the propagation pass (e.g., a builtin or a struct whose fields are all resolved).
@@ -519,9 +527,9 @@ local function propagate_type_sizes(out)
 		for name, entry in pairs(reg) do
 			if entry.byte_size == nil then
 				local resolved = resolve_typedef_byte_size(name, reg, {}, 1)
-				if resolved ~= nil then
+				if    resolved ~= nil then
 					entry.byte_size = resolved
-					any_change = true
+					any_change      = true
 				end
 			end
 		end
@@ -802,6 +810,11 @@ local BYTE_x          = 0x78   -- 'x'
 local BYTE_X          = 0x58   -- 'X'
 local BYTE_OPEN_BRACE = 0x7B   -- '{'
 local BYTE_CLOSE_BRACE= 0x7D   -- '}'
+local BYTE_SLASH      = 0x2F   -- '/'
+local BYTE_STAR       = 0x2A   -- '*'
+local BYTE_SPACE      = 0x20   -- ' '
+local BYTE_TAB        = 0x09   -- '\t'
+local BYTE_CR         = 0x0D   -- '\r'
 
 -- Maximum chain depth when resolving `R_*_Code` symbol RHS references.
 -- Eight hops is enough for any production chain (R_TapePtr_Code -> R_T8_Code -> ...).
@@ -825,10 +838,48 @@ local function hex_digit_value(b)
 	return nil
 end
 
+-- Read one trailing C-comment that appears immediately after `pos` in `body`,
+-- skipping horizontal whitespace and newlines first. Used by `parse_enum_entry` to
+-- recover the `atom_auto_reg:` / `phase_auto_reg:` scope annotation embedded by
+-- the `atom_auto_reg` / `phase_auto_reg` macros' RHS expansion
+-- (`R_<Sym> = R_<Sym>_Code /* atom_auto_reg: <scope> */`).
+-- Handles both block (`/* ... */`) and line (`// ...`) forms.
+-- Returns the comment text (without delimiters), or nil if no comment is adjacent.
+local function read_trailing_cmt_after(body, pos)
+	local body_len = #body
+	while pos <= body_len do
+		local b = body:byte(pos)
+		if b == BYTE_SPACE or b == BYTE_TAB or b == BYTE_NEWLINE or b == BYTE_CR then
+			pos = pos + 1
+		elseif b == BYTE_SLASH then
+			local b2 = body:byte(pos + 1)
+			if b2 == BYTE_STAR then
+				-- Block comment /* ... */
+				local i = pos + 2
+				while i < body_len do
+					if body:byte(i) == BYTE_STAR and body:byte(i + 1) == BYTE_SLASH then
+						return body:sub(pos + 2, i - 1)
+					end
+					i = i + 1
+				end
+				return nil  -- unterminated; treat as no comment
+			elseif b2 == BYTE_SLASH then
+				-- Line comment // ... (strip the trailing newline)
+				local end_pos = duffle.find_byte(body, BYTE_NEWLINE, pos + 2) or (body_len + 1)
+				return body:sub(pos + 2, end_pos - 1)
+			end
+			return nil
+		else
+			return nil
+		end
+	end
+	return nil
+end
+
 --- Parse a decimal/negative-decimal/hex integer literal starting at byte position `start`.
 --- Returns (value, end_pos) on success, or (nil, start) on failure / no match.
 --- Accepts: 12, -1, 0, 0x10, 0X1F, -0x10.
---- @param text string
+--- @param text  string
 --- @param start integer
 --- @return integer|nil, integer
 --- Implementation note: this is a plain assignment (not `local function`) 
@@ -952,9 +1003,9 @@ end
 --- Always saves the raw RHS text into `code_macro_bodies` (for cross-source fallback during chain resolution),
 --- then (if resolvable) stores the resolved integer code into `code_macros` keyed by the macro name.
 --- `directive_start` points at the `#` byte. The function is silent on non-matching directives, the caller skips the line in any case.
---- @param source string
+--- @param source          string
 --- @param directive_start integer  -- byte position of `#`
---- @param code_macros table        -- out._code_macros / ctx.shared._code_macros
+--- @param code_macros     table    -- out._code_macros / ctx.shared._code_macros
 --- @param code_macro_bodies table  -- out._code_macro_bodies / ctx.shared._code_macro_bodies
 local function try_extract_code_macro(source, directive_start, code_macros, code_macro_bodies)
 	local rest = duffle.skip_ws_and_cmt(source, directive_start + 1)
@@ -985,8 +1036,8 @@ end
 --- Populates `code_macros` with resolved integer codes AND `code_macro_bodies` with raw RHS text
 --- (used by the chain walker as cross-source fallback during pass 1b in `M.run`); ignores everything else.
 --- Used by `M.run` pass 1a to build the cross-source `_code_macros` + `_code_macro_bodies` registries before pass 1b resolves chains.
---- @param source      string
+--- @param source            string
---- @param code_macros table
+--- @param code_macros       table
 --- @param code_macro_bodies table
 local function scan_source_pre_pass(source, code_macros, code_macro_bodies)
 	local pos     = 1
@@ -1043,7 +1094,7 @@ local function parse_enum_atom_type_default(body, pos)
 	if pos > #body then return nil, 0, pos end
 	-- Bare `atom_type` word with word-bounding on both sides.
 	local ident, ident_end = duffle.read_ident(body, pos)
-	if ident ~= "atom_type" then return nil, 0, pos end
+	if    ident ~= "atom_type" then return nil, 0, pos end
 	if pos > 1 then
 		local prev = body:byte(pos - 1)
 		if duffle.is_alnum_byte(prev) then return nil, 0, pos end
@@ -1054,19 +1105,19 @@ local function parse_enum_atom_type_default(body, pos)
 	end
 	-- Expect `( ... )` immediately after.
 	local open_pos = duffle.skip_ws_and_cmt(body, ident_end)
-	if open_pos > #body or body:sub(open_pos, open_pos) ~= "(" then return nil, 0, pos end
+	if    open_pos > #body or body:sub(open_pos, open_pos) ~= "(" then return nil, 0, pos end
 	local inner, after_close = duffle.read_parens(body, open_pos)
 	-- Reject any trailing tokens past the close paren other than comma / close-brace (next enum entry / end of enum).
 	local residue = duffle.skip_ws_and_cmt(body, after_close)
-	if residue <= #body then
+	if    residue <= #body then
 		local rbyte = body:byte(residue)
 		if rbyte ~= BYTE_COMMA and rbyte ~= BYTE_CLOSE_BRACE then return nil, 0, pos end
 	end
 	-- Parse the type chain inside the parens (e.g. `V4_S2*` -> ("V4_S2", 1)).
-	local type_name, depth, after_chain = parse_type_chain(inner, 1)
+	local  type_name, depth, after_chain = parse_type_chain(inner, 1)
 	if not type_name then return nil, 0, pos end
 	local end_check = duffle.skip_ws_and_cmt(inner, after_chain)
-	if end_check <= #inner then return nil, 0, pos end
+	if    end_check <= #inner then return nil, 0, pos end
 	return type_name, depth, duffle.skip_ws_and_cmt(body, after_close)
 end
 
@@ -1098,11 +1149,11 @@ end
 ---
 --- Diagnostic-only path: a following `(...)` is recorded as an invalid parenthesized-form marker so the annotation rule can emit a precise "parenthesized form" diagnostic.
 --- The parenthesized form stays diagnostic; the bare form alone carries the runtime stamp.
---- @param source string
+--- @param source    string
---- @param pos integer
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
 --- @return integer -- source cursor position to resume from
 local function parse_dbg_skip_marker(source, pos, ident_end, line_of, out)
 	local marker_kind = source:sub(pos, ident_end - 1)
@@ -1131,18 +1182,58 @@ local function parse_dbg_skip_marker(source, pos, ident_end, line_of, out)
 	return marker_end
 end
 
+--- Parse `atom_auto_reg(<atom>, R_<Sym>)` and `phase_auto_reg(<phase>, R_<Sym>)` markers.
+---
+--- The macros expand to `sym = sym##_Code` per their definition in dsl.atom.h.
+--- After preprocessing, the marker renders as a full enum entry of the form `R_<Sym> = R_<Sym>_Code,`.
+--- This parser detects the macro invocation site, extracts `(scope_name, sym)`, and stores it
+--- in the per-source table (atom_auto_regs or phase_auto_regs) under the scope's name.
+---
+--- @param source    string
+--- @param pos       integer
+--- @param ident_end integer
+--- @param line_of   fun(pos: integer): integer
+--- @param out       SourceScan
+--- @return integer
+local function parse_auto_reg_marker(source, pos, ident_end, line_of, out)
+	local marker_kind = source:sub(pos, ident_end - 1)  -- "atom_auto_reg" or "phase_auto_reg"
+	local scope_kind  = marker_kind == "atom_auto_reg" and "atom" or "phase"
+
+	local inner, after_paren = read_parens_after(source, ident_end)
+	if not inner then return after_paren end
+
+	local args = duffle.split_top_level_commas(inner)
+	local scope_name = args[1] and duffle.trim(args[1]) or nil
+	local sym        = args[2] and duffle.trim(args[2]) or nil
+
+	-- Filter: only accept `R_<Sym>` form (matches `^R_[%w_]+$`).
+	if scope_name and sym and sym:match("^R_[%w_]+$") then
+		if scope_kind == "atom" then
+			out.atom_auto_regs = out.atom_auto_regs or {}
+			out.atom_auto_regs[scope_name] = out.atom_auto_regs[scope_name] or {}
+			out.atom_auto_regs[scope_name][sym] = sym
+		else
+			out.phase_auto_regs = out.phase_auto_regs or {}
+			out.phase_auto_regs[scope_name] = out.phase_auto_regs[scope_name] or {}
+			out.phase_auto_regs[scope_name][sym] = sym
+		end
+	end
+
+	return after_paren
+end
+
 -- Parse `atom_dbg_reg_default(R_X, <type>...)`;
 -- the second argument may be a `Type` or `Type*`/`Type**` chain. Records in `out.types[R_X]`.
 local function parse_atom_dbg_reg_default(source, pos, ident_end, line_of, out)
-	local inner, after_paren = read_parens_after(source, ident_end)
+	local  inner, after_paren = read_parens_after(source, ident_end)
 	if not inner then return after_paren end
 	local args = duffle.split_top_level_commas(inner)
-	if #args < 1 then
+	if   #args < 1 then
 		-- Annotation pass surfaces this; we still consume the marker.
 		return after_paren
 	end
-	local reg_name  = duffle.trim(args[1])
+	local reg_name         = duffle.trim(args[1])
-	local type_part = args[2] or "void"
+	local type_part        = args[2] or "void"
 	local type_name, depth = parse_type_chain(type_part, 1)
 	if not type_name then type_name, depth = duffle.trim(type_part), 0 end
 	out.types[reg_name] = {
@@ -1161,23 +1252,23 @@ local function parse_atom_dbg_reg_default(source, pos, ident_end, line_of, out)
 end
 
 --- Parse: `MipsAtom_(<name>) [atom_info(<binds>, <reads>, <writes>)] { <body> }`
---- @param source string
+--- @param source    string
---- @param pos integer
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
 --- @return integer
 local function parse_mips_atom(source, pos, ident_end, line_of, out)
-	local inner, after_paren, open_paren = read_parens_after(source, ident_end)
+	local  inner, after_paren, open_paren = read_parens_after(source, ident_end)
 	if not inner then return after_paren end
 
 	local raw_name = duffle.read_ident(inner, 1)
 
 	-- Lookahead for atom_info(...) between `)` and `{`. Captures sub-calls; updates brace search start.
-	local brace_search_pos = after_paren
+	local brace_search_pos     = after_paren
-	local lookahead        = duffle.skip_ws_and_cmt(source, after_paren)
+	local lookahead            = duffle.skip_ws_and_cmt(source, after_paren)
 	local look_ident, look_end = duffle.read_ident(source, lookahead)
-	if look_ident == "atom_info" then
+	if    look_ident == "atom_info" then
 		local info_open = duffle.skip_ws_and_cmt(source, look_end)
 		if source:sub(info_open, info_open) == "(" then
 			local info_inner, info_after = duffle.read_parens(source, info_open)
@@ -1221,7 +1312,7 @@ local function parse_mips_atom(source, pos, ident_end, line_of, out)
 		end
 	end
 
-	local body, after_brace, body_off = find_body_braces(source, brace_search_pos, open_paren + 1)
+	local  body, after_brace, body_off = find_body_braces(source, brace_search_pos, open_paren + 1)
 	if not body then return after_brace end
 	if raw_name and raw_name ~= "" then
 		register_atom(out, "atom", line_of(pos), raw_name, body, body_off, raw_name, pos, after_paren, source)
@@ -1231,20 +1322,20 @@ local function parse_mips_atom(source, pos, ident_end, line_of, out)
 end
 
 --- Parse: `MipsAtomComp_(<name>) { <body> }`
---- @param source string
+--- @param source    string
---- @param pos integer
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
 --- @return integer
 local function parse_mips_atom_comp(source, pos, ident_end, line_of, out)
-	local inner, after_paren, open_paren = read_parens_after(source, ident_end)
+	local  inner, after_paren, open_paren = read_parens_after(source, ident_end)
 	if not inner then return after_paren end
 
 	local  raw_name = duffle.read_ident(inner, 1)
 	if not raw_name then return open_paren + 1 end
 
-	local body, after_brace, body_off = find_body_braces(source, after_paren, open_paren + 1)
+	local  body, after_brace, body_off = find_body_braces(source, after_paren, open_paren + 1)
 	if not body then return after_brace end
 	local name = strip_ac_prefix(raw_name)
 	register_atom(out, "comp_bare", line_of(pos), name, body, body_off, raw_name, pos, after_paren, source)
@@ -1253,14 +1344,14 @@ local function parse_mips_atom_comp(source, pos, ident_end, line_of, out)
 end
 
 --- Parse: `MipsAtomComp_Proc_(<name>, { <body> })` — body is inside the LAST `{` in args.
---- @param source string
+--- @param source    string
---- @param pos integer
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
 --- @return integer
 local function parse_mips_atom_comp_proc(source, pos, ident_end, line_of, out)
-	local inner, after_paren, open_paren = read_parens_after(source, ident_end)
+	local  inner, after_paren, open_paren = read_parens_after(source, ident_end)
 	if not inner then return after_paren end
 
 	-- Find the LAST `{` in inner (the body brace, not any potential embedded braces in expressions).
@@ -1276,22 +1367,71 @@ local function parse_mips_atom_comp_proc(source, pos, ident_end, line_of, out)
 	local body, close_pos = duffle.read_braces(inner, last_brace_pos)
 	if close_pos > #inner + 1 then return after_paren end
 
-	local raw_name = inner:match("^%s*([%w_]+)") or "?"
+	-- The component name is derived from the preceding function declaration
-	local name = strip_ac_prefix(raw_name)
+	-- (`FI_ Slice_MipsCode ac_X(...)`), not from the first macro arg (which
+	-- is now `ab`). The backward walk finds the function decl before open_paren.
+	local raw_name = duffle.find_function_decl_for(source, open_paren, SLICE_MIPS_CODE_LEN)
+	if not raw_name then raw_name = "?" end
+	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
 end
 
---- Parse: `MipsCode code_<name> { <body> }` (raw atom form — offsets pass only).
+--- Parse: `MipsAtom_Proc_(<name>, <abuilder>, { <body> })` — body is inside the LAST `{` in args.
---- @param source string
+--- Per Task 12.10: full support for the runtime-proc atom form. Registers the atom
---- @param pos integer
+--- with kind `"atom_proc"` so offsets.lua / components.lua can emit
+---   * `mac_<name>` aliases in `gen/macs.h` (the components pass)
+---   * `atom_offset__X__Y` defs in `gen/offsets.h` (the offsets pass)
+--- The atom name is the FIRST ident of the args (the second arg `ab` is the
+--- atom-builder, not the name). Unlike `MipsAtomComp_Proc_`, there is no `ac_`
+--- prefix on the symbol — `MipsAtom_Proc_` is the runtime-proc wrapper, so the
+--- symbol IS the bare atom name (e.g. `normalize_v3s4`, not `ac_normalize_v3s4`).
+--- @param source    string
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
+--- @return integer
+local function parse_mips_atom_proc(source, pos, ident_end, line_of, out)
+	local  inner, after_paren, open_paren = read_parens_after(source, ident_end)
+	if not inner then return after_paren end
+
+	-- Find the LAST `{` in inner (the body brace, not any potential embedded braces in expressions).
+	local last_brace_pos = nil
+	for search_pos = #inner, 1, -1 do
+		if inner:sub(search_pos, search_pos) == "{" then last_brace_pos = search_pos; break end
+	end
+	if not last_brace_pos then return after_paren end
+
+	-- Use duffle.read_braces to find the matching close brace.
+	-- Uses `read_balanced` for delimiter-depth tracking.
+	-- If close_pos is past the end of inner, the brace didn't match (malformed input); skip.
+	local body, close_pos = duffle.read_braces(inner, last_brace_pos)
+	if close_pos > #inner + 1 then return after_paren end
+
+	-- The atom name is derived from the preceding function declaration
+	-- (`internal MipsAtom* X_proc(...)`), not from the first macro arg (which
+	-- is now `aa`). The backward walk finds the function decl before open_paren
+	-- and strips the `_proc` suffix.
+	local raw_name = duffle.find_atom_proc_decl_for(source, open_paren, MIPS_ATOM_PTR_LEN)
+	if not raw_name then raw_name = "?" end
+	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, "atom_proc", line_of(pos), name, body, body_off, raw_name, pos, after_paren, source)
+
+	return after_paren
+end
+
+--- Parse: `MipsCode code_<name> { <body> }` (raw atom form — offsets pass only).
+--- @param source    string
+--- @param pos       integer
+--- @param ident_end integer
+--- @param line_of   fun(pos: integer): integer
+--- @param out       SourceScan
 --- @return integer
 local function parse_mips_code(source, pos, ident_end, line_of, out)
 	local  next_pos               = duffle.skip_ws_and_cmt(source, ident_end)
@@ -1300,8 +1440,8 @@ local function parse_mips_code(source, pos, ident_end, line_of, out)
 		return ident_end
 	end
 
-	local  atom_name = next_ident:sub(6)
+	local  atom_name                   = next_ident:sub(6)
-	local body, after_brace, body_off = find_body_braces(source, next_after, ident_end)
+	local  body, after_brace, body_off = find_body_braces(source, next_after, ident_end)
 	if not body then return after_brace end
 	register_raw_atom(out, line_of(pos), atom_name, body, body_off, atom_name, pos)
 
@@ -1320,13 +1460,13 @@ end
 ---     pointer_depth = 0 
 --- 	} -- byte_size + per-field offset/byte_size set by the propagation pass.
 --- Also populates `out.binds[]` IFF `name:sub(1, 6) == "Binds_"`.
---- @param body string
+--- @param body    string
---- @param name string
+--- @param name    string
---- @param pos integer
+--- @param pos     integer
 --- @param line_of fun(pos: integer): integer
---- @param out SourceScan
+--- @param out     SourceScan
 local function register_struct_type(body, name, pos, line_of, out)
-	local fields = parse_struct_body_fields(body)
+	local fields     = parse_struct_body_fields(body)
 	local source_pos = line_of(pos)
 	out.type_name_registry[name] = {
 		name          = name,
@@ -1352,11 +1492,11 @@ end
 --- Register an Enum_ entry in type_name_registry.
 --- Local helper for parse_typedef_binds. Captures the underlying type (1st arg of `Enum_(<underlying>, <name>)`) and the body fields.
 --- @param underlying string
---- @param name string
+--- @param name       string
---- @param body string
+--- @param body       string
---- @param pos integer
+--- @param pos        integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of    fun(pos: integer): integer
---- @param out SourceScan
+--- @param out        SourceScan
 local function register_enum_type(underlying, name, body, pos, line_of, out)
 	local fields = parse_enum_body_fields(body)
 	out.type_name_registry[name] = {
@@ -1375,18 +1515,18 @@ end
 --- Captures the underlying type ident (LHS of `typedef <type> <alias>;`) and exposes it through the registry.
 --- The propagation pass follows the underlying_type chain to resolve byte_size.
 --- @param underlying string
---- @param name string
+--- @param name       string
---- @param pos integer
+--- @param pos        integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of    fun(pos: integer): integer
---- @param out SourceScan
+--- @param out        SourceScan
 local function register_typedef_alias(underlying, name, pos, line_of, out)
 	out.type_name_registry[name] = {
-		name           = name,
+		name            = name,
-		kind           = "typedef",
+		kind            = "typedef",
 		underlying_type = underlying,
-		source_line    = line_of(pos),
+		source_line     = line_of(pos),
-		source_file    = out._source_file,
+		source_file     = out._source_file,
-		pointer_depth  = 0,
+		pointer_depth   = 0,
 	}
 end
 
@@ -1396,29 +1536,29 @@ 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.
 ---       Strips TSet_ wrapper; adds to type_name_registry (kind="typedef") with underlying_type=<type>.
 ---
 --- All four shapes also attach an "unrelated" debug-skip marker (the existing behavior — typedef declarations don't carry atom_dbg_skip).
---- @param source string
+--- @param source    string
---- @param pos integer
+--- @param pos       integer
 --- @param ident_end integer
---- @param line_of fun(pos: integer): integer
+--- @param line_of   fun(pos: integer): integer
---- @param out SourceScan
+--- @param out       SourceScan
---- @return integer
+--- @return          integer
 local function parse_typedef_binds(source, pos, ident_end, line_of, out)
-	local after_typedef = duffle.skip_ws_and_cmt(source, ident_end)
+	local  after_typedef = duffle.skip_ws_and_cmt(source, ident_end)
-	local id2, id2_end  = duffle.read_ident(source, after_typedef)
+	local  id2, id2_end  = duffle.read_ident(source, after_typedef)
 	if not id2 then return ident_end end
 
 	-- ── Shape 1: `typedef Struct_(<name>) { <body> } <alias>;` ────────────
 	if id2 == "Struct_" then
-		local inner, after_paren, open_paren = read_parens_after(source, id2_end, id2_end)
+		local  inner, after_paren, open_paren = read_parens_after(source, id2_end, id2_end)
 		if not inner then return id2_end end
-		local name               = duffle.trim(inner)
+		local name = duffle.trim(inner)
 
 		local body, after_brace = find_body_braces(source, after_paren, open_paren + 1)
 		if not body then return after_brace end
@@ -1428,7 +1568,7 @@ local function parse_typedef_binds(source, pos, ident_end, line_of, out)
 
 	-- ── Shape 2: `typedef Enum_(<underlying>, <name>) { <body> } <alias>;`
 	elseif id2 == "Enum_" then
-		local inner, after_paren, open_paren = read_parens_after(source, id2_end, id2_end)
+		local  inner, after_paren, open_paren = read_parens_after(source, id2_end, id2_end)
 		if not inner then return id2_end end
 		-- Split `inner` on the first top-level comma into (<underlying>, <name>).
 		local args = duffle.split_top_level_commas(inner)
@@ -1436,7 +1576,7 @@ local function parse_typedef_binds(source, pos, ident_end, line_of, out)
 		local underlying = duffle.trim(args[1])
 		local name       = duffle.trim(args[2])
 
-		local body, after_brace = find_body_braces(source, after_paren, open_paren + 1)
+		local  body, after_brace = find_body_braces(source, after_paren, open_paren + 1)
 		if not body then return after_brace end
 		register_enum_type(underlying, name, body, pos, line_of, out)
 		attach_debug_skip_marker(out, "unrelated")
@@ -1469,11 +1609,10 @@ local function parse_typedef_binds(source, pos, ident_end, line_of, out)
 
 	-- Shape 4 (TSet_ at id2 position): no preceding underlying span.
 	if id2 == "TSet_" then
-		local inner, after_paren = read_parens_after(source, id2_end, id2_end)
+		local  inner, after_paren = read_parens_after(source, id2_end, id2_end)
 		if not inner then return id2_end end
 		local tset_name = duffle.trim(inner)
-		-- Empty underlying span is acceptable; the TSet_ wrapper itself
+		-- Empty underlying span is acceptable; the TSet_ wrapper itself encodes the alias identity (per the duffle TSet_ convention).
-		-- encodes the alias identity (per the duffle TSet_ convention).
 		register_typedef_alias("", tset_name, pos, line_of, out)
 		attach_debug_skip_marker(out, "unrelated")
 		return after_paren
@@ -1491,13 +1630,13 @@ local function parse_typedef_binds(source, pos, ident_end, line_of, out)
 	while scan < semi_pos do
 		scan = duffle.skip_ws_and_cmt(source, scan)
 		if scan >= semi_pos then break end
-		local id, id_end = duffle.read_ident(source, scan)
+		local  id, id_end = duffle.read_ident(source, scan)
 		if not id then
 			scan = scan + 1
 		elseif id == "TSet_" then
 			-- Shape 4 (TSet_ at non-id2 position): grab the parenthesized argument.
 			local inner, after_paren = read_parens_after(source, id_end, id_end)
-			if inner then
+			if    inner then
 				tset_arg     = duffle.trim(inner)
 				tset_arg_end = after_paren
 				tset_pos     = scan
@@ -1607,6 +1746,16 @@ local function parse_enum_entry(source, body, body_offset, line_of, out, entry_n
 	local value, value_end = parse_enum_value(body, after_ws, out)
 	if value == nil then return value_start end
 
+	-- Capture the trailing C-comment (if any) before `skip_ws_and_cmt` discards it.
+	-- The `atom_auto_reg(<scope>, <sym>)` macro expands to `R_<Sym> = R_<Sym>_Code /* atom_auto_reg: <scope> */`,
+	-- so the scope name lives in the comment after the RHS value. Routes through `out.atom_entry_comments`
+	-- for downstream `parse_enum` to split into `out.atom_auto_regs` / `out.phase_auto_regs`.
+	local trailing_cmt = read_trailing_cmt_after(body, value_end)
+	if trailing_cmt then
+		out.atom_entry_comments = out.atom_entry_comments or {}
+		out.atom_entry_comments[entry_name] = trailing_cmt
+	end
+
 	local after_value       = duffle.skip_ws_and_cmt(body, value_end)
 	local has_atom_reg, end_after_atom_reg = check_bare_atom_reg(body, after_value)
 
@@ -1662,15 +1811,24 @@ local function parse_enum_body(source, body, body_offset, line_of, out)
 		else
 			local entry_name, name_end = duffle.read_ident(body, pos)
 			if entry_name then
-				local after_name = duffle.skip_ws_and_cmt(body, name_end)
+				-- In-enum `atom_auto_reg(<scope>, R_<Sym>)` / `phase_auto_reg(<scope>, R_<Sym>)` markers:
-				if body:byte(after_name) == BYTE_EQUAL then
+				-- the C preprocessor expands them to `R_<Sym> = R_<Sym>_Code /* atom_auto_reg: <scope> */`,
-					local new_pos = parse_enum_entry(
+				-- but the metaprogram reads source-as-written so we must dispatch the parser here too.
-						source, body, body_offset, line_of, out,
+				-- Mirrors the top-level `DECL_PARSERS` entry for `atom_auto_reg` / `phase_auto_reg`.
-						entry_name, pos, after_name + 1
+				if entry_name == "atom_auto_reg" or entry_name == "phase_auto_reg" then
-					)
+					local new_pos = parse_auto_reg_marker(body, pos, name_end, line_of, out)
-					if new_pos > pos then pos = new_pos else pos = after_name + 1 end
+					if new_pos > pos then pos = new_pos else pos = name_end end
 				else
-					pos = name_end
+					local after_name = duffle.skip_ws_and_cmt(body, name_end)
+					if body:byte(after_name) == BYTE_EQUAL then
+						local new_pos = parse_enum_entry(
+							source, body, body_offset, line_of, out,
+							entry_name, pos, after_name + 1
+						)
+						if new_pos > pos then pos = new_pos else pos = after_name + 1 end
+					else
+						pos = name_end
+					end
 				end
 			else
 				pos = pos + 1
@@ -1700,6 +1858,25 @@ local function parse_enum(source, pos, ident_end, line_of, out)
 	if not body then return after_brace end
 	parse_enum_body(source, body, body_off, line_of, out)
 
+	-- Route `atom_auto_reg:` / `phase_auto_reg:` markers discovered in trailing C-comments
+	-- into the per-source `atom_auto_regs` / `phase_auto_regs` projections.
+	-- Pattern matches the RHS expansion `R_<Sym> = R_<Sym>_Code /* <kind>_auto_reg: <scope> */`
+	-- emitted by the `atom_auto_reg` / `phase_auto_reg` macros in dsl.atom.h.
+	for entry_name, cmt_text in pairs(out.atom_entry_comments or {}) do
+		local atom_scope = cmt_text:match("atom_auto_reg:%s*([%w_]+)")
+		if atom_scope then
+			out.atom_auto_regs = out.atom_auto_regs or {}
+			out.atom_auto_regs[atom_scope] = out.atom_auto_regs[atom_scope] or {}
+			out.atom_auto_regs[atom_scope][entry_name] = entry_name
+		end
+		local phase_scope = cmt_text:match("phase_auto_reg:%s*([%w_]+)")
+		if phase_scope then
+			out.phase_auto_regs = out.phase_auto_regs or {}
+			out.phase_auto_regs[phase_scope] = out.phase_auto_regs[phase_scope] or {}
+			out.phase_auto_regs[phase_scope][entry_name] = entry_name
+		end
+	end
+
 	return after_brace
 end
 
@@ -1713,12 +1890,18 @@ end
 
 local DECL_PARSERS = {
 	MipsAtom_                  = parse_mips_atom,
+	MipsAtom_Proc_             = parse_mips_atom_proc,
 	MipsAtomComp_              = parse_mips_atom_comp,
 	MipsAtomComp_Proc_         = parse_mips_atom_comp_proc,
 	-- `atom_dbg_skip` is the only debug-skip parser entry. Every other
 	-- identifier follows the ordinary unrelated-token path; there is no alias.
 	atom_dbg_skip              = parse_dbg_skip_marker,
 	atom_dbg_reg_default       = parse_atom_dbg_reg_default,
+	-- `atom_auto_reg(atom, R_<Sym>)` and `phase_auto_reg(phase, R_<Sym>)` populate per-source
+	-- `out.atom_auto_regs` / `out.phase_auto_regs`; the cross-source merge lands in
+	-- `corpus.atom_auto_regs` / `corpus.phase_auto_regs` (first-wins).
+	atom_auto_reg              = parse_auto_reg_marker,
+	phase_auto_reg             = parse_auto_reg_marker,
 	MipsCode                   = parse_mips_code,
 	typedef                    = parse_typedef_binds,
 	_Pragma                    = parse_pragma_macro,
@@ -1753,6 +1936,14 @@ local function scan_source(source, source_file, code_macros, code_macro_bodies)
 		debug_skip_markers = {},
 		types             = {},
 		atom_views        = {},
+		-- Per-source projection for `atom_auto_reg(<atom>, R_<Sym>)` markers.
+		-- Each entry is keyed by atom_name; the inner table maps `R_<Sym>` -> `R_<Sym>` (raw LHS sym).
+		-- Merged cross-source into `corpus.atom_auto_regs` (first-wins).
+		atom_auto_regs    = {},
+		-- Per-source projection for `phase_auto_reg(<phase>, R_<Sym>)` markers.
+		-- Each entry is keyed by phase_label; the inner table maps `R_<Sym>` -> `R_<Sym>` (raw LHS sym).
+		-- Merged cross-source into `corpus.phase_auto_regs` (first-wins).
+		phase_auto_regs   = {},
 		line_of           = line_of,
 		-- Source-derived register-alias registry (atom_reg opt-in entries).
 		-- Keys are full R_* idents (never stripped); see parse_enum / parse_enum_body.
@@ -1957,7 +2148,7 @@ local function merge_named_with_sites(registry, name, new_entry, site, collision
 		registry[name].sites = { site }
 		return
 	end
-	local existing = registry[name]
+	local existing  = registry[name]
 	local new_shape = shape_fn(new_entry)
 	local old_shape = shape_fn(existing)
 	if new_shape == old_shape and new_shape ~= "" then
@@ -1992,6 +2183,8 @@ local function merge_corpus_registries(corpus)
 	corpus.atom_ctxs               = corpus.atom_ctxs               or {}
 	corpus.atom_phases             = corpus.atom_phases             or {}
 	corpus.atom_infos              = corpus.atom_infos              or {}
+	corpus.atom_auto_regs          = corpus.atom_auto_regs          or {}
+	corpus.phase_auto_regs         = corpus.phase_auto_regs         or {}
 	corpus.collisions              = corpus.collisions              or {}
 
 	-- Replace the existing corpus collections with empty tables so a re-run on the same corpus produces identical state (deterministic merge).
@@ -2035,7 +2228,7 @@ local function merge_corpus_registries(corpus)
 					corpus.collisions, "binds", bind_shape)
 			end
 
-			-- atoms_by_name: MipsAtom_(name) + MipsAtomComp_(name) + MipsAtomComp_Proc_(name).
+			-- atoms_by_name: MipsAtom_(name) + MipsAtom_Proc_(name) + MipsAtomComp_(name) + MipsAtomComp_Proc_(name).
 			-- Each atom carries `{line, name, body, body_off, kind, raw_name, ...}`.
 			-- Duplicate atom names across sources are first-wins + collision; see the atom_infos block below for the evidence list.
 			for _, atom_entry in ipairs(scan.atoms or {}) do
@@ -2070,6 +2263,22 @@ local function merge_corpus_registries(corpus)
 					corpus.collisions, "phase", phase_shape)
 			end
 
+			-- atom_auto_regs: keyed by atom scope name; each carries a `{R_<Sym> = R_<Sym>}` map.
+			-- Per-source entries are simple inner maps (no body / no shape comparison); first-wins suffices.
+			for atom_scope, syms in pairs(scan.atom_auto_regs or {}) do
+				if corpus.atom_auto_regs[atom_scope] == nil then
+					corpus.atom_auto_regs[atom_scope] = syms
+				end
+			end
+
+			-- phase_auto_regs: keyed by phase label; each carries a `{R_<Sym> = R_<Sym>}` map.
+			-- Per-source entries are simple inner maps (no body / no shape comparison); first-wins suffices.
+			for phase_label, syms in pairs(scan.phase_auto_regs or {}) do
+				if corpus.phase_auto_regs[phase_label] == nil then
+					corpus.phase_auto_regs[phase_label] = syms
+				end
+			end
+
 			-- atom_infos: ALWAYS append every record in source/declaration order.
 			-- Duplicates are preserved so the annotation pass can flag them via `check_unique_annotation`;
 			-- The merge is purely order-preserving.

scripts/passes/static_analysis.lua

@@ -1,20 +1,17 @@
 --- passes/static_analysis.lua — Per-atom static-analysis checks.
----
+--- Ownership: `ctx.shared.corpus` canonical merged registry; per-source fallback synthesis is rejected.
---- Ownership: `ctx.shared.corpus` is the canonical merged registry; per-source fallback synthesis is rejected.
 --- `atom.paths` supplies the emitted and analysis projections consumed by this pass.
 ---
 --- Per-atom rules:
 ---   1. transfer_hazards: A single forward walker (`analyze_hardware_relations`) reads `atom.paths.word_events` once per atom.
----      For each emitted word event it (a) inspects pending CPU/COP0/COP2/GTE relations against the event as CONSUMER 
+---      For each emitted word event it (a) inspects pending CPU / COP0 / COP2 / GTE relations against the event as CONSUMER 
 ---      (recording a hazard on `atom.paths.hazards` when the producer→consumer gap is below the required retire-slot count), 
----      (b) applies the event's GPR value effects (`duffle.INSTRUCTION_GPR_EFFECTS`) to `atom.paths.forward_state.gpr_values`, 
+---      (b) Applies the event's GPR value effects (`duffle.INSTRUCTION_GPR_EFFECTS`) to `atom.paths.forward_state.gpr_values`, 
 ---      applies bounded constant propagation, and stages matching relation rows as PRODUCERS (with `destination_match` filters, e.g. for the IRGB fan-out).
 ---      The `transfer_hazards` CHECK_RULES reader projects `atom.paths.hazards` into per-atom findings.
----      The reader does NOT re-walk source; this is the per-check purity contract.
+---      The reader does NOT re-walk source. The walker runs once per atom before the per-atom dispatch; the reader runs inside the same dispatch.
----      The walker runs once per atom before the per-atom dispatch; the reader runs inside the same dispatch.
 ---   2. control_transfer_delay_slot_use: For every emitted branch/jump/call encoder in `duffle.CONTROL_TRANSFER_DELAY_SLOT_POLICIES` 
----      (the six `branch_*` encoders plus `jump` / `jump_reg` / `jump_link` / `call_reg` / `call_addr`),
+---      (the six `branch_*` encoders plus `jump` / `jump_reg` / `jump_link` / `call_reg` / `call_addr`), inspect the next emitted event in `atom.paths.word_events`.
----      inspect the next emitted event in `atom.paths.word_events`.
 ---      Emit an `info`-severity finding when the successor is `nop` or absent (the next emitted word IS the hardware delay slot).
 ---      `jump_reg(R_AtomJmp)` is suppressed by policy (the fixed `mac_yield()` handshake).
 ---      `nop2` needs no special case: emission-model emits two `nop` events for it, so the first expansion is the hardware delay slot.
@@ -42,8 +39,8 @@
 --- `── Info` section renders finding-level info between `── Warnings` and the per-atom cycle counts.
 ---
 --- The structural handshake checks (`mac_yield_uniformity`, `hazard_nop_use`, `control_transfer_delay_slot_use`) skip atoms/components with `debug_skip == true`.
---- The `atom_dbg_skip` marker designates runtime-helper declarations whose structure is fixed by the tape runtime (e.g. `tape_exit`, `ac_yield`).
+--- `atom_dbg_skip` marker designates runtime-helper declarations whose structure is fixed by the tape runtime (e.g. `tape_exit`, `ac_yield`).
---- Flagging them as "missing mac_yield" or "BD slot is redundant" is signal noise, not a logic failure. 
+--- Flagging them as "missing mac_yield" or "BD slot is redundant".
 --- Other checks (transfer_hazards, gpu_portstore_shape, abi_handoff, enum_alias_membership, …) still apply to debug_skip declarations because real hazards / typos can still surface in them.
 ---
 --- The orchestrator (`ps1_meta.lua`) wires this module in via the PASSES table:
@@ -101,10 +98,10 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
 --- @field sources            SourceFile[]
@@ -121,43 +118,42 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 --- @field outputs  table[]
 --- @field errors   table[]
 --- @field warnings table[]
---- @field info     table[]   -- finding-level info (kind == "info"); distinct from per-source scanned/cycles summary rows
+--- @field info     table[]   -- Finding-level info (kind == "info"); distinct from per-source scanned/cycles summary rows
 
 --- @alias AtomName    string  -- lower_snake_case atom nameMacroName   string  -- lower_snake_case macro identifier
 --- @alias CheckName   string  -- "transfer_hazards" | "control_transfer_delay_slot_use" | "mac_yield_uniformity" | "yield_load_tail_pairing" | "abi_handoff" | "gpu_portstore_shape" | "per_atom_cycle_budget" | "enum_alias_membership" | "atom_type_consistency" | "binds_no_substruct_deref"
 
 --- @class AtomBody
---- @field line     integer  -- source line of the atom declaration
+--- @field line     integer  -- Source line of the atom declaration
---- @field name     AtomName -- atom name (e.g. "cube_g4_face")
+--- @field name     AtomName -- Atom name (e.g. "cube_g4_face")
---- @field body     string   -- the brace-delimited body (without the braces)
+--- @field body     string   -- Brace-delimited body (without the braces)
 --- @field body_off integer  -- char offset of body[1] in source
 --- @field kind     string   -- "atom" | "comp_bare" | "comp_proc"
 
 --- @class Token
---- @field tok      string     -- the raw token text (trimmed)
+--- @field tok      string     -- Raw token text (trimmed)
---- @field line     integer    -- source line of the token's start
+--- @field line     integer    -- Source line of the token's start
---- @field ident    string|nil -- the leading ident of the token (if any)
+--- @field ident    string|nil -- Leading ident of the token (if any)
 --- @field kind     string     -- "n_words" | "mac_yield" | "gte_cmdw" | "mac_format" | "mac_gte_store" | "mac_insert_ot_tag" | "atom_label" | "atom_offset" | "other"
 
 --- @class Finding 
---- @field line     integer   -- source line of the finding
+--- @field line     integer   -- Source line of the finding
---- @field atom     AtomName  -- the atom this finding is for (or "")
+--- @field atom     AtomName  -- Atom this finding is for (or "")
---- @field check    CheckName -- the check identifier
+--- @field check    CheckName -- Check identifier
 --- @field kind     string    -- "error" | "warning" | "info"
---- @field msg      string    -- the finding message
+--- @field msg      string    -- Finding message
 
 --- @class AtomAnalysis
 --- @field atom       AtomBody
---- @field tokens     Token[]   -- the tokens in the atom body, annotated
+--- @field tokens     Token[]   -- Tokens in the atom body, annotated
---- @field findings   Finding[] -- findings for this atom
+--- @field findings   Finding[] -- Findings for this atom
---- @field total_cycles integer -- sum of token cycle costs
+--- @field total_cycles integer -- Sum of token cycle costs
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Per-word-event helpers
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Pick the source-line field that best represents "where in the user's source file is this word?".
---
 -- `word_events` (populated by `passes/emission_model.lua::stamp_root_provenance`) carry four line fields:
 --   * `call_line`  — physical line in the ROOT atom's source (the line of the `mac_X(...)` call site that triggered this emission, or `body_line` for direct words in the atom body)
 --   * `body_line`  — physical line in the body containing the emitted word (the atom body for direct words; the component body for words expanded inside `mac_X(...)`)
@@ -169,8 +165,7 @@ local OUTPUT_EXTENSION = ".static_analysis.txt"
 -- The user editing their atom body expects the line to point at THEIR source — i.e. the line where `mac_yield()`
 -- was called (e.g. `hello_gte_tape.c:35`). That line is `call_line`.
 --
--- For direct words in the atom body (no invocation wrapping them), `call_line == body_line` already,
+-- For direct words in the atom body (no invocation wrapping them), `call_line == body_line` already, so `call_line` works for both cases.
--- so `call_line` works for both cases.
 local function line_for_word_event(ev)
     if ev == nil then return 0 end
     return ev.call_line or ev.body_line or ev.line or ev.def_line or 0
@@ -193,46 +188,44 @@ end
 --
 -- The classification is stored on `atom.paths.tok_class` as an array indexed by token index (1..#tokens). 
 -- Each entry has:
---   ident              — the leading identifier (e.g. "load_word", "gte_cmdw_rtpt", "nop", "mac_yield")
+--   ident                 — the leading identifier (e.g. "load_word", "gte_cmdw_rtpt", "nop", "mac_yield")
---   nop_words          — 0 / 1 / 2 (for "nop" / "nop2" / anything else)
+--   nop_words             — 0 / 1 / 2 (for "nop" / "nop2" / anything else)
---   nop_prefix         — consecutive nop words ending just BEFORE this token (forward-pass pre-compute;
+--   nop_prefix            — consecutive nop words ending just BEFORE this token (forward-pass pre-compute; makes preceding-nop lookup O(N))
---                        makes preceding-nop lookup O(N))
+--   is_yield              — true if this token is `mac_yield` or `mac_yield(...)`
---   is_yield           — true if this token is `mac_yield` or `mac_yield(...)`
+--   is_atom_label         — true if this token is `atom_label(name)`; label_name has the name
---   is_atom_label      — true if this token is `atom_label(name)`; label_name has the name
+--   is_branch             — true if this token is `branch_*(...)` OR an unconditional-jump-with-offset (`jump(off)` / `call_addr(off)`); branch_label has the target label or false
---   is_branch          — true if this token is `branch_*(...)` OR an unconditional-jump-with-offset (`jump(off)` / `call_addr(off)`); branch_label has the target label or false
 --   is_unconditional_jump — true if this token is `jump` or `call_addr` (BD slot + single successor — taken only; no fall-through).
 --                           Mutually exclusive with the conditional-branch semantics; combined with `is_branch` above.
---   is_terminal_jump   — true if this token is `jump_reg` / `call_reg` / `jump_link` (transfers control OUT of the current atom; the `mac_yield()` handshake ends in `jump_reg(R_AtomJmp), nop`). 
+--   is_terminal_jump      — true if this token is `jump_reg` / `call_reg` / `jump_link` (transfers control OUT of the current atom; the `mac_yield()` handshake ends in `jump_reg(R_AtomJmp), nop`). 
---                        No offset field — `atom_offset` is invalid here. Terminates the current path in the CFG.
+--                           No offset field — `atom_offset` is invalid here. Terminates the current path in the CFG.
---   is_load            — true if this token starts with any of: load_word, load_half, load_half_u, load_byte,
+--   is_load               — true if this token starts with any of: load_word, load_half, load_half_u, load_byte,
---                        load_byte_u, gte_lw, gte_lwc2. These all have MIPS load-delay semantics (the
+--                           load_byte_u, gte_lw, gte_lwc2. These all have MIPS load-delay semantics (the destination register is volatile for 1 word after the load).
---                        destination register is volatile for 1 word after the load).
+--   is_store_word         — true if this token starts with `store_word(`
---   is_store_word      — true if this token starts with `store_word(`
 --
 -- Checks that need the leading ident use `tok_class.ident` instead of re-matching the token string.
 -- Checks that need "how many nops before token i" use `tok_class.nop_prefix` instead of walking backwards.
 
 --- @class TokClass
---- @field ident              string          -- leading identifier
+--- @field ident              string           -- lLading identifier
---- @field nop_words          integer         -- 0/1/2
+--- @field nop_words          integer          -- 0 / 1 / 2
---- @field nop_prefix         integer         -- consecutive nop words before this token
+--- @field nop_prefix         integer          -- Consecutive nop words before this token
 --- @field is_yield           boolean
 --- @field is_atom_label      boolean
---- @field label_name         string|nil      -- for atom_label(name)
+--- @field label_name         string|nil       -- For atom_label(name)
---- @field is_branch          boolean         -- conditional branch OR unconditional-jump-with-offset
+--- @field is_branch          boolean          -- Conditional branch OR unconditional-jump-with-offset
---- @field is_unconditional_jump boolean      -- `jump` / `call_addr` only
+--- @field is_unconditional_jump boolean       -- `jump` / `call_addr` only
---- @field is_terminal_jump   boolean         -- `jump_reg` / `call_reg` / `jump_link` only
+--- @field is_terminal_jump   boolean          -- `jump_reg` / `call_reg` / `jump_link` only
---- @field branch_label       string|false|nil -- for branch_*(..., atom_offset(F, label)) OR jump/call_addr
+--- @field branch_label       string|false|nil -- For branch_*(..., atom_offset(F, label)) OR jump/call_addr
---- @field is_load            boolean         -- load_word | load_half | load_half_u | load_byte | load_byte_u | gte_lw | gte_lwc2
+--- @field is_load            boolean          -- load_word | load_half | load_half_u | load_byte | load_byte_u | gte_lw | gte_lwc2
 --- @field is_store_word      boolean
---- @field mac_format_shape   string|nil      -- "f3" / "g4" etc. for mac_format_X_color; nil otherwise
+--- @field mac_format_shape   string|nil       -- "f3" / "g4" etc. for mac_format_X_color; nil otherwise
---- @field is_gte_store       boolean         -- ident matches `mac_gte_store_<shape>`
+--- @field is_gte_store       boolean          -- Ident matches `mac_gte_store_<shape>`
---- @field is_ot_tag          boolean         -- ident matches `mac_insert_ot_tag_<shape>`
+--- @field is_ot_tag          boolean          -- Ident matches `mac_insert_ot_tag_<shape>`
---- @field writes_r_prim_cursor boolean       -- store_word targeting R_PrimCursor
+--- @field writes_r_prim_cursor boolean        -- store_word targeting R_PrimCursor
---- @field reads_r_tape_ptr   boolean         -- any token referencing R_TapePtr
+--- @field reads_r_tape_ptr   boolean          -- Any token referencing R_TapePtr
---- @field o_arg1             string|nil      -- first arg of O_(<a>, <b>) captures; nil for non-O_ tokens
+--- @field o_arg1             string|nil       -- First arg of O_(<a>, <b>) captures; nil for non-O_ tokens
---- @field o_arg2             string|nil      -- second arg of O_(<a>, <b>) captures
+--- @field o_arg2             string|nil       -- Second arg of O_(<a>, <b>) captures
---- @field s_arg1             string|nil      -- arg of S_(<a>) captures; nil for non-S_ tokens
+--- @field s_arg1             string|nil       -- Arg of S_(<a>) captures; nil for non-S_ tokens
 
 -- The set of MIPS instruction idents that have a load-delay slot.
 -- Per MIPS I R3000A: `lw`, `lh`, `lhu`, `lb`, `lbu`, `lwc2` (gte_lw).
@@ -263,8 +256,21 @@ local BRANCH_PATTERN          = "^branch_[%w_]+%s*%("
 -- The C preprocessor expands it BEFORE the metaprogram sees the source, but for source-level metadata consistency we still match it here and classify it as a branch_equal.
 -- This keeps `consuming_encoder` canonical for any downstream tooling that consults the metadata field.
 local JUMP_REL_PATTERN      = "^jump_rel%s*%("
-local UNCOND_JUMP_PATTERN   = "^%f[%w](jump|call_addr)%f[%W]"
+local UNCOND_JUMP_PATTERNS = {
-local TERMINAL_JUMP_PATTERN = "^%f[%w](jump_reg|call_reg|jump_link)%f[%W]"
+    "^%f[%w]jump%f[%W]",
+    "^%f[%w]call_addr%f[%W]",
+}
+local TERMINAL_JUMP_PATTERNS = {
+    "^%f[%w]jump_reg%f[%W]",
+    "^%f[%w]call_reg%f[%W]",
+    "^%f[%w]jump_link%f[%W]",
+}
+local function matches_any(tok, patterns)
+    for i = 1, #patterns do
+        if tok:match(patterns[i]) then return true end
+    end
+    return false
+end
 
 local function classify_tokens(tokens)
 	local n       = #tokens
@@ -308,13 +314,13 @@ local function classify_tokens(tokens)
 			-- Both encode a 16-bit signed relative word offset.
 			is_branch    = true
 			branch_label = tok:match("atom_offset%s*%([^,]+,%s*([%w_]+)%s*%)") or false
-		elseif tok:match(UNCOND_JUMP_PATTERN) then
+		elseif matches_any(tok, UNCOND_JUMP_PATTERNS) then
 			-- Unconditional absolute jump / call: `jump(off)` / `call_addr(off)`.
 			-- One immediate offset field; can carry an `atom_offset(F, T)` marker (the offsets pass dispatches on `consuming_encoder` — see `passes/offsets.lua::compute_offsets`).
 			is_branch              = true
 			is_unconditional_jump = true
 			branch_label           = tok:match("atom_offset%s*%([^,]+,%s*([%w_]+)%s*%)") or false
-		elseif tok:match(TERMINAL_JUMP_PATTERN) then
+		elseif matches_any(tok, TERMINAL_JUMP_PATTERNS) then
 			-- Register-form jump / call: no offset field; `atom_offset` is invalid here (the offsets pass will error if one is supplied).
 			-- Transfers control OUT of the current atom — the CFG treats this as a path terminator.
 			is_terminal_jump = true
@@ -343,7 +349,7 @@ local function classify_tokens(tokens)
 			is_atom_label          = is_atom_label,
 			label_name             = label_name,
 			is_branch              = is_branch,
-			is_unconditional_jump = is_unconditional_jump,
+			is_unconditional_jump  = is_unconditional_jump,
 			is_terminal_jump       = is_terminal_jump,
 			branch_label           = branch_label,
 			is_load                = is_load,
@@ -401,15 +407,13 @@ end
 -- therefore counts ONLY words strictly between the producer and the consumer.
 -- ─────────────────────────────────────────────────────────────────────────
 
--- True iff `consumer_event` is a GTE command (gte_cmdw_* or one of the human-readable aliases
+-- True iff `consumer_event` is a GTE command (gte_cmdw_* or one of the human-readable aliases mapped in `duffle.GTE_COMMAND_ALIASES`).
--- mapped in `duffle.GTE_COMMAND_ALIASES`). Used by the LWC2 retirement-regime dispatch in the
+-- Used by the LWC2 retirement-regime dispatch in the forward walker: a GTE-command consumer can read the LWC2 result in the very next slot
--- forward walker: a GTE-command consumer can read the LWC2 result in the very next slot (the GTE
+-- (the GTE pipeline latches the LWC2 data); any other consumer must observe the standard MIPS load delay (gap >= 1).
--- pipeline latches the LWC2 data); any other consumer must observe the standard MIPS load delay
--- (gap >= 1).
 local function is_gte_command(consumer_event)
 	local tok = consumer_event.encoder or consumer_event.ident or ""
 	if tok:sub(1, 9) == "gte_cmdw_" then return true end
-	local aliases = duffle.GTE_COMMAND_ALIASES or {}
+	local  aliases = duffle.GTE_COMMAND_ALIASES or {}
 	return aliases[tok] ~= nil
 end
 
@@ -422,7 +426,7 @@ local function is_cop2_consumer_of(consumer_event, destination, producer_rel)
 	-- not used by MTC2/CTC2 today because the "consumer" is a GTE command and its reads are not operand positions.)
 	local args = consumer_event.args or {}
 	for _, pos in ipairs(args) do
-		if pos == destination then return true end
+		if   pos == destination then return true end
 	end
 	-- Match via the command's input set: the consumer encoder resolves to a `gte_cmdw_*`
 	-- short form whose `duffle.GTE_COMMAND_INPUTS` entry includes the destination (or a fan-out target).
@@ -448,7 +452,7 @@ local function is_cop2_consumer_of(consumer_event, destination, producer_rel)
 end
 
 -- True iff `consumer_event` reads the GPR operand at any position the destination register occupies.
--- The read-position lookup consults `duffle.OPERAND_READ_POSITIONS` for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
+-- read_pos lookup consults `duffle.OPERAND_READ_POSITIONS` for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
 local function is_gpr_consumer_of(consumer_event, destination)
 	local consumer_token = consumer_event.encoder or consumer_event.ident
 	local read_pos       = duffle.OPERAND_READ_POSITIONS or {}
@@ -576,13 +580,31 @@ local function evaluate_gpr_value_rule(rule, ev_args, gpr_values)
 		return shift_left_u4(immediate % 0x10000, 16)
 	end
 
-	local source = nil
+	-- Encoders that take `R_0` implicitly (e.g. `li_s(rt, imm)` which is `add_ui(rt, R_0, imm)`) have a non-GPR operand at the source position.
+	-- Fall back to R_0 = 0.
+	-- The implicit-R_0 macros also use a different immediate position (e.g. `li_s`'s `add_ui` rule has source = 2 / immediate = 3
+	-- but the macro takes 2 args); when the configured immediate position is out of bounds.
+	-- Fall back instead to scanning the macro's args for the first integer literal and use that as the immediate.
+	local source = 0
 	if rule.source then
-		source = constant_for_operand(gpr_values, ev_args[rule.source])
+		if is_gpr_operand(ev_args[rule.source]) then
-		if source == nil then return nil end
+			source = constant_for_operand(gpr_values, ev_args[rule.source])
+			if source == nil then return nil end
+		end
+		-- Non-GPR at source position = implicit R_0; source stays 0.
+	end
+	local immediate = nil
+	if rule.immediate and ev_args[rule.immediate] ~= nil then
+		immediate = parse_integer_literal(ev_args[rule.immediate])
+		if immediate == nil then return nil end
+	elseif rule.immediate then
+		-- Immediate position out of bounds: scan for the first integer literal in the args.
+		for _, arg in ipairs(ev_args) do
+			immediate = parse_integer_literal(arg)
+			if immediate ~= nil then break end
+		end
+		if immediate == nil then return nil end
 	end
-	local immediate = rule.immediate and parse_integer_literal(ev_args[rule.immediate]) or nil
-	if rule.immediate and immediate == nil then return nil end
 	if     operation == "add_ui"      then return wrap_u4(      source + sign_extend_i16(immediate))
 	elseif operation == "or_i"        then return bit_binary(   source, immediate % 0x10000, "or")
 	elseif operation == "and_i"       then return bit_binary(   source, immediate % 0x10000, "and")
@@ -718,8 +740,8 @@ local function consume_cu2_transition(atom, event, ev_word, forward)
 	local transition = forward.cu2_transition
 	if not transition then return end
 
-	local gap = ev_word - transition.producer_word - 1
+	local gap        = ev_word - transition.producer_word - 1
-	local target = transition.target_state
+	local target     = transition.target_state
 	local event_line = line_for_word_event(event)
 	if target == "unknown" then
 		append_cu2_finding(atom, event, forward, transition, gap, "info", "unknown",
@@ -752,8 +774,7 @@ local function consume_cu2_transition(atom, event, ev_word, forward)
 	else
 		append_cu2_finding(atom, event, forward, transition, gap,
 			"error", "exact",
-			string.format(
+			string.format("%s at line %d: COP2 unavailable after SR.CU2 was disabled"
-				"%s at line %d: COP2 unavailable after SR.CU2 was disabled"
 				.. " (gap=%d, required=%d)",
 				atom.name, event_line,
 				gap, transition.required))
@@ -831,8 +852,8 @@ local function analyze_hardware_relations(atom)
 			local is_match = false
 			if semantic == "MTC2" or semantic == "CTC2" or semantic == "LWC2_to_GTE" or semantic == "LWC2_to_other" then
 				-- Consumer is a GTE command whose input set contains the producer's COP2 destination (or a fan-out target).
-				-- LWC2_to_GTE  — GTE-command consumer: gap = 0 OK (the pipeline latches the LWC2 result).
+				-- LWC2_to_GTE   — GTE-command consumer: gap = 0 OK (the pipeline latches the LWC2 result).
-				-- LWC2_to_other — non-GTE consumer: standard load delay applies.
+				-- LWC2_to_other — non-GTE consumer:     standard load delay applies.
 				if relation.id == "lwc2_to_gte_command" then
 					is_match = is_gte_command(ev) and is_cop2_consumer_of(ev, prod.destination, relation)
 				elseif relation.id == "lwc2_to_other_consumer" then
@@ -1266,21 +1287,26 @@ local function check_hazard_nop_use(atom, _pipe_ctx, findings)
 				else
 					-- Track the slot_kind so the BD-separation case can assert the mac_yield handshake is still suppressed.
 					local slot_kind = "plain"
-					-- MIPS load-delay slot: a `load_*` wrote a register in the previous slot, and the result
+					-- MIPS load-delay slot: a `load_*` wrote a register in the previous slot, and the result is unavailable for 1 cycle.
-					-- is unavailable for 1 cycle. This `nop` is structurally required; classifying it as
+					-- This `nop` is structurally required; classifying it as `modeled-required` is the correct signal
-					-- `modeled-required` is the correct signal (removing it would make the following
+					-- (removing it would make the following instruction read the OLD value of the loaded register, a load-use hazard).
-					-- instruction read the OLD value of the loaded register, a load-use hazard). The
+					-- The `load_delay_violations` check (Concern 3) catches the actual read-side error; here we suppress the `modeled-redundant` misclassification.
-					-- `load_delay_violations` check (Concern 3) catches the actual read-side error; here
-					-- we suppress the `modeled-redundant` misclassification.
 					-- The set of load instructions mirrors the LOAD_INSTRUCTION_IDENTS in `check_load_delay_slots`.
-					local load_idents    = { load_word = true, load_half = true, load_half_u = true,
+					local load_idents = {
-					                         load_byte = true, load_byte_u = true, gte_lw = true, gte_lwc2 = true }
+						 load_word   = true,
+						 load_half   = true,
+						 load_half_u = true,
+					   load_byte   = true,
+						 load_byte_u = true,
+						 gte_lw      = true,
+						 gte_lwc2    = true 
+					}
 					local is_load_delay = load_idents[prev_ident] == true
-					if is_load_delay then
+					if    is_load_delay then
 						-- Determine the destination register from the load's `writes` field.
-						local prev_writes   = gpr_effects[prev_ident] and gpr_effects[prev_ident].writes or {}
+						local prev_writes = gpr_effects[prev_ident] and gpr_effects[prev_ident].writes or {}
-						local prev_args     = prev_ev.args or {}
+						local prev_args   = prev_ev.args or {}
-						local load_dest     = prev_writes[1] and prev_args[prev_writes[1]] or "<load-destination>"
+						local load_dest   = prev_writes[1] and prev_args[prev_writes[1]] or "<load-destination>"
 						findings[#findings + 1] = {
 							check                = "hazard_nop_use",
 							kind                 = "info",
@@ -1438,17 +1464,21 @@ end
 --- The register becomes non-volatile again at word N+2 (the load has retired), OR sooner if a non-load instruction overwrites the register
 --- (the overwriter's write is the fresh producer; the load's value is shadowed and never observed by any reader).
 ---
---- Runtime-helper atoms / components (`debug_skip == true`) are exempt: their internal load-then-use sequences
+--- Runtime-helper atoms / components (`debug_skip == true`) are exempt from some checks, but load-delay
---- are part of the fixed handshake (e.g. `ac_load_tri_indices` loads into R_T0..R_T2, but those are caller-supplied).
+--- safety applies to their emitted instructions as well.
 ---
 --- The walker reads `duffle.OPERAND_READ_POSITIONS[event.encoder]` to determine which args are read-source
 --- (the destination of a load is in `writes`, not `reads` — see `duffle.INSTRUCTION_GPR_EFFECTS`).
---- The check is purely structural; it does not consult the GPR-value lattice (no constant propagation needed for load-delay detection — the volatility window is unconditional).
+--- The check is purely structural; it does not consult the GPR-value lattice
+--- (no constant propagation needed for load-delay detection — the volatility window is unconditional).
 local function check_load_delay_slots(atom, pipe_ctx, findings)
-	if atom.kind ~= "atom" then return end
+	-- The load-delay check applies to every atom and component body, including debug-skipped components (`ac_*` and `atom_dbg_skip MipsAtom_(...)`).
-	local events = atom.paths.word_events or {}
+	-- The `atom_dbg_skip` marker controls debugger stepping, not instruction safety.
-	if #events == 0 then return end
+	-- `atom_proc` atoms have full bodies with loads that need delay slots, so the check applies to them too.
-	if is_runtime_helper(atom) then return end
+	local p = atom.paths or {}
+	if atom.kind ~= "atom" and atom.kind ~= "atom_proc" then return end
+	local events = p.word_events or {}
+	if   #events == 0 then return end
 
 	local gpr_effects = duffle.INSTRUCTION_GPR_EFFECTS or {}
 	local read_positions = duffle.OPERAND_READ_POSITIONS or {}
@@ -1550,6 +1580,8 @@ local function check_mac_yield_uniformity(atom, pipe_ctx, findings)
 	if is_runtime_helper(atom) then return end
 	-- Per-kind semantics:
 	--   MipsAtom_          (baked atom): exactly 1 mac_yield at the end of the body. Control transfer is the atom's job.
+	--   MipsAtom_Proc_     (runtime-proc atom): exactly 1 mac_yield at the end of the body. Same as baked atom;
+	--                      the proc IS the atom; the runtime call to `atombuilder_unroll` doesn't introduce a parent atom.
 	--   MipsAtomComp_      (bare static-array component): ZERO mac_yield.
 	--                      The component is invoked from inside an atom body; the parent atom does the yield.
 	--   MipsAtomComp_Proc_ (procedural component): ZERO mac_yield.
@@ -1573,7 +1605,7 @@ local function check_mac_yield_uniformity(atom, pipe_ctx, findings)
 		return atom.line + line_in_body[tokens[idx].rel]
 	end
 
-	if atom.kind == "atom" then
+	if atom.kind == "atom" or atom.kind == "atom_proc" then
 		-- Baked atom: exactly 1 yield at the end.
 		if count == 0 then
 			findings[#findings + 1] = {
@@ -1618,6 +1650,7 @@ local function check_mac_yield_uniformity(atom, pipe_ctx, findings)
 		-- The parent atom does the yield.
 		-- A yield inside a component would either be dead code (bare) or prematurely terminate the function (proc).
 		-- Both are bugs.
+		-- `atom_proc` atoms are NOT components; they're runtime-proc atoms that own their own yield (handled in the `if` branch above).
 		if count > 0 then
 			findings[#findings + 1] = {
 				atom  = atom.name,
@@ -1649,7 +1682,7 @@ end
 --- Per-atom. Runtime-helper atoms (`debug_skip`) are exempt.
 --- Takes `(atom, pipe_ctx, findings)`; `pipe_ctx` is unused.
 local function check_yield_load_tail_pairing(atom, _pipe_ctx, findings)
-	if atom.kind ~= "atom" then return end
+	if atom.kind ~= "atom" and atom.kind ~= "atom_proc" then return end
 	if is_runtime_helper(atom) then return end
 
 	local tokens       = atom.paths.tokens
@@ -1661,21 +1694,39 @@ local function check_yield_load_tail_pairing(atom, _pipe_ctx, findings)
 		return atom.line + line_in_body[tokens[idx].rel]
 	end
 
-	-- ── Rule 1: every `mac_yield_load()` must be in a branch BD-slot.
+	-- ── Rule 1: every `mac_yield_load()` must be in a branch BD-slot, OR sit between two `atom_label`s (natural fall-through load pattern).
+	-- When the pattern is satisfied, the check stays silent; only violations emit findings.
 	for tok_idx = 1, n do
 		local c = tc[tok_idx]
 		if c.ident == "mac_yield_load" then
-			if tok_idx < 2 or not tc[tok_idx - 1].is_branch then
+			local prev_tc = (tok_idx >= 2) and tc[tok_idx - 1] or nil
-				local prev_ident = (tok_idx >= 2) and (tc[tok_idx - 1].ident or "?") or "<none>"
+			-- Look for the next `atom_label()` token (skip `atom_offset` markers; check immediately-adjacent first).
-				findings[#findings + 1] = {
+			local next_label_tc = (tok_idx + 1 <= n) and tc[tok_idx + 1] or nil
-					atom  = atom.name,
+			if    next_label_tc and next_label_tc.ident ~= "atom_label" then
-					line  = tok_idx >= 2 and line_for(tok_idx) or atom.line,
+				next_label_tc = nil
-					check = "yield_load_tail_pairing",
+				for j = tok_idx + 1, n do
-					kind  = "error",
+					local t = tc[j]
-					msg   = string.format(
+					if t.ident == "atom_label" then
-						"%s at line %d has `mac_yield_load()` at word %d but the previous token is `%s`, not a branch — `mac_yield_load()` must fill a branch BD-slot."
+						next_label_tc = t
-						, atom.name, tok_idx >= 2 and line_for(tok_idx) or atom.line, tok_idx, prev_ident),
+						break
-				}
+					end
+				end
+			end
+			local  natural_fallthrough = prev_tc and prev_tc.is_atom_label and next_label_tc ~= nil
+			if not natural_fallthrough then
+				if tok_idx < 2 or not prev_tc.is_branch then
+					local prev_ident = prev_tc and (prev_tc.ident or "?") or "<none>"
+					local next_ident = next_label_tc and (next_label_tc.ident .. "(" .. (next_label_tc.label_name or "?") .. ")") or "<no following label>"
+					findings[#findings + 1] = {
+						atom  = atom.name,
+						line  = tok_idx >= 2 and line_for(tok_idx) or atom.line,
+						check = "yield_load_tail_pairing",
+						kind  = "error",
+						msg   = string.format(
+							"%s at line %d has `mac_yield_load()` at word %d but the previous token is `%s`, not a branch — and the next `atom_label()` token is `%s` — `mac_yield_load()` must fill a branch BD-slot or sit between two `atom_label`s for the natural fall-through load."
+							, atom.name, tok_idx >= 2 and line_for(tok_idx) or atom.line, tok_idx, prev_ident, next_ident),
+					}
+				end
 			end
 		end
 	end
@@ -1850,9 +1901,9 @@ end
 ---   - Atoms containing a `mac_<name>(...)` call whose `name` is not registered in `pipe_ctx.components_by_name` emit a "new macro;
 ---     Not in corpus.components" advisory — the auto-derivation returned nil for that name.
 ---
---- Applies only to `kind = "atom"` (baked atoms). Components don't emit full primitives.
+--- Applies only to `kind = "atom"` or `kind = "atom_proc"` (full-atom bodies). Components don't emit full primitives.
 local function check_gpu_portstore_shape(atom, pipe_ctx, findings)
-	if atom.kind ~= "atom" then return end
+	if atom.kind ~= "atom" and atom.kind ~= "atom_proc" then return end
 	local tokens         = atom.paths.tokens
 	local line_in_body   = atom.paths.line_in_body
 	local tc             = atom.paths.tok_class
@@ -2020,8 +2071,9 @@ local function analyze_atom_paths(atom, pipe_ctx)
 						succ[#succ + 1] = label_pos + 1
 					end
 				end
-				-- For literal-offset jumps (label == false), the target is a non-tracked address; conservatively omit.
+				-- For literal-offset jumps (label == false), control transfers out unconditionally.
-				return succ, nil
+				-- Treat as a terminator so the path is recorded (NOT as a silent fall-through to the next token, which is unreachable in this atom's execution).
+				return {}, tok_idx
 			end
 			-- Conditional branch: BD slot absorbed; two successors — fall-through (tok_idx+2) + taken (if known).
 			if tok_idx + 2 <= n then
@@ -2037,9 +2089,11 @@ local function analyze_atom_paths(atom, pipe_ctx)
 			-- Return (succ, nil), the second value is the terminator marker (nil = not a terminator).
 			return succ, nil
 		end
-		-- Normal token: just the next one
+		-- Normal token: just the next one.
+		-- The final ordinary word of the body has no successor and terminates the path;
+		-- record it as an implicit endpoint so the cycle budget for non-yield components is not silently zeroed.
 		if tok_idx + 1 <= n then return { tok_idx + 1 }, nil end
-		return {}, nil
+		return {}, tok_idx
 	end
 
 	-- DFS through all paths. Track the current cycle sum, a visited set scoped to the current path (to detect loops), and a count of paths.
@@ -2334,6 +2388,274 @@ end
 
 
 -- ════════════════════════════════════════════════════════════════════════════
+-- ════════════════════════════════════════════════════════════════════════════
+-- GTE control-register alias + RT-diagonal + TR-naming helpers and checks
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Resolve a `gte_cr_<Alias>` ident to its alias-group entry, or nil if the alias
+--- is in a distinct-slot group (or the alias name is not a known C2 control-register alias).
+--- Reads `M.GTE_CR_ALIAS_GROUPS` from `duffle.lua`.
+local function find_alias_pair_for(alias_name, duffle)
+	local groups = (duffle and duffle.GTE_CR_ALIAS_GROUPS) or {}
+	for _, group in ipairs(groups) do
+		for _, name in ipairs(group[2] or {}) do
+			if name == alias_name then return group end
+		end
+	end
+	return nil
+end
+
+-- True iff `c` (a TokClass entry) is a CPU→COP2 control-register transfer
+-- (`gte_mv_to_ctrl_r` / `gte_mv_from_ctrl_r`).
+local function is_ctrl_r_transfer(c)
+	if c == nil then return false end
+	return c.ident == "gte_mv_to_ctrl_r" or c.ident == "gte_mv_from_ctrl_r"
+end
+
+-- Resolve a token's source line. The per-token `line` is the body-relative
+-- line; `atom.line` is the source line of the atom declaration; `line_in_body`
+-- (atom.paths) maps a body-relative line to its source line. The arithmetic
+-- `atom.line + line_in_body[tok.rel] - 1` matches the convention used by
+-- check_abi_handoff and check_control_transfer_delay_slot_use elsewhere.
+local function atom_body_token_source_line(atom, token, line_in_body)
+	if line_in_body == nil or token == nil or token.rel == nil then
+		return atom.line or 0
+	end
+	local body_line = line_in_body[token.rel]
+	if body_line == nil then return atom.line or 0 end
+	return (atom.line or 0) + body_line - 1
+end
+
+-- Check #N: gte_cr_alias_writes
+-- Fires one warning per atom per alias-group when the atom body touches two
+-- distinct aliases from the same group. Aliases within a group write to the
+-- same C2 control-register slot on real silicon; cross-alias writes inside
+-- one atom body silently clobber each other.
+--
+-- Severity: warning. Build continues. The libgte outer-product convention
+-- uses only RT-row aliases (which are NOT in `M.GTE_CR_ALIAS_GROUPS`), so
+-- the canonical convention does not trigger this check.
+local function check_gte_cr_alias_writes(atom, pipe_ctx, findings)
+	local groups = pipe_ctx.gte_cr_alias_groups or {}
+	if not next(groups) then return end
+
+	local tokens = atom.paths and atom.paths.tokens or {}
+	local tc     = atom.paths and atom.paths.tok_class or {}
+	local line_in_body = atom.paths and atom.paths.line_in_body
+	if not next(tokens) then return end
+
+	-- Build a per-group set of (alias, source_line) pairs touched in this atom body.
+	-- Walks every token; when the token is a ctrl-r transfer, the alias is at
+	-- position tok_idx + 2 (rt, alias, [imm-or-arg]). The pre-classified
+	-- `tc` table tells us whether the token is a ctrl-r transfer and what its
+	-- source line is.
+	local touched = {}
+	for tok_idx, token in ipairs(tokens) do
+		local c = tc[tok_idx]
+		if is_ctrl_r_transfer(c) and tokens[tok_idx + 2] then
+			local alias = tokens[tok_idx + 2].tok
+			local group = find_alias_pair_for(alias, pipe_ctx.duffle)
+			if group then
+				touched[group[1]] = touched[group[1]] or {}
+				touched[group[1]][#touched[group[1]] + 1] = {
+					alias = alias,
+					line  = atom_body_token_source_line(atom, token, line_in_body),
+				}
+			end
+		end
+	end
+
+	-- Fire one warning per group touched with 2+ distinct aliases.
+	for slot, hits in pairs(touched) do
+		local seen = {}
+		local distinct = {}
+		for _, h in ipairs(hits) do
+			if not seen[h.alias] then
+				seen[h.alias] = true
+				distinct[#distinct + 1] = h
+			end
+		end
+		if #distinct >= 2 then
+			local aliases = {}
+			for _, d in ipairs(distinct) do aliases[#aliases + 1] = d.alias end
+			findings[#findings + 1] = {
+				atom  = atom.name or "",
+				line  = distinct[1].line,
+				check = "gte_cr_alias_writes",
+				kind  = "warning",
+				msg   = string.format(
+					"atom '%s' touches %d aliases that share C2[%d]: %s; verify the intent"
+					, atom.name or "", #distinct, slot, table.concat(aliases, ", ")),
+			}
+		end
+	end
+end
+
+-- Check #N+1: rtdiagonal_completeness
+-- Fires one info per atom body when the bare `gte_cmdw_mvmva` macro is used.
+-- The bare macro encodes only the cmd field; the canonical libgte-2-pass
+-- shape uses `gte_cmdw_mvmva_c11_pass2_exact = 0x4A49E012` (gte.h:430).
+--
+-- Severity: info by default. Escalates to warning when
+-- `GTE_RT_DIAGONAL_STRICT=1` env var is set (CI / production builds).
+--
+-- The bare macro IS the right call for the canonical libgte outer-product
+-- convention, so this is an opt-out hint rather than a hard warning.
+local function check_rtdiagonal_completeness(atom, _pipe_ctx, findings)
+	local tokens = atom.paths and atom.paths.tokens or {}
+	local tc     = atom.paths and atom.paths.tok_class or {}
+	local line_in_body = atom.paths and atom.paths.line_in_body
+	if not next(tokens) then return end
+	local strict = os.getenv("GTE_RT_DIAGONAL_STRICT") == "1"
+	for tok_idx, token in ipairs(tokens) do
+		local c = tc[tok_idx]
+		if c and c.ident == "gte_cmdw_mvmva" then
+			findings[#findings + 1] = {
+				atom  = atom.name or "",
+				line  = atom_body_token_source_line(atom, token, line_in_body),
+				check = "rtdiagonal_completeness",
+				kind  = strict and "warning" or "info",
+				msg   = string.format(
+					"atom '%s' uses the bare gte_cmdw_mvmva macro; "
+					.. "the canonical libgte-2-pass shape is gte_cmdw_mvmva_c11_pass2_exact = 0x4A49E012 "
+					.. "(gte.h:430). The bare macro does not encode RT23/RT31/RT32/RT33; "
+					.. "for a full 3x3 matrix, use the dedicated literal or hand-build via enc_gte_*()."
+					, atom.name or ""),
+			}
+		end
+	end
+end
+
+-- Check #N+2: gte_cr_TR_naming
+-- Fires one info per atom body when a `gte_cr_TR[XYZ]` alias is used.
+-- Translation-vector registers are the only 3-letter-suffix C2 aliases
+-- (`TRX/TRY/TRZ`); an agent who reads `TRX` might typo it as `RT_X` or
+-- `RTX0` and either get a compile error (best case) or a build that
+-- links but routes the `ctc2` write to the wrong C2 slot.
+--
+-- Severity: info. The convention is correct; this is a documentation-pointer check.
+local function check_gte_cr_TR_naming(atom, _pipe_ctx, findings)
+	local tokens = atom.paths and atom.paths.tokens or {}
+	local tc     = atom.paths and atom.paths.tok_class or {}
+	local line_in_body = atom.paths and atom.paths.line_in_body
+	if not next(tokens) then return end
+	local touched = false
+	local first_line = 0
+	for tok_idx, token in ipairs(tokens) do
+		local c = tc[tok_idx]
+		if c and c.ident and c.ident:match("^gte_cr_TR[XYZ]$") then
+			touched = true
+			if first_line == 0 then
+				first_line = atom_body_token_source_line(atom, token, line_in_body)
+			end
+		end
+	end
+	if touched then
+		findings[#findings + 1] = {
+			atom  = atom.name or "",
+			line  = first_line,
+			check = "gte_cr_TR_naming",
+			kind  = "info",
+			msg   = string.format(
+				"atom '%s' uses gte_cr_TR[XYZ]; translation-vector registers are the only "
+				.. "3-letter-suffix C2 aliases (TRX/TRY/TRZ). See docs/gte_reference.md §"
+				.. "\"The `gte_cmdw_mvmva_c11_pass2_exact` literal\" for the libgte outer-product "
+				.. "convention that uses these names."
+				, atom.name or ""),
+		}
+	end
+end
+
+-- check_immediate_field_width — flags integer literals passed to instruction
+-- macros that exceed the immediate field width. Reads `IMMEDIATE_FIELD_WIDTHS`
+-- from duffle.lua. Only fires on parseable integer literals; register names,
+-- O_(...) offsets, atom_offset(...) markers, and enum tokens are skipped.
+local function check_immediate_field_width(atom, pipe_ctx, findings)
+	local widths = duffle.IMMEDIATE_FIELD_WIDTHS or {}
+	local events = atom.paths and atom.paths.word_events or {}
+	local line_for_word_event = pipe_ctx.line_for_word_event
+	for _, ev in ipairs(events) do
+		local ev_ident = ev.encoder or ev.ident or "?"
+		local rules = widths[ev_ident]
+		if rules then
+			local ev_args = ev.args or {}
+			local ev_line = line_for_word_event and line_for_word_event(ev) or atom.line
+			for _, rule in ipairs(rules) do
+				local arg_str = ev_args[rule.arg]
+				if arg_str then
+					local value = parse_integer_literal(arg_str)
+					if value then
+						local width = rule.width
+						local is_signed = rule.signed == true
+						-- parse_integer_literal returns a U4-wrapped value in [0, 2^32).
+						-- For signed fields, re-interpret the high bit as the sign.
+						local signed_value = value
+						if is_signed and value >= 0x80000000 then
+							signed_value = value - 0x100000000
+						end
+						local lo, hi
+						if is_signed then
+							lo = -(bit.lshift(1, width - 1))
+							hi =  bit.lshift(1, width - 1) - 1
+						else
+							lo = 0
+							hi = bit.lshift(1, width) - 1
+						end
+						-- For unsigned fields, a negative C literal (high bit set in U4)
+						-- is valid if the low `width` bits fit — IMM_MASK truncates it.
+						-- Flag as a warning (code smell), not an error.
+						local check_value = is_signed and signed_value or value
+						local field_max = bit.lshift(1, width) - 1
+						local low_bits_fit = (value % (bit.lshift(1, width))) == value or (is_signed and signed_value >= lo and signed_value <= hi)
+						if is_signed then
+							if signed_value < lo or signed_value > hi then
+								findings[#findings + 1] = {
+									check = "immediate_field_width",
+									kind  = "error",
+									atom  = atom.name,
+									line  = ev_line,
+									msg   = string.format(
+										"%s: immediate %d at arg %d overflows %d-bit %s field (valid %d..%d)",
+										ev_ident, signed_value, rule.arg, width,
+										"signed", lo, hi),
+								}
+							end
+						else
+							-- Unsigned field: check if the low `width` bits exceed the field.
+							-- A negative C literal (U4 >= 0x80000000) whose low bits fit is
+							-- valid but a code smell — warn, don't error.
+							local low_bits = value % (bit.lshift(1, width))
+							if value > field_max then
+								if value >= 0x80000000 and low_bits <= field_max then
+									findings[#findings + 1] = {
+										check = "immediate_field_width",
+										kind  = "warning",
+										atom  = atom.name,
+										line  = ev_line,
+										msg   = string.format(
+											"%s: negative immediate %d at arg %d on unsigned %d-bit field (truncated to %d by IMM_MASK)",
+											ev_ident, signed_value, rule.arg, width, low_bits),
+									}
+								else
+									findings[#findings + 1] = {
+										check = "immediate_field_width",
+										kind  = "error",
+										atom  = atom.name,
+										line  = ev_line,
+										msg   = string.format(
+											"%s: immediate %d at arg %d overflows %d-bit unsigned field (valid 0..%d)",
+											ev_ident, value, rule.arg, width, field_max),
+									}
+								end
+							end
+						end
+					end
+				end
+			end
+		end
+	end
+end
+
 -- CHECK_RULES — data-driven check dispatch (Muratori: data over control flow)
 -- ════════════════════════════════════════════════════════════════════════════
 
@@ -2349,20 +2671,24 @@ end
 -- This is the plex pattern: the iteration is in ONE place (validate), the variation is in DATA (this table).
 
 local CHECK_RULES = {
-	{ name = "transfer_hazards",               per_atom   = check_transfer_hazards               },
+	{ name = "transfer_hazards",                per_atom   = check_transfer_hazards                },
-	{ name = "gte_input_latch",                per_atom   = check_gte_input_latch                },
+	{ name = "gte_input_latch",                 per_atom   = check_gte_input_latch                 },
-	{ name = "gte_role_mismatch",              per_atom   = check_gte_role_mismatch              },
+	{ name = "gte_role_mismatch",               per_atom   = check_gte_role_mismatch               },
-	{ name = "hazard_nop_use",                 per_atom   = check_hazard_nop_use                 },
+	{ name = "hazard_nop_use",                  per_atom   = check_hazard_nop_use                  },
-	{ name = "control_transfer_delay_slot_use",per_atom   = check_control_transfer_delay_slot_use},
+	{ name = "control_transfer_delay_slot_use", per_atom   = check_control_transfer_delay_slot_use },
-	{ name = "load_delay_violation",          per_atom   = check_load_delay_slots             },
+	{ name = "load_delay_violation",            per_atom   = check_load_delay_slots                },
-	{ name = "mac_yield_uniformity",           per_atom   = check_mac_yield_uniformity           },
+	{ name = "mac_yield_uniformity",            per_atom   = check_mac_yield_uniformity            },
-	{ name = "yield_load_tail_pairing",       per_atom   = check_yield_load_tail_pairing       },
+	{ name = "yield_load_tail_pairing",         per_atom   = check_yield_load_tail_pairing         },
-	{ name = "abi_handoff",                    per_atom   = check_abi_handoff                    },
+	{ name = "abi_handoff",                     per_atom   = check_abi_handoff                     },
-	{ name = "gpu_portstore_shape",            per_atom   = check_gpu_portstore_shape            },
+	{ name = "gpu_portstore_shape",             per_atom   = check_gpu_portstore_shape             },
-	{ name = "per_atom_cycle_budget",          per_atom   = check_per_atom_cycle_budget          },
+	{ name = "per_atom_cycle_budget",           per_atom   = check_per_atom_cycle_budget           },
-	{ name = "enum_alias_membership",          per_source = check_enum_alias_membership          },
+	{ name = "gte_cr_alias_writes",             per_atom   = check_gte_cr_alias_writes             },
-	{ name = "atom_type_consistency",          per_source = check_atom_type_consistency          },
+	{ name = "rtdiagonal_completeness",         per_atom   = check_rtdiagonal_completeness         },
-	{ name = "binds_no_substruct_deref",       per_source = check_binds_no_substruct_deref       },
+	{ name = "gte_cr_TR_naming",                per_atom   = check_gte_cr_TR_naming                },
+	{ name = "immediate_field_width",           per_atom   = check_immediate_field_width           },
+	{ name = "enum_alias_membership",           per_source = check_enum_alias_membership           },
+	{ name = "atom_type_consistency",           per_source = check_atom_type_consistency           },
+	{ name = "binds_no_substruct_deref",        per_source = check_binds_no_substruct_deref        },
 }
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -2397,12 +2723,17 @@ local function build_corpus_pipe_ctx(ctx)
 		atoms_by_name           = corpus.atoms_by_name           or {},
 		-- Per-component metadata (cycle_cost + gp0_contrib) auto-derived from the original
 		-- `MipsAtomComp_` body by `passes/components.lua::compute_components_metadata`.
-		-- Keyed by bare name (e.g. `format_f3_color`, `gte_store_f3`); the `mac_` prefix at call sites is stripped before lookup. 
+		-- Keyed by bare name (e.g. `format_f3_color`, `gte_store_f3`); the `mac_` prefix at call sites is stripped before lookup.
 		components_by_name      = corpus.components              or {},
 		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
 		atom_infos_list         = corpus.atom_infos              or {},
 		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
 		collisions              = corpus.collisions              or {},
+		-- GTE control-register alias groups (from `duffle.GTE_CR_ALIAS_GROUPS`).
+		-- The three new per_atom checks (gte_cr_alias_writes, rtdiagonal_completeness,
+		-- gte_cr_TR_naming) read from this view. `duffle` is exposed alongside so
+		-- `find_alias_pair_for` can resolve alias → group without a separate registry.
+		gte_cr_alias_groups     = duffle.GTE_CR_ALIAS_GROUPS or {},
 	}
 end
 
@@ -2521,28 +2852,27 @@ local function validate(ctx, src, corpus_pipe_ctx)
 		-- Hazard readers (transfer_hazards) populate `f.check`, `f.relation_id`, `f.semantic`, `f.direction`, `f.producer_destination`, `f.gap`, `f.required`, `f.evidence_confidence`, etc.;
 		-- Copying them through keeps the per-severity bucket schema compatible with the renderer while making the diagnostic payload queryable.
 		local payload = {
-			line = f.line,
+			line                 = f.line,
-			msg  = f.msg,
+			msg                  = f.msg,
-			check = f.check,
+			check                = f.check,
-			atom = f.atom,
+			atom                 = f.atom,
-			source = f.source,
+			source               = f.source,
-			relation_id       = f.relation_id,
+			relation_id          = f.relation_id,
-			semantic          = f.semantic,
+			semantic             = f.semantic,
-			direction         = f.direction,
+			direction            = f.direction,
 			producer_destination = f.producer_destination,
-			producer_word     = f.producer_word,
+			producer_word        = f.producer_word,
-			producer_line     = f.producer_line,
+			producer_line        = f.producer_line,
-			producer_source   = f.producer_source,
+			producer_source      = f.producer_source,
-			consumer_word     = f.consumer_word,
+			consumer_word        = f.consumer_word,
-			consumer_token    = f.consumer_token,
+			consumer_token       = f.consumer_token,
-			gap               = f.gap,
+			gap                  = f.gap,
-			required          = f.required,
+			required             = f.required,
-			evidence_confidence = f.evidence_confidence,
+			evidence_confidence  = f.evidence_confidence,
-			evidence_source   = f.evidence_source,
+			evidence_source      = f.evidence_source,
 		}
-		-- Preserve relation fields such as target_state and status_register,
+		-- Preserve relation fields such as target_state and status_register status_value,
-		-- status_value, and future policy metadata) without making the binner
+		-- and future policy metadata) without making the binner another semantic walker.
-		-- another semantic walker.
 		for key, value in pairs(f) do
 			if payload[key] == nil then payload[key] = value end
 		end
@@ -2624,7 +2954,7 @@ function M.run(ctx)
 	-- Build the corpus-wide pipe_ctx ONCE per pass run.
 	-- The pipe_ctx is shared across every validate() invocation in this M.run so cross-source visibility is constant.
 	local corpus_pipe_ctx = build_corpus_pipe_ctx(ctx)
-	local corpus = ctx.shared.corpus
+	local corpus          = ctx.shared.corpus
 
 	-- Aggregate per-DIRECTORY (per-module).
 	-- One static_analysis.txt per source-directory, emitted only if the directory contains at least one atom.
@@ -2676,8 +3006,8 @@ function M.run(ctx)
 
 		-- Aggregate per-dir errors/warnings/info into the orchestrator totals.
 		-- Hoisted out of any per-dir file-emit so `report.lua` can drop the on-disk file emitter without losing the cross-module rollup.
-		for _, e in ipairs(dir_errors)   do errors  [#errors   + 1] = e end
+		for _, e  in ipairs(dir_errors)   do errors  [#errors   + 1] = e end
-		for _, w in ipairs(dir_warnings) do warnings[#warnings + 1] = w end
+		for _, w  in ipairs(dir_warnings) do warnings[#warnings + 1] = w end
 		for _, i_ in ipairs(dir_info)     do info    [#info     + 1] = i_ end
 		-- (No per-dir emit: per-module findings are stashed on `corpus.static_analysis_results` above. 
 		-- `report.lua` reads that projection to render `<module>.atom_meta_report.md` without re-running validate().)

scripts/passes/word_count_eval.lua

@@ -93,7 +93,7 @@ end
 --- Load the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts`.
 --- Generated `.macs.h` files are OUTPUT artifacts and are NOT scanned as inputs.
 --- Current component counts are computed and inserted by `passes/components.lua`
---- after the components pass iterates `corpus.source_order` and writes each source's `<dir_basename>.macs.h` file.
+--- after the components pass iterates `corpus.source_order` and writes each source-directory's `gen/macs.h` file.
 ---
 --- Contract:
 ---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).

scripts/ps1_meta.lua

@@ -16,7 +16,7 @@
 
 -- Bootstrap: load `duffle_paths.lua` via this script's own path.
 -- Use `arg[0]` when this file is the entry script (`arg[0]` ends in "ps1_meta.lua");
--- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path, not ours).
+-- fall back to `debug.getinfo(1, "S").source` when this file is being dofile()'d or require()'d (in which case `arg[0]` is the *caller's* path).
 -- That single statement: (a) sets `package.path` + `package.cpath`, (b) at the bottom returns `require("duffle")`.
 -- So the dofile's return value is the duffle module.
 local _is_entry_script = arg and arg[0] and arg[0]:match("ps1_meta%.lua$") ~= nil
@@ -54,45 +54,44 @@ local PASS_FLAG_DISPATCH_KEY   = "__pass__"
 -- ════════════════════════════════════════════════════════════════════════════
 
 --- @class PassDescriptor
---- @field module string       -- module name passed to require()
+--- @field module string       -- Module name passed to require()
 --- @field kind   string       -- "shared" | "header-output" | "validation" | "diagnostic" | "report"
 ---                            -- Report severity is independent from process exit policy (see PASS_KIND_STOP_ON_ERROR).
---- @field deps   string[]     -- names of upstream passes
+--- @field deps   string[]     -- Names of upstream passes
---- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of
+--- @field groups string[]?    -- OPTIONAL build-phase groups this pass is a root of (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
----                            -- (e.g. { "pre-link" }, { "post-link" }); absent ⇒ dependency-only
 
 --- @class SourceFile
---- @field path     string  -- absolute path to the source file
+--- @field path     string  -- Absolute path to the source file
---- @field text     string  -- the full source text
+--- @field text     string  -- Full source text
---- @field dir      string  -- the directory containing the source
+--- @field dir      string  -- Directory containing the source
---- @field basename string  -- filename without extension
+--- @field basename string  -- Filename without extension
 
 --- @class PassCtx
---- @field metadata_path      string                 -- path to word_count.metadata.h
+--- @field metadata_path      string                 -- Path to word_count.metadata.h
---- @field shared             table                  -- cross-pass shared state
+--- @field shared             table                  -- Cross-pass shared state
---- @field shared.corpus      table                  -- canonical authored-source/project projection
+--- @field shared.corpus      table                  -- Authored-source/project projection
---- @field out_root           string                 -- output root (e.g. "build/gen")
+--- @field out_root           string                 -- Output root (e.g. "build/gen")
 --- @field project_root       string                 -- PS1 repository root
 --- @field flags              table                  -- CLI flags + per-pass stash
---- @field verbose            boolean                -- if true, log diagnostic info
+--- @field verbose            boolean                -- If true, log diagnostic info
 
 --- @class Finding
---- @field line integer  -- source line (or 0 for pass-level)
+--- @field line integer  -- Source line (or 0 for pass-level)
---- @field msg  string   -- finding message
+--- @field msg  string   -- Finding message
 
 --- @class PassResult
---- @field outputs  PassOutputEntry[] -- emitted file paths
+--- @field outputs  PassOutputEntry[] -- Emitted file paths
---- @field errors   Finding[]         -- build-stops (per-pass kind policy)
+--- @field errors   Finding[]         -- Build-stops (per-pass kind policy)
---- @field warnings Finding[]         -- informational
+--- @field warnings Finding[]         -- Informational
 
 --- @class ParsedArgs
---- @field requested_set string[]   -- pass names to run (explicit --all expanded)
+--- @field requested_set string[]   -- Pass names to run (explicit --all expanded)
---- @field sources       string[]   -- exact --source values, retained in CLI order
+--- @field sources       string[]   -- Exact --source values, retained in CLI order
 --- @field unity_root    string|nil -- --unity-root value; mutually exclusive with sources
 --- @field metadata      string     -- --metadata value
 --- @field out_root      string     -- --out-root value (default "build/gen")
 --- @field project_root  string     -- PS1 repository root (derived from metadata by default)
---- @field verbose       boolean    -- if true, log diagnostic info
+--- @field verbose       boolean    -- If true, log diagnostic info
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- PASSES Table
@@ -119,6 +118,12 @@ local PASSES = {
 		kind   = "header-output",
 		deps   = {"scan-source", "word-counts"},
 	},
+	auto_reg = {
+		module = "passes.auto_reg",
+		kind   = "header-output",
+		deps   = {"components"},
+		groups = { "pre-link" },
+	},
 	["emission-model"] = {
 		module = "passes.emission_model",
 		kind   = "validation",
@@ -138,7 +143,7 @@ local PASSES = {
 	["static-analysis"] = {
 		module = "passes.static_analysis",
 		-- "diagnostic" — every `error`/`warning` finding is written to the report file;
-		-- the orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
+		-- The orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
 		-- Report severity is independent from process exit policy.
 		kind   = "diagnostic",
 		deps   = {"scan-source", "word-counts", "components", "emission-model"},
@@ -163,12 +168,12 @@ local PASSES = {
 }
 
 -- ────────────────────────────────────────────────────────────────────────────
--- Phase-root selection: derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
+-- Phase-root selection: Derive the sorted set of roots belonging to a named build-phase group, then append them to `args.requested_set`.
 -- topo_sort closes the transitive deps from there; dispatch_passes runs every resolved pass without phase-filtering.
 -- ────────────────────────────────────────────────────────────────────────────
 
---- @param group_name string  -- the build-phase group ("pre-link" | "post-link")
+--- @param group_name string  -- Build-phase group ("pre-link" | "post-link")
---- @return string[]          -- sorted root pass names belonging to that group
+--- @return string[]          -- Sorted root pass names belonging to that group
 local function roots_for_group(group_name)
 	local names = {}
 	for name, pass in pairs(PASSES) do
@@ -206,7 +211,7 @@ end
 -- Report severity is independent from process exit policy.
 -- A "diagnostic" pass still writes every `error`/`warning` finding into its report file,
 -- but `report_validation_errors` returns early for non-stopping kinds, so nothing is printed to stderr and the orchestrator does not exit non-zero.
--- Adding a new pass kind requires listing it here explicitly; an unknown kind must not silently fall back to "true".
+-- Adding a new pass kind requires listing it here explicitly; An unknown kind must not silently fall back to "true".
 local PASS_KIND_STOP_ON_ERROR = {
 	["shared"]        = false,
 	["header-output"] = true,
@@ -216,8 +221,7 @@ local PASS_KIND_STOP_ON_ERROR = {
 }
 
 -- Closed set of CLI flags -> pass names.
--- Per-pass flags (e.g. --word-counts) live here; phase flags (--pre-link, --post-link, --all)
+-- Per-pass flags (e.g. --word-counts); phase flags (--pre-link, --post-link, --all) are within FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
--- live in FLAG_HANDLERS because they own side effects or invoke group-derivation logic.
 -- dwarf-injection is *also* a per-pass opt-in flag, but its selection + opt-in state are both owned by the explicit FLAG_HANDLERS entry below
 -- (it sets args.flags.dwarf_injection and appends "dwarf-injection" to requested_set), so it is intentionally absent from this table.
 local PASS_FLAG_TO_NAME = {
@@ -246,7 +250,6 @@ end
 
 -- Per-flag handlers. Each handler takes (args, argv, arg_idx) and returns the new arg_idx (so multi-arg flags like --source FILE advance it). 
 -- Returning nil + os.exit() handles termination flags (--help). 
-
 local FLAG_HANDLERS = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
@@ -272,9 +275,9 @@ PASS_FLAGS:
   Or pick any subset:
     --scan-source         Scan sources into the fat SourceScan payload
     --word-counts         Load metadata.h + scan for existing .macs.h
-    --components          Generate <module>/gen/<basename>.macs.h
+    --components          Generate <srcdir>/gen/macs.h (per-directory aggregation)
     --validate            Run atom annotation DSL validation
-    --offsets             Generate <module>/gen/<basename>.offsets.h
+    --offsets             Generate <srcdir>/gen/offsets.h (per-directory aggregation)
     --atoms-source-map    Generate <basename>.atoms.sourcemap.txt per source
     --dwarf-injection     [opt-in] Select the post-link dwarf-injection pass + set the opt-in flag. Requires --elf.
     --static-analysis     Static analysis: GTE pipeline-fill, mac_yield, ABI handoff, cycle budget
@@ -317,8 +320,7 @@ local function require_flag_value(argv, arg_idx, flag)
 	local next_known = type(value) == "string"
 		and (FLAG_HANDLERS[value] ~= nil or PASS_FLAG_TO_NAME[value] ~= nil)
 	if value == nil or next_known then
-		io.stderr:write("ps1_meta: " .. flag .. " requires "
+		io.stderr:write("ps1_meta: " .. flag .. " requires " .. FLAG_VALUE_NAMES[flag] .. "\n")
-			.. FLAG_VALUE_NAMES[flag] .. "\n")
 		os.exit(EXIT_INTERNAL_ERROR)
 	end
 	return value, arg_idx + 1
@@ -328,12 +330,10 @@ 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["--verbose"] = function(args) args.verbose = true 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
@@ -490,14 +490,13 @@ end
 --- @param args ParsedArgs
 --- @return PassCtx
 local function build_ctx(args)
-	local normalized_project_root = duffle.normalize_path(args.project_root)
+	local normalized_project_root  = duffle.normalize_path(args.project_root)
-	local project_root            = normalized_project_root
+	local project_root             = normalized_project_root
 	local project_root_is_absolute = normalized_project_root:match("^%a:/")
 		or normalized_project_root:sub(1, 2) == "//"
 		or normalized_project_root:sub(1, 1) == "/"
 	if not project_root_is_absolute then
-		-- canonical_path_key validates ordinary relative paths and rejects
+		-- canonical_path_key validates ordinary relative paths and rejects drive-relative paths before the absolute-path rewrite is performed.
-		-- drive-relative paths before the absolute-path rewrite is performed.
 		duffle.canonical_path_key(normalized_project_root)
 		project_root = duffle.normalize_path(duffle.to_absolute_path(normalized_project_root))
 	else
@@ -511,8 +510,7 @@ local function build_ctx(args)
 			project_root = project_root,
 		})
 		if not ok_resolve then
-			io.stderr:write("ps1_meta: cannot resolve --unity-root "
+			io.stderr:write("ps1_meta: cannot resolve --unity-root " .. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
-				.. tostring(args.unity_root) .. ": " .. tostring(resolved) .. "\n")
 			os.exit(EXIT_INTERNAL_ERROR)
 		end
 		resolution = resolved
@@ -528,8 +526,7 @@ local function build_ctx(args)
 			local path = duffle.normalize_path(input_path)
 			local key_ok, key_or_error = pcall(duffle.canonical_path_key, path)
 			if not key_ok then
-				error("ps1_meta: invalid --source " .. input_path .. ": "
+				error("ps1_meta: invalid --source " .. input_path .. ": " .. tostring(key_or_error), 0)
-					.. tostring(key_or_error), 0)
 			end
 			local file = io.open(path, "r")
 			if not file then
@@ -627,9 +624,7 @@ local function topo_sort(passes, requested_set)
 		changed = false
 		for name, _ in pairs(needed) do
 			local pass = passes[name]
-			if not pass then
+			if not pass then error("unknown pass '" .. name .. "' requested") end
-				error("unknown pass '" .. name .. "' requested")
-			end
 			for _, dep in ipairs(pass.deps) do
 				if not needed[dep] then
 					needed[dep] = true

scripts/update_deps.ps1

@@ -14,16 +14,21 @@ $url_armips     = 'https://github.com/Kingcom/armips.git'
 $url_pcsx_redux = 'https://github.com/grumpycoders/pcsx-redux.git'
 $url_psyq_iwyu  = 'https://github.com/johnbaumann/psyq_include_what_you_use.git'
 $url_lpeg       = 'https://github.com/roberto-ieru/LPeg.git'
+# $url_mkpsxiso   = 'https://github.com/Lameguy64/mkpsxiso.git'
+
+$url_mkpsxiso_win64 = 'https://github.com/Lameguy64/mkpsxiso/releases/download/v2.30/mkpsxiso-2.30-win64.zip'
 
 $path_armips     = join-path $path_toolchain 'armips'
 $path_pcsx_redux = join-path $path_toolchain 'pcsx-redux'
 $path_psyq_iwyu  = join-path $path_toolchain 'psyq_iwyu'
 $path_lpeg       = join-path $path_toolchain 'lpeg'
+$path_mkpsxiso   = join-path $path_toolchain 'mkpsxiso'
 
 clone-gitrepo $path_armips     $url_armips
 clone-gitrepo $path_lpeg       $url_lpeg
 clone-gitrepo $path_pcsx_redux $url_pcsx_redux
 clone-gitrepo $path_psyq_iwyu  $url_psyq_iwyu
+# clone-gitrepo $path_mkpsxiso   $url_mkpsxiso
 
 $path_armips_build = join-path $path_armips 'build'
 verify-path $path_armips_build
@@ -56,7 +61,120 @@ if (-not $msbuild_exe) {
 }
 
 $path_pcsx_sln = join-path $path_pcsx_redux 'vsprojects\pcsx-redux.sln'
-&	$msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
+
+# ════════════════════════════════════════════════════════════════════════════
+# NuGet restore — required before MSBuild.
+# pcsx-redux's .vcxproj files use the legacy packages.config style with
+# hardcoded `<Import Project="..\packages\{id}.{ver}\...">` directives.
+# MSBuild's `/t:Restore` won't fetch missing packages here (the local
+# packages\ dir is checked but no package-source lookup happens), and
+# `dotnet restore` errors on packages.config projects, so we walk every
+# packages.config, parse out the <package id version/> entries, and pull
+# any missing .nupkg directly from api.nuget.org's flat container.
+# ════════════════════════════════════════════════════════════════════════════
+$path_pcsx_packages = join-path $path_pcsx_redux 'vsprojects\packages'
+$nuget_flat_container = 'https://api.nuget.org/v3-flatcontainer'
+
+# Collect required (id, version) pairs from every packages.config.
+$required_packages = @{}
+Get-ChildItem -Path (join-path $path_pcsx_redux 'vsprojects') -Filter 'packages.config' -Recurse -ErrorAction SilentlyContinue |
+    ForEach-Object {
+        [xml]$xml = Get-Content -LiteralPath $_.FullName -Raw
+        foreach ($pkg in $xml.packages.package) {
+            $key = '{0}|{1}' -f $pkg.id, $pkg.version
+            $required_packages[$key] = @{ id = $pkg.id; version = $pkg.version }
+        }
+    }
+
+# Ensure the packages root exists.
+if (-not (Test-Path -LiteralPath $path_pcsx_packages)) {
+    New-Item -ItemType Directory -Path $path_pcsx_packages -Force | Out-Null
+}
+
+# Download anything missing. Skip the package entirely if its dir already has
+# any contents (the legacy packages.config style means the targets file
+# location varies per package — `luajit.native` puts it at build/native/,
+# `glfw` puts it elsewhere — so we can't probe a specific path; just check
+# whether the dir is non-empty).
+Add-Type -AssemblyName System.IO.Compression.FileSystem
+foreach ($pkg in $required_packages.Values) {
+    $pkgDir = Join-Path $path_pcsx_packages ('{0}.{1}' -f $pkg.id, $pkg.version)
+    if ((Test-Path -LiteralPath $pkgDir) -and `
+        (@(Get-ChildItem -LiteralPath $pkgDir -Recurse -ErrorAction SilentlyContinue).Count -gt 0)) {
+        continue
+    }
+    $url = '{0}/{1}/{2}/{1}.{2}.nupkg' -f $nuget_flat_container, $pkg.id, $pkg.version
+    $nupkg = Join-Path $pkgDir ('{0}.{1}.nupkg' -f $pkg.id, $pkg.version)
+    New-Item -ItemType Directory -Path $pkgDir -Force | Out-Null
+    Write-Host "Fetching NuGet package: $($pkg.id) $($pkg.version)"
+    try {
+        Invoke-WebRequest -Uri $url -OutFile $nupkg -UseBasicParsing -ErrorAction Stop
+        [System.IO.Compression.ZipFile]::ExtractToDirectory($nupkg, $pkgDir)
+        Remove-Item -LiteralPath $nupkg -Force
+    } catch {
+        $msg = $_.Exception.Message
+        if ($msg -match '404') {
+            Write-Host "  Not on nuget.org (vendored?) — skipping $url"
+        } else {
+            Write-Warning "Failed to fetch $url — $msg"
+        }
+        if (Test-Path -LiteralPath $nupkg) { Remove-Item -LiteralPath $nupkg -Force }
+    }
+}
+
+# ════════════════════════════════════════════════════════════════════════════
+# isoffi.lua size guard — `core.vcxproj` #includes src/core/isoffi.lua into
+# luaiso.cc via the `-- lualoader, R"EOF(...)EOF"` trick. The raw string
+# literal between R"EOF(-- and -- )EOF" must stay under ~16,379 bytes or
+# MSVC (19.44) fails with C2026 (its actual raw-string limit is 16,384,
+# minus 5 bytes for the `-- lualoader, ` prefix). If the upstream file
+# grows past that, trim it: remove license header, trailing whitespace,
+# blank separators, inline comments, and shrink 4-space indent to 2-space.
+# Idempotent — only writes when the raw string exceeds the limit.
+# ════════════════════════════════════════════════════════════════════════════
+$path_isoffi = join-path $path_pcsx_redux 'src\core\isoffi.lua'
+if (Test-Path -LiteralPath $path_isoffi) {
+    $content = Get-Content -LiteralPath $path_isoffi -Raw -Encoding utf8
+    $startMarker = $content.IndexOf('R"EOF(--')
+    $endMarker   = $content.IndexOf('-- )EOF"')
+    $literalLen  = if ($startMarker -ge 0 -and $endMarker -gt $startMarker) {
+        $endMarker - ($startMarker + 8)
+    } else { -1 }
+    # Effective MSVC raw-string limit for the lualoader prefix is 16379 bytes.
+    if ($literalLen -gt 16379) {
+        Write-Host "isoffi.lua raw string is $literalLen bytes (>16379); trimming for MSVC C2026 limit."
+        $lines     = $content -split "`n"
+        $markerIdx = -1
+        for ($i = 0; $i -lt $lines.Length; $i++) {
+            if ($lines[$i] -match '^-- \)EOF"') { $markerIdx = $i; break }
+        }
+        $newLines = @()
+        for ($i = 0; $i -lt $lines.Length; $i++) {
+            $lineNum = $i + 1
+            $line    = $lines[$i]
+            # Keep the first line and the EOF-marker line untouched.
+            if ($i -eq 0 -or $i -eq $markerIdx) { $newLines += $line; continue }
+            # Drop the GPL license header (lines 2-17).
+            if ($lineNum -ge 2 -and $lineNum -le 17) { continue }
+            # Drop blank separator lines.
+            if ($line -match '^\s*$') { continue }
+            # Drop trailing whitespace.
+            $line = $line -replace '\s+$', ''
+            # Drop inline comments (anything from `--` to end of line).
+            $line = $line -replace '\s*--.*$', ''
+            # Shrink 4-space indent to 2-space.
+            $line = $line -replace '^(    )', '  '
+            if ($line -match '^\s*$') { continue }
+            $newLines += $line
+        }
+        ($newLines -join "`n") | Out-File -LiteralPath $path_isoffi -Encoding utf8 -NoNewline
+        $newLen = ((Get-Content -LiteralPath $path_isoffi -Raw -Encoding utf8) `
+            -replace '.*R"EOF\(--', '' -replace '-- \)EOF".*', '').Length
+        Write-Host "isoffi.lua trimmed: $literalLen -> $newLen bytes of raw string content."
+    }
+}
+
+& $msbuild_exe $path_pcsx_sln /p:Configuration=Release /p:Platform=x64 /p:PlatformToolset=v143 /m /v:minimal
 
 # Locate luajit via scoop. `luajit.exe` is on PATH via scoop's shim;
 # we use `scoop prefix` to find the install root for the include dir (needed to compile lpeg against luajit's headers).
@@ -70,8 +188,8 @@ if (-not $luajit_prefix -or -not (Test-Path (Join-Path $luajit_prefix 'bin/luaji
 # Discover the luajit include dir by globbing `include/luajit-*`.
 # This avoids hardcoding a specific version (e.g. `luajit-2.1`).
 $luajit_include_root = Join-Path $luajit_prefix 'include'
-$lua_inc_dir = Get-ChildItem -Path $luajit_include_root -Directory -Filter 'luajit-*' -ErrorAction SilentlyContinue |
+$lua_inc_dir         = Get-ChildItem -Path $luajit_include_root -Directory -Filter 'luajit-*' -ErrorAction SilentlyContinue |
-	Select-Object -First 1 -ExpandProperty FullName
+Select-Object -First 1 -ExpandProperty FullName
 if (-not $lua_inc_dir) {
 	write-error "No 'luajit-*' include dir found under '$luajit_include_root'. The scoop luajit install may be broken."
 	exit 1
@@ -90,7 +208,7 @@ $lpeg_compile_args = @(
 	'-o', 'lpeg.dll'
 ) + $lpeg_sources + @('-lluajit-5.1')
 push-location $path_lpeg
-&	gcc @lpeg_compile_args
+& gcc @lpeg_compile_args
 pop-location
 
 # ════════════════════════════════════════════════════════════════════════════
@@ -101,8 +219,8 @@ pop-location
 
 $path_lfs = join-path $path_toolchain 'lfs'
 verify-path $path_lfs
-$lfs_src = join-path $path_pcsx_redux 'third_party\luafilesystem\src\lfs.c'
+$lfs_src        = join-path $path_pcsx_redux 'third_party\luafilesystem\src\lfs.c'
-$lfs_dll = join-path $path_lfs 'lfs.dll'
+$lfs_dll        = join-path $path_lfs 'lfs.dll'
 $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 & gcc -O2 -shared "-I$lua_inc_dir" -o $lfs_dll $lfs_src $lfs_dll_import
 
@@ -112,6 +230,17 @@ $lfs_dll_import = join-path $luajit_lib_dir 'libluajit-5.1.dll.a'
 # ════════════════════════════════════════════════════════════════════════════
 
 $path_openbios = join-path $path_pcsx_redux 'src\mips\openbios'
+
+# Wipe stale *.dep files across src\mips. These cache absolute paths to the
+# GCC headers directory; if the toolchain was upgraded (e.g. v14.2.0 → v16.1.0)
+# Make reads the stale paths and aborts with "no rule to make target .../stddef.h".
+# `make clean` in openbios only clears its own dir — subdirs like
+# common/crt0/, modplayer/, and shell/ keep their stale .dep files. Easier to
+# just delete the lot before each build than to teach every Makefile about
+# deepclean recursion.
+Get-ChildItem -Path (join-path $path_pcsx_redux 'src\mips') -Recurse -Filter '*.dep' -ErrorAction SilentlyContinue |
+    ForEach-Object { Remove-Item -LiteralPath $_.FullName -Force }
+
 push-location $path_openbios
 &	make clean
 &	make