starting to optimize

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

.gitignore

@@ -17,3 +17,6 @@ toolchain/PSn00bSDK
 .vscode/settings.json
 toolchain/lfs
 toolchain/lpeg
+
+scratch
+toolchain/libpsn00b

.vscode/launch.json

@@ -4,7 +4,7 @@
 	// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
 	"version": "0.2.0",
 	"configurations": [
-		{
+{
 			"name": "Debug: Hello Psy-Q!",
 			"type": "gdb",
 			"request": "attach",
@@ -12,6 +12,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -20,10 +24,17 @@
 			"osx": {
 				"gdbpath": "gdb"
 			},
-			"executable": "${workspaceRoot}/build/hello_psyq.elf",
+			"executable": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
-				"load hello_psyq.elf",
+				"load hello_gte.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
 				"tbreak main",
 				"continue"
 			]
@@ -36,30 +47,10 @@
 			"remote": true,
 			"cwd": "${workspaceRoot}/build",
 			"valuesFormatting": "parseText",
-			"stopAtConnect": true,
-			"gdbpath": "gdb-multiarch",
-			"windows": {
-				"gdbpath": "gdb-multiarch.exe"
-			},
-			"osx": {
-				"gdbpath": "gdb"
-			},
-			"executable": "${workspaceRoot}/build/hello_gpu.elf",
-			"autorun": [
-				"monitor reset shellhalt",
-				"load hello_gpu.elf",
-				"tbreak main",
-				"continue"
-			]
-		},
-		{
-			"name": "Debug: Hello GTE Psy-Q!",
-			"type": "gdb",
-			"request": "attach",
-			"target": "localhost:3333",
-			"remote": true,
-			"cwd": "${workspaceRoot}/build",
-			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
 			"stopAtConnect": true,
 			"gdbpath": "gdb-multiarch",
 			"windows": {
@@ -69,12 +60,89 @@
 				"gdbpath": "gdb"
 			},
 			"executable": "${workspaceRoot}/build/hello_gte.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
 			"autorun": [
 				"monitor reset shellhalt",
 				"load hello_gte.elf",
 				"tbreak main",
 				"continue"
 			]
+		},
+		{
+			"name": "Debug: Hello GTE!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${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": "${workspaceRoot}/build/hello_gte.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_gte.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_gte.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"tbreak main",
+				"continue"
+			]
+		},
+		{
+			"name": "Debug: Hello Joypad!",
+			"type": "gdb",
+			"request": "attach",
+			"target": "localhost:3333",
+			"remote": true,
+			"cwd": "${workspaceRoot}",
+			"valuesFormatting": "parseText",
+			"registerLimit": "1-32",
+			"frameFilters": false,
+			"showDevDebugOutput": false,
+			"printCalls": false,
+			"stopAtConnect": true,
+			"gdbpath": "gdb-multiarch",
+			"windows": {
+				"gdbpath": "gdb-multiarch.exe"
+			},
+			"osx": {
+				"gdbpath": "gdb"
+			},
+			"executable": "${workspaceRoot}/build/hello_joypad.dwarf-injected.elf",
+			"setupCommands": [
+				{ "text": "set mi-async off" },
+				{ "text": "set remotetimeout 0" },
+				{ "text": "set logging file build/gen/hello_joypad.gdb.log" },
+				{ "text": "set logging redirect on" }
+			],
+			"autorun": [
+				"monitor reset shellhalt",
+				"load build/hello_joypad.dwarf-injected.elf",
+				"source scripts/gdb/gdb_tape_atoms.gdb",
+				"tbreak main",
+				"continue"
+			]
 		}
 	]
 }
+

code/duffle/atom_dsl.h

@@ -3,34 +3,18 @@
  * ============================================================================
  *
  * ATOM DSL: Annotation layer for tape atoms (lottes_tape.h).
+ * The metaprogram (scripts/passes/annotation.lua) reads source-as-written and validates:
+ *   - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X), atom_reads(...), atom_writes(...)) in any order and are optional.
+ *   - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a real Binds_* struct declaration.
+ *   - atom word-counts in word_counts.metadata.h match the body's actual .word count.
  *
- *   WHAT THIS HEADER IS
- *   -------------------
- *   The metaprogram (scripts/passes/annotation.lua) reads source-as-written
- *   and validates:
- *     - atom_info(...) shape: up to three sub-calls (atom_bind(Binds_X),
- *       atom_reads(...), atom_writes(...)) in any order. All optional.
- *       (No phase token for now; phases may be reintroduced later.)
- *     - rbind atoms (atom_info(..., atom_bind(Binds_X), ...)) reference a
- *       real Binds_* struct declaration.
- *     - wave-context positions only reference the canonical 4-register
- *       set: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
- *     - atom word-counts in word_counts.metadata.h agree with the body's
- *       actual .word count.
- *
- *   WHY A PURE MACRO (atom_info, atom_bind, atom_reads, atom_writes, atom_label)
- *   -----------------------------------------------------------------
- *   Each of these expands to a C comment or to nothing. The C preprocessor
- *   strips them to whitespace. The metaprogram reads the literal token from
- *   source-as-written, NOT from the preprocessed output. This means:
- *     - the C compiler does no work for them (no __attribute__, no
- *       _Pragma, no asm side-effects)
- *     - they can never silently drift from the metaprogram's view
- *       (the metaprogram re-reads the source on every build)
- *     - the annotation is invisible to the linker, debugger, and IDE
+ * Pure macro anntation.
+ * ---------------
+ * Don't want to constraint the macro usage to some attribute placment constraint, etc, don't want ot dela with the compiler.
+ * atom_info, atom_bind, atom_reads, atom_writes, atom_label, atom_dbg_skip each expand to a C comment or to nothing
+ * (C preprocessor strips them to whitespace).
  *
  * ============================================================================
- *
  *  Usage:
  *      MipsAtom_(cube_tri) atom_info(
  *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
@@ -60,30 +44,15 @@
  *
  *  Annotation rules
  *  ----------------
- *      1. atom_info(...) is OPTIONAL. Most atoms have no annotation.
- *         Atoms without atom_info are silently skipped by the metaprogram.
- *
- *      2. If present, atom_info takes up to three sub-calls, all
- *         order-independent within the arg list:
- *           - atom_bind(Binds_X) (optional; only for rbind atoms)
- *           - atom_reads(...) (optional; wave-context registers)
- *           - atom_writes(...) (optional; wave-context registers)
- *
- *      3. atom_bind(Binds_X) pins the ABI-struct shape -- the metaprogram
- *         cross-references Binds_X against the
- *         `typedef struct Binds_X { ... } Binds_X;` declaration.
- *
- *      4. atom_reads(...) and atom_writes(...) args are wave-context
- *         registers: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
- *         Closed set. GTE / SP / DMA / I/O state is declared in source
- *         comments, not in atom_reads/atom_writes.
- *
- *      5. atom_label(name) is an anchor -- the macro is empty in C; the
- *         metaprogram records the marker at the current pos for offset
- *         calculation.
- *
- *      6. atom_offset(F, T) is resolved by gen/atom_offsets.h, generated
- *         from the atom_label markers.
+ *      1. atom_info(...) is OPTIONAL. Atoms without atom_info are silently skipped by the metaprogram.
+ *      2. If present, atom_info takes up to three sub-calls, all order-independent within the arg list:
+ *           - atom_bind(Binds_X)
+ *           - atom_reads(...)
+ *           - atom_writes(...)
+ *      3. atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *      4. atom_reads(...) and atom_writes(...): Used to to check if registers are used correctly in macros: R_PrimCursor / R_FaceCursor / R_VertBase / R_OtBase.
+ *      5. atom_label(name: Utilize with atom_offset as a target location.
+ *      6. atom_offset(F, T): Resolved by gen/atom_offsets.h, generated from the atom_label markers. Calculated during the offset pass of the lua  metaprogram.
  */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -92,64 +61,74 @@
 #endif
 
 /* ============================================================================
- * WAVE-CONTEXT REGISTERS -- canonical register set for the tape wave model.
- *
- *   R_PrimCursor  output pointer into the prim arena (next OT entry to write)
- *   R_FaceCursor  input pointer into the face array (next face to consume)
- *   R_VertBase    base pointer into the vertex arena (this wave's vertices)
- *   R_OtBase      base pointer into the ordering table (this wave's OT slot)
- *
- * Closed set. If your atom needs to touch GTE / SP / DMA / other side state,
- * declare it at the source level as you normally would -- but DO NOT put
- * those registers in atom_reads/atom_writes.
- *
- * ============================================================================*/
-
-/* ============================================================================
- * atom_reads(...) / atom_writes(...) -- wave-context register list
- *
- *   atom_reads(R_PrimCursor, R_FaceCursor)
- *         ->  (R_PrimCursor, R_FaceCursor)   // comma-evaluated, discarded
- *
- * The macro produces a comma-evaluated expression that the C compiler
- * silently discards (it sits in an unused arg position -- the result is
- * never bound). The Lua tool pattern-matches the "atom_reads(...)" /
- * "atom_writes(...)" token to extract the list.
- *
- * You can have at most one atom_reads(...) and at most one atom_writes(...)
- * in an atom_info(...) call. To declare multiple disjoint sets (rare), just
- * declare the union -- the metaprogram doesn't track which reads need which
- * writes at this granularity.
+ * atom_reads(...) / atom_writes(...)
  *
+ * Used during the static analysis pass of the metaprogram to do
  * ============================================================================*/
 #define atom_reads(...)  (__VA_ARGS__)
 #define atom_writes(...) (__VA_ARGS__)
 
+/* ----------------------------------------------------------------------------
+ * 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.
+ * The C preprocessor strips it to a comment so no runtime symbol is created; the Lua scanner reads the bare token.
+ * ----------------------------------------------------------------------------*/
+#define atom_reg /* atom_reg: opt the preceding enum entry into the DWARF registry */
+
 /* ============================================================================
- * ATOM ANNOTATION MACROS
- *
- *   atom_info -- single unified annotation. OPTIONAL. Most atoms have none.
- *
+ *   atom_info :
  *      MipsAtom_(cube_tri) atom_info(
  *         atom_reads (R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
  *       , atom_writes(R_PrimCursor, R_FaceCursor)
  *      ){ ... };
  * 
- *   Shape (sub-args order-independent; all optional):
- *     - atom_bind(Binds_X):  at most one; pins the ABI-struct shape
- *     - atom_reads(...):     at most one; comma-list of wave-context registers
- *     - atom_writes(...):    at most one; comma-list of wave-context registers
- *
- *   No phase token for now. The metaprogram doesn't check ordering across
- *   atoms -- phases (init / bind / setup / work / commit / terminate) will
- *   be reintroduced when ordering checks are added.
- *
- *   The macro expands to a C comment (or to nothing). The C compiler does
- *   no work. The metaprogram reads the source-as-written directly.
- *
+ *  - atom_bind(Binds_X): metaprogram cross-references Binds_X against the `typedef struct Binds_X { ... } Binds_X;` declaration.
+ *  - atom_reads(...):    comma-list of registers
+ *  - atom_writes(...):   comma-list of registers
  * ============================================================================*/
 #define atom_info(...)  /* atom_info(__VA_ARGS__) */
 
+/* ----------------------------------------------------------------------------
+ * DEBUG SOURCE-STEP MARKER
+ *
+ * Place `atom_dbg_skip` (BARE) before a MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_.
+ * The following declaration kind determines whether the marker selects a whole atom or a component inline view.
+ * The source scanner associates the marker with that declaration; placement diagnostics are handled by the annotation pass.
+ *
+ * Example:
+ *     atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+ *     atom_dbg_skip MipsAtomComp_(ac_yield) { ... };
+ *     atom_dbg_skip MipsAtomComp_Proc_(ac_format_f3_color, { ... });
+ * ----------------------------------------------------------------------------*/
+#define atom_dbg_skip /* atom_dbg_skip: skip the following atom or component source view */
+
+/* ----------------------------------------------------------------------------
+ * Typed-view annotations (Registry for DWARF RR_<R_X> chain resolution)
+ *   atom_type(<T>) -- overloaded:
+ *      (a) enum-site default: `R_Foo = R_Tn, atom_reg atom_type(T)` 
+ *      Sets the per-alias default typed view in the register_alias_registry.
+ *      Consumed by the DWARF chain step (e) when no per-atom atom_ctx / atom_phase / atom_type callsite provides a stronger resolution.
+ *      (b) callsite override: `atom_reads(R_Foo atom_type(T), ...)` Overrides the per-alias default for THIS atom only.
+ *      Last-write-wins per R_Name; conflict -> error.
+ *   atom_ctx(<atom_name>) -- atom-info sub-call:
+ *       Propagate another atom's atom.rbind.fields (its Binds_* typed fields) into THIS atom's typed-view resolution.
+ *       The named atom must be an rbind atom (have `atom_bind(Binds_X)` in its `atom_info`).
+ *       Used as the escape hatch when atom_phase is not the natural correlation.
+ *   atom_phase(<label>) -- atom-info sub-call: 
+ *       Free-form C-identifier label for grouping atoms.
+ *       Within a phase, the FIRST atom in source-order that owns its own atom.rbind provides 
+ *       the Binds_* field types used by all other atoms in the same phase.
+ *       The preferred correlation mechanism; atom_ctx is the escape hatch for non-natural cases.
+ *
+ * All three expand to C comments
+ * (the bare-token convention matching `atom_reg` and `atom_dbg_skip`).
+ * The Lua scanner reads the bare tokens in source-as-written; the C preprocessor strips them.
+ * ----------------------------------------------------------------------------*/
+#define atom_type(T)        /* atom_type: associate <T> with the preceding enum entry (enum site) or this register (atom-info site) */
+#define atom_ctx(atom_name) /* atom_ctx: propagate <atom_name>'s Binds_* field types into this atom's typed views */
+#define atom_phase(label)   /* atom_phase: tag this atom with <label> for grouped typed-view resolution */
+
 /* ----------------------------------------------------------------------------
  * atom_bind(Binds_X) -- rbind sub-call of atom_info
  *
@@ -158,13 +137,7 @@
  *       , atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
  *   ){ ... };
  *
- * The Binds_X MUST be a typedef'd type (declared via
- * `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
- * The Lua tool cross-references this. Missing struct = error.
- *
- * atom_bind is a SUB-CALL of atom_info, not a standalone annotation macro.
- *
- * The macro expands to a C comment. The metaprogram reads source-as-written.
+ * The Binds_X MUST be a typedef'd type (declared via `typedef struct Binds_X { ... } Binds_X;` somewhere in the source).
  * ----------------------------------------------------------------------------*/
 #define atom_bind(binds_struct)  /* atom_bind(binds_struct) */
 
@@ -177,25 +150,11 @@
  *
  *   atom_offset(culling, bounds_chk)         ← resolved by gen/.offsets.h
  *
- * The metaprogram generates gen/atom_offsets.h with one
- *   #define atom_offset__culling__bounds_chk  ((target - branch_pos - 1))
- * per atom_offset(F, T) call. The preprocessor then expands your 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.
- * This catches:
- *   - typo in atom_label (no anchor → metaprogram doesn't emit the macro)
- *   - .offsets.h not regenerated after body edits
- *   - body edit that broke the offset math (recompile + retest picks it up
- *     in CPU emulator)
+ * The metaprogram generates gen/atom_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.
  * ============================================================================*/
-
 #define atom_offset(F, T) atom_offset_ ## F ## _ ## T
-/* atom_label is a pure annotation for the metaprogram's offset calculations.
- * The macro expands to a C comment, so the C preprocessor strips it to
- * whitespace — NO instruction word is emitted in the asm. The metaprogram
- * still recognises the literal `atom_label(name)` token in source and
- * records the marker at the current pos. */
+// atom_label is a pure annotation for the metaprogram's offset calculations.
 #define atom_label(name) /* atom_label anchor: name */

code/duffle/dsl.h

@@ -135,16 +135,17 @@ enum { false = 0, true  = 1, true_overflow, };
 
 typedef void Proc_(VoidFn) (void);
 
-#define kilo(n)         (C_(U4, n) << 10)
-#define mega(n)         (C_(U4, n) << 20)
-#define giga(n)         (C_(U4, n) << 30)
-#define tera(n)         (C_(U4, n) << 40)
-#define null             C_(U4,    0)
-#define nullptr          C_(void*, 0)
-#define O_(type, field) (C_(U4, & C_(type*,0)->field))
+#define kilo(n)                (C_(U4, n) << 10)
+#define mega(n)                (C_(U4, n) << 20)
+#define giga(n)                (C_(U4, n) << 30)
+#define tera(n)                (C_(U4, n) << 40)
 
-#define OT_(field)       O_(typeof_ptr(& field), filed))
-#define S_(data)         C_(U4, sizeof(data))
+#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), filed))
+#define S_(data)                C_(U4, sizeof(data))
 
 #define sop_1(op,a,b) C_(U1, s1_(a) op s1_(b))
 #define sop_2(op,a,b) C_(U2, s2_(a) op s2_(b))
@@ -218,3 +219,6 @@ IA_ void assert(U8 cond) { if(cond){return;} else{debug_trap(); ms_exit_process(
 #endif
 #pragma endregion Debug
 #endif
+
+#define GCC_OPTIMIZATION_DISABLE _Pragma("GCC push_options") _Pragma("GCC optimize(\"O0\")")
+#define GCC_OPTIMIZATION_ENABLE  _Pragma("GCC pop_options")

code/duffle/gcc_asm.h

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

code/duffle/gen/duffle.macs.h

@@ -9,6 +9,12 @@
 #define WORD_COUNT(name, count)  enum { words_##name = (count) };
 #endif
 
+/* atom_dbg_skip */
+/* ---------------------------------------------------------------------------
+ *  MACRO ATOM Components (Reusable Assembly Components)
+ *  These do NOT yield. They are expanded inline inside Tape Atoms.
+ * ---------------------------------------------------------------------------*/
+// The 'Yield' sequence for Tape Atoms (mac_yield).
 #define mac_yield(...) \
 	load_word(R_AtomJmp, R_TapePtr, 0) \
 ,	add_ui_self(         R_TapePtr, S_(MipsCode)) \
@@ -16,6 +22,7 @@
 ,	nop
 WORD_COUNT(mac_yield, 4)
 
+/* atom_dbg_skip */
 /* Words: 3; Loads 3 S2 indices from the face array */
 #define mac_load_tri_indices(...) \
 	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)) \
@@ -23,6 +30,7 @@ WORD_COUNT(mac_yield, 4)
 ,	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2))
 WORD_COUNT(mac_load_tri_indices, 3)
 
+/* atom_dbg_skip */
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
 #define mac_gte_load_tri_verts(...) \
 	shift_lleft(R_AT, R_T0, v3s2_byteoff) \
@@ -75,23 +83,26 @@ WORD_COUNT(mac_insert_ot_tag_f3, 11)
 ,	store_word(  R_AT, R_T1,         O_(PolyTag,code))        /* OrderingTable[OTZ] = PrimCursor */
 WORD_COUNT(mac_insert_ot_tag_g4, 11)
 
+/* atom_dbg_skip */
 #define mac_pack_color_word(off, cmd, r, g, b) \
 	load_upper_i(R_AT, (cmd) << 8  | (b)) \
-,	or_i_self(   R_AT, ((g)   << 8) | (r)) \
+,	or_i_self(   R_AT, ((g)  << 8) | (r)) \
 ,	store_word(  R_AT, R_PrimCursor, (off))
 WORD_COUNT(mac_pack_color_word, 3)
 
+/* atom_dbg_skip */
 #define mac_format_f3_color(r, g, b) \
 	mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b)
 WORD_COUNT(mac_format_f3_color, 3)
 
+/* atom_dbg_skip */
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices to the F3.
  * PIPELINE: post-RTPT (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). */
-#define mac_gte_store_f3_post_rtpt(...) \
+#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_post_rtpt, 3)
+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) \
@@ -100,27 +111,24 @@ WORD_COUNT(mac_gte_store_f3_post_rtpt, 3)
 ,	mac_pack_color_word(O_(Poly_G4,c3), 0,               r3,g3,b3)
 WORD_COUNT(mac_format_g4_color, 12)
 
+/* atom_dbg_skip */
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
  * G4 triangle portion to p0/p1/p2.
  * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
- * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
- * three registers aligned with v0/v1/v2 you must store before RTPS).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-#define mac_gte_store_g4_p012_post_rtpt_pre_rtps(...) \
+ * 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_post_rtpt_pre_rtps, 3)
+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 — DO NOT read SXY0 here, that's the bug this name
- * prevents).
+ * 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_post_rtps(...) \
+#define mac_gte_store_g4_p3(...) \
 	gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3))
-WORD_COUNT(mac_gte_store_g4_p3_post_rtps, 1)
+WORD_COUNT(mac_gte_store_g4_p3, 1)
 

code/duffle/gp.h

@@ -1,7 +1,6 @@
 /* ============================================================================
  *  duffle DSL Suffix Conventions
  *  ============================================================================
- *
  *  Every mnemonic in this header follows the same suffix grammar:
  *
  *  Primitive commands: gp0_cmd_poly_f3 = 0x20    (byte opcode)
@@ -26,8 +25,7 @@
  *    0. Opcode IDs             gp0_cmd_poly_f3 = 0x20
  *
  *  Vendor mnemonics (gte_mtc2, gte_mfc2, etc.) are NOT in this header.
- *  They live in the opt-in `gp_vendor_sym.h` for users who prefer the
- *  PSYQ-style names.
+ *  They live in the opt-in `gp_vendor_sym.h` for users who prefer the PSYQ-style names.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -41,29 +39,28 @@
 /* ============================================================================
  *  Hardware MMIO Addresses
  *  ============================================================================
- *
- *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2
- *  0x1F800000+. GP0 (offset 0x10) is the data port (commands + params).
+ *  PSX GPU has two 32-bit ports in the I/O register region at KSEG2 0x1F800000+.
+ *  GP0 (offset 0x10) is the data port (commands + params).
  *  GP1 (offset 0x14) is the control port (status, ctrl writes).
  * ============================================================================ */
 /* IO base address (KSEG2 0x1F800000+ for the I/O register region).
- * The 16-bit upper half `IO_BASE_ADDR_HI16` is the form used by
- * tape-side macros that pin a register to hold the IO base and access
- * ports via offsets — `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
+ * The 16-bit upper half `IO_BASE_ADDR_HI16` is the form used by tape-side macros that pin a register
+ * to hold the IO base and access ports via offsets:
+ *   `lui $reg, 0x1F80` (1 word) then `sw $data, GPIO_PORT*_OFFSET($reg)` (1 word). 
  * Mirrors the `IO_BASE_ADDR equ 0x1F80` + `gpio_port0 equ 0x1810` pattern from graphics_hello/gp.s. */
 enum {
-    IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
-    IO_BASE_ADDR_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
+	IO_BASE_ADDR      = 0x1F800000,  /* full 32-bit I/O region base */
+	IO_BASE_ADDR_HI16 = 0x1F80,      /* fits in a single `lui $reg, 0x1F80` */
 
-    /* Offsets from IO_BASE_ADDR to each port. Used by tape-side macros
-     * that pin a register to IO_BASE_ADDR and access ports via offsets:
-     *   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
-     *   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
-    GPIO_PORT0_OFFSET = 0x1810,
-    GPIO_PORT1_OFFSET = 0x1814,
+	/* Offsets from IO_BASE_ADDR to each port. Used by tape-side macros
+		* that pin a register to IO_BASE_ADDR and access ports via offsets:
+		*   sw $data, GPIO_PORT0_OFFSET($io_base)   ; write GP0
+		*   sw $data, GPIO_PORT1_OFFSET($io_base)   ; write GP1 */
+	GPIO_PORT0_OFFSET = 0x1810,
+	GPIO_PORT1_OFFSET = 0x1814,
 
-    HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
-    HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
+	HW_GP0_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT0_OFFSET,
+	HW_GP1_ADDR = (IO_BASE_ADDR_HI16 << 16) | GPIO_PORT1_OFFSET,
 };
 
 #define HW_GP0 C_(U4 V_*, HW_GP0_ADDR)
@@ -75,77 +72,71 @@ enum {
 /* ============================================================================
  *  GP0 command byte constants + Layer 1 (GPU bitfield shifts)
  *  ============================================================================
- *
  *  8-bit GP0 opcodes (the upper byte of a primitive's first word). These are the BYTE only.
- *  The layer-1 bitfield-layout constants live in the same enum block
- *  so the encoder can reference them by name. 
  *  NO macro body past this point uses a raw shift or raw mask.
- *  Every shift/width/mask is named here, named once. 
  *  Mirrors the OPCODE_SHIFT / RS_SHIFT / REG_MASK convention from mips.h.
  * ============================================================================ */
 enum {
-    gp0_cmd_Nop             = 0x00,
+	gp0_cmd_Nop             = 0x00,
 
-    /* Cache management */
-    gp0_cmd_ClearCache      = 0x01,
-    gp0_cmd_FillVram        = 0x02,
-    gp0_cmd_CopyVram        = 0x80,
-    gp0_cmd_CopyVramChained = 0x81,
-    gp0_cmd_ReadVram        = 0xC0,
+	/* Cache management */
+	gp0_cmd_ClearCache      = 0x01,
+	gp0_cmd_FillVram        = 0x02,
+	gp0_cmd_CopyVram        = 0x80,
+	gp0_cmd_CopyVramChained = 0x81,
+	gp0_cmd_ReadVram        = 0xC0,
 
-    /* Polygons */
-    gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
-    gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
-    gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
-    gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
-    gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
-    gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
-    gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
-    gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
+	/* Polygons */
+	gp0_cmd_poly_f3         = 0x20,  /* Flat Triangle           */
+	gp0_cmd_poly_ft3        = 0x24,  /* Flat Textured Triangle  */
+	gp0_cmd_poly_g3         = 0x30,  /* Gouraud Triangle        */
+	gp0_cmd_poly_gt3        = 0x34,  /* Gouraud Textured Tri    */
+	gp0_cmd_poly_f4         = 0x28,  /* Flat Quad               */
+	gp0_cmd_poly_ft4        = 0x2C,  /* Flat Textured Quad      */
+	gp0_cmd_poly_g4         = 0x38,  /* Gouraud Quad            */
+	gp0_cmd_poly_gt4        = 0x3C,  /* Gouraud Textured Quad   */
 
-    /* Lines */
-    gp0_cmd_line_f2         = 0x40,
-    gp0_cmd_line_g2         = 0x50,
+	/* Lines */
+	gp0_cmd_line_f2         = 0x40,
+	gp0_cmd_line_g2         = 0x50,
 
-    /* Sprites + Tiles + Rects */
-    gp0_cmd_sprt_1          = 0x64,
-    gp0_cmd_sprt_8          = 0x74,
-    gp0_cmd_sprt_16         = 0x7C,
-    gp0_cmd_tile_1          = 0x60,
-    gp0_cmd_tile_8          = 0x68,
-    gp0_cmd_tile_16         = 0x70,
+	/* Sprites + Tiles + Rects */
+	gp0_cmd_sprt_1          = 0x64,
+	gp0_cmd_sprt_8          = 0x74,
+	gp0_cmd_sprt_16         = 0x7C,
+	gp0_cmd_tile_1          = 0x60,
+	gp0_cmd_tile_8          = 0x68,
+	gp0_cmd_tile_16         = 0x70,
 
-    /* State setters (not drawing primitives; set render context). */
-    gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
-    gp0_cmd_SetTextureWindow     = 0xE2,
-    gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
-    gp0_cmd_SetDrawArea_BotRight = 0xE4,
-    gp0_cmd_SetDrawOffset        = 0xE5,
-    gp0_cmd_SetMaskBit           = 0xE6,
+	/* State setters (not drawing primitives; set render context). */
+	gp0_cmd_DrawModeSetting      = 0xE1,  /* TPage / draw-mode (semi-trans, dither, etc.) */
+	gp0_cmd_SetTextureWindow     = 0xE2,
+	gp0_cmd_SetDrawArea_TopLeft  = 0xE3,
+	gp0_cmd_SetDrawArea_BotRight = 0xE4,
+	gp0_cmd_SetDrawOffset        = 0xE5,
+	gp0_cmd_SetMaskBit           = 0xE6,
 
-    /* bitfield shifts / widths / masks ----
-     * 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,
+	/* bitfield shifts / widths / masks ----
+		* 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_blue_shift  = 16, gp0_color_blue_width  = 8, gp0_color_blue_mask  = 0xFF,
-    gp0_color_green_shift =  8, gp0_color_green_width = 8, gp0_color_green_mask = 0xFF,
-    gp0_color_red_shift   =  0, gp0_color_red_width   = 8, gp0_color_red_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_blue_shift  = 16, gp0_color_blue_width  = 8, gp0_color_blue_mask  = 0xFF,
+	gp0_color_green_shift =  8, gp0_color_green_width = 8, gp0_color_green_mask = 0xFF,
+	gp0_color_red_shift   =  0, gp0_color_red_width   = 8, gp0_color_red_mask   = 0xFF,
 };
 
 /* ============================================================================
  *  Layer 1.5 (per-field encoders) + Layer 2 (composite) + Layer 3 (semantic GP0 word builders)
  *  ============================================================================
- *
- *  Layer 1.5 encoders take one field's value, mask it to its own width,
- *  and shift it to its own position. 
+ *  Layer 1.5 encoders take one field's value, mask it to its own width, and shift it to its own position. 
  *  Mirrors `enc_op` / `enc_rs` / `enc_rt` in mips.h and `enc_gte_sf` / `enc_gte_mx` in gte.h. 
  *  Layer-2 composite encoders OR the per-field encoders together; layer-3 semantic macros delegate to the composites.
  *  No raw shifts or magic numbers in any macro body below this point.
@@ -178,73 +169,73 @@ enum {
 #define gp0_word_poly_gt4(r,g,b) enc_color_word(gp0_cmd_poly_gt4, (r),(g),(b))
 
 /* Cache management — bare-cmd words (no color/range payload). */
-#define gp0_word_clear_cache() enc_gp0_cmd_word(gp0_cmd_ClearCache)
-#define gp0_word_fill_vram()   enc_gp0_cmd_word(gp0_cmd_FillVram)
-#define gp0_word_copy_vram()   enc_gp0_cmd_word(gp0_cmd_CopyVram)
-#define gp0_word_read_vram()   enc_gp0_cmd_word(gp0_cmd_ReadVram)
+#define gp0_word_clear_cache()       enc_gp0_cmd_word(gp0_cmd_ClearCache)
+#define gp0_word_fill_vram()         enc_gp0_cmd_word(gp0_cmd_FillVram)
+#define gp0_word_copy_vram()         enc_gp0_cmd_word(gp0_cmd_CopyVram)
+#define gp0_word_read_vram()         enc_gp0_cmd_word(gp0_cmd_ReadVram)
+
+/* NOP — bare-cmd word (no effect; used as DR_ENV padding). */
+#define gp0_word_nop()               enc_gp0_cmd_word(gp0_cmd_Nop)
 
 /* ============================================================================
  *  GP1 command byte constants + Layer 1 (display-mode + range + draw-area bitfield shifts)
  *  ============================================================================
- *
- *  GP1 status bits are read from HW_GP1; ctrl writes use GP1 commands
- *  packed into 32-bit words (cmd byte in the upper 8 bits via
- *  `enc_gp0_cmd(cmd)` — never a raw shift).
+ *  GP1 status bits are read from HW_GP1;
+ *  ctrl writes use GP1 commands packed into 32-bit words 
+ *  (cmd byte in the upper 8 bits via `enc_gp0_cmd(cmd)`).
  * ============================================================================ */
 enum {
-    gp1_cmd_Reset                  = 0x00,
-    gp1_cmd_ResetCmdBuffer         = 0x01,
-    gp1_cmd_AcknowledgeIRQ         = 0x02,
-    gp1_cmd_DisplayEnable          = 0x03,
-    gp1_cmd_DMADirection           = 0x04,
-    gp1_cmd_StartDisplayArea       = 0x05,
-    gp1_cmd_HorizontalDisplayRange = 0x06,
-    gp1_cmd_VerticalDisplayRange   = 0x07,
-    gp1_cmd_DisplayMode            = 0x08,
-    /* Note: GP1 only has commands 0x00..0x08. 
-		 * The state-setter commands (SetTextureWindow, * SetDrawArea*, 
-		 * SetDrawOffset, SetMaskBit) live in the GP0 enum as * 0xE1..0xE6. 
-		 * DrawArea word builders are below as GP0s * macros 
-		 * (since they emit GP0 commands). */
+	gp1_cmd_Reset                  = 0x00,
+	gp1_cmd_ResetCmdBuffer         = 0x01,
+	gp1_cmd_AcknowledgeIRQ         = 0x02,
+	gp1_cmd_DisplayEnable          = 0x03,
+	gp1_cmd_DMADirection           = 0x04,
+	gp1_cmd_StartDisplayArea       = 0x05,
+	gp1_cmd_HorizontalDisplayRange = 0x06,
+	gp1_cmd_VerticalDisplayRange   = 0x07,
+	gp1_cmd_DisplayMode            = 0x08,
+	/* Note: GP1 only has commands 0x00..0x08. 
+		* The state-setter commands (SetTextureWindow, * SetDrawArea*, SetDrawOffset, SetMaskBit) 
+		* live in the GP0 enum as * 0xE1..0xE6.
+		* DrawArea word builders are below as GP0s * macros (since they emit GP0 commands). */
 
-    /* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
-     * Bit positions match the encoder shifts below; values are the
-     * *payload* bits only (the cmd byte is OR'd in by enc_gp1_disp_mode_word). */
-    gp1_disp_HRes_256   = 0x0,
-    gp1_disp_HRes_320   = 0x1,
-    gp1_disp_HRes_512   = 0x2,
-    gp1_disp_HRes_640   = 0x3,
-    gp1_disp_VRes_240   = 0x0,
-    gp1_disp_VRes_480   = 0x1,
-    gp1_disp_Color15    = 0x0,
-    gp1_disp_Color24    = 0x1,
-    gp1_disp_VInterlace = 0x1,
+	/* ---- Display-mode payload flags (per PSX-SPX §"GP1 Display Mode").
+		* Bit positions match the encoder shifts below; values are the *payload* bits only (cmd byte is OR'd in by enc_gp1_disp_mode_word). */
+	gp1_disp_HRes_256   = 0x0,
+	gp1_disp_HRes_320   = 0x1,
+	gp1_disp_HRes_512   = 0x2,
+	gp1_disp_HRes_640   = 0x3,
+	gp1_disp_VRes_240   = 0x0,
+	gp1_disp_VRes_480   = 0x1,
+	gp1_disp_Color15    = 0x0,
+	gp1_disp_Color24    = 0x1,
+	gp1_disp_VInterlace = 0x1,
 
-    /* ---- Layer 1: GP1 display-mode + range + draw-area shifts/masks ---- */
-    gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2, gp1_disp_hres_mask      = 0x3,
-    gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1, gp1_disp_vres_mask      = 0x1,
-    gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1, gp1_disp_color_mask     = 0x1,
-    gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1, gp1_disp_interlace_mask = 0x1,
+	/* ---- Layer 1: GP1 display-mode + range + draw-area shifts/masks ---- */
+	gp1_disp_hres_shift      =  0, gp1_disp_hres_width      = 2, gp1_disp_hres_mask      = 0x3,
+	gp1_disp_vres_shift      =  2, gp1_disp_vres_width      = 1, gp1_disp_vres_mask      = 0x1,
+	gp1_disp_color_shift     =  4, gp1_disp_color_width     = 1, gp1_disp_color_mask     = 0x1,
+	gp1_disp_interlace_shift =  5, gp1_disp_interlace_width = 1, gp1_disp_interlace_mask = 0x1,
 
-    /* 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_x2_shift =  0, gp1_hrange_x2_width = 12, gp1_hrange_x2_mask = 0xFFF,
+	/* 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_x2_shift =  0, gp1_hrange_x2_width = 12, gp1_hrange_x2_mask = 0xFFF,
 
-    /* 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_y2_shift =  0, gp1_vrange_y2_width = 10, gp1_vrange_y2_mask = 0x3FF,
+	/* 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_y2_shift =  0, gp1_vrange_y2_width = 10, gp1_vrange_y2_mask = 0x3FF,
 
-    /* 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) */
-    gp1_draw_x_shift =  0, gp1_draw_x_width = 10, gp1_draw_x_mask = 0x3FF,
-    gp1_draw_y_shift = 10, gp1_draw_y_width = 10, gp1_draw_y_mask = 0x3FF,
+	/* 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) */
+	gp1_draw_x_shift =  0, gp1_draw_x_width = 10, gp1_draw_x_mask = 0x3FF,
+	gp1_draw_y_shift = 10, gp1_draw_y_width = 10, gp1_draw_y_mask = 0x3FF,
 };
 
 /* ---- 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_vres(v)      (((v) & gp1_disp_vres_mask)     << gp1_disp_vres_shift)
-#define enc_gp1_disp_color(c)     (((c) & gp1_disp_color_mask)    << gp1_disp_color_shift)
-#define enc_gp1_disp_interlace(i) (((i) & gp1_disp_interlace_mask << gp1_disp_interlace_shift)
+#define enc_gp1_disp_hres(h)      (((h) & gp1_disp_hres_mask)      << gp1_disp_hres_shift)
+#define enc_gp1_disp_vres(v)      (((v) & gp1_disp_vres_mask)      << gp1_disp_vres_shift)
+#define enc_gp1_disp_color(c)     (((c) & gp1_disp_color_mask)     << gp1_disp_color_shift)
+#define enc_gp1_disp_interlace(i) (((i) & gp1_disp_interlace_mask) << gp1_disp_interlace_shift)
 
 #define enc_gp1_hrange_x1(x1) (((x1) & gp1_hrange_x1_mask) << gp1_hrange_x1_shift)
 #define enc_gp1_hrange_x2(x2) (((x2) & gp1_hrange_x2_mask) << gp1_hrange_x2_shift)
@@ -259,12 +250,16 @@ enum {
 #define enc_gp1_vrange_word(y1, y2)        (enc_gp0_cmd(gp1_cmd_VerticalDisplayRange)   | enc_gp1_vrange_y1(y1) | enc_gp1_vrange_y2(y2))
 
 /* ---- Layer 2: GP0 state-setter composite encoders ----
- * GP0(0xE3) SetDrawArea top-left and GP0(0xE4) SetDrawArea bottom-right
- * both use the same X/Y 10-bit signed payload as GP1 DisplayRange. */
+ * GP0(0xE3) SetDrawArea top-left and GP0(0xE4) SetDrawArea bottom-right both use the same X/Y 10-bit signed payload as GP1 DisplayRange. */
 #define enc_gp0_draw_area_tl_word(x, y)    (enc_gp0_cmd(gp0_cmd_SetDrawArea_TopLeft)  | enc_gp1_draw_x(x) | enc_gp1_draw_y(y))
 #define enc_gp0_draw_area_br_word(x, y)    (enc_gp0_cmd(gp0_cmd_SetDrawArea_BotRight) | enc_gp1_draw_x(x) | enc_gp1_draw_y(y))
 
 /* ---- Layer 3: GP1 semantic word builders ---- */
+#define gp1_word_Reset()             enc_gp0_cmd_word(gp1_cmd_Reset)
+#define gp1_word_ResetCmdBuffer()    enc_gp0_cmd_word(gp1_cmd_ResetCmdBuffer)
+#define gp1_word_AcknowledgeIRQ()    enc_gp0_cmd_word(gp1_cmd_AcknowledgeIRQ)
+#define gp1_word_StartDisplayArea()  enc_gp0_cmd_word(gp1_cmd_StartDisplayArea)
+
 #define gp1_word_display_enable(on)                         (enc_gp0_cmd(gp1_cmd_DisplayEnable) | ((on) & 1))
 #define gp1_word_display_disable()                          gp1_word_display_enable(0)
 #define gp1_word_display_mode_320x240_15bit_ntsc            enc_gp1_disp_mode_word(gp1_disp_HRes_320, gp1_disp_VRes_240, gp1_disp_Color15, 0)
@@ -282,7 +277,6 @@ enum {
 /* ============================================================================
  *  Pre-baked GPU state words
  *  ============================================================================
- *
  *  Common command words for boot-time GPU init and standard display configurations.
  * ============================================================================ */
 
@@ -290,31 +284,36 @@ enum {
 #define gp1_word_display_enabled   enc_gp0_cmd_word(gp1_cmd_DisplayEnable)
 #define gp1_word_display_disabled (enc_gp0_cmd_word(gp1_cmd_DisplayEnable) | 1)
 
+#define gp1_word_DisplayOn()       gp1_word_display_enable(0)
+#define gp1_word_DisplayOff()      gp1_word_display_enable(1)
+
 /* ---- DMA direction (2-bit payload on DMADirection cmd 0x04) ---- */
 enum {
-    gp1_dma_dir_Off            = 0,
-    gp1_dma_dir_FIFO           = 1,
-    gp1_dma_dir_CPU_to_GPU     = 2,
-    gp1_dma_dir_GPUREAD_to_CPU = 3,
+	gp1_dma_dir_Off            = 0,
+	gp1_dma_dir_FIFO           = 1,
+	gp1_dma_dir_CPU_to_GPU     = 2,
+	gp1_dma_dir_GPUREAD_to_CPU = 3,
 };
 #define gp1_word_dma_direction(dir) (enc_gp0_cmd(gp1_cmd_DMADirection) | ((dir) & 0x3))
+#define gp1_word_dma_to_gpu()       gp1_word_dma_direction(gp1_dma_dir_CPU_to_GPU)
+#define gp1_word_dma_read_cpu()     gp1_word_dma_direction(gp1_dma_dir_GPUREAD_to_CPU)
 
 /* ---- Standard display ranges (NTSC + PAL pre-baked) ---- */
 /* Horizontal range values are in video clock units (8 units/pixel); vertical range values are scanline numbers. */
 enum {
-    /* NTSC horizontal range: X1=608, X2=3168 */
-    gp1_hrange_NTSC_x1 = 0x260,
-    gp1_hrange_NTSC_x2 = 0xC60,
-    /* PAL horizontal range (same as NTSC for most CRTs) */
-    gp1_hrange_PAL_x1  = 0x260,
-    gp1_hrange_PAL_x2  = 0xC60,
+	/* NTSC horizontal range: X1=608, X2=3168 */
+	gp1_hrange_NTSC_x1 = 0x260,
+	gp1_hrange_NTSC_x2 = 0xC60,
+	/* PAL horizontal range (same as NTSC for most CRTs) */
+	gp1_hrange_PAL_x1  = 0x260,
+	gp1_hrange_PAL_x2  = 0xC60,
 
-    /* NTSC vertical range: Y1=24, Y2=264 */
-    gp1_vrange_NTSC_y1 = 24,
-    gp1_vrange_NTSC_y2 = 264,
-    /* PAL vertical range: Y1=24, Y2=504 */
-    gp1_vrange_PAL_y1  = 24,
-    gp1_vrange_PAL_y2  = 504,
+	/* NTSC vertical range: Y1=24, Y2=264 */
+	gp1_vrange_NTSC_y1 = 24,
+	gp1_vrange_NTSC_y2 = 264,
+	/* PAL vertical range: Y1=24, Y2=504 */
+	gp1_vrange_PAL_y1  = 24,
+	gp1_vrange_PAL_y2  = 504,
 };
 
 #define gp1_word_horizontal_range_ntsc enc_gp1_hrange_word(gp1_hrange_NTSC_x1, gp1_hrange_NTSC_x2)
@@ -325,14 +324,49 @@ enum {
 /* ---- Draw-mode setting (TPage / draw-area allowance) ---- */
 /* The "drawing enabled" word is the standard post-init state. */
 enum {
-    gp0_DrawMode_DrawToDispBit = 10,
+	/* Per psx-spx, the standard 0xE1 layout has dfe at bit 10. But libpsyx's PutDrawEnv
+	 * uses bit 19 (in the "unused" 14-23 range) for dfe in the DR_ENV code[0] — and the
+	 * PSX hardware honors bit 19 in the DR_ENV context (not bit 10). So we need a
+	 * separate bit definition for the DR_ENV-specific DrawMode. */
+	gp0_DrawMode_DrawToDispBit         = 10,   // standard psx-spx bit 10 (dfe)
+	gp0_DrawMode_DR_ENV_DrawToDispBit  = 19,   // libpsyx DR_ENV code[0] (dfe in DR_ENV context)
+	gp0_DrawMode_DR_ENV_isbgBit        = 19,   // libpsyx uses bit 19 for isbg too
 };
 #define gp0_word_draw_mode_drawing_allowed (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit))
 
-/* ---- DrawArea pre-baked at origin (0,0) and full screen (320x240) ---- */
+/* DR_ENV-specific DrawMode variants (libpsyx SetDrawEnv layout).
+ * The DR_ENV is a 16-word packet emitted at boot by gp_screen_init's ac_put_draw_env_demo
+ * atom component. Within the DR_ENV, the 0xE1 command is reused in three different bit
+ * configurations:
+ *   code[0] = `gp0_word_draw_mode_drawing_allowed` (dfe=1; standard post-init state)
+ *   code[6] = `gp0_word_dr_env_bg_color_cmd(isbg, r, g, b)` (initial-bg-color path)
+ *   code[7] = `gp0_word_dr_env_draw_mode(isbg)` (isbg-flag path)
+ * Bits 0-23 of the 0xE1 word are the payload; bits 24-31 are the cmd byte (0xE1). */
+#define gp0_word_dr_env_bg_color_cmd(isbg, r, g, b) (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0) | enc_gp0_color_r(r) | enc_gp0_color_g(g) | enc_gp0_color_b(b))
+#define gp0_word_dr_env_draw_mode(isbg)             (enc_gp0_cmd(gp0_cmd_DrawModeSetting) | (1 << gp0_DrawMode_DrawToDispBit) | ((isbg) ? gp0_dr_env_isbg_bit : 0))
+
+/* State-setter bare-cmd words (no immediate payload; the GPU uses the current state machine already programmed). */
+#define gp0_word_set_texture_window() enc_gp0_cmd_word(gp0_cmd_SetTextureWindow)
+#define gp0_word_set_draw_offset()    enc_gp0_cmd_word(gp0_cmd_SetDrawOffset)
+#define gp0_word_set_mask_bit()       enc_gp0_cmd_word(gp0_cmd_SetMaskBit)
+
+/* DR_ENV code[5] Mask (0xE6 cmd + isbg bit). The isbg bit is set so the GPU knows the auto-clear path is active (paired with code[6] + code[7]). */
+#define gp0_word_dr_env_mask() (gp0_word_set_mask_bit() | gp0_dr_env_isbg_bit)
+
+/* DR_ENV pre-baked constants (libpsyx PutDrawEnv layout).
+ * DR_ENV is a 16-word packet: tag = (length << 24) | addr, where length = 15 (15 code words follow) and addr = 0 (chain to nothing). */
+enum {
+	PolyTag_len_bits  =  8,
+	PolyTag_addr_bits = 24,
+
+	gp0_dr_env_tag      = (15 << 24) | 0x00FFFFFF,
+	gp0_dr_env_isbg_bit = (1 << gp0_DrawMode_DR_ENV_isbgBit),
+};
+
+/* ---- DrawArea at origin (0,0) and full screen (320x240) ---- */
 #define gp0_word_draw_area_top_left_origin      enc_gp0_draw_area_tl_word(0, 0)
-#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(320, 240)
-#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(640, 480)
+#define gp0_word_draw_area_bottom_right_320x240 enc_gp0_draw_area_br_word(319, 239)
+#define gp0_word_draw_area_bottom_right_640x480 enc_gp0_draw_area_br_word(639, 479)
 
 #pragma endregion GPU Ports & Commands
 
@@ -343,9 +377,9 @@ enum {
  *  Read from HW_GP1; the lower bits are DMA-block-size (variable-width).
  * ============================================================================ */
 enum {
-    gp1_Status_BitReady          = 31,
-    gp1_Status_BitSendingDMA     = 25,
-    gp1_Status_DMABlockSizeShift =  0,
+	gp1_Status_BitReady          = 31,
+	gp1_Status_BitSendingDMA     = 25,
+	gp1_Status_DMABlockSizeShift =  0,
 };
 
 #define gp1_status_is_ready()       ((HW_GP1[0] >> gp1_Status_BitReady)      & 1)
@@ -356,7 +390,6 @@ enum {
 /* ============================================================================
  *  Primitive structs (8 polygon variants + tag)
  *  ============================================================================
- *
  *  Each struct follows the GPU-documented memory layout for the corresponding primitive command. 
  *  The PolyTag is the OT-link header; the rest of the struct is the primitive's body.
  *
@@ -372,10 +405,10 @@ typedef Struct_(RGB8) { B1 r; B1 g; B1 b; };
 #define rgb8(r,g,b) ((RGB8){r,g,b})
 
 /* ---------- PolyTag (the OT-link header; 1 word) ---------- */
-enum {
-    PolyTag_len_bits  =  8,
-    PolyTag_addr_bits = 24,
-};
+// enum {
+// 	PolyTag_len_bits  =  8,
+// 	PolyTag_addr_bits = 24,
+// };
 typedef Struct_(PolyTag) {
 	union {
 		U4 code;
@@ -390,8 +423,8 @@ typedef Struct_(PolyTag) {
  * No raw C-style casts. RHS values are assumed to be `U4` — caller passes a `U4` directly. */
 #define set_len(tag,v)  (C_(PolyTag_R,tag)->len  = u4_(v))
 #define set_addr(tag,v) (C_(PolyTag_R,tag)->addr = u4_(v))
-/* `set_code` is no longer in the new PolyTag design — the code byte lives
- *  in the primitive body (e.g. `((Poly_F3*)(p))->code`), not in the tag.
+/* `set_code` is no longer in the new PolyTag design — the code byte lives in the primitive body
+ * (e.g. `((Poly_F3*)(p))->code`), not in the tag.
  *  Use the typed primitive structs (Poly_F3, Poly_G4, etc.) and the `set_poly_*` setters,
  *  which set both the tag's length and the code. */
 #define get_len(tag)  C_(U4,C_(PolyTag_R,tag)->len)
@@ -399,95 +432,95 @@ typedef Struct_(PolyTag) {
 
 /* ---------- Poly_F3 (Flat Triangle; 5 words) ---------- */
 typedef Struct_(Poly_F3) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
-        A3_V2_S2 points;
-    };
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	union {
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; };
+		A3_V2_S2 points;
+	};
 };
 
 /* ---------- Poly_F4 (Flat Quad; 6 words) ---------- */
 typedef Struct_(Poly_F4) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    union {
-        struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
-        A4_V2_S2 points;
-    };
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	union {
+		struct { V2_S2 p0; V2_S2 p1; V2_S2 p2; V2_S2 p3; };
+		A4_V2_S2 points;
+	};
 };
 
 /* ---------- Poly_G3 (Gouraud Triangle; 7 words) ---------- */
 typedef Struct_(Poly_G3) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;
 };
 
 /* ---------- Poly_G4 (Gouraud Quad; 9 words) ---------- */
 typedef Struct_(Poly_G4) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p3;
 };
 
 /* ---------- Poly_FT3 (Flat Textured Triangle; placeholder layout) ---------- */
 /* TODO(Ed): verify the textured-variant layout against PSX-SPX when needed. */
 typedef Struct_(Poly_FT3) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    U4   tpage;
-    U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	U4   tpage;
+	U4   clut;
+	V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p2; U1 u2; U1 v2;
 };
 
 /* ---------- Poly_FT4 (Flat Textured Quad) ---------- */
 typedef Struct_(Poly_FT4) {
-    U4   tag;
-    RGB8 color;
-    B1   code;
-    U4   tpage;
-    U4   clut;
-    V2_S2 p0; U1 u0; U1 v0;
-    V2_S2 p1; U1 u1; U1 v1;
-    V2_S2 p2; U1 u2; U1 v2;
-    V2_S2 p3; U1 u3; U1 v3;
+	U4   tag;
+	RGB8 color;
+	B1   code;
+	U4   tpage;
+	U4   clut;
+	V2_S2 p0; U1 u0; U1 v0;
+	V2_S2 p1; U1 u1; U1 v1;
+	V2_S2 p2; U1 u2; U1 v2;
+	V2_S2 p3; U1 u3; U1 v3;
 };
 
 /* ---------- Poly_GT3 (Gouraud Textured Triangle) ---------- */
 typedef Struct_(Poly_GT3) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;
-    U4    tpage;
-    U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;
+	U4    tpage;
+	U4    clut;
+	V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp2; U1 u2; U1 v2;
 };
 
 /* ---------- Poly_GT4 (Gouraud Textured Quad) ---------- */
 typedef Struct_(Poly_GT4) {
-    U4    tag; RGB8 c0; B1 code;
-    V2_S2 p0;  RGB8 c1; B1 pad1;
-    V2_S2 p1;  RGB8 c2; B1 pad2;
-    V2_S2 p2;  RGB8 c3; B1 pad3;
-    V2_S2 p3;
-    U4    tpage;
-    U4    clut;
-    V2_S2 tp0; U1 u0; U1 v0;
-    V2_S2 tp1; U1 u1; U1 v1;
-    V2_S2 tp2; U1 u2; U1 v2;
-    V2_S2 tp3; U1 u3; U1 v3;
+	U4    tag; RGB8 c0; B1 code;
+	V2_S2 p0;  RGB8 c1; B1 pad1;
+	V2_S2 p1;  RGB8 c2; B1 pad2;
+	V2_S2 p2;  RGB8 c3; B1 pad3;
+	V2_S2 p3;
+	U4    tpage;
+	U4    clut;
+	V2_S2 tp0; U1 u0; U1 v0;
+	V2_S2 tp1; U1 u1; U1 v1;
+	V2_S2 tp2; U1 u2; U1 v2;
+	V2_S2 tp3; U1 u3; U1 v3;
 };
 
 /* ---------- Primitive setters (C-level) ----------
@@ -511,7 +544,6 @@ typedef Struct_(Poly_GT4) {
 /* ============================================================================
  *  Texture Page (TPage) bit layout
  *  ============================================================================
- *
  *  The TPage data word sent via GP0(0x2X) has:
  *    bits  0..3   = texture page X    (4 bits, 64-px units, 0..16)
  *    bit   4      = texture page Y    (1 bit,  64-px units, 0/1)
@@ -523,26 +555,29 @@ typedef Struct_(Poly_GT4) {
  *    bits 12..31  = reserved (zero)
  * ============================================================================ */
 enum {
-    /* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
-    gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4, gp0_tpage_x_mask            = 0xF,
-    gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1, gp0_tpage_y_mask            = 0x1,
-    gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2, gp0_tpage_semi_trans_mask   = 0x3,
-    gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2, gp0_tpage_color_depth_mask  = 0x3,
-    gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1, gp0_tpage_dither_mask       = 0x1,
-    gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1, gp0_tpage_draw_to_disp_mask = 0x1,
-    gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1, gp0_tpage_tex_disable_mask  = 0x1,
+	/* ---- Layer 1: TPage bitfield shifts / widths / masks ---- */
+	gp0_tpage_x_shift            =  0, gp0_tpage_x_width            =  4, gp0_tpage_x_mask            = 0xF,
+	gp0_tpage_y_shift            =  4, gp0_tpage_y_width            =  1, gp0_tpage_y_mask            = 0x1,
+	gp0_tpage_semi_trans_shift   =  5, gp0_tpage_semi_trans_width   =  2, gp0_tpage_semi_trans_mask   = 0x3,
+	gp0_tpage_color_depth_shift  =  7, gp0_tpage_color_depth_width  =  2, gp0_tpage_color_depth_mask  = 0x3,
+	gp0_tpage_dither_shift       =  9, gp0_tpage_dither_width       =  1, gp0_tpage_dither_mask       = 0x1,
+	gp0_tpage_draw_to_disp_shift = 10, gp0_tpage_draw_to_disp_width =  1, gp0_tpage_draw_to_disp_mask = 0x1,
+	gp0_tpage_tex_disable_shift  = 11, gp0_tpage_tex_disable_width  =  1, gp0_tpage_tex_disable_mask  = 0x1,
 
-    /* TPage color-depth payload values (NOT bit positions — these go in
-     * the 2-bit field at gp0_tpage_color_depth_shift). */
-    gp0_tpage_color_4bpp  = 0x0,
-    gp0_tpage_color_8bpp  = 0x1,
-    gp0_tpage_color_16bpp = 0x2,
+	/* TPage color-depth payload values (NOT bit positions — these go in
+	 * the 2-bit field at gp0_tpage_color_depth_shift). */
+	gp0_tpage_color_4bpp  = 0x0,
+	gp0_tpage_color_8bpp  = 0x1,
+	gp0_tpage_color_16bpp = 0x2,
 
-    /* TPage semi-transparency mode payload values (NOT bit positions). */
-    gp0_tpage_semi_trans_none  = 0x0,
-    gp0_tpage_semi_trans_alpha = 0x1,
-    gp0_tpage_semi_trans_add   = 0x2,
-    gp0_tpage_semi_trans_sub   = 0x3,
+	/* Default TPage value libpsyx's SetDefDrawEnv writes (matches the `li v1, 10; sh v1, 20(v0)` sequence at C11_only.elf:0x8001273C). */
+	gp0_tpage_default    = 10,
+
+	/* TPage semi-transparency mode payload values (NOT bit positions). */
+	gp0_tpage_semi_trans_none  = 0x0,
+	gp0_tpage_semi_trans_alpha = 0x1,
+	gp0_tpage_semi_trans_add   = 0x2,
+	gp0_tpage_semi_trans_sub   = 0x3,
 };
 
 /* ---- Layer 1.5: TPage per-field encoders. Mirrors enc_gte_sf/mx/v in gte.h. ---- */
@@ -556,26 +591,25 @@ enum {
 
 /* ---- 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) \
-    (enc_gp0_tpage_x(x)                       \
-   | enc_gp0_tpage_y(y)                       \
-   | enc_gp0_tpage_semi_trans(semi_trans)     \
-   | enc_gp0_tpage_color_depth(color_depth)   \
-   | enc_gp0_tpage_dither(dither)             \
-   | enc_gp0_tpage_draw_to_disp(draw_to_disp) \
-   | enc_gp0_tpage_tex_disable(tex_disable))
+	 (enc_gp0_tpage_x(x)                       \
+	| enc_gp0_tpage_y(y)                       \
+	| enc_gp0_tpage_semi_trans(semi_trans)     \
+	| enc_gp0_tpage_color_depth(color_depth)   \
+	| enc_gp0_tpage_dither(dither)             \
+	| enc_gp0_tpage_draw_to_disp(draw_to_disp) \
+	| enc_gp0_tpage_tex_disable(tex_disable))
 
 typedef Struct_(TexturePage) { U4 raw; };
 
 /* ---- Layer 3: TPage semantic word builder ---- */
 #define gp0_word_tpage(x, y, semi_trans, color_depth, dither, draw_to_disp, tex_disable) \
-    enc_gp0_tpage_word((x), (y), (semi_trans), (color_depth), (dither), (draw_to_disp), (tex_disable))
+	enc_gp0_tpage_word((x), (y), (semi_trans), (color_depth), (dither), (draw_to_disp), (tex_disable))
 #pragma endregion TPage
 
 #pragma region CLUT
 /* ============================================================================
  *  CLUT (Color Look-Up Table) semantics
  *  ============================================================================
- * 
  *  CLUT is loaded into VRAM by sending a GP0 command whose payload is:
  *    bits  0..5   = Y in 16-px units (palette row)
  *    bits  6..14  = X in 16-px units (palette column)
@@ -583,12 +617,12 @@ 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 ---- */
-    gp0_clut_y_shift = 0, gp0_clut_y_width = 6, gp0_clut_y_mask = 0x3F,
-    gp0_clut_x_shift = 6, gp0_clut_x_width = 9, gp0_clut_x_mask = 0x1FF,
-    /* CLUT-load cmd-byte variants — the upper byte of the GP0 word. */
-    gp0_clut_cmd_Load4bpp = 0x20,
-    gp0_clut_cmd_Load8bpp = 0x25,
+	/* ---- Layer 1: CLUT bitfield shifts / widths / masks ---- */
+	gp0_clut_y_shift = 0, gp0_clut_y_width = 6, gp0_clut_y_mask = 0x3F,
+	gp0_clut_x_shift = 6, gp0_clut_x_width = 9, gp0_clut_x_mask = 0x1FF,
+	/* 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 ---- */
@@ -608,7 +642,6 @@ enum {
 /* ============================================================================
  *  TIM file format constants and headers
  *  ============================================================================
- *
  *  TIM (Sony .TIM texture image) file structure:
  *    +0x00  U4 file_id (always 0x10 = TIM magic)
  *    +0x04  U4 version (always 0x00 for v1)
@@ -626,31 +659,30 @@ enum {
  *         +0x06 U2 px_height
  *         +0x08 ... pixel data
  *
- *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that
- *  emits the necessary GP0 commands. Stoppped for now at the
- *  struct + enum level.
+ *  Future?: add `tim_load_to_vram(tim_ptr, vram_addr)` that emits the necessary GP0 commands.
+ *  Stoppped for now at the struct + enum level.
  * ============================================================================ */
 enum {
-    tim_file_id_magic = 0x10,
-    tim_type_4bpp     = 0x00,
-    tim_type_8bpp     = 0x01,
-    tim_type_16bpp    = 0x02,
-    tim_type_32bpp    = 0x03,
-    tim_type_mixed    = 0x04,
-    tim_flag_has_clut = 0x08,
+	tim_file_id_magic = 0x10,
+	tim_type_4bpp     = 0x00,
+	tim_type_8bpp     = 0x01,
+	tim_type_16bpp    = 0x02,
+	tim_type_32bpp    = 0x03,
+	tim_type_mixed    = 0x04,
+	tim_flag_has_clut = 0x08,
 };
 
 typedef Struct_(TIM_Header) {
-    U4 file_id;       /* always 0x10 = "TIM" magic */
-    U4 version;       /* ignored; always 0 */
-    U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
+	U4 file_id;       /* always 0x10 = "TIM" magic */
+	U4 version;       /* ignored; always 0 */
+	U4 flags;         /* bits 0..2 = type, bit 3 = has_clut */
 };
 typedef Struct_(TIM_SectionHeader) {
-    U4 section_length;  /* bytes in this section including this header */
-    U2 org_x;           /* origin in VRAM */
-    U2 org_y;
-    U2 width;           /* width in pixels */
-    U2 height;          /* height in pixels */
+	U4 section_length;  /* bytes in this section including this header */
+	U2 org_x;           /* origin in VRAM */
+	U2 org_y;
+	U2 width;           /* width in pixels */
+	U2 height;          /* height in pixels */
 };
 #pragma endregion TIM File Format
 
@@ -659,35 +691,24 @@ typedef Struct_(TIM_SectionHeader) {
  *  Tape-side GPU operations (NOT in this header)
  *  ============================================================================
  *
- *  No `mac_gp0_send` or related macros live in gp.h. Rationale: the
- *  Lottes tape model uses OT-DMA for primitive submission, so atom bodies
- *  write to main RAM (the OT/primitive buffer) and to GTE state — never
- *  directly to the GPU ports at 0x1F801810 / 0x1F801814. See
- *  `mac_format_f3_color`, `mac_insert_ot_tag`, `mac_gte_store_f3` in
- *  lottes_tape.h for the patterns atom bodies actually use.
+ *  No `mac_gp0_send` or related macros live in gp.h.
+ *  Rationale: the Lottes tape model uses OT-DMA for primitive submission, so atom bodies write to main RAM (the OT/primitive buffer)
+ *  and to GTE state — never directly to the GPU ports at 0x1F801810 / 0x1F801814.
+ *  See `mac_format_f3_color`, `mac_insert_ot_tag`, `mac_gte_store_f3` in lottes_tape.h for the patterns atom bodies actually use.
  *
- *  If a feature need arises requires tape-side GPU port writes (e.g. DMA-kick to
- *  start GPU consumption of the OT, VBlank sync via GP1 status poll),
- *  the right home is `lottes_tape.h` alongside the rest of the `mac_*`
- *  family — the encoder infrastructure is already in place:
+ *  If a feature need arises requires tape-side GPU port writes 
+ *  (e.g. DMA-kick to start GPU consumption of the OT, VBlank sync via GP1 status poll),
+ *  the right home is `lottes_tape.h` alongside the rest of the `mac_*` family:
+ *    1. The caller pins a register to hold the IO base, e.g. register U4 r_io rgcc(R_T4) = IO_BASE_ADDR;
+ *       The compiler emits `lui R_T4, IO_BASE_ADDR_HI16` outside the atom body (in the C prologue before tape_run).
+ *    2. The atom body uses `store_word(R_data, R_T4, GPIO_PORT0_OFFSET)` to write to GP0, and `store_word(R_data, R_T4, GPIO_PORT1_OFFSET)`
+ *       to write to GP1. Both are preprocessor-encodable because R_T4 is a fixed register and the GPIO_PORT*_OFFSET constants
+ *       fit in the `sw`'s 16-bit signed offset field. No placeholder-pun, no asm constraints, no hidden register choice.
+ *       Same pattern as the old graphics_hello/hello_gp_routines.s `reg_io_offset`/`gcmd_push` convention.
  *
- *    1. The caller pins a register to hold the IO base, e.g.
- *         register U4 r_io rgcc(R_T4) = IO_BASE_ADDR;
- *       The compiler emits `lui R_T4, IO_BASE_ADDR_HI16` outside the
- *       atom body (in the C prologue before tape_run).
- *
- *    2. The atom body uses `store_word(R_data, R_T4, GPIO_PORT0_OFFSET)`
- *       to write to GP0, and `store_word(R_data, R_T4, GPIO_PORT1_OFFSET)`
- *       to write to GP1. Both are preprocessor-encodable because R_T4 is
- *       a fixed register and the GPIO_PORT*_OFFSET constants fit in the
- *       `sw`'s 16-bit signed offset field. No placeholder-pun, no asm
- *       constraints, no hidden register choice. Same pattern as the
- *       old graphics_hello/hello_gp_routines.s `reg_io_offset`/`gcmd_push`
- *       convention.
- *
- *  This mirrors the existing tape-side wave-context discipline: the
- *  caller binds the IO-base register via `rgcc()`, the macro assumes
- *  the binding is in effect, and the encoding falls out at preprocessor
- *  time. No additional GPU-domain macro layer required.
+ *  This mirrors the existing tape-side wave-context discipline: 
+ *  the caller binds the IO-base register via `rgcc()`, the macro assumes the binding is in effect,
+ *  and the encoding falls out at preprocessor time. 
+ *  No additional GPU-domain macro layer required.
  * ============================================================================ */
 #pragma endregion Tape-Side Macros

code/duffle/gp_vendor_sym.h

@@ -2,10 +2,8 @@
  *  duffle DSL — GPU Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the PSYQ-style CamelCase aliases for the canonical duffle GPU
- *  primitive setters and OT operations. The duffle snake_case names are
- *  primary; this header is for users who prefer the PSYQ SDK function
- *  names from the legacy C API.
+ *  Provides the PSYQ-style CamelCase aliases for the duffle GPU primitive setters and OT operations.
+ *  The duffle snake_case names are primary; this header is for users who prefer the PSYQ SDK function names from the legacy C API.
  *
  *  USAGE:   #include "duffle/gp_vendor_sym.h"   // after gp.h
  *
@@ -23,15 +21,11 @@
  *    OT operations:
  *      AddPrim(ot, p)  -> orderingtbl_add_primitive(ot, p)
  *
- *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and
- *  descriptive; no vendor alias is provided for them.
- *
- *  The vendor mnemonics are NOT registered with the duffle word-count
- *  metadata (word_counts.metadata.h). They expand to the duffle canonical
- *  macros which DO have word-count entries (the ones emitted by
- *  mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13
- *  (objdump byte-identical) holds.
+ *  The gp0_cmd_* / gp1_cmd_* byte constants are already short and descriptive; no vendor alias is provided for them.
  *
+ *  The vendor mnemonics are NOT registered with the duffle word-count metadata (word_counts.metadata.h).
+ *  They expand to the duffle macros which DO have word-count entries
+ *  (the ones emitted by mac_format_f3_color / mac_gte_store_f3 / etc.). Verification: V13 (objdump byte-identical) holds.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES

code/duffle/gte.h

@@ -17,9 +17,8 @@
  *            gte_lw_v0_xy(base) (gte + lw + v0 + xy)
  *            load_upper_i       (load-upper + immediate, unique verb)
  *
- *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are
- *  NOT in this header. They live in the opt-in `gte_vendor_sym.h` for
- *  users who prefer the textbook MIPS assembly mnemonics.
+ *  Vendor mnemonics (gte_mtc2, gte_mfc2, gte_lwc2, gte_swc2, etc.) are NOT in this header.
+ *  They are in the opt-in `gte_vendor_sym.h` for users who prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -34,20 +33,16 @@
  *  gte.h — Geometry Transformation Engine (COP2) for the PS1
  * ============================================================================
  *
- *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word`
- *  constants from C. No GCC inline-assembly string syntax in the code body.
+ *  Hand-rolled DSL for emitting GTE/MIPS instruction words as raw `.word` constants from C.
+ *  No GCC inline-assembly string syntax in the code body.
  *
  *  STYLE NOTES
  *  -----------
- *  - Per-field encoders are named `enc_gte_<field>(value)` and each one
- *    self-masks its argument before shifting. Mirrors the `enc_op / enc_rs
- *    / enc_rt / ...` family in mips.h.
- *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of
- *    the per-field encoders, plus the COP2/CO base.
- *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined
- *    for the common cases so call sites read like assembly source.
- *  - All register/field values are enums (not `#define`s) so they show up
- *    in debugger symbol tables and IDE autocomplete.
+ *  - Per-field encoders are named `enc_gte_<field>(value)` and each one self-masks its argument before shifting.
+ *    Mirrors the `enc_op / enc_rs / enc_rt / ...` family in mips.h.
+ *  - The composite `enc_gte_cmdw(sf, mx, v, cv, lm, cmd)` is a flat OR of the per-field encoders, plus the COP2/CO base.
+ *  - Pre-baked shortcuts (`gte_cmd_rtpt`, `gte_cmd_rtps`, …) are defined for the common cases so call sites read like assembly source.
+ *  - All register/field values are enums (not `#define`s) so they show up in debugger symbol tables and IDE autocomplete.
  *
  *  SEE ALSO
  *  --------
@@ -58,8 +53,7 @@
 
 /* --- GTE Data Registers (Coprocessor 2) ---
  * Preprocessor-visible integer ids for the COP2 data register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
  * Same pattern as the GPR `_Code` set in mips.h. */
 #define C2_VXY0_Code  0
 #define C2_VZ0_Code   1
@@ -192,10 +186,8 @@ enum {
 
 /* --- GTE Control Register Indices (for ctc2/cfc2) ---
  * Preprocessor-visible integer ids for the COP2 control register file.
- * Each enum value is bound to a parallel `_Code` `#define` so the
- * preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
- * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23
- * are reserved/unused on real hardware, so there's a gap. */
+ * Each enum value is bound to a parallel `_Code` `#define` so the preprocessor can stringify the integer (for `reg_str`/`rgcc` paths).
+ * Same pattern as the GPR `_Code` set in mips.h. Note: indices 21-23 are reserved/unused on real hardware, so there's a gap. */
 #define gte_cr_RT11_Code  0
 #define gte_cr_RT12_Code  1  /* packed with RT13 in bits 16..31 */
 #define gte_cr_RT13_Code  2  /* packed with RT22 in bits 16..31 */
@@ -223,8 +215,9 @@ enum {
 #define gte_cr_RFC_Code  27
 #define gte_cr_GFC_Code  28
 #define gte_cr_BFC_Code  29
-#define gte_cr_OFX_Code  30
-#define gte_cr_OFY_Code  31
+#define gte_cr_OFX_Code  24
+#define gte_cr_OFY_Code  25
+#define gte_cr_H_Code    26
 
 enum {
 	gte_cr_RT11 = gte_cr_RT11_Code, gte_cr_RT12 = gte_cr_RT12_Code, gte_cr_RT13 = gte_cr_RT13_Code,
@@ -246,21 +239,16 @@ enum { _C2_OPS_ = 0
 
 /* COP2 transfer sub-opcodes (5-bit field in the `rs` slot of enc_gte_tx).
  *
- * Spans the 2x2 {From, To} × {Data, Control} register classes that the
- * GTE exposes:
- *
+ * Spans the 2x2 {From, To} × {Data, Control} register classes that the GTE exposes:
  *   bit 1 (0x02): register class — 0 = data,  1 = control
  *   bit 2 (0x04): direction      — 0 = read,  1 = write
  *
- * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit
- * numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
- * (which target the data register file on any coprocessor). They are
- * re-aliased here so the four-way table reads like the spec mnemonics
- * (MFC2 / CFC2 / MTC2 / CTC2) and so the encoding lives next to its
- * only consumer (this header).
+ * The values 0x00 (sub_mfc2) and 0x04 (sub_mtc2) are the same 5-bit numbers as the general MIPS `cop_mf` / `cop_mt` defined in mips.h
+ * (which target the data register file on any coprocessor).
+ * They are re-aliased here so the four-way table reads like the spec mnemonics (MFC2 / CFC2 / MTC2 / CTC2)
+ * and so the encoding lives next to its only consumer (this header).
  *
- * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2)
- * live in gte_vendor_sym.h. */
+ * Vendor mnemonic aliases (gte_mfc2 / gte_mtc2 / gte_cfc2 / gte_ctc2) live in gte_vendor_sym.h. */
 enum { _C2_TX_SUBS_ = 0
 	, sub_mfc2 = 0x00 /* MFC2: Move From Coprocessor 2 data   reg */
 	, sub_cfc2 = 0x02 /* CFC2: Copy From Coprocessor 2 ctrl   reg */
@@ -270,11 +258,11 @@ enum { _C2_TX_SUBS_ = 0
 
 /* COP2 (GTE) Transfer Format: mfc2 / cfc2 / mtc2 / ctc2 rt, rd
  * Layout: [op_cop2:6][sub:5][rt:5][rd:5][0:11]
- *   - sub: one of sub_mfc2 / sub_cfc2 / sub_mtc2 / sub_ctc2
- *   - rt:  GPR source/dest
- *   - rd:  COP2 register index (0..31):
- *            data class → C2_VXY0_Code..C2_LZCR_Code (gte_in_v0_xy..gte_math_accum2 aliases)
- *            ctrl class → gte_cr_RT11_Code..gte_cr_OFY_Code */
+ *  - sub: one of sub_mfc2 / sub_cfc2 / sub_mtc2 / sub_ctc2
+ *  - rt:  GPR source/dest
+ *  - rd:  COP2 register index (0..31):
+ *    data class → C2_VXY0_Code..C2_LZCR_Code (gte_in_v0_xy..gte_math_accum2 aliases)
+ *    ctrl class → gte_cr_RT11_Code..gte_cr_OFY_Code */
 #define enc_gte_tx(sub, rt, rd) (enc_op(op_cop2) | enc_rs(sub) | enc_rt(rt) | enc_rd(rd))
 
 
@@ -314,8 +302,8 @@ enum { _C2_TX_SUBS_ = 0
  * `swc2` is redundant when we're already inside the `gte_` namespace.
  *   gte_lw rt, base, off  →  lwc2 rt, off(base)
  *   gte_sw rt, base, off  →  swc2 rt, off(base)
- * For the typical user-facing vector-level load (xy + z as two
- * instructions), use the higher-level `gte_load_vN` macros below. */
+ * For the typical user-facing vector-level load (xy + z as two instructions),
+ * use the higher-level `gte_load_vN` macros below. */
 #define gte_lw(rt, base, off) enc_gte_lw(rt, base, off)
 #define gte_sw(rt, base, off) enc_gte_sw(rt, base, off)
 
@@ -323,13 +311,12 @@ enum { _C2_TX_SUBS_ = 0
  * Opcode is always MIPS_OP_COP2, RS is always 1 (CO).
  * The lower 25 bits are the GTE-specific command payload.
  *
- * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs`
- * pattern in mips.h: each one self-masks and shifts its own field, so a
- * caller can build up a GTE command piece by piece (handy for state-driven
- * MVMVA emitters that vary one field at a time).
+ * The granular `enc_gte_<field>(x)` macros below mirror the `enc_op`/`enc_rs` pattern in mips.h:
+ * Each one self-masks and shifts its own field, so a caller can build up a GTE command piece by piece
+ * (handy for state-driven MVMVA emitters that vary one field at a time).
  *
- * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command
- * word in one go. It just ORs the per-field encoders together. */
+ * `ENC_GTE_CMD` is the all-in-one convenience for emitting a full command word in one go.
+ * It just ORs the per-field encoders together. */
 #define gte_cmd_base (enc_op(op_cop2) | (1 << 25))
 
 /* Per-field encoders. Each one does (value & mask) << shift on its own. */
@@ -353,41 +340,35 @@ enum { _C2_TX_SUBS_ = 0
 
 /* GTE command words for the common cases.
  *
- * These are pure compile-time integer constants — the C compiler
- * constant-folds them into `.word` directives in .rodata. Use them
- * inside `asm_inline(...)` blocks (see `gte_rtpt` below for the
- * canonical idiom).
+ * These are pure compile-time integer constants — the C compiler constant-folds them into `.word` directives in .rodata.
+ * Use them inside `asm_inline(...)` blocks (see `gte_rtpt` below for the idiom).
  *
  * Decomposition (per the `enc_gte_<field>` definitions above):
  *   gte_cmdw_<name> = gte_cmd_base | enc_gte_cmd(<cmd>)
  
- * The SF/MX/V/CV/LM fields are all zero in the common cases (standard
- * rotation-matrix, no scaling factor, V0 vector, translation vector,
- * no clamp), so the only varying bits are the `cmd` field.
+ * The SF / MX / V / CV / LM fields are all zero in the common cases 
+ * (standard rotation-matrix, no scaling factor, V0 vector, translation vector, no clamp),
+ * so the only varying bits are the `cmd` field.
  *
- * Naming follows the file's convention: `gte_cmd_*` is the raw
- * 6-bit `cmd` field id, `gte_cmdw_*` is the fully-encoded 32-bit
- * instruction word ready to drop into a `.word` directive.
+ * Naming convention:
+ * - `gte_cmd_*` : Raw 6-bit `cmd` field id 
+ * - `gte_cmdw_* : 32-bit instruction word ready to drop into a `.word` directive.
  *
  * --------------------------------------------------------------------------
  *  PsyQ-compatibility note (RTPS/RTPT):
- *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and
- *  RTPS as `cop2 0x0180001`. Both have `0x20` set in the upper-reserved
- *  region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
- *  Per psx-spec these bits are reserved/must-be-zero, but the real GTE
- *  hardware and PCSX-Redux's GTE model both IGNORE them on these two
- *  commands (the perspective divide happens regardless of `sf`).
+ *  The original Sony PsyQ `inline_n.h` ships RTPT as `cop2 0x0280030` and RTPS as `cop2 0x0180001`.
+ *  Both have `0x20` set in the upper-reserved region (bit 21) AND `sf=1` (bit 19) — i.e. the "no division" flag.
+ *  Per psx-spec these bits are reserved/must-be-zero,
+ *  but the real GTE hardware and PCSX-Redux's GTE model both IGNORE them on these two commands
+ *  (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 — the floor's screen
- *  coordinates come out as raw projection-of-rotation (Z never
- *  divided), `nclip` ends up wrong, and the triangle is culled.
+ *  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 — 
+ *  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.
- *  NCLIP/OP/MVMVA stay spec-clean — their reserved bits really are
- *  zero in the original PsyQ source.
+ *  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.
+ *  NCLIP / OP / MVMVA stay spec-clean — their reserved bits really are zero in the original PsyQ source.
  * --------------------------------------------------------------------------
  */
 #define gte_cmdw_psyq_compat  (1u << 21 | enc_gte_sf(gte_sf_integer))
@@ -420,20 +401,16 @@ enum { _C2_TX_SUBS_ = 0
 
 /**
  * @brief Loads a single SVECTOR to GTE vector register V0
- *
  * @details Loads values from an SVECTOR struct to GTE data registers C2_VXY0
  * (XY at offset 0) and C2_VZ0 (Z at offset 4) using `lwc2`.
  *
- * Uses string-style GCC inline asm with `%0` substitution because the
- * base register `r0` is a runtime GPR chosen by the compiler — it cannot
- * be encoded into a static `.word` constant.
+ * Uses string-style GCC inline asm with `%0` substitution because the base register `r0` is a runtime GPR chosen by the compiler.
+ * It cannot be encoded into a static `.word` constant.
  *
- * Usage:
- *   asm_gte_load_v0(svector_ptr);
+ * Usage: asm_gte_load_v0(svector_ptr);
  */
 
 /* lwc2 encoding helpers parameterized on the base GPR.
- *
  * gte_lw_v0_xy(base)  → lwc2 $0,  0(base)   ; C2_VXY0
  * gte_lw_v0_z(base)   → lwc2 $1,  4(base)   ; C2_VZ0
  * gte_lw_v1_xy(base)  → lwc2 $2,  0(base)   ; C2_VXY1
@@ -442,8 +419,7 @@ enum { _C2_TX_SUBS_ = 0
  * gte_lw_v2_z(base)   → lwc2 $5,  4(base)   ; C2_VZ2
  *
  * `base` is the GPR number to bake into the .word constant's `rs` field.
- * These are pure compile-time integers; the C compiler constant-folds
- * them into .word directives. */
+ * These are pure compile-time integers; the C compiler constant-folds them into .word directives. */
 
 enum {
 	GTE_Z_Offset = 4
@@ -458,26 +434,21 @@ enum {
 
 /* 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
+ * 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).
  *
  * The caller MUST bind `r_ptr` to that same GPR via a register variable:
  *     register V3_S2* p_in_12 __asm__("$12") = my_ptr;
  *     gte_load_v0(p_in_12, R_T4);   // R_T4 = 12, base is $12
  *
- * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register
- * `p_in_12` can live in), which is exactly the register the .word
- * constants expect. A `"$12"` clobber would conflict with the
- * register-variable binding ("asm specifier for variable conflicts
- * with asm clobber list"), so we omit it. The other ABI-clobbers
- * ($2/$8/$9/$31) stay because the GTE instructions don't touch
- * caller-saved GPRs but the kernel does treat them as volatile.
+ * Then `"r"(r_ptr)` inside the asm binds to $12 (the only register `p_in_12` can live in),
+ * which is exactly the register the .word constants expect.
+ * A `"$12"` clobber would conflict with the register-variable binding ("asm specifier for variable conflicts with asm clobber list"), so we omit it.
+ * The other ABI-clobbers ($2/$8/$9/$31) stay because the GTE instructions don't touch caller-saved GPRs but the kernel does treat them as volatile.
  *
  * WHICH REGISTER TO PICK
  * ----------------------
- * Any caller-saved GPR is safe. Recommended default for an RTPT-style
- * 3-pointer pipeline:
+ * Any caller-saved GPR is safe. Recommended default for an RTPT-style 3-pointer pipeline:
  *   gte_load_v0(p0, R_T4);  // $12
  *   gte_load_v1(p1, R_T5);  // $13
  *   gte_load_v2(p2, R_T6);  // $14
@@ -490,8 +461,7 @@ enum {
  *   clobbers section   : "$2", "$8", ..., "memory"      (from asm_clobber)
  *   3 colons total, GCC-legal. No string-syntax mnemonics in the .word body.
  *
- * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that
- * starts the clobbers section. */
+ * The `asm_clobber(...)` helper from gcc_asm.h prepends the colon that starts the clobbers section. */
 #define gte_load_v0(r_ptr, base) asm volatile(       \
 	asm_words( gte_lw_v0_xy(base), gte_lw_v0_z(base) ) \
 	asm_rpins,   r_use(r_ptr)                          \
@@ -510,12 +480,10 @@ enum {
 	asm_clobber: rlit(R_V0), rlit(R_T0), rlit(R_T1), rlit(R_RA), clb_mem_drain \
 )
 
-/* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — the canonical prelude to gte_cmd_rtpt.
- *
- * Loads all three GTE input vectors (6 words) from three separate pointers,
- * one per GTE vector register, each loaded from its own base GPR. Caller
- * must bind each `pN` to `bN` via a register variable.
+/* gte_load_v0v1v2(p0, p1, p2, b0, b1, b2) — prelude to gte_cmd_rtpt.
  *
+ * Loads all three GTE input vectors (6 words) from three separate pointers, one per GTE vector register,
+ * each loaded from its own base GPR. Caller must bind each `pN` to `bN` via a register variable.
  *   register V3_S2* p0 rgcc(R_T4) = verts[0].ptr;  // → __asm__("$12")
  *   register V3_S2* p1 rgcc(R_T5) = verts[1].ptr;  // → __asm__("$13")
  *   register V3_S2* p2 rgcc(R_T6) = verts[2].ptr;  // → __asm__("$14")
@@ -534,29 +502,20 @@ enum {
 
 /**
  * @brief Rotate, Translate and Perspective Triple (23 cycles)
- *
- * @details Performs rotation, translation and perspective calculation of three
- * vertices at once. The equation performed is the same as gte_rtps() only
- * repeated three times for each vertex. The result of the first vertex is
- * stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then
- * C2_SXY2.
+ * @details Performs rotation, translation and perspective calculation of three vertices at once.
+ * The equation performed is the same as gte_rtps() only repeated three times for each vertex.
+ * The result of the first vertex is stored in GTE data register C2_SXY0, the second vector in C2_SXY1 then C2_SXY2.
  *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency — the GTE
- *      takes ~8 cycles per perspective divide, and the nops let any
- *      preceding lwc2/swc2 retire before RTPT starts reading its
- *      inputs from V0/V1/V2.
- *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the
- *      pre-baked encoders above) — `0x0280030` decoded as
- *      `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields
- *      zero (standard rotation, no scaling, V0 vector, translation
- *      vector, no clamp).
+ *   1. Two `nop` words fill the COP2 pipeline latency — the GTE takes ~8 cycles per perspective divide, 
+ *      and the nops let any preceding lwc2/swc2 retire before RTPT starts reading its inputs from V0/V1/V2.
+ *   2. The RTPT command word itself is `gte_cmdw_rtpt` (see the pre-baked encoders above) — 
+ *      `0x0280030` decoded as `op_cop2` | CO(1) | cmd=RTPT, with all SF/MX/V/CV/LM fields zero
+ *      (standard rotation, no scaling, V0 vector, translation vector, no clamp).
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register — those have to be saved by the caller if
- * they need to survive across the call (RTPT writes SXY0..2, SZ0..3,
- * OTZ, MAC0..3, IR0..3, etc.).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI)
+ * plus the standard "memory" barrier. Does not clobber any COP2 data/control register — 
+ * those have to be saved by the caller if they need to survive across the call (RTPT writes SXY0..2, SZ0..3, OTZ, MAC0..3, IR0..3, etc.).
  */
 #define gte_rtpt()                        \
 	asm volatile(                           \
@@ -572,32 +531,24 @@ enum {
 
 /**
  * @brief Normal clipping (8 cycles)
- *
- * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for
- * backface culling. If the value of C2_MAC0 is negative, the coordinates are
- * inverted and thus the triangle is back facing.
+ * @details Computes the sign of three screen coordinates (C2_SXY0-2) used for backface culling.
+ * If the value of C2_MAC0 is negative, the coordinates are inverted and thus the triangle is back facing.
  *
  * The following equation is performed when executing this GTE command:
- *
  *     MAC0 = SX0*SY1 + SX1*SY2 + SX2*SY0 - SX0*SY2 - SX1*SY0 - SX2*SY1
- *
  * Encoder-style emission (no inline-asm strings in the code body):
- *   1. Two `nop` words fill the COP2 pipeline latency - the GTE
- *      pipeline takes a few cycles per op, and the nops let any
- *      preceding lwc2/swc2/RTPT retire before NCLIP starts reading
- *      its inputs from SXY0/SXY1/SXY2.
- *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the
- *      pre-baked encoders above) - `0x01400006` decoded as
- *      `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields
- *      zero. NCLIP is spec-clean in the original PsyQ source
- *      (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat`
- *      quirk), so `gte_cmdw_nclip` does NOT OR in any reserved bits.
+ *   1. Two `nop` words fill the COP2 pipeline latency
+ *      - the GTE pipeline takes a few cycles per op, and the nops let any preceding
+ *      lwc2/swc2/RTPT retire before NCLIP starts reading its inputs from SXY0/SXY1/SXY2.
+ *   2. The NCLIP command word itself is `gte_cmdw_nclip` (see the pre-baked encoders above)
+ *      - `0x01400006` decoded as  `op_cop2` | CO(1) | cmd=NCLIP, with all SF/MX/V/CV/LM fields zero.
+ *      NCLIP is spec-clean in the original PsyQ source (unlike RTPS/RTPT which carry the `gte_cmdw_psyq_compat` quirk),
+ *      so `gte_cmdw_nclip` does NOT OR in any reserved bits.
  *
- * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel
- * ABI) plus the standard "memory" barrier. Does not clobber any COP2
- * data/control register - those have to be saved by the caller if
- * they need to survive across the call (NCLIP writes MAC0 only; it
- * is purely a sign-of-double-product computation on SXY0..2).
+ * Clobbers the caller-saved GPRs via `clbr_volatile_gprs` (per the kernel ABI) plus the standard "memory" barrier.
+ * Does not clobber any COP2 data/control register.
+ * Those have to be saved by the caller if they need to survive across the call (NCLIP writes MAC0 only;
+ * it is purely a sign-of-double-product computation on SXY0..2).
  */
 #define gte_nclip()                       \
 	asm volatile(                           \
@@ -623,13 +574,10 @@ enum {
 		"cop2 0x0158002D;")
 
 /* asm_gte_matrix_set_rotation(r0)
+ * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix control registers (RT11..RT22, indices 0..4) via ctc2.
  *
- * Loads the 3x3 rotation matrix at `r0` into the GTE's rotation-matrix
- * control registers (RT11..RT22, indices 0..4) via ctc2.
- *
- * Memory layout at r0: five contiguous 32-bit words (offsets 0..16),
- * each holding two packed 16-bit matrix elements. The first 1.5 rows
- * of a standard PSX SDK MATRIX struct (where each row is laid out as
+ * Memory layout at r0: five contiguous 32-bit words (offsets 0..16), each holding two packed 16-bit matrix elements.
+ * The first 1.5 rows of a standard PSX SDK MATRIX struct (where each row is laid out as
  * [RT_xx, RT_xy] | [RT_xz, pad] | ...).
  *
  * Generated MIPS (mirrors the source macro):
@@ -644,27 +592,22 @@ enum {
  *   ctc2 $13,  $3         ; → C2_RT21
  *   ctc2 $14,  $4         ; → C2_RT22
  *
- * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a
- * register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
- * instructions to read from the right base. The `"r"(r0)` constraint
- * alone doesn't force a specific GPR — it just lets GCC pick one.
- * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field
- * of each lw, so the lw instructions will only do the right thing
- * if $12/$13/$14 hold the matrix base at runtime.
+ * Same contract as gte_load_v0: caller MUST bind `r0` to $12 via a register variable (`rgcc(R_T4)`) for the `lw $12, off(...)`
+ * instructions to read from the right base. The `"r"(r0)` constraint alone doesn't force a specific GPR — it just lets GCC pick one.
+ * The .word constants here bake R_T4/R_T5/R_T6 into the `rs` field of each lw, so the lw instructions will 
+ * only do the right thing if $12 / $13 / $14 hold the matrix base at runtime.
  *
  *   M3_S2* m = ...;
  *   register M3_S2* m_in_12 rgcc(R_T4) = m;
  *   asm_gte_matrix_set_rotation(m_in_12);
  *
- * We clobber $12/$13/$14 (the ones we use as scratch inside the
- * inline asm) plus the system clobbers; we don't clobber `r0` because
- * the `rgcc` binding already says "this variable lives in $12".
+ * We clobber $12/$13/$14 (the ones we use as scratch inside the inline asm)
+ * plus the system clobbers; we don't clobber `r0` because the `rgcc` binding already says "this variable lives in $12".
  *
- * WARNING: Incomplete by design. The source macro only writes RT11..RT22
- * (5 of 9 rotation elements); RT23 and the entire RT3x row are left
- * untouched. Real libpsn00b SetRotMatrix writes all 9. Use only when the
- * GTE's remaining rotation entries are already correct, or you will
- * get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
+ * WARNING: Incomplete by design. The source macro only writes RT11..RT22 (5 of 9 rotation elements);
+ * RT23 and the entire RT3x row are left untouched.
+ * Real libpsn00b SetRotMatrix writes all 9. Use only when the GTE's remaining rotation entries are already correct,
+ * or you will get stale-RT2x/RT3x artifacts in RTPS/RTPT/MVMVA output.
  */
 #define asm_gte_matrix_set_rotation(r0) \
 	asm volatile(                         \

code/duffle/gte_vendor_sym.h

@@ -2,10 +2,8 @@
  *  duffle DSL — GTE Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2
- *  instructions as thin aliases to the canonical duffle macros in gte.h.
- *  The duffle names are primary; this header is for users who prefer
- *  the textbook mnemonics.
+ *  Provides the textbook MIPS assembly mnemonics for the GTE/COP2 instructions as thin aliases to the duffle macros in gte.h.
+ *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/gte_vendor_sym.h"   // after gte.h
  *

code/duffle/lottes_tape.h

@@ -10,78 +10,91 @@
 #	include "gen/duffle.offsets.h"
 #endif
 
-typedef U4 const MipsCode;
+typedef U4 const MipsCode; // Underlying type to mips asm words.
 typedef Slice_(MipsCode);
-typedef Slice_MipsCode MipsAtom;
 
-#define MipsAtom_(sym) MipsCode tmpl(code,sym) [] align_(4) =
+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) =
 
-// Bare form: file-scope declaration with hardcoded body.
 // 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) =
 
-// Function form: function-body block that returns a MipsAtom slice.
 // Used for components with value-args (e.g., ac_format_f3_color).
-//   FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
+//   FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })
 // expands to:
-//   FI_ MipsAtom ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
+//   FI_ Slice_MipsCode ac_X(args) { MipsCode ac_X[] align_(4) = { body }; return slice_from_array(MipsCode, ac_X); }
 #define MipsAtomComp_Proc_(sym, ...) { MipsCode sym [] align_(4) = __VA_ARGS__; return slice_from_array(MipsCode, sym); }
 
 // Auto-generated component macros (<module>/gen/<dir>/<dir>.macs.h) are included manually by the unity build.
 
 /* Register aliases */
 enum {
-	R_AtomJmp    = R_T9,
-	R_TapePtr    = R_T8,  /* The Instruction Stream Pointer */
-	R_InCursor   = R_T4,  /* Input data cursor */
+	R_AtomJmp  = R_T8 atom_reg,  /* debug-visible; tape yield handshake scratch */
+	R_TapePtr  = R_T9 atom_reg,  /* The Instruction Stream Pointer */
+/* Stringification codes for the GCC inline assembler clobber lists. */
+#define R_AtomJmp_Code R_T8_Code
+#define R_TapePtr_Code R_T9_Code
 
-	R_PrimCursor = R_T7,  /* VRAM output cursor (primitive buffer) */
-	R_FaceCursor = R_T4,  /* Input data cursor (indices/faces) */
-	R_VertBase   = R_T5,  /* Base address of the vertex array */
-	R_OtBase     = R_T6,  /* Base address of the Ordering Table */
+// R_InCursor = R_T4,
+// #define R_InCursor_Code    R_T4_Code
 
-/* Stringification codes for the GCC inline assembler clobber lists */
-#define R_TapePtr_Code     R_T8_Code
-#define R_InCursor_Code    R_T4_Code
+// Reserved Registers (Callee-saved):
+// - R_T9: Holds the Tape Ptr which we need to increment
+//         If we hit a wall with register allocations we can clobber V0 & V1 (return values), defering as opt-in by user.
+// - R_RA: Not sure??
+// Needed by ac_yield but can be used as atom scratch:
+// - R_T8: Will be used as the atom jump register.
 
-#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
+// All allocatable registers for mips atoms:
+	R_TScratchVolatile = R_AT, // This one is reserved for psuedo instructions, but you can technically use it.
+	R_TScratch0  = R_T0,
+	R_TScratch1  = R_T1,
+	R_TScratch2  = R_T2,
+	R_TScratch3  = R_T3,
+	R_TScratch4  = R_T4,
+	R_TScratch5  = R_T5,
+	R_TScratch6  = R_T6,
+	R_TScratch7  = R_T7,
+	R_TScratch8 =  R_T8,
+	R_TScratch10 = R_V0,
+	R_TScratch11 = R_V1,
+	// Note(Ed): We can technically clobber these, but don't unless we hit a bottleneck.
+	// R_TScratch12 = R_A0,
+	// R_TScratch13 = R_A1,
+	// 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
 /* ---------------------------------------------------------------------------
  *  TAPE DRIVE ABI & REGISTER ALIASES (the enum moved earlier; see below)
  * ---------------------------------------------------------------------------*/
+typedef Slice_(MipsAtom); typedef Slice_MipsAtom Tape;
 
 /* The 'Exit' Atom */
-MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
+atom_dbg_skip MipsAtom_(tape_exit) { jump_reg(rret_addr), nop };
 
+//TODO(Ed): Do we backup R_S0-7 here? Have it in a heavier tape run as a opt-in? Same with V0-1 and A0-3?
 /* Generalized Tape Engine Runner */
-FI_ void tape_run(Slice_U4 tape) { register U4* tp rgcc(R_TapePtr) = tape.ptr; asm volatile(
+FI_ void tape_run(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u4_r(tape.ptr); asm volatile(
 	asm_words(
-			add_ui(     R_SP, R_SP, -MipsStackAlignment) /* Allocate stack space */
-		, store_word( R_RA, R_SP,           0)         /* Safely backup $ra to the stack */
-		, load_word(  R_AtomJmp, R_TapePtr, 0)         /* Bootstrap the first jump */
-		, add_ui_self(R_TapePtr, S_(MipsCode))         /* Advance tape */
-		, call_reg(   R_AtomJmp)                       /* jalr $t9 */
-		, nop                                          /* Branch delay slot */
-		, load_word(  R_RA, R_SP, 0)                   /* Restore $ra from stack */
-		, add_ui_self(R_SP, MipsStackAlignment)        /* Deallocate stack space */
+		  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(tp)
+	asm_rpins, r_use(tape_ptr)
 	asm_clobber: 
-			rlit(R_AT)
-		, rlit(R_V0), rlit(R_V1)
-		, rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3)
-		/* Tell GCC the tape engine owns and destroys the workspace registers */
-		, rlit(R_PrimCursor), rlit(R_FaceCursor), rlit(R_VertBase), rlit(R_OtBase)
-		, rlit(R_T9)
-		, clb_mem_drain 
+		rlit(R_AT),
+		rlit(R_V0), rlit(R_V1),
+		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 
 ); }
 
 typedef Relative_(FArena) Struct_(TapeBuilder) { U4 ptr; U4 capacity; U4 used; };
@@ -89,13 +102,17 @@ FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start
 FI_ TapeBuilder tb_make_old(             FArena* arena) { return (TapeBuilder){ arena->start, 0 }; }
 FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }; }
 
-#define tb_emit_(tb, atom) tb_emit(tb, tmpl(code,atom))
 FI_ void tb_emit(TapeBuilder* tb, MipsCode* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
+#define tb_emit_(atom)        tb_emit(& tb, atom)
+#define tb_data_(field, data) tb_data(& tb, u4_(data))
 
-FI_ Slice_U4 tb_end  (TapeBuilder* tb) { tb_emit(tb,code_tape_exit); return (Slice_U4){ C_(U4*,tb->ptr), tb->used }; }
-FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Slice_U4){ C_(U4*,tb.ptr),  tb.used }; }
-#define tb_scope(tb) for(U4 tbs_once=0;tbs_once==0;++tbs_once,tb_emit(tb,code_tape_exit))
+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
 
@@ -106,22 +123,32 @@ FI_ Slice_U4 tb_slice(TapeBuilder  tb) {                             return (Sli
  * ---------------------------------------------------------------------------*/
 
 // The 'Yield' sequence for Tape Atoms (mac_yield). 
-MipsAtomComp_(ac_yield) {
+atom_dbg_skip MipsAtomComp_(ac_yield) {
 	load_word(R_AtomJmp, R_TapePtr, 0),
 	add_ui_self(         R_TapePtr, S_(MipsCode)),
-	jump_reg( R_AtomJmp),
-	nop,
+	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 */
-MipsAtomComp_(ac_load_tri_indices) {
+atom_dbg_skip MipsAtomComp_(ac_load_tri_indices) {
 	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
 	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
 	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
 };
 
 /* Words: 18; Translates indices to vertex addresses and pushes them to GTE  */
-MipsAtomComp_(ac_gte_load_tri_verts) {
+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),
@@ -159,22 +186,21 @@ MipsAtomComp_(ac_insert_ot_tag_g4) {
 
 /* Words: 3; Emits one (cmd|color) word to R_PrimCursor at the given
  * byte offset. Internal helper used by the *_format_*_color macros. */
-FI_ MipsAtom ac_pack_color_word(U4 off, U4 cmd, U1 r, U1 g, U1 b)
-MipsAtomComp_Proc_(ac_pack_color_word, {
+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)),
+	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).
- * Migrated from hello_gte_tape.c; takes RGB form per the Phase 3 convention. */
-FI_ MipsAtom ac_format_f3_color(U1 r, U1 g, U1 b)
-MipsAtomComp_Proc_(ac_format_f3_color, { mac_pack_color_word(O_(Poly_F3,color), gp0_cmd_poly_f3, r, g, b) })
+ * 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). */
-MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
+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)),
@@ -182,7 +208,7 @@ MipsAtomComp_(ac_gte_store_f3_post_rtpt) {
 
 /* Words: 12; Emits the four (code|color) words of a Poly_G4.
  * Args: rN,gN,bN are 8-bit RGB byte values for each of the 4 vertices. */
-FI_ MipsAtom ac_format_g4_color(
+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,
@@ -197,24 +223,19 @@ MipsAtomComp_Proc_(ac_format_g4_color, {
 /* Words: 3; Stores the 3 transformed (V2_S2 screen) vertices of the
  * G4 triangle portion to p0/p1/p2.
  * PIPELINE: post-RTPT, pre-RTPS (SXY0=v0.screen, SXY1=v1.screen, SXY2=v2.screen). 
- * MUST be called BEFORE V3-RTPS, otherwise SXY0/1/2
- * get overwritten with v3 (RTPS writes only to SXY2, but to keep the
- * three registers aligned with v0/v1/v2 you must store before RTPS).
- * The macro name declares the pipeline position; check #6 (GTE state-
- * machine validation) verifies the call site matches the declaration. */
-MipsAtomComp_(ac_gte_store_g4_p012_post_rtpt_pre_rtps) {
+ * 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 — DO NOT read SXY0 here, that's the bug this name
- * prevents).
+ * 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.
  */
-MipsAtomComp_(ac_gte_store_g4_p3_post_rtps) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
+atom_dbg_skip MipsAtomComp_(ac_gte_store_g4_p3) { gte_sw(C2_SXY2, R_PrimCursor, O_(Poly_G4,p3)) };
 
 #pragma endregion Macro Atom Components
 
@@ -237,12 +258,11 @@ FI_ void atombuilder_unroll(MipsAtomBuilder_R ab, Slice_MipsCode_R code) {
 
 // When done authoring, utilize this to cap-off the atom
 FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
-	LP_ MipsAtom_(yield) { mac_yield() };
-	mem_copy(ab->start, u4_(code_yield), S_(code_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(code_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) (MipsAtom){ab.start, ab.used}
+#define mipsatom_from_builder(ab) (Slice_MipsCode){ab.start, ab.used}
 
 #pragma endregion Mips Atom Builder
 
@@ -269,8 +289,7 @@ internal MipsAtom_(mips_flush_icache) {
 	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),                        // jalr  $t0, $ra 
-	nop,                                                 // BD slot 
+	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) 
@@ -297,68 +316,4 @@ internal MipsAtom_(set_gte_world) atom_info(
 	mac_yield()
 };
 
-/* DIAGNOSTIC 1: Pure tape loop test */
-internal MipsAtom_(diag_yield) { mac_yield() };
-
-// TODO(Ed): Reduce magic numbers/offsets
-/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
-internal MipsAtom_(diag_color) {
-	store_word(  R_0, R_T7, 0), 
-	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
-	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
-	store_word(  R_AT, R_T7, 4),
-	
-	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
-	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
-	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
-
-	add_ui(          R_T1, R_0,  10),
-	shift_lleft_self(R_T1,       S_(U4)/2), 
-	add_u_self(      R_T1,       R_T6),         
-	
-	load_word(   R_AT, R_T1, 0),        
-	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
-	store_word(  R_AT, R_T7, 0),       
-	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
-	or_u_self(  R_AT, R_V0),          
-	store_word( R_AT, R_T1, 0),       
-
-	add_ui(R_T7, R_T7, 20),          
-
-	mac_yield()
-};
-
-// TODO(Ed): Reduce magic numbers/offsets
-/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
-internal MipsAtom_(diag_gte) {
-	/* Load 3 indices */
-	load_half_u(R_T0, R_T4, 0),
-	load_half_u(R_T1, R_T4, 2),
-	load_half_u(R_T2, R_T4, 4),
-
-	/* Load Vertices into GTE */
-	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
-
-	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
-
-	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
-	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
-	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
-
-	/* Run Math */
-	nop2, gte_cmdw_rtpt,
-	nop2, gte_cmdw_nclip,
-	nop2,
-
-	/* Advance Face Cursor and Yield */
-	add_ui(R_T4, R_T4, 8),
-
-	mac_yield()
-};
-
 #pragma endregion Baked Mips Atoms

code/duffle/math.h

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

code/duffle/memory.h

@@ -67,8 +67,8 @@ typedef Slice_(B1);
 #define slice_end(slice)   ((slice).ptr + (slice).len)
 #define S_slice(s)         ((s).len * S_((s).ptr[0]))
 
-#define slice_ut(ptr,len)  slice_ut_(u4_(ptr), u4_(len))
-#define slice_ut_arr(a)    slice_ut_(u4_(a),   S_(a))
+#define slice_ut(ptr,len)  slice_ut_(u4_(ptr),     u4_(len))
+#define slice_ut_arr(a)    slice_ut_(u4_(a),       S_(a))
 #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)
@@ -103,7 +103,7 @@ FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
 	arena->used     = 0;
 }
 FI_ FArena farena_make(Slice mem) { FArena a; farena_init(& a, mem); return a; }
-I_ Slice   farena_push(FArena_R arena, U4 amount, Opt_farena o) {
+I_  Slice  farena_push(FArena_R arena, U4 amount, Opt_farena o) {
 	if (amount == 0) { return (Slice){}; }
 	U4 desired   = amount * (o.type_width == 0 ? 1 : o.type_width);
 	U4 to_commit = align_pow2(desired, o.alignment ?  o.alignment : MEM_ALIGNMENT_DEFAULT);

code/duffle/mips.h

@@ -1,38 +1,28 @@
 /* ============================================================================
  *  duffle DSL Suffix Conventions
  *  ============================================================================
- *
  *  Every mnemonic in this header follows the same suffix grammar:
- *
- *    _i        Immediate value (16-bit constant operand). Combine with
- *              _u or _s (single-letter modifier + type combined): add_ui,
- *              add_si. Examples: add_ui, add_si, and_i, or_i, xor_i,
- *              load_upper_i. and_i is sign-agnostic (andi zero-extends).
- *              load_upper_i is a unique verb; _i is the immediate marker,
- *              not a modifier+type combination.
- *
- *    _u        Unsigned (no-overflow, no-sign-extension). R-type
- *              arithmetic examples: add_u, sub_u, mult_u, div_u. I-type
- *              (combined with _i): add_ui.
- *
- *    _s        Signed (overflow-traps, sign-extends). R-type: add_s,
- *              sub_s, mult_s, div_s, set_lt_s. I-type (combined with _i):
- *              add_si.
+ *    _i: Immediate value (16-bit constant operand). 
+ *        Combine with _u or _s (single-letter modifier + type combined): add_ui, add_si.
+ *        Examples: add_ui, add_si, and_i, or_i, xor_i, load_upper_i. and_i is sign-agnostic (andi zero-extends).
+ *        load_upper_i is a unique verb; _i is the immediate marker, not a modifier+type combination.
+ *    _u: Unsigned (no-overflow, no-sign-extension).
+ *        R-type arithmetic examples: add_u, sub_u, mult_u, div_u. I-type (combined with _i): add_ui.
+ *    _s: Signed (overflow-traps, sign-extends). 
+ *        R-type: add_s, sub_s, mult_s, div_s, set_lt_s. I-type (combined with _i): add_si.
  *
  *  --- Shift family (R-type): verb-modifier-direction ---
- *    The shift macros use `shift_<modifier><direction>`. Modifier is
- *    the single letter `l` (logical) or `a` (arithmetic). Direction
- *    is the word `left` or `right`. Combined: `_lleft`, `_lright`,
- *    `_aright`. Examples: shift_lleft( rd, rt, shamt) (= sll)
- *                         shift_lright(rd, rt, shamt) (= srl)
- *                         shift_aright(rd, rt, shamt) (= sra)
- *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical
- *    to logical-left, so use shift_lleft for that case)
+ *    The shift macros use `shift_<modifier><direction>`.
+ *    Modifier is the single letter `l` (logical) or `a` (arithmetic).
+ *    Direction is the word `left` or `right`. Combined: `_lleft`, `_lright`, `_aright`.
+ *    Examples: shift_lleft( rd, rt, shamt) (= sll)
+ *              shift_lright(rd, rt, shamt) (= srl)
+ *              shift_aright(rd, rt, shamt) (= sra)
+ *    (no `_aleft`; MIPS has no `sla` — arithmetic-left is bit-identical to logical-left, so use shift_lleft for that case)
  *
  *  --- Jump/Call family ---
- *    Simple jumps keep the original short names: jump (j), jump_reg
- *    (jr), jump_link (jalr rs, rd). The jump-and-link-to variants
- *    (jal, jalr rs with default $ra) get the `call_` verb instead:
+ *    Simple jumps keep the original short names: jump (j), jump_reg (jr), jump_link (jalr rs, rd).
+ *    The jump-and-link-to variants (jal, jalr rs with default $ra) get the `call_` verb instead:
  *    call_addr (jal), call_reg (jalr rs, default $ra).
  *    Examples: jump(off)         (= j)
  *              jump_reg(rs)      (= jr)
@@ -40,32 +30,22 @@
  *              call_reg(rs)      (= jalr rs, default $ra)
  *              call_addr(off)    (= jal)
  *
- *    _r        Register marker — used only when the register type needs
- *              disambiguation (e.g., GTE data register vs control
- *              register). NOT used in plain R-type arithmetic (the
- *              R-type is implicit). Examples: gte_mv_to_data_r,
- *              gte_mv_to_ctrl_r.
+ *    _r: Register marker — used only when the register type needs disambiguation (e.g., GTE data register vs control register).
+ *        NOT used in plain R-type arithmetic (the R-type is implicit). Examples: gte_mv_to_data_r, gte_mv_to_ctrl_r.
+ *    _self:     Destination equals one source operand.
+ *               Examples: add_ui_self (I-type, to self), add_u_self (R-type, to self).
+ *    _mv_to_:   Direction: data flows into X.
+ *               Example: gte_mv_to_data_r, gte_mv_to_ctrl_r.
+ *    _mv_from_: Direction: data flows out of X.
+ *               Example: gte_mv_from_data_r, gte_mv_from_ctrl_r.
+ *    _str:      String-form — emits inline-asm string instead of `.word`.
+ *               Example: gte_rtpt_asm_str.
+ *    _2w / _1w: Word count of the emitted sequence.
+ *               Example: load_imm_2w.
  *
- *    _self     Destination equals one source operand.
- *              Examples: add_ui_self (I-type, to self),
- *                        add_u_self (R-type, to self).
- *
- *    _mv_to_   Direction: data flows into X.
- *              Example: gte_mv_to_data_r, gte_mv_to_ctrl_r.
- *
- *    _mv_from_ Direction: data flows out of X.
- *              Example: gte_mv_from_data_r, gte_mv_from_ctrl_r.
- *
- *    _str      String-form — emits inline-asm string instead of `.word`.
- *              Example: gte_rtpt_asm_str.
- *
- *    _2w / _1w Word count of the emitted sequence.
- *              Example: load_imm_2w.
- *
- *    _cop2     RESERVED — DO NOT USE in macro names. The `gte_` namespace
- *              prefix already implies coprocessor 2. Use `c2` only in:
- *                (a) integer opcode enums (op_lwc2 = 0x32, op_swc2 = 0x3A)
- *                (b) vendor-mnemonic macro aliases (gte_mtc2, gte_mfc2)
+ *    _cop2: RESERVED — DO NOT USE in macro names. The `gte_` namespace prefix already implies coprocessor 2. Use `c2` only in:
+ *      (a) integer opcode enums (op_lwc2 = 0x32, op_swc2 = 0x3A)
+ *      (b) vendor-mnemonic macro aliases (gte_mtc2, gte_mfc2)
  *
  *  Primitive commands: gp0_cmd_poly_f3 = 0x20 (byte opcode)
  *  Packed 32-bit cmd:  gp0_word_poly_f3(r, g, b) (32-bit, shifted)
@@ -80,9 +60,8 @@
  *            gte_lw_v0_xy(base) (gte + lw + v0 + xy)
  *            load_upper_i       (load-upper + immediate, unique verb)
  *
- *  Vendor mnemonics (sll, srl, sra, jr, j, jal, jalr) are NOT in this
- *  header. They live in the opt-in `mips_vendor_sym.h` for users who
- *  prefer the textbook MIPS assembly mnemonics.
+ *  Vendor mnemonics (sll, srl, sra, jr, j, jal, jalr) are NOT in this header.
+ *  They live in the opt-in `mips_vendor_sym.h` for users who prefer the textbook MIPS assembly mnemonics.
  * ============================================================================ */
 
 #ifdef INTELLISENSE_DIRECTIVES
@@ -98,19 +77,17 @@ enum {
 /* ============================================================================
  *  REGISTER INTEGER IDS (preprocessor-visible)
  * ============================================================================
- *  Every R_* enum below has a parallel R_*_Code `#define` so that the
- *  preprocessor can stringify the integer (e.g. for asm clobber lists and
- *  register-variable declarations via `rgcc(R_X)`). The enum value is
- *  bound to the `#define` so the two forms cannot drift apart.
+ *  Every R_* enum below has a parallel R_*_Code `#define` so that the preprocessor can stringify the integer
+ *  (e.g. for asm clobber lists and register-variable declarations via `rgcc(R_X)`).
+ *  The enum value is bound to the `#define` so the two forms cannot drift apart.
  *
- *  Only registers that get stringified need a `_Code` form; the rest are
- *  plain enum values. If you need to add a new one, follow the pattern:
+ *  Only registers that get stringified need a `_Code` form; the rest are plain enum values.
+ *  If you need to add a new one, follow the pattern:
  *      #define R_T7_Code 15
- *      R_T7 = R_T7_Code,        // in the enum
+ *      R_T7 = R_T7_Code,     // in the enum
  *
- *  User code should always reference the enum form (`R_T4`) at arithmetic
- *  sites and let `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify
- *  cases — never write the bare number `12`.
+ *  User code should always reference the enum form (`R_T4`) at arithmetic sites and let
+ *  `rlit(R_T4_Code)` / `rgcc(R_T4)` handle the stringify cases — never write the bare number `12`.
  * ============================================================================ */
 #define R_0_Code    0
 #define R_AT_Code   1
@@ -225,7 +202,6 @@ enum {
 	/* 2F: N/A */
 	// , op_lwc0
 
-
 	// , op_load_addr  = op_la
 	// , op_load_imm   = op_li
 	, op_jump       = op_j
@@ -327,15 +303,15 @@ enum { _BitOffsets = 0
  * Argument order matches the MIPS assembly syntax:
  *   dest-first, then source operands, then immediate last.
  *
- *   load_word(rt, base, off)   → lw   rt, off(base)
- *   store_word(rt, base, off)  → sw   rt, off(base)
- *   add_ui(rt, rs, imm)        → addiu rt, rs, imm
- *   shift_lleft(rd, rt, shamt)   → sll   rd, rt, shamt
- *   shift_lright(rd, rt, shamt)  → srl   rd, rt, shamt
- *   shift_aright(rd, rt, shamt)  → sra   rd, rt, shamt
- *   jump_reg(rs)               → jr    rs
- *   jump_link(rs, rd)          → jalr  rs        (link in rd, default $ra)
- *   nop                        → sll   $0, $0, 0
+ *   load_word(rt, base, off)    → lw   rt, off(base)
+ *   store_word(rt, base, off)   → sw   rt, off(base)
+ *   add_ui(rt, rs, imm)         → addiu rt, rs, imm
+ *   shift_lleft(rd, rt, shamt)  → sll   rd, rt, shamt
+ *   shift_lright(rd, rt, shamt) → srl   rd, rt, shamt
+ *   shift_aright(rd, rt, shamt) → sra   rd, rt, shamt
+ *   jump_reg(rs)                → jr    rs
+ *   jump_link(rs, rd)           → jalr  rs        (link in rd, default $ra)
+ *   nop                         → sll   $0, $0, 0
  */
 #define load_word(rt, base, off)   enc_i(op_lw,    (base), (rt), (off))
 #define load_byte(rt, base, off)   enc_i(op_lb,    (base), (rt), (off))
@@ -360,10 +336,10 @@ enum { _BitOffsets = 0
 
 /* Logic Opcodes */
 
-#define and_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_and)
-#define or_u(rd, rs, rt)           enc_r(op_special, (rs), (rt), (rd), 0, fc_or)
-#define xor_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_xor)
-#define nor_u(rd, rs, rt)          enc_r(op_special, (rs), (rt), (rd), 0, fc_nor)
+#define and_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_and)
+#define or_u(rd, rs, rt)         enc_r(op_special, (rs), (rt), (rd), 0, fc_or)
+#define xor_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_xor)
+#define nor_u(rd, rs, rt)        enc_r(op_special, (rs), (rt), (rd), 0, fc_nor)
 
 #define or_u_self(rd_rs, rt)       enc_r(op_special, (rd_rs), (rt), (rd_rs), 0, fc_or)
 
@@ -404,12 +380,9 @@ enum { _BitOffsets = 0
  * 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
+ * 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
@@ -441,7 +414,7 @@ enum { _BitOffsets = 0
 #define mov_to_low(rs)             enc_r(op_special, (rs), R_0, R_0, 0, fc_mtlo)
 
 /* --- Atomic branches (no pseudos like bgt/bge; compose with slt_* + branch_ne) ---
- * branch_equal  rs, rt, off → beq   rs, rt, off
+ * branch_equal   rs, rt, off → beq   rs, rt, off
  * branch_ne      rs, rt, off → bne   rs, rt, off
  * branch_lt_zero rs, off     → bltz  rs, off
  * branch_gt_zero rs, off     → bgtz  rs, off
@@ -463,7 +436,7 @@ enum { _BitOffsets = 0
 /* --- Shift-amount alias (matches the gas convention `\p3 = shamt`) --- */
 #define shift_amount(rd, rt, n)     shift_lleft(rd, rt, n)
 
-/* nop — canonical sll $0, $0, 0 */
+/* nop — sll $0, $0, 0 */
 #define nop  shift_lleft(rdiscard, rdiscard, 0)
 #define nop2 nop, nop
 
@@ -471,23 +444,16 @@ enum { _BitOffsets = 0
 #define load_imm_1w_s0(rt, imm) add_si((rt)), R_0, (imm))
 
 /* load_imm_2w — unconditional 2-word `li` form: `lui` + (ori | addi).
- *
- * Granular companion to `load_imm`: skips the compile-time range checks
- * and always emits 2 .words. Use this when:
+ * Granular companion to `load_imm`: skips the compile-time range checks and always emits 2 .words. Use this when:
  *   - you know `imm` is > 0xFFFF (otherwise you're wasting a word), OR
- *   - `imm` is not a compile-time constant and you want predictable
- *     2-word emission without the `__builtin_constant_p` branches.
+ *   - `imm` is not a compile-time constant and you want predictable 2-word emission without the `__builtin_constant_p` branches.
  *
  * The lo16 strategy is still chosen at expansion time on the lo half:
- *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui
- *                              already cleared bits 15..0)
- *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the
- *                              intended bit pattern)
+ *   lo16 in 0x0000..0x7FFF  →  addi (sign-ext is harmless, the lui already cleared bits 15..0)
+ *   lo16 in 0x8000..0xFFFF  →  ori  (zero-extends to preserve the intended bit pattern)
  *
- * For situations where you need to bypass even this choice (e.g. to
- * force a specific encoding for a known discontiguous high/low pair),
+ * For situations where you need to bypass even this choice (e.g. to force a specific encoding for a known discontiguous high/low pair),
  * see `load_imm_2w_ori_forced` and `load_imm_2w_addi_forced` below.
- *
  * Statement-level (not expression-level): emits its own `asm volatile(...)`.
  */
 #define load_imm_2w(rt, imm) do {                         \
@@ -518,9 +484,8 @@ enum { _BitOffsets = 0
 } while (0)
 
 /* load_imm_2w_addi_forced — force the `lui` + `addi` form regardless of lo16 sign.
- * Use when you know sign-extension is fine (e.g. lo16 is treated as
- * signed downstream) and you want a smaller effective instruction
- * (the assembler/MIPS hardware will sign-extend the imm16). */
+ * Use when you know sign-extension is fine (e.g. lo16 is treated as signed downstream) 
+ * and you want a smaller effective instruction (the assembler/MIPS hardware will sign-extend the imm16). */
 #define load_imm_2w_addi_forced(rt, imm) do {   \
 	/*U4 _li2a_imm_ = (U4)(imm);*/                \
 	asm volatile(asm_words(                       \
@@ -532,23 +497,17 @@ enum { _BitOffsets = 0
 
 /* load_imm rt, imm — true `li` semantics (assembler `li` pseudo)
  *
- * Dispatches at compile time on the immediate's range, picking the
- * smallest single-instruction form when possible:
- *
- *   imm in   0 ..  0x7FFF  →  addi  rt, $0, imm   (1 word)
- *   imm in 0x8000 .. 0xFFFF →  ori   rt, $0, imm   (1 word; sign-bit must be zeroed)
- *   imm in  0x10000 ..  0xFFFFFFFF  →  lui + (ori | addi)  (2 words)
- *
- * Statement-level (not expression-level): the macro emits its own
- * `asm volatile(...)` block with 1 or 2 .word constants. Callers can
- * group multiple `load_imm` calls in a single volatile by using the
- * lower-level encoders directly:
+ * Dispatches at compile time on the immediate's range, picking the smallest single-instruction form when possible:
+ *   imm in 0       .. 0x7FFF     →  addi  rt, $0, imm   (1 word)
+ *   imm in 0x8000  .. 0xFFFF     →  ori   rt, $0, imm   (1 word; sign-bit must be zeroed)
+ *   imm in 0x10000 .. 0xFFFFFFFF →  lui + (ori | addi)  (2 words)
  *
+ * Statement-level (not expression-level): the macro emits its own `asm volatile(...)` block with 1 or 2 .word constants.
+ * Callers can group multiple `load_imm` calls in a single volatile by using the lower-level encoders directly:
  *      load_imm(R_T4, 0x12345678);          // emits 2 .words
  *
- * Falls back to a 2-word form if `imm` is not a compile-time constant,
- * but that path is unusual (load_imm is most useful with literal
- * addresses and magic numbers). */
+ * Falls back to a 2-word form if `imm` is not a compile-time constant, but that path is unusual
+ * (load_imm is most useful with literal addresses and magic numbers). */
 #define load_imm(rt, imm) do {                                     \
 	if (cexpr_(imm) && ((imm) <= 0x7FFFU)) {                         \
 		/* Small positive: addi rt, $0, imm */                         \
@@ -588,9 +547,8 @@ enum { _BitOffsets = 0
 
 
 /* Standard clobber list for pure-MIPS asm volatile blocks: caller-saved
- * GPRs that the kernel treats as volatile (v0/v1/t0/t1/ra) plus the
- * "memory" barrier. The register ids are passed through `rlit` so
- * the R_*_Code `#define`s are stringified into "$N" at expansion time. */
+ * GPRs that the kernel treats as volatile (v0/v1/t0/t1/ra) plus the "memory" barrier.
+ * The register ids are passed through `rlit` so the R_*_Code `#define`s are stringified into "$N" at expansion time. */
 #define clbr_volatile_gprs rlit(R_V0), rlit(R_T0), rlit(R_T1), rlit(R_RA), clb_mem_drain
 
 #define asm_mips_flush_icache() asm volatile( asm_words( \

code/duffle/mips_vendor_sym.h

@@ -2,9 +2,8 @@
  *  duffle DSL — MIPS Vendor Mnemonics (opt-in)
  *  ============================================================================
  *
- *  Provides the textbook MIPS assembly mnemonics as thin aliases to the
- *  canonical duffle macros in mips.h. The duffle names are primary; this
- *  header is for users who prefer the textbook mnemonics.
+ *  Provides the textbook MIPS assembly mnemonics as thin aliases to the duffle macros in mips.h.
+ *  The duffle names are primary; this header is for users who prefer the textbook mnemonics.
  *
  *  USAGE:   #include "duffle/mips_vendor_sym.h"   // after mips.h
  *

code/duffle/pad.h

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

code/duffle/word_count.metadata.h

@@ -5,11 +5,10 @@
 // Format: WORD_COUNT(MACRO_NAME, COUNT)
 // 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 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.
+// 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 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.
 //
 // To regenerate: hand-count the instructions in each macro definition.
 // (You'll only need to do this once per macro — they don't change often.)
@@ -23,6 +22,7 @@ WORD_COUNT(call_reg,                         1)
 WORD_COUNT(call_addr,                        1)
 WORD_COUNT(branch_le_zero,                   1)
 WORD_COUNT(branch_equal,                     1)
+WORD_COUNT(branch_ne,                        1)
 WORD_COUNT(add_ui,                           1)
 WORD_COUNT(set_lt_u,                         1)
 WORD_COUNT(set_lt_s,                         1)
@@ -30,7 +30,9 @@ WORD_COUNT(set_lt_si,                        1)
 WORD_COUNT(set_lt_ui,                        1)
 WORD_COUNT(load_word,                        1)
 WORD_COUNT(load_half_u,                      1)
+WORD_COUNT(load_byte_u,                      1)
 WORD_COUNT(store_word,                       1)
+WORD_COUNT(store_byte,                       1)
 WORD_COUNT(add_ui_self,                      1)
 WORD_COUNT(add_u_self,                       1)
 WORD_COUNT(add_u,                            1)
@@ -38,6 +40,7 @@ WORD_COUNT(or_i,                             1)
 WORD_COUNT(or_i_self,                        1)
 WORD_COUNT(or_u,                             1)
 WORD_COUNT(or_u_self,                        1)
+WORD_COUNT(nor_u,                            1)
 WORD_COUNT(shift_lleft,                      1)
 WORD_COUNT(shift_lleft_self,                 1)
 WORD_COUNT(shift_lright,                     1)

code/gte_hello/hello_gte_tape.c

@@ -1,161 +0,0 @@
-#ifdef INTELLISENSE_DIRECTIVES
-#	include "duffle/gen/duffle.macs.h"
-#	include "duffle/gen/duffle.offsets.h"
-#	include "duffle/atom_dsl.h"
-#	include "duffle/lottes_tape.h"
-#	include "duffle/word_count.metadata.h"
-#	include "gen/gte_hello.offsets.h"
-#	include "hello_gte.h"
-#endif
-
-#pragma region MACs (Mips Atom components)
-
-
-
-#pragma endregion MACs
-
-#pragma region Baked Atoms
-
-typedef Struct_(Binds_CubeTri) {
-	U4     PrimCursor;
-	V4_S2* FaceCursor;
-	V3_S2* VertBase;
-	U4*    OtBase;
-};
-internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri)
-,	atom_reads(R_TapePtr)
-,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-){
-	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)), 
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)), 
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)), 
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)), 
-	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
-	mac_yield()
-};
-
-/* ============================================================================
- *  cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
- * ============================================================================
- *  Reads 4 indices from R_FaceCur (V4_S2 = 8 bytes), loads 4 vertices into
- *  the GTE, runs the PsyQ RotAverageNclip4 sequence, and renders a Poly_G4.
- */
-internal 
-MipsAtom_(cube_g4_face) atom_info(
-		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
-		atom_writes(R_PrimCursor, R_FaceCursor)
-){
-	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
-	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
-	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
-	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
-
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-	nop2, gte_cmdw_nclip,
-
-	nop2,  gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
-	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
-
-	store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
-	mac_format_g4_color(
-		/* c0 magenta */ 0xFF, 0x00, 0xFF,
-		/* c1 yellow  */ 0xFF, 0xFF, 0x00,
-		/* c2 cyan    */ 0x00, 0xFF, 0xFF,
-		/* c3 green   */ 0x00, 0xFF, 0x00),
-	mac_gte_store_g4_p012_post_rtpt_pre_rtps(),
-
-	shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
-	load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
-	gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
-
-	nop2, gte_cmdw_rotate_translate_perspective_single,
-	mac_gte_store_g4_p3_post_rtps(),
-
-	nop2, gte_cmdw_avg_sort_z4,
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),
-
-	add_ui(      R_AT, R_0,  OrderingTbl_Len),
-	set_lt_u(    R_AT, R_T1, R_AT),
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
-	mac_insert_ot_tag_g4(),
-
-atom_label(cube_g4_face_exit)
-	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
-	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
-	mac_yield()
-};
-
-typedef Struct_(Binds_FloorTri) {
-	U4     PrimCursor;
-	V3_S2* FaceCursor;
-	V3_S2* VertBase;
-	U4*    OtBase;
-};
-internal 
-MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri)
-	, atom_reads(R_TapePtr)
-	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-){
-	/* Pop 4 arguments from the tape directly into the workspace registers */
-	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)), 
-	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)), 
-	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)), 
-	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)), 
-	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
-	mac_yield()
-};
-
-internal 
-MipsAtom_(floor_f3_face) atom_info(
-	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
-	, atom_writes(R_PrimCursor, R_FaceCursor)
-) {
-	mac_load_tri_indices(  R_T0, R_T1, R_T2),
-	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
-	nop2, gte_cmdw_rotate_translate_perspective_triple,
-	nop2, gte_cmdw_nclip,
- 
-	/* Culling (Branch forward if Backface) */
-	nop2, gte_mv_from_data_r(R_T0, C2_MAC0),
-	nop,
-	branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop,
-	/* Format Primitive */
-	mac_format_f3_color(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
-	mac_gte_store_f3_post_rtpt(),
-	
-	/* Calculate Depth */
-	nop2, gte_avg_sort_z3,            
-	nop2, gte_mv_from_data_r(R_T1, C2_OTZ),      
-	/* Bounds Check OTZ < 2048 (Branch forward to skip insertion) */
-	add_ui(      R_AT, R_0,  OrderingTbl_Len),   
-	set_lt_u(    R_AT, R_T1, R_AT), 
-	branch_equal(R_AT, R_0,  atom_offset(bounds_chk, floor_f3_face_exit)), nop, 
-	/* Insert into Ordering Table Linked List */
-	mac_insert_ot_tag_f3(),
-	add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
-		// Note(Ed): No bounds checking, should be checked before atom runs.
-
-/* Advance Input Cursor & Yield (Both branch targets land here) */
-atom_label(floor_f3_face_exit) 
-	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
-	mac_yield()
-};
-
-typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
-internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
-	, atom_reads( R_TapePtr, R_PrimCursor)
-	, atom_writes(R_TapePtr)
-){
-	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
-	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),      
-	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
-	/* Calculate byte offset and store directly back to RAM */
-	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
-	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
-	mac_yield()
-};
-
-#pragma endregion Baked Atoms

code/hello_gte/gen/gte_hello.offsets.h

@@ -5,20 +5,20 @@
 #pragma region hello_gte_tape
 
 
-// --- atom: cube_g4_face (87 words) ---
+// --- atom: cube_g4_face (77 words) ---
 
-#define _atom_offset_cull_cube_g4_face_exit           48
-#define _atom_offset_bounds_chk_cube_g4_face_exit     12
+#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
 
 enum {
     atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
     atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
 };
 
-// --- atom: floor_f3_face (66 words) ---
+// --- atom: floor_f3_face (58 words) ---
 
-#define _atom_offset_culling_floor_f3_face_exit       29
-#define _atom_offset_bounds_chk_floor_f3_face_exit    13
+#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,

code/hello_gte/gen/hello_gte.offsets.h

@@ -0,0 +1,29 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\hello_gte\hello_gte.tape.c
+#pragma once
+
+#pragma region hello_gte.tape
+
+
+// --- atom: cube_g4_face (77 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+#pragma endregion hello_gte.tape
+

code/hello_gte/hello_gte.c

@@ -1,6 +1,6 @@
-#include "stdio.h"
+#include <stdio.h>
 #include <stdlib.h>
-#include "assert.h"
+#include <assert.h>
 // #include "libgpu.h"
 // #include "libetc.h"
 // #include "libgte.h"
@@ -20,10 +20,10 @@
 #include "duffle/lottes_tape.h"
 #include "duffle/word_count.metadata.h"
 
-#	include "gen/gte_hello.offsets.h"
+#	include "gen/hello_gte.offsets.h"
 #include "hello_gte.h"
 
-#include "hello_gte_tape.c"
+#include "hello_gte.tape.c"
 
 typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
 typedef Array_(OrderingTable_Buffer, 2);
@@ -99,8 +99,13 @@ typedef Struct_(Ent_Floor) {
 	A2_V3_S2 faces;
 };
 
-enum { scratchpad_size = 1024, };
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
 typedef Struct_(SMemory) {
+	U4 MemTape[MemTape_Len];
+
 	DoubleBuffer            screen_buf;
 	A2_OrderingTable_Buffer ordering_tbl;
 	PrimitiveArena          primitives;
@@ -117,7 +122,7 @@ global SMemory smem;
 extern SMemory smem;
 
 // TODO(Ed):
-FI_ U4* spad_warm(MipsAtom atom) {
+FI_ U4* spad_warm(Slice_MipsCode atom) {
 	return nullptr;
 }
 
@@ -182,6 +187,7 @@ void gp_display_frame(DoubleBuffer* screen_buf, S4* active_buf_id, U4* ordering_
 void render(void) {
 }
 
+GCC_OPTIMIZATION_DISABLE
 void update(PrimitiveArena* pa, U4* ordering_buf) 
 {
 	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
@@ -207,6 +213,8 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 	A2_S2 p;    //???
 	S4 flag; //????
 
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
 	// Draw Cube
 	if (0)
 	{
@@ -259,9 +267,8 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
 
-		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
-		TapeBuilder tb = tb_make_old(&tape_arena); tb_scope(& tb) {
-			tb_emit(& tb, code_rbind_cube_g4_face);
+		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));
@@ -269,10 +276,10 @@ void update(PrimitiveArena* pa, 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, code_cube_g4_face);
+				tb_emit(& tb, cube_g4_face);
 			}
 
-			tb_emit(& tb, code_sync_primitive_arena);
+			tb_emit(& tb, sync_primitive_arena);
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}
@@ -344,48 +351,40 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
 		U4 prim_cursor = prim_base + pa->used;
 
-		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris.
+		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris?
 		// The tape atoms in-flight should not need to care.
 
 		// Prepare the tape. (Push protocol to tape)
-		LP_ U4 mem_temp_tape[512];
-		TapeBuilder tb = tb_make(slice_ut_arr(mem_temp_tape)); tb_scope(& tb) {
-			tb_emit(& tb, code_set_gte_world);
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, set_gte_world);
 				tb_data(& tb, u4_(& smem.tform_world));
 
-			tb_emit(& tb, code_rbind_floor_f3_face);
+			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, code_floor_f3_face);
+				tb_emit(& tb, floor_f3_face);
 			}
-			// After code_floor_f3_face iterations complete, the primitive arena's used counter needs updating.
-			tb_emit(& tb, code_sync_primitive_arena);
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
 				tb_data(& tb, u4_(& pa->used));
 				tb_data(& tb, prim_base);
 		}
-
 		tape_run(tb_slice(tb));// Fire off the tape.
 
 		// C-side state (pa->used) has already been updated by the tape!
 		smem.floor.rot.y += 5;
 	}
 	// --- TAPE DIAGNOSTICS ---
-	if (1)
+	if (0)
 	{
 		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
 		TapeBuilder tb = tb_make_old(& tape_arena); tb_scope(& tb) {
 			// Skip set_gte_world atom for diagnostics to isolate the triangle loop
 			for (U4 i = 0; i < Floor_num_faces; i++) {
-				// =======================================================
-				// SWAP EMIT TO TEST DIFFERENT PARTS OF THE PIPELINE:
-				// =======================================================
-				// 1. code_diag_yield -> Tests Tape Engine jump logic
-				// 2. code_diag_color -> Tests OT and Prim Arena memory
-				// 3. code_diag_gte   -> Tests Vertex arrays and GTE Math
 				// tb_emit(& tb, code_diag_yield);
 				// tb_emit(& tb, code_diag_color);
 				// tb_emit(& tb, code_diag_gte); 
@@ -394,9 +393,9 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 		B1* prim_cursor = (B1*)r_(pa->buf)[smem.active_buf_id] + pa->used;
 		tape_run(tb_slice(tb));
 		pa->used = (U4)prim_cursor - (U4)r_(pa->buf)[smem.active_buf_id];
-		smem.floor.rot.y += 5;
 	}
 }
+GCC_OPTIMIZATION_ENABLE
 
 int main(void)
 {

code/hello_gte/hello_gte.tape.c

@@ -0,0 +1,218 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/atom_dsl.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/word_count.metadata.h"
+#	include "gen/hello_gte.offsets.h"
+#	include "hello_gte.h"
+#endif
+
+#pragma region MACs (Mips Atom components)
+
+
+
+#pragma endregion MACs
+
+#pragma region Baked Atoms
+
+/* DIAGNOSTIC 1: Pure tape loop test */
+internal MipsAtom_(diag_yield) { mac_yield() };
+
+/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
+internal MipsAtom_(diag_color) {
+	store_word(  R_0, R_T7, 0), 
+	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
+	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
+	store_word(  R_AT, R_T7, 4),
+	
+	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
+	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
+
+	add_ui(          R_T1, R_0,  10),
+	shift_lleft_self(R_T1,       S_(U4)/2), 
+	add_u_self(      R_T1,       R_T6),         
+	
+	load_word(   R_AT, R_T1, 0),        
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
+	store_word(  R_AT, R_T7, 0),       
+	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
+	or_u_self(  R_AT, R_V0),          
+	store_word( R_AT, R_T1, 0),       
+
+	add_ui(R_T7, R_T7, 20),          
+
+	mac_yield()
+};
+
+/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
+internal MipsAtom_(diag_gte) {
+	/* Load 3 indices */
+	load_half_u(R_T0, R_T4, 0),
+	load_half_u(R_T1, R_T4, 2),
+	load_half_u(R_T2, R_T4, 4),
+
+	/* Load Vertices into GTE */
+	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
+
+	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
+
+	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+
+	/* Run Math */
+	nop2, gte_cmdw_rtpt,
+	nop2, gte_cmdw_nclip,
+	nop2,
+
+	/* Advance Face Cursor and Yield */
+	add_ui(R_T4, R_T4, 8),
+
+	mac_yield()
+};
+
+typedef Struct_(Binds_CubeTri) {
+	U4     PrimCursor;
+	V4_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal MipsAtom_(rbind_cube_g4_face) atom_info(atom_bind(Binds_CubeTri), atom_phase(cube_g4)
+,	atom_reads(R_TapePtr)
+,	atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_CubeTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_CubeTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_CubeTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_CubeTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_CubeTri)),
+	mac_yield()
+};
+
+ // cube_g4_face — Draw one cube face (Gouraud-shaded quad) via the GTE tape pipeline
+internal
+MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),
+		atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase),
+		atom_writes(R_PrimCursor, R_FaceCursor)
+){
+	load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+	load_half_u(R_T1, R_FaceCursor, 1 * S_(S2)),
+	load_half_u(R_T2, R_FaceCursor, 2 * S_(S2)),
+	load_half_u(R_T3, R_FaceCursor, 3 * S_(S2)),
+
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
+		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+			mac_insert_ot_tag_g4(),
+			mac_format_g4_color(
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+){
+	/* Pop 4 arguments from the tape directly into the workspace registers */
+	load_word(R_PrimCursor, R_TapePtr, O_(Binds_FloorTri,PrimCursor)),
+	load_word(R_FaceCursor, R_TapePtr, O_(Binds_FloorTri,FaceCursor)),
+	load_word(R_VertBase,   R_TapePtr, O_(Binds_FloorTri,VertBase)),
+	load_word(R_OtBase,     R_TapePtr, O_(Binds_FloorTri,OtBase)),
+	add_ui_self(            R_TapePtr, S_(Binds_FloorTri)),
+	mac_yield()
+};
+
+// atom_dbg_skip
+internal
+MipsAtom_(floor_f3_face) atom_info(atom_phase(floor_f3)
+	, atom_reads( R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase)
+	, atom_writes(R_PrimCursor, R_FaceCursor)
+) {
+	mac_load_tri_indices(  R_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
+		/* Format Primitive */
+		mac_gte_store_f3(),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < 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(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(),                 /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+#pragma endregion Baked Atoms

code/hello_joypad/gen/hello_joypad.macs.h

@@ -0,0 +1,71 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.tape.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)
+
+#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, 48)
+

code/hello_joypad/gen/hello_joypad.offsets.h

@@ -0,0 +1,73 @@
+// Auto-generated by ps1_meta.lua (passes/offsets.lua) — DO NOT EDIT
+// Source: C:\projects\Pikuma\ps1\code\hello_joypad\hello_joypad.tape.c
+#pragma once
+
+#pragma region hello_joypad.tape
+
+
+// --- atom: cube_g4_face (77 words) ---
+
+#define _atom_offset_cull_cube_g4_face_exit           42
+#define _atom_offset_bounds_chk_cube_g4_face_exit     24
+
+enum {
+    atom_offset_cull_cube_g4_face_exit = _atom_offset_cull_cube_g4_face_exit,
+    atom_offset_bounds_chk_cube_g4_face_exit = _atom_offset_bounds_chk_cube_g4_face_exit,
+};
+
+// --- atom: floor_f3_face (58 words) ---
+
+#define _atom_offset_culling_floor_f3_face_exit       25
+#define _atom_offset_bounds_chk_floor_f3_face_exit    16
+
+enum {
+    atom_offset_culling_floor_f3_face_exit = _atom_offset_culling_floor_f3_face_exit,
+    atom_offset_bounds_chk_floor_f3_face_exit = _atom_offset_bounds_chk_floor_f3_face_exit,
+};
+
+// --- atom: pad_bios_snapshot (78 words) ---
+
+#define _atom_offset_snap_root_skip_disconnected      8
+#define _atom_offset_disconnected_snap_end            60
+#define _atom_offset_case_2_id_dispatch               8
+#define _atom_offset_pending_snap_end                 50
+#define _atom_offset_id_dispatch_try_analog_stick     11
+#define _atom_offset_id_dispatch_snap_end             37
+#define _atom_offset_try_analog_stick_try_analog_pad  12
+#define _atom_offset_analog_stick_snap_end            23
+#define _atom_offset_try_analog_pad_try_unsupported   11
+#define _atom_offset_analog_pad_snap_end              9
+
+enum {
+    atom_offset_snap_root_skip_disconnected = _atom_offset_snap_root_skip_disconnected,
+    atom_offset_disconnected_snap_end = _atom_offset_disconnected_snap_end,
+    atom_offset_case_2_id_dispatch = _atom_offset_case_2_id_dispatch,
+    atom_offset_pending_snap_end = _atom_offset_pending_snap_end,
+    atom_offset_id_dispatch_try_analog_stick = _atom_offset_id_dispatch_try_analog_stick,
+    atom_offset_id_dispatch_snap_end = _atom_offset_id_dispatch_snap_end,
+    atom_offset_try_analog_stick_try_analog_pad = _atom_offset_try_analog_stick_try_analog_pad,
+    atom_offset_analog_stick_snap_end = _atom_offset_analog_stick_snap_end,
+    atom_offset_try_analog_pad_try_unsupported = _atom_offset_try_analog_pad_try_unsupported,
+    atom_offset_analog_pad_snap_end = _atom_offset_analog_pad_snap_end,
+};
+
+// --- atom: pad_apply_input (60 words) ---
+
+#define _atom_offset_dpad_left_exit_dpad_left         6
+#define _atom_offset_dpad_right_exit_dpad_right       6
+#define _atom_offset_dead_zone_low_check_dead_low_active  8
+#define _atom_offset_dead_zone_high_check_dead_high_active  15
+#define _atom_offset_dead_zone_skip_exit_stick        23
+#define _atom_offset_end_low_exit_stick               11
+
+enum {
+    atom_offset_dpad_left_exit_dpad_left = _atom_offset_dpad_left_exit_dpad_left,
+    atom_offset_dpad_right_exit_dpad_right = _atom_offset_dpad_right_exit_dpad_right,
+    atom_offset_dead_zone_low_check_dead_low_active = _atom_offset_dead_zone_low_check_dead_low_active,
+    atom_offset_dead_zone_high_check_dead_high_active = _atom_offset_dead_zone_high_check_dead_high_active,
+    atom_offset_dead_zone_skip_exit_stick = _atom_offset_dead_zone_skip_exit_stick,
+    atom_offset_end_low_exit_stick = _atom_offset_end_low_exit_stick,
+};
+
+#pragma endregion hello_joypad.tape
+

code/hello_joypad/hello_joypad.c

@@ -0,0 +1,545 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+// #include "libgpu.h"
+// #include "libetc.h"
+// #include "libgte.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/gen/duffle.macs.h"
+#	include "duffle/gen/duffle.offsets.h"
+#include "duffle/atom_dsl.h"
+#include "duffle/lottes_tape.h"
+#include "duffle/word_count.metadata.h"
+
+#include "psyq.h"
+
+#	include "gen/hello_joypad.macs.h"
+#	include "gen/hello_joypad.offsets.h"
+#include "hello_joypad.h"
+
+#include "psyq.c"
+#include "hello_joypad.tape.c"
+
+
+typedef U4 OrderingTable_Buffer[OrderingTbl_Len];
+typedef Array_(OrderingTable_Buffer, 2);
+
+typedef B1 PrimitiveBuffer[PrimitiveBuff_Len];
+typedef Array_(PrimitiveBuffer, 2);
+typedef Struct_(PrimitiveArena) {
+	A2_PrimitiveBuffer buf;
+	U4                 used;
+};
+
+#define Cube_num_verts 8
+typedef Array_(V3_S2, Cube_num_verts);
+#define Cube_num_faces 6
+typedef Array_(V4_S2, Cube_num_faces);
+I_ void ent_cube128_init(A8_V3_S2* verts, A6_V4_S2* faces) {
+	LP_ A8_V3_S2 baked_verts = (A8_V3_S2) {
+		{ -128, -128, -128 },
+		{  128, -128, -128 },
+		{  128, -128,  128 },
+		{ -128, -128,  128 },
+		{ -128,  128, -128 },
+		{  128,  128, -128 },
+		{  128,  128,  128 },
+		{ -128,  128,  128 }
+	};
+	LP_ A6_V4_S2 baked_faces = (A6_V4_S2) {
+		{ 3, 2, 0, 1 },
+		{ 0, 1, 4, 5 },
+		{ 4, 5, 7, 6 },
+		{ 1, 2, 5, 6 },
+		{ 2, 3, 6, 7 },
+		{ 3, 0, 7, 4 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A8_V3_S2) );
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A6_V4_S2) );
+	return;
+}
+typedef Struct_(Ent_Cube) {
+	V3_S4 accel;
+	V3_S4 vel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A8_V3_S2 verts;
+	A6_V4_S2 faces;
+};
+
+#define Floor_num_verts 4
+typedef Array_(V3_S2, Floor_num_verts);
+#define Floor_num_faces 2
+typedef Array_(V3_S2, Floor_num_faces);
+I_ void ent_floor_init(A4_V3_S2* verts, A2_V3_S2* faces) {
+	LP_ A4_V3_S2 baked_verts = (A4_V3_S2) {
+		{ -900, 0, -900 },
+		{ -900, 0,  900 },
+		{  900, 0, -900 },
+		{  900, 0,  900 },
+	};
+	LP_ A2_V3_S2 baked_faces = (A2_V3_S2) {
+		{ 0, 1, 2 },
+		{ 1, 3, 2 },
+	};
+	mem_copy(u4_(verts), u4_(& baked_verts), S_(A4_V3_S2));
+	mem_copy(u4_(faces), u4_(& baked_faces), S_(A2_V3_S2));
+};
+typedef Struct_(Ent_Floor) {
+	V3_S4 accel;
+	V3_S4 pos;
+	V3_S4 scale;
+	V3_S2 rot;
+	A4_V3_S2 verts;
+	A2_V3_S2 faces;
+};
+
+
+enum { 
+	Scratchpad_Len = 1024, 
+	MemTape_Len    = 512,
+};
+typedef Struct_(SMemory) {
+	U4 MemTape[MemTape_Len];
+
+	DoubleBuffer            screen_buf;
+	A2_OrderingTable_Buffer ordering_tbl;
+	PrimitiveArena          primitives;
+	S4                      active_buf_id;
+
+	M3_S2 tform_world;
+
+	Ent_Cube  cube;
+	Ent_Floor floor;
+
+	PadBiosRaw pad_raw[2];
+	PadState   pad[2];
+
+	U4_V scratchpad; // d-cache
+};
+global SMemory smem;
+extern SMemory smem;
+
+I_ B1* prim__alloc(U4 type_width, Str8 type_name) {
+	gknown PrimitiveArena* pa  = & smem.primitives;
+	gknown B1*             buf = (B1*) r_(smem.primitives.buf)[smem.active_buf_id];
+	assert(pa->used + type_width < PrimitiveBuff_Len);
+	B1* next  = buf + pa->used;
+	pa->used += type_width;
+	return next;
+}
+#define prim_alloc(type) (type*)prim__alloc(S_(type), slit( stringify(type)))
+
+/* Uses ONE 8-byte frame allocated via the compiler's standard prologue.
+ * The 4 wasted-arg words for B(12h) InitPAD2 live at [SP+0..15] but are not explicitly allocated.
+ * The compiler handles the MIPS O32 "wasted stack" convention for us by treating the B-call as a 4-arg call.
+ *
+ * The buffer pointers are passed as arguments so the compiler keeps them in callee-saved registers;
+ * The B(12h) asm volatile block does NOT clobber those registers (it clobbers only the volatile GPRs + the B-table arg registers explicitly).
+ * The C-level writes after the call re-load the pointers from their callee-saved homes.
+ *
+ * The clobber list for both B-calls names the full BIOS destroy set documented in kernelbios.md:167-174 (R1..R15, R24..R25, R31, HI/LO).
+ * The kernel-ABI "volatile GPRs" subset is clb_system; the rest of the destroy set is enumerated explicitly here. */
+NI_ void pad_bios_init_start(PadBiosRaw* raw0, PadBiosRaw* raw1)
+{
+	/* Pin raw0 + raw1 to $a0 + $a1 via rgcc; the B(12h) call uses these directly.
+	 * The `(void)` casts mark them as unread after the call so the compiler doesn't need to move them back. */
+	register PadBiosRaw* p0 rgcc(R_A0) = raw0;
+	register PadBiosRaw* p1 rgcc(R_A1) = raw1;
+	(void)p0; (void)p1;
+
+	// TODO(Ed): Properly annotate the raw values in the inline asm instructions.
+	// Use enums.
+
+	/* B(12h) InitPAD2(raw0, 0x22, raw1, 0x22)
+	 *   $a0 = raw0 (rgcc-bound; survives the sequence below)
+	 *   $a1 = raw1 (preserved into $a2 before $a1 is overwritten)
+	 *   $a2 = raw1 (moved from $a1; survives $a1's overwrite)
+	 *   $a3 = 0x22 (immediate)
+	 *   $t1 = 0x12 (function number)
+	 *   $t2 = 0xB0 (BIOS B-table address) */
+	asm volatile(
+		asm_words(
+			or_u(    rarg_2, rarg_1, rdiscard), /* $a2 = $a1 = raw1 */
+			add_ui(  rarg_1, rdiscard, 0x22),   /* $a1 = 0x22 */
+			add_ui(  rarg_3, rdiscard, 0x22),   /* $a3 = 0x22 */
+			add_ui(  rtmp_1, rdiscard, 0x12),   /* $t1 = 0x12 */
+			add_ui(  rtmp_2, rdiscard, 0xB0),   /* $t2 = 0xB0 */
+			call_reg(rtmp_2),                   /* jalr $t2, $ra */
+			nop                                 /* BD slot */
+		)
+		asm_rpins, r_use(p0), r_use(p1)
+		asm_clobber: 
+			rlit(R_AT), 
+			rlit(R_V0), rlit(R_V1), 
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+
+	/* The C-level writes re-load the pointers via the parameter names and write 0xFF to each 
+	 * buffer's status byte to mark the initial-state hazard documented in kernelbios.md:1621-1624. */
+	u1_v(raw0)[0] = 0xFF;
+	u1_v(raw1)[0] = 0xFF;
+
+	/* B(13h) StartPAD2() — no args. The BIOS preserves $sp. */
+	asm volatile(
+		asm_words(
+			add_ui(  rtmp_1, rdiscard, 0x13), /* $t1 = 0x13 */
+			add_ui(  rtmp_2, rdiscard, 0xB0), /* $t2 = 0xB0 (re-load) */
+			call_reg(rtmp_2),                 /* jalr $t2, $ra */
+			nop                               /* BD slot */
+		)
+		asm_clobber:
+			rlit(R_AT),
+			rlit(R_V0), rlit(R_V1),
+			rlit(R_T0), rlit(R_T1), rlit(R_T2), rlit(R_T3), rlit(R_T4),
+			rlit(R_T5), rlit(R_T6), rlit(R_T7), rlit(R_T8), rlit(R_T9),
+			rlit(R_RA),
+			clb_mem_drain
+	);
+}
+
+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) 
+{
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape));
+
+	if (0) // Pad Input (dead — kept for the source-as-written record; references the deleted `pad_state` field)
+	{
+		(void)Pad_Left; (void)Pad_Right; /* suppress unused-token warnings */
+		if (false) {
+			smem.cube.rot.y  += 30;
+			smem.floor.rot.y += 5;
+		}
+		if (false) {
+			smem.cube.rot.y  -= 30;
+			smem.floor.rot.y -= 5;
+		}
+	}
+	if (1) // Pad Input (Tape version)
+	{
+		tb.used = 0; tb_scope_run(& tb) {
+			/* BIOS-owned polling: per-frame snapshot of both ports. */
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw, & smem.pad_raw[0]);
+				tb_data_(state, & smem.pad[0]);
+			tb_emit_(pad_bios_snapshot);
+				tb_data_(raw,   & smem.pad_raw[1]);
+				tb_data_(state, & smem.pad[1]);
+			/* Per-frame rotation apply: consume pad[0].buttons + pad[0].left_x */
+			tb_emit_(pad_apply_input);
+				tb_data_(state,     & smem.pad[0]);
+				tb_data_(cube_rot,  & smem.cube.rot);
+				tb_data_(floor_rot, & smem.floor.rot);
+		}
+	}
+
+	orderingtbl_clear_reverse(ordering_buf, OrderingTbl_Len);
+
+	// Update the position based on acceleration and velocity
+	gknown V3_S4_R pos = & smem.cube.pos;
+	gknown V3_S4_R vel = & smem.cube.vel;
+	gknown V3_S4_R acc = & smem.cube.accel;
+	add_v3s4(vel, acc[0]);
+	add_v3s4_fp(pos, vel[0]);
+	// vel->x += acc->x;
+	// vel->y += acc->y;
+	// vel->z += acc->z;
+	// pos->x += vel->x;
+	// pos->y += vel->y;
+	// pos->z += vel->z;
+
+	if (pos->y + 150 > smem.floor.pos.y) vel->y *= -1;
+
+	// Prep
+	S4 nclip = 0;
+	S4 orderingtbl_z = 0;
+	A2_S2 p;    //???
+	S4 flag; //????
+
+
+	// Draw Cube
+	if (0)
+	{
+		m3s2_rotation   (& smem.cube.rot,    & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.cube.pos);
+		m3s2_scale      (& smem.tform_world, & smem.cube.scale);
+		// gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+		for (U4 face_id = 0; face_id < Cube_num_faces; face_id += 1)
+		{
+			Poly_G4* quad = prim_alloc(Poly_G4); set_poly_g4(quad);
+			quad->c0 = rgb8(255,   0, 255);
+			quad->c1 = rgb8(255, 255,   0);
+			quad->c2 = rgb8(  0, 255, 255);
+			quad->c3 = rgb8(  0, 255,   0);
+
+			V4_S2* face = & smem.cube.faces[face_id];
+			V3_S2* p0   = & smem.cube.verts[face->x];
+			V3_S2* p1   = & smem.cube.verts[face->y];
+			V3_S2* p2   = & smem.cube.verts[face->z];
+			V3_S2* p3   = & smem.cube.verts[face->w];
+
+			nclip = rtp_avg_nclip_a4_v3s2(
+				p0, p1, p2, p3,
+				& quad->p0, & quad->p1, & quad->p2, & quad->p3,
+				& p, & orderingtbl_z, & flag
+			);
+			if (nclip <= 0) {
+				continue;
+			}
+
+			if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+				orderingtbl_add_primitive(ordering_buf[orderingtbl_z], quad);
+			}
+		}
+		// smem.cube.rot.x +=  6;
+		// smem.cube.rot.y +=  8;
+		// smem.cube.rot.z += 12;
+		smem.cube.rot.y += 30;
+	}
+	// Draw cube (tape method) - two triangles per face
+	if (1)
+	{
+		m3s2_rotation   (& smem.cube.rot,    & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.cube.pos);
+		m3s2_scale      (& smem.tform_world, & smem.cube.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, rbind_cube_g4_face);
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.cube.faces));
+				tb_data(& tb, u4_(smem.cube.verts));
+				tb_data(& tb, u4_(ordering_buf));
+
+			for (U4 i = 0; i < Cube_num_faces; i++) {
+				// Two triangles per quad face: (x,y,z) and (x,z,w)
+				tb_emit(& tb, cube_g4_face);
+			}
+
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));
+
+		// smem.cube.rot.y += 30;
+	}
+	// Draw Floor
+	if (0)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+		gte_matrix_set_rotation   (& smem.tform_world);
+		gte_matrix_set_translation(& smem.tform_world);
+		for (U4 face_id = 0; face_id < Floor_num_faces; face_id += 1)
+		{
+			Poly_F3* tri = prim_alloc(Poly_F3); set_poly_f3(tri);
+			tri->color   = rgb8(255, 255, 255);
+
+			V3_S2* face = & smem.floor.faces[face_id];
+			register V3_S2* p0 rgcc(R_T4) = & smem.floor.verts[face->x];
+			register V3_S2* p1 rgcc(R_T5) = & smem.floor.verts[face->y];
+			register V3_S2* p2 rgcc(R_T6) = & smem.floor.verts[face->z];
+
+			gte_load_v0(p0, R_T4);
+			/*
+			asm volatile( ".word " "%0" ", %1" : : 
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_xy & REG_MASK) << RT_SHIFT) | (0            & IMM_MASK)),
+				"i"(((op_lwc2 & OPCODE_MASK) << OPCODE_SHIFT) | ((R_T4 & REG_MASK) << RS_SHIFT) | ((gte_in_v0_z  & REG_MASK) << RT_SHIFT) | (GTE_Z_Offset & IMM_MASK)), 
+				"r"(p0) : 
+				"$2", "$8", "$9", "$31", "memory" 
+			);
+			*/
+			gte_load_v1(p1, R_T5);
+			gte_load_v2(p2, R_T6);
+
+			gte_rtpt();
+			gte_nclip();
+			gte_stotz(& nclip);
+
+			// nclip = rtp_avg_nclip_a3_v3s2(p0, p1, p2
+			// 	, & tri->p0, & tri->p1, & tri->p2
+			// 	, & p, & orderingtbl_z, & flag
+			// );
+			// if (nclip <= 0) {
+			// 	continue;
+			// }
+
+			if (nclip > 0 ) {
+				gte_stsxy3(& tri->p0, & tri->p1, & tri->p2);
+				gte_avsz3();
+				gte_stotz(& orderingtbl_z);
+
+				if ((orderingtbl_z > 0) && (orderingtbl_z < OrderingTbl_Len)) {
+					orderingtbl_add_primitive(ordering_buf[orderingtbl_z], tri);
+				}
+			}
+		}
+		smem.floor.rot.y += 5;
+	}
+	// Draw floor tape method
+	if (1)
+	{
+		m3s2_rotation   (& smem.floor.rot,   & smem.tform_world);
+		m3s2_translation(& smem.tform_world, & smem.floor.pos);
+		m3s2_scale      (& smem.tform_world, & smem.floor.scale);
+
+		U4 prim_base   = u4_(pa->buf[smem.active_buf_id]);
+		U4 prim_cursor = prim_base + pa->used;
+
+		// TODO(Ed): We should do a bounds check beforehand to confirm pa can hold all tris?
+		// The tape atoms in-flight should not need to care.
+
+		// Prepare the tape. (Push protocol to tape)
+		tb.used = 0; tb_scope(& tb) {
+			tb_emit(& tb, set_gte_world);
+				tb_data(& tb, u4_(& smem.tform_world));
+
+			tb_emit(& tb, rbind_floor_f3_face);
+			// TODO(Ed): Just use a single context struct ref
+				tb_data(& tb, prim_cursor);
+				tb_data(& tb, u4_(smem.floor.faces));
+				tb_data(& tb, u4_(smem.floor.verts));
+				tb_data(& tb, u4_(ordering_buf));
+			for (U4 i = 0; i < Floor_num_faces; i++) {
+				tb_emit(& tb, floor_f3_face);
+			}
+			// After floor_f3_face iterations complete, the primitive arena's used counter needs updating.
+			tb_emit(& tb, sync_primitive_arena);
+				tb_data(& tb, u4_(& pa->used));
+				tb_data(& tb, prim_base);
+		}
+		tape_run(tb_slice(tb));// Fire off the tape.
+
+		// C-side state (pa->used) has already been updated by the tape!
+		// smem.floor.rot.y += 5;
+	}
+	// --- TAPE DIAGNOSTICS ---
+	if (0)
+	{
+		LP_ U4 mem_temp_tape[512]; FArena tape_arena; farena_init(& tape_arena, slice_ut_arr(mem_temp_tape));
+		TapeBuilder tb = tb_make_old(& tape_arena); tb_scope(& tb) {
+			// Skip set_gte_world atom for diagnostics to isolate the triangle loop
+			for (U4 i = 0; i < Floor_num_faces; i++) {
+				// tb_emit(& tb, code_diag_yield);
+				// tb_emit(& tb, code_diag_color);
+				// tb_emit(& tb, code_diag_gte); 
+			}
+		}
+		B1* prim_cursor = (B1*)r_(pa->buf)[smem.active_buf_id] + pa->used;
+		tape_run(tb_slice(tb));
+		pa->used = (U4)prim_cursor - (U4)r_(pa->buf)[smem.active_buf_id];
+	}
+}
+GCC_OPTIMIZATION_ENABLE
+
+void render(void) {
+}
+
+int main(void)
+{
+	smem = (SMemory){0};
+	smem.scratchpad = C_(U4_V, 0x1F800000);
+	// smem.primitives.used = 0;
+	// smem.active_buf_id   = 0;
+	/*Persistent Entity Setup*/{
+		ent_cube128_init(& smem.cube.verts, & smem.cube.faces); {
+			Ent_Cube* cube = & smem.cube;
+			cube->rot    = v3s2(0, 0, 0);
+			cube->scale  = v3s4_fp_one();
+			cube->accel  = v3s4(0, 1, 0);
+			cube->pos    = v3s4(0, -400, 1800);
+		}
+		ent_floor_init(& smem.floor.verts, & smem.floor.faces); {
+			Ent_Floor* floor = & smem.floor;
+			floor->rot   = v3s2(0, 0, 0);
+			floor->pos   = v3s4(0, 450, 1800);
+			floor->scale = v3s4_fp_one();
+		}
+	}
+	TapeBuilder tb = tb_make(slice_ut_arr(smem.MemTape)); {
+		reset_graph(0);
+		/* Direct BIOS: poll both ports during VBlank. */
+		pad_bios_init_start(& smem.pad_raw[0], & smem.pad_raw[1]);
+		/* Pinned registers for the GPU init atom. */
+		register U4*           io_base_addr rgcc(R_IO_BaseAddr) = u4_r(IO_BASE_ADDR);
+		register DoubleBuffer* screen_buf   rgcc(R_ScreenBuf)   = & smem.screen_buf;
+		tb.used = 0; tb_scope_run(& tb) {
+			tb_emit(& tb, screen_env_init);
+			tb_emit(& tb, gp_screen_init);
+		}
+	}
+	while (1) {
+		gknown S4* active_buf_id  = & smem.active_buf_id;
+		gknown U4* ordering_buf   = r_(smem.ordering_tbl)[active_buf_id[0]];
+		gknown PrimitiveArena* pa = & smem.primitives;
+		update(pa, ordering_buf);
+		render();
+		gp_display_frame(& smem.screen_buf, active_buf_id, ordering_buf, pa);
+	};
+	return 0;
+}

code/hello_joypad/hello_joypad.h

@@ -0,0 +1,26 @@
+#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
+};
+
+enum {
+	ScreenRes_X = 320,
+	ScreenRes_Y = 240,
+	ScreenZ     = 320,
+	ScreenRes_CenterX = (ScreenRes_X >> 1),
+	ScreenRes_CenterY = (ScreenRes_Y >> 1),
+};
+
+enum {
+	fp_one = (1 << 12),
+};
+
+#define v3s4_fp_one() v3s4(fp_one, fp_one, fp_one)

code/hello_joypad/hello_joypad.tape.c

@@ -0,0 +1,705 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#	include "duffle/gen/duffle.macs.h"
+#	include "duffle/gen/duffle.offsets.h"
+#	include "duffle/atom_dsl.h"
+#	include "duffle/lottes_tape.h"
+#	include "duffle/mips.h"
+#	include "duffle/gte.h"
+#	include "duffle/gp.h"
+#	include "duffle/pad.h"
+#	include "duffle/word_count.metadata.h"
+#	include "psyq.h"
+#	include "gen/hello_joypad.offsets.h"
+#	include "gen/hello_joypad.macs.h"
+#	include "hello_joypad.h"
+#endif
+
+#pragma region MACs (Mips Atom components)
+
+FI_ Slice_MipsCode ac_load_v2s2(U4 rs_x, U4 rs_y, U4 r_base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_load_v2s2, {
+	load_half( rs_x, r_base, O_(V3_S2,x)),
+	load_half( rs_y, r_base, O_(V3_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_v2s2(U4 rt_x, U4 rt_y, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_v2s2, {
+	store_half(rt_x, base, offset + O_(V2_S2,x)),
+	store_half(rt_y, base, offset + O_(V2_S2,y)),
+})
+
+FI_ Slice_MipsCode ac_store_rects2(U4 rt_x, U4 rt_y, U4 rt_width, U4 rt_height, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rects2, {
+	store_half(rt_x,      base, offset + O_(Rect_S2,x)),
+	store_half(rt_y,      base, offset + O_(Rect_S2,y)),
+	store_half(rt_width,  base, offset + O_(Rect_S2,width)),
+	store_half(rt_height, base, offset + O_(Rect_S2,height)),
+})
+
+FI_ Slice_MipsCode ac_store_rgb8(U1 rr, U1 rg, U1 rb, U4 base, U4 offset) atom_dbg_skip MipsAtomComp_Proc_(ac_store_rgb8, {
+	store_byte(rr, base, offset + O_(DrawEnv,initial_bg_color.r)),
+	store_byte(rg, base, offset + O_(DrawEnv,initial_bg_color.g)),
+	store_byte(rb, base, offset + O_(DrawEnv,initial_bg_color.b)),	
+})
+
+FI_ Slice_MipsCode ac_gcmd_push(U4 cmd, U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_gcmd_push, {
+	load_upper_i(reg_transfer, cmd >> 16),
+	or_i_self(   reg_transfer, cmd & 0xFFFF),
+	store_word(  reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_put_disp_env(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_disp_env, {
+	// Emits 5 GP0 commands for buffer 0 (display_area = (0,0,320,240)).
+	// Sequence per libpsyx PutDispEnv: DrawArea TL → DrawArea BR → Mask → DrawArea TL → DrawArea BR
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_mask_bit(),                 reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_top_left_origin,      reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240, reg_transfer, reg_base, port),
+})
+
+FI_ Slice_MipsCode ac_put_draw_env(U4 reg_transfer, U4 reg_base, U2 port)
+MipsAtomComp_Proc_(ac_put_draw_env, {
+	/*
+	* ORIGIN: each code word corresponds to the EXACT value libpsyx's PutDrawEnv function would compute for the same DrawEnv settings.
+	* References:
+	*   - libpsyx source: `toolchain/psyq-4_7/lib/libgpu.a` (binary, function `PutDrawEnv`)
+	*   - PSX-SPX doc:     https://problemkaputt.de/psx-spx.htm#gputdrawingcommands
+	*   - PSYQ SDK:        `setdrawenv` / `makelongdr_env` source
+	*   - NOCASH PSX spec: §"GP0(E1h) Draw Mode setting" through §"DR_ENV"
+	*
+	* The 16-word format is documented in the PSYQ SDK manual and on NOCASH's PSX-spec.txt. The libpsyx reference is at:
+	*   ./toolchain/psyq-4_7/lib/libgpu.a
+	*   (binary; the PutDrawEnv implementation builds the 16-word DR_ENV from the user's DRAWENV struct and emits it via GP0 GPU commands.)
+	*
+	* Word indices (libpsyx PutDrawEnv / SetDrawEnv order):
+	*   tag      = (length << 24) | addr                — 16-word packet (1 tag + 15 code)
+	*   code[0]  = DrawMode (dfe=1, dtd=0, tpage=0)     — must come first per libpsyx
+	*   code[1]  = TextureWindow (tw=(0,0))             — bare-cmd word; GPU uses current state
+	*   code[2]  = DrawArea top-left (clip.x=0, clip.y=240)
+	*   code[3]  = DrawArea bottom-right (clip.x+w=320, clip.y+h=480)
+	*   code[4]  = DrawOffset (ofs=(0,0))               — bare-cmd word
+	*   code[5]  = Mask (dtd=0, dfe=1, isbg=1)          — 0xE6 cmd + isbg bit
+	*   code[6]  = Initial-bg-color (isbg=1, r=7, g=7, b=7)
+	*   code[7]  = DrawMode (isbg=1, tpage=0)           — re-asserts DrawMode with isbg
+	*   code[8..10]  = padding (NOP)                    — 3 words to fill the packet
+	*   code[11..12] = TextureWindow bottom-right       — defaults to (0,0,0,0)
+	*   code[13..14] = padding (NOP)                    — completes the 16-word packet
+	*/
+	mac_gcmd_push(gp0_dr_env_tag,                            reg_transfer, reg_base, port), /* tag (length=15 << 24, addr=0) — packet header for the DR_ENV sequence. The GPU needs this to recognize the next 15 words as a DR_ENV packet and trigger the isbg auto-clear. */
+	mac_gcmd_push(gp0_word_draw_mode_drawing_allowed,        reg_transfer, reg_base, port), /* code[0]  DrawMode (dfe=1, dtd=0, tpage=0) */
+	mac_gcmd_push(gp0_word_set_texture_window(),             reg_transfer, reg_base, port), /* code[1]  TextureWindow (tw=(0,0)) */
+	mac_gcmd_push(enc_gp0_draw_area_tl_word(0, ScreenRes_Y), reg_transfer, reg_base, port), /* code[2]  DrawArea top-left (clip.x=0, clip.y=ScreenRes_Y=240) */
+	mac_gcmd_push(gp0_word_draw_area_bottom_right_320x240,   reg_transfer, reg_base, port), /* code[3]  DrawArea bottom-right (clip.x+w=320, clip.y+h=480) */
+
+	mac_gcmd_push(gp0_word_set_draw_offset(),               reg_transfer, reg_base, port), /* code[4]  DrawOffset (ofs=(0,0)) — bare-cmd word; the GPU uses the current state machine. */
+	mac_gcmd_push(gp0_word_dr_env_mask(),                   reg_transfer, reg_base, port), /* code[5]  Mask (dtd=0, dfe=1, isbg=1) — 0xE6 cmd + isbg bit. */
+	mac_gcmd_push(gp0_word_dr_env_bg_color_cmd(1, 7, 7, 7), reg_transfer, reg_base, port), /* code[6]  Initial-bg-color + auto-clear (isbg=1, r=7, g=7, b=7). */
+	mac_gcmd_push(gp0_word_dr_env_draw_mode(1),             reg_transfer, reg_base, port), /* code[7]  Re-assert DrawMode with isbg=1 (isbg-flag set; the 0xE1 cmd byte plus isbg only). */
+
+	/* code[8..10]  Padding (NOP — GPU discards; the DR_ENV requires 16 words total). */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+
+	/* code[11..12]  TextureWindow bottom-right (tw.x+tw.w=0, tw.y+tw.h=0) — libpsyx emits twice. */
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_set_texture_window(), reg_transfer, reg_base, port),
+
+	/* code[13..14]  Padding (NOP) — completes the 16-word packet. */
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port),
+})
+
+#pragma endregion MACs
+
+#pragma region Baked Atoms
+
+/* DIAGNOSTIC 1: Pure tape loop test */
+internal MipsAtom_(diag_yield) { mac_yield() };
+
+/* DIAGNOSTIC 2: Pure memory test (No GTE). Draws a fixed cyan triangle. */
+internal MipsAtom_(diag_color) {
+	store_word(  R_0, R_T7, 0), 
+	load_upper_i(R_AT, gp0_cmd_poly_f3 << 8 | 0xFF), /* High: MipsCode Poly_F3(0x20) + Color B:FF */
+	or_i_self(   R_AT, 0xFF00),                      /* Low:  Color G:FF, R:00 (Cyan) */
+	store_word(  R_AT, R_T7, 4),
+	
+	/* Fake coordinates - Swapped winding order to prevent GPU culling! */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 8),  /* (16, 16) */
+	load_upper_i(R_AT, 0x0050), or_i_self(R_AT, 0x0010), store_word(R_AT, R_T7, 12), /* (80, 16) */
+	load_upper_i(R_AT, 0x0010), or_i_self(R_AT, 0x0050), store_word(R_AT, R_T7, 16), /* (16, 80) */
+
+	add_ui(          R_T1, R_0,  10),
+	shift_lleft_self(R_T1,       S_(U4)/2), 
+	add_u_self(      R_T1,       R_T6),         
+	
+	load_word(   R_AT, R_T1, 0),        
+	load_upper_i(R_V0, (S_(Poly_F3)/S_(U4) - S_(PolyTag)/S_(U4)) << PolyTag_len_bits),
+	store_word(  R_AT, R_T7, 0),       
+	shift_lleft(R_AT, R_T7, S_(PolyTag_len_bits)), shift_lright(R_AT, R_AT, S_(PolyTag_len_bits)),         
+	or_u_self(  R_AT, R_V0),          
+	store_word( R_AT, R_T1, 0),       
+
+	add_ui(R_T7, R_T7, 20),          
+
+	mac_yield()
+};
+
+/* DIAGNOSTIC 3: Pure GTE test (No Memory Writes) */
+internal MipsAtom_(diag_gte) {
+	/* Load 3 indices */
+	load_half_u(R_T0, R_T4, 0),
+	load_half_u(R_T1, R_T4, 2),
+	load_half_u(R_T2, R_T4, 4),
+
+	/* Load Vertices into GTE */
+	shift_lleft( R_AT, R_T0, 3), add_u(    R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY0), gte_mv_to_data_r(R_V1, C2_VZ0),
+
+	shift_lleft( R_AT, R_T1, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY1), gte_mv_to_data_r(R_V1, C2_VZ1),
+
+	shift_lleft(R_AT, R_T2, 3), add_u(R_AT, R_AT, R_T5),
+	load_word(R_V0, R_AT, 0), load_word(R_V1, R_AT, 4),
+	gte_mv_to_data_r(R_V0, C2_VXY2), gte_mv_to_data_r(R_V1, C2_VZ2),
+
+	/* Run Math */
+	nop2, gte_cmdw_rtpt,
+	nop2, gte_cmdw_nclip,
+	nop2,
+
+	/* Advance Face Cursor and Yield */
+	add_ui(R_T4, R_T4, 8),
+
+	mac_yield()
+};
+
+enum {
+	R_ScreenX   = R_T5 atom_reg atom_type(U2),
+	R_ScreenY   = R_T6 atom_reg atom_type(U2),
+	R_ScreenBuf = R_T7 atom_reg, /* Caller-pinned: & smem.screen_buf */
+#define R_ScreenBuf_Code R_T7_Code
+};
+//screen_env_init. Mirrors the libpsyx's SetDefDispEnv + SetDefDrawEnv + the manual enable_auto_clear / initial_bg_color writes.
+internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
+,	atom_reads(R_T0, R_ScreenX, R_ScreenY, R_ScreenBuf)
+, atom_writes(R_T0, R_ScreenX, R_ScreenY)
+) {
+	/* display[0] = (0, 0, 320, 240); rest of struct zeroed. */
+	add_ui(R_ScreenX, R_0, ScreenRes_X), add_ui(R_ScreenY, R_0, ScreenRes_Y),
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area.width) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)       + OA_(DoubleBuffer,display,0)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace)   + OA_(DoubleBuffer,display,0)),
+
+	/* display[1] = (0, 240, 320, 240); rest of struct zeroed. */
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DisplayEnv,display_area) + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,screen)     + OA_(DoubleBuffer,display,1)),
+	store_word(R_0, R_ScreenBuf, O_(DisplayEnv,vinterlace) + OA_(DoubleBuffer,display,1)),
+
+	mac_store_rects2(R_0, R_ScreenY, R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area)         + OA_(DoubleBuffer,draw,0)), /* draw[0].clip_area = (0, 240, 320, 240). C11's SetDefDrawEnv writes clip.y = y_arg. */
+	mac_store_v2s2(  R_0, R_ScreenY,                       R_ScreenBuf, O_(DrawEnv,drawing_offset[0]) + OA_(DoubleBuffer,draw,0)), /* draw[0].drawing_offset[0] = (0, 240); C11 passes y_arg as ofs. */
+
+	mac_store_v2s2(R_ScreenX, R_ScreenY, R_ScreenBuf, O_(DrawEnv,clip_area.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_window = (0, 0, 0, 0); two word-zeroes cover the full 8-byte tw field. */
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,0)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,0)),
+
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,drawing_offset[0].x)  + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.x)     + OA_(DoubleBuffer,draw,1)),
+	store_word(R_0, R_ScreenBuf, O_(DrawEnv,texture_window.width) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].texture_page = 10 (gp0_tpage_default). C11 SetDefDrawEnv at C11_only.elf:0x8001273C writes the same 0x0A. . */
+	add_ui(R_T0, R_0, gp0_tpage_default), 
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,0)),
+	store_half(R_T0, R_ScreenBuf, O_(DrawEnv,texture_page) + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0] control bytes: flag_dither=1, flag_draw_on_display=1 (the dfe bit per psx-spx; libpsyx sets it via `SetDefDrawEnv`'s conditional at C11_only.elf:0x80012728), enable_auto_clear=1. Each byte is named;
+	 * the previous `store_word(R_0, ..., +20)` overwrote all four with zero. */
+	add_ui(R_T0, R_0, 1),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,0)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_dither)          + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,flag_draw_on_display) + OA_(DoubleBuffer,draw,1)),
+	store_byte(R_T0, R_ScreenBuf, O_(DrawEnv,enable_auto_clear)    + OA_(DoubleBuffer,draw,1)),
+
+	/* draw[0].initial_bg_color = (r=7, g=7, b=7). */
+	add_ui(R_T0, R_0, 7), 
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,0)),
+	mac_store_rgb8(R_T0,R_T0,R_T0, R_ScreenBuf, O_(DrawEnv,initial_bg_color) + OA_(DoubleBuffer,draw,1)),
+
+	mac_yield(),
+};
+
+enum {
+	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+#define R_IO_BaseAddr_Code R_T4_Code
+};
+internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads(R_IO_BaseAddr)) {
+	store_word(R_0, R_IO_BaseAddr, GPIO_PORT1_OFFSET),                                   /* GP1(00h) Reset */
+	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(01h) ClearFIFO */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(02h) AckIRQ */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(03h) Display ON */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(04h) DMADirection=2 (CPU→GPU). libpsyx's per-frame PutDrawEnv/DrawOTag use DMA2; without this the DMA queue never drains. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),  /* GP1(05h) StartDisplayArea (X=0, Y=0) */
+
+	/* GP1: DisplayMode + Display Ranges */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+
+	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
+	load_upper_i(R_T5, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_T5, gte_cr_OFX_Code),
+	load_upper_i(R_T5, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_T5, gte_cr_OFY_Code),
+
+	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
+	add_ui(R_T5, R_0, ScreenZ), gte_mv_to_ctrl_r(R_T5, gte_cr_H_Code),
+
+	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
+	mac_gcmd_push(gp1_word_DisplayOn(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_yield(),
+};
+
+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_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // required cpu -> gte delay slot
+	gte_cmdw_nclip,
+
+	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
+	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)), nop,
+		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
+		shift_lleft(R_AT, R_T3, v3s2_byteoff), add_u(R_AT, R_AT, R_VertBase),
+		load_word(R_V0, R_AT, O_(V3_S2, x)),   load_word(R_V1, R_AT, O_(V3_S2, z)),
+		gte_mv_to_data_r(R_V0, C2_VXY0),       gte_mv_to_data_r(R_V1, C2_VZ0),
+
+		mac_gte_store_g4_p012(),
+		gte_cmdw_rotate_translate_perspective_single,
+		mac_gte_store_g4_p3(),
+
+		gte_cmdw_avg_sort_z4,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		add_ui(      R_AT, R_0,  OrderingTbl_Len),
+		set_lt_u(    R_AT, R_T1, R_AT),
+
+		branch_equal(R_AT, R_0,  atom_offset(bounds_chk, cube_g4_face_exit)), nop,
+			mac_insert_ot_tag_g4(),
+			mac_format_g4_color(
+				/* c0 magenta */ 0xFF, 0x00, 0xFF,
+				/* c1 yellow  */ 0xFF, 0xFF, 0x00,
+				/* c2 cyan    */ 0x00, 0xFF, 0xFF,
+				/* c3 green   */ 0x00, 0xFF, 0x00),
+		// end: branch(bounds_chk)
+// end: branch(cull)
+
+atom_label(cube_g4_face_exit)
+	add_ui_self(R_PrimCursor, S_(Poly_G4)),     /* 9 words = Poly_G4 */
+	add_ui_self(R_FaceCursor, S_(S2) * 4),      /* 4 × S2 = 8 bytes */
+	mac_yield()
+};
+
+typedef Struct_(Binds_FloorTri) {
+	U4     PrimCursor;
+	V3_S2* FaceCursor;
+	V3_S2* VertBase;
+	U4*    OtBase;
+};
+internal
+MipsAtom_(rbind_floor_f3_face) atom_info(atom_bind(Binds_FloorTri), atom_phase(floor_f3)
+	, atom_reads(R_TapePtr)
+	, atom_writes(R_PrimCursor, R_FaceCursor, R_VertBase, R_OtBase, 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_T0, R_T1, R_T2),
+	mac_gte_load_tri_verts(R_T0, R_T1, R_T2),
+	nop2, gte_cmdw_rotate_translate_perspective_triple, // 2 nops retire the final cpu -> gte writes before RTPT
+	gte_cmdw_nclip,
+
+	/* Culling (Branch forward if Backface) */
+	gte_mv_from_data_r(R_T0, C2_MAC0),
+	nop, branch_le_zero(R_T0, atom_offset(culling, floor_f3_face_exit)), nop, // required gte -> cpu load-delay slot. 
+		/* Format Primitive */
+		mac_gte_store_f3(),
+
+		/* Calculate Depth */
+		gte_avg_sort_z3,
+		gte_mv_from_data_r(R_T1, C2_OTZ),
+		/* Bounds Check OTZ < 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(0xFF, 0xFF, 0xFF),  // RGB-form (R=FF, G=FF, B=FF = white)
+			mac_insert_ot_tag_f3(),                 /* Insert into Ordering Table Linked List */
+			add_ui_self(R_PrimCursor, S_(Poly_F3)), /* Advance Prim Cursor (5 words) */
+				// Note(Ed): No bounds checking, should be checked before atom runs.
+		// end: branch(bounds_chk)
+	// end: branch(culling)
+
+/* Advance Input Cursor & Yield (Both branch targets land here) */
+atom_label(floor_f3_face_exit)
+	add_ui_self(R_FaceCursor, S_(S2) * 4),  /* Advance Face Cursor (4 * S2 = 8 bytes) */
+	mac_yield()
+};
+
+typedef Struct_(Binds_SyncPrimitiveArena) { U4 used; U4 cursor; };
+internal MipsAtom_(sync_primitive_arena) atom_info(atom_bind(Binds_SyncPrimitiveArena)
+	, atom_reads( R_TapePtr, R_PrimCursor)
+	, atom_writes(R_TapePtr)
+){
+	load_word(R_AT, R_TapePtr, O_(Binds_SyncPrimitiveArena,used)),
+	load_word(R_T0, R_TapePtr, O_(Binds_SyncPrimitiveArena,cursor)),
+	add_ui_self(    R_TapePtr, S_(Binds_SyncPrimitiveArena)),
+	/* Calculate byte offset and store directly back to RAM */
+	sub_u(     R_T0, R_PrimCursor, R_T0), // R_T0    = R_PrimCursor - binds.cursor
+	store_word(R_T0, R_AT, 0),            // R_AT[0] = R_T0
+	mac_yield()
+};
+
+/* ----- pad_bios_snapshot -----
+ * Per-frame snapshot of one BIOS pad buffer into PadState.
+ * Decoder (branch ladder on raw[0] status + raw[1] id):
+ *   1. raw[0] == 0xFF         -> Disconnected (buttons=0, axes=0x80)
+ *   2. raw[0]==0 && raw[1]==0 -> Pending      (buttons=0, axes=0x80)
+ *   3. raw[1] == 0x41         -> Digital      (buttons normalized; axes=0x80)
+ *   4. raw[1] == 0x53         -> AnalogStick  (buttons normalized; axes from raw[4..7])
+ *   5. raw[1] in 0x7x         -> AnalogPad    (buttons normalized; axes from raw[4..7])
+ *   6. else                   -> Unsupported  (buttons=0, axes=0x80)
+ *
+ * Buttons normalization: byte_swap16((~raw_buttons) & 0xFFFF).
+ *   raw_buttons    = load_half_u(raw, 2) = raw[2] | (raw[3] << 8).
+ *   byte_swap16(x) = (x >> 8) | (x << 8); nor(x, R_0) = ~x. store_half truncates to 16 bits so the upper-16 mask is implicit in the store.
+ *
+ * Register use (atom-local; no wave-context touched):
+ *   R_T0 = raw base      (kept throughout; axes loads read raw[4..7] from R_T0)
+ *   R_T1 = state base    (kept throughout; all stores go through R_T1)
+ *   R_T2 = raw[0] status (alive across the disc/pending/id dispatch, then dead)
+ *   R_T3 = raw[1] id     (alive across the id dispatch, then dead)
+ *   R_T4 = scratch       (shifts, compares, immediate loads, store values)
+ *   R_T5 = scratch       (parallel lui+ori for the 0x80808080 axes constant + byte-swap target)
+ */
+enum {
+	R_PadRaw    = R_T0 atom_reg atom_type(U1),
+	R_PadState  = R_T1 atom_reg,
+	R_RawStatus = R_T2 atom_reg,
+	R_RawId     = R_T3 atom_reg,
+};
+typedef Struct_(Binds_PadBiosSnapshot) {
+    PadBiosRaw* raw;
+    PadState*   state;
+};
+internal MipsAtom_(pad_bios_snapshot) atom_info(atom_bind(Binds_PadBiosSnapshot)
+, atom_reads( R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+, atom_writes(R_PadRaw, R_PadState, R_RawStatus, R_RawId, R_T4, R_T5, R_TapePtr)
+) {
+	/* === Bind consumption: T0 = raw, T1 = state, advance R_TapePtr by 8. */
+	load_word(R_PadRaw,   R_TapePtr, O_(Binds_PadBiosSnapshot,raw)),
+	load_word(R_PadState, R_TapePtr, O_(Binds_PadBiosSnapshot,state)),
+	add_ui_self(          R_TapePtr, S_(Binds_PadBiosSnapshot)),
+
+	/* === Read raw[0] (status) + raw[1] (id) */
+	load_byte_u(R_RawStatus, R_PadRaw, 0),
+	load_byte_u(R_RawId,     R_PadRaw, 1),
+
+atom_label(snap_root) /* === Case 1: Disconnected (status == 0xFF). */
+	add_ui(R_T4, R_0, 0xFF), branch_ne(R_RawStatus, R_T4, atom_offset(snap_root, skip_disconnected)),
+	/* BD-slot: pre-compute PadStatus_Disconnected. Branch reads R_T4=0xFF in EX before this WB completes.
+	 * If branch NOT taken (fall through to pending/id_dispatch), R_T4 is overwritten by the next case body's add_ui — harmless. */
+
+atom_label(disconnected) /* === Disconnected body. */
+	/* R_T4 = PadStatus_Disconnected from snap_root BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	branch_equal(R_0, R_0, atom_offset(disconnected, snap_end)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(disconnected, snap_end)), nop,
+atom_label(skip_disconnected)
+
+	/* === Case 2: Pending (status == 0 && id == 0)
+	 * Combined check: if (status | id) != 0 then skip to id_dispatch.
+	 * Falls through to the Pending case only when both are zero. */
+	or_u_self(R_RawStatus, R_RawId), branch_ne(R_RawStatus, R_0, atom_offset(case_2, id_dispatch)),
+	/* BD-slot: pre-compute PadStatus_Pending. Branch reads R_RawStatus in EX before this WB completes.
+	 * If branch NOT taken (fall through to id_dispatch), R_T4 is overwritten by the digital/analog body add_ui — harmless. */
+
+atom_label(pending) /* === Pending body */
+	/* R_T4 = PadStatus_Pending from case_2 BD-slot. */
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4,    R_PadState, O_(PadState,left_x)),
+	store_byte(  R_RawId, R_PadState, O_(PadState,id)),
+	branch_equal(R_0, R_0, atom_offset(pending, snap_end)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(pending, snap_end)), nop,
+
+atom_label(id_dispatch) /* === Case 3-6: ID dispatch */
+	add_ui(R_T4, R_0, 0x41), branch_ne(R_RawId, R_T4, atom_offset(id_dispatch, try_analog_stick)),
+	/* BD-slot: pre-compute PadStatus_Digital. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_stick), R_T4 is overwritten by the analog body add_ui. */
+
+	/* === Digital body (status, buttons normalize, axes=0x80, id, branch. */
+	/* R_T4 = PadStatus_Digital from id_dispatch BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),
+	/* Fill R_T4's load-delay slot with the 0x80808080 axes constant into R_T5
+	 * (R_T5 is dead on this path; it's only consumed at the analog_pad range check). */
+	load_upper_i(R_T5, 0x8080), or_i_self(R_T5, 0x8080),
+	nor_u(      R_T4, R_T4, R_0), /* raw_buttons is already in host bit order; no swap needed */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	store_word(  R_T5, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0x41),
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+
+	branch_equal(R_0, R_0, atom_offset(id_dispatch, snap_end)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(id_dispatch, snap_end)), nop,
+
+atom_label(try_analog_stick) /* === Case 4: AnalogStick (id == 0x53)*/
+	add_ui(R_T4, R_0, 0x53), branch_ne(R_RawId, R_T4, atom_offset(try_analog_stick, try_analog_pad)),
+	/* BD-slot: pre-compute PadStatus_AnalogStick. Branch reads R_RawId in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_analog_pad), R_T4 is overwritten by the analog_pad body add_ui. */
+
+atom_label(analog_stick) /* === AnalogStick body
+	* Axes are loaded as two halfwords: raw[6..7] → left_xy (sh at offset 8), raw[4..5] → right_xy (sh at offset 10).
+	* R_T5 holds left_xy / id-value in turn (it's dead on this path — only consumed at the analog_pad range check). */
+	/* R_T4 = PadStatus_AnalogStick from try_analog_stick BD-slot. */
+	store_word(  R_T4, R_PadState, O_(PadState,status)),
+	load_half_u( R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u( R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot (doesn't read R_T4) */
+	nor_u(       R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half(  R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u( R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half(  R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half(  R_T4, R_PadState, O_(PadState,right_x)),
+	add_ui(      R_T5, R_0, 0x53),                        /* R_T5 = id value (clobbers left_xy, already stored) */
+	store_byte(  R_T5, R_PadState, O_(PadState,id)),
+	branch_equal(R_0, R_0, atom_offset(analog_stick, snap_end)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(analog_stick, snap_end)), nop,
+
+atom_label(try_analog_pad) /* === Case 5-6: AnalogPad (id & 0xF0 == 0x70) */
+	and_i(    R_T4, R_RawId, 0xF0),
+	add_ui(   R_T5, R_0,     0x70),
+	branch_ne(R_T4, R_T5, atom_offset(try_analog_pad, try_unsupported)),
+	/* BD-slot: pre-compute PadStatus_AnalogPad. Branch reads R_T4 in EX before this WB completes.
+	 * If branch NOT taken (fall through to try_unsupported), R_T4 is overwritten by the unsupported body add_ui. */
+
+atom_label(analog_pad) /* === AnalogPad body
+	* Same shape as AnalogStick with AnalogPad status. R_T5 holds left_xy (it's dead on this path). */
+	/* R_T4 = PadStatus_AnalogPad from try_analog_pad BD-slot. */
+	store_word( R_T4, R_PadState, O_(PadState,status)),
+	load_half_u(R_T4, R_PadRaw,   2 * S_(U1)),           /* R_T4 = raw_buttons */
+	load_half_u(R_T5, R_PadRaw,   6 * S_(U1)),           /* R_T5 = left_xy; fills R_T4's load-delay slot */
+	nor_u(      R_T4, R_T4, R_0),                        /* R_T4 = ~raw_buttons */
+	store_half( R_T4, R_PadState, O_(PadState,buttons)),
+	load_half_u(R_T4, R_PadRaw,   4 * S_(U1)),           /* R_T4 = right_xy; fills R_T5's load-delay slot */
+	store_half( R_T5, R_PadState, O_(PadState,left_x)),  /* R_T5 settled, store left_xy */
+	store_half( R_T4, R_PadState, O_(PadState,right_x)),
+	store_byte( R_RawId, R_PadState, O_(PadState,id)),
+
+	branch_equal(R_0, R_0, atom_offset(analog_pad, snap_end)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(analog_pad, snap_end)), nop,
+
+atom_label(try_unsupported) /* === Case 7: Unsupported — fall through from the AnalogPad range-check miss. */
+	add_ui(    R_T4, R_0, PadStatus_Unsupported),
+	store_word(R_T4, R_PadState, O_(PadState,status)),
+	store_half(R_0,  R_PadState, O_(PadState,buttons)),
+	/* axes = 0x80808080 (centered) — single sw writes the 4-byte axes block at offset 8 (left_x, left_y, right_x, right_y). */
+	load_upper_i(R_T4, 0x8080), or_i_self(R_T4, 0x8080),
+	store_word(  R_T4, R_PadState, O_(PadState,left_x)),
+	add_ui(      R_T4, R_0, 0xFF),                      /* 0xFF sentinel: "unknown id" */
+	store_byte(  R_T4, R_PadState, O_(PadState,id)),
+	/* Fall through to snap_end. */
+
+atom_label(snap_end)
+	mac_yield(),
+};
+
+/* ----- pad_apply_input -----
+ * Reads pad[0].buttons + pad[0].left_x; 
+ * Applies the input-semantics deltas to cube_rot.y + floor_rot.y:
+ *   - D-pad Left:  cube_rot.y += 30, floor_rot.y += 5
+ *   - D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5
+ *   - Analog stick X (dead zone 0x70..0x90):
+ *       cube delta  = (0x80 - left_x) >> 2  (range approx -32..+32)
+ *       floor delta = (0x80 - left_x) >> 5  (range approx -4..+4)
+ *   - D-pad + analog deltas add when used together.
+ *
+ * Convention:
+ *   pad_state = 0 means no buttons active. 
+ *   The fail-safe zero-button value flows through unchanged, so a disconnected/fresh pad produces no rotation.
+ *   The branch_le_zero pattern below matches the existing pad_input_demo convention (atom body lines 248/257).
+ *
+ * Signed-delta trick:
+ * load_byte_u zero-extends left_x to 32 bits; sub_u from 0x80 wraps to a SIGNED two's-complement value in the negative range;
+ * shift_aright (sra) then correctly sign-extends the shift for both positive (left_x < 0x80) and negative (left_x > 0x80) cases.
+ * Digital pads publish left_x = 0x80 → delta = 0 → no rotation, so the analog step is naturally a no-op for digital controllers.
+ */
+typedef Struct_(Binds_PadApplyInput) {
+    PadState* state;
+    V3_S2*    cube_rot;
+    V3_S2*    floor_rot;
+};
+enum {
+	R_PadStateT5 = R_T5 atom_reg,
+	R_CubeRot    = R_T1 atom_reg,
+	R_FloorRot   = R_T2 atom_reg,
+};
+internal MipsAtom_(pad_apply_input) atom_info(atom_bind(Binds_PadApplyInput)
+, atom_reads(R_T0, R_CubeRot, R_FloorRot, R_T3, R_T4, R_PadStateT5, R_TapePtr)
+, atom_writes(     R_CubeRot, R_FloorRot)
+) {
+	/* Pop Binds from tape (state, cube_rot, floor_rot) */
+	load_word(R_PadStateT5, R_TapePtr, O_(Binds_PadApplyInput,state)),
+	load_word(R_CubeRot,    R_TapePtr, O_(Binds_PadApplyInput,cube_rot)),
+	load_word(R_FloorRot,   R_TapePtr, O_(Binds_PadApplyInput,floor_rot)),
+	add_ui_self(            R_TapePtr, S_(Binds_PadApplyInput)),
+
+	/* Load pad[0].buttons into R_T0. */
+	load_word(R_T0, R_PadStateT5, O_(PadState,buttons)), nop,
+	// Note(Ed): Potential op with delay slot?
+
+	/* D-pad Left: cube_rot.y += 30, floor_rot.y += 5.  */
+	and_i(R_T3, R_T0, pad0_(Pad_Left)), branch_le_zero(R_T3, atom_offset(dpad_left, exit_dpad_left)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, 30),
+		add_si(    R_T3, R_T3, 5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_left)
+
+	/* D-pad Right: cube_rot.y -= 30, floor_rot.y -= 5. */
+	and_i(R_T3, R_T0, pad0_(Pad_Right)), branch_le_zero(R_T3, atom_offset(dpad_right, exit_dpad_right)),
+		load_half( R_T4, R_CubeRot,  O_(V3_S2,y)),   /* BD-slot */
+		load_half( R_T3, R_FloorRot, O_(V3_S2,y)),
+		add_si(    R_T4, R_T4, -30),
+		add_si(    R_T3, R_T3, -5),
+		store_half(R_T4, R_CubeRot,  O_(V3_S2,y)),
+		store_half(R_T3, R_FloorRot, O_(V3_S2,y)),
+	atom_label(exit_dpad_right)
+
+	/* Analog left-stick X: dead zone 0x70..0x90.
+	 * Cube delta = (0x80 - left_x) >> 2; floor delta = (0x80 - left_x) >> 5. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)),
+
+	/* Dead-zone check: skip analog if left_x in [0x70, 0x90] inclusive. Outside dead zone on LOW side: left_x < 0x70 (strictly).
+	 * set_lt_u(R_T4, R_T3, R_T4=0x70) → R_T4 = (left_x < 0x70) ? 1 : 0. */
+	add_ui(R_T4, R_0, 0x70), set_lt_u(R_T4, R_T3, R_T4), branch_ne(R_T4, R_0, atom_offset(dead_zone_low_check, dead_low_active)),
+	add_ui(R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_low_active */
+
+atom_label(dead_check_upper)
+	/* left_x >= 0x70 → check upper bound. */
+	load_byte_u(R_T3, R_PadStateT5, O_(PadState,left_x)), /* reload */
+	add_ui(     R_T4, R_0, 0x90),
+
+	/* R_T4 = (0x90 < left_x) ? 1 : 0 → (left_x > 0x90) ? 1 : 0 */
+	set_lt_u(R_T4, R_T4, R_T3), branch_ne(R_T4, R_0, atom_offset(dead_zone_high_check, dead_high_active)),
+	add_ui(  R_T4, R_0, 0x80), /* BD-slot: pre-load 0x80 for dead_high_active */
+	branch_equal(R_0, R_0, atom_offset(dead_zone_skip, exit_stick)), nop,
+	/* Fall-through = left_x in [0x70, 0x90] (dead zone); skip analog entirely. */
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(dead_zone_skip, exit_stick)), nop,
+
+atom_label(dead_low_active)
+	/* R_T3 = left_x (from line 632 lbu; not clobbered between dead_zone_low_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * 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)),
+
+	branch_equal(R_0, R_0, atom_offset(end_low, exit_stick)), nop,
+	// TODO(Ed): Lua metaprogram: Support jump instruction here..
+	// jump(atom_offset(end_low, exit_stick)), nop,
+
+atom_label(dead_high_active)
+	/* R_T3 = left_x (from line 641 lbu in dead_check_upper; not clobbered between dead_zone_high_check branch + its BD-slot `add_ui R_T4, 0x80`).
+	 * 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(exit_stick)
+	mac_yield(),
+};
+
+#pragma endregion Baked Atoms

code/hello_joypad/psyq.c

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

code/hello_joypad/psyq.h

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

code/hello_joypad/sio.tape.c

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

readme.md

@@ -7,12 +7,12 @@ A rest from the usual.
 ## Dependencies
 
 I will be programming from a Windows 11 machine (may eventually try this on the Steam Deck...):
-![system_info](./docs/assets/system_info.png)
 
 [armips](https://github.com/Kingcom/armips)
 
 * Supports doing bare-metal assembly for the ps1
 * `scoop install armips` or just clone and build..
+* Was used early in the course. Now I just use an macro asm dsl in C11.
 
 [luajit-2.1](https://github.com/LuaJIT/LuaJIT.git)
 
@@ -27,6 +27,13 @@ scoop install luajit
 
 * Lua is slow (even jitted) so this helps.
 
+[lfs (LuaFileSystem)](https://github.com/lunarmodules/luafilesystem)
+
+* Native directory enumeration + `mkdir` for the build scripts.
+* Used by `passes/word_count_eval.lua :: scan_dir` (native walk vs. `dir /b /s` subprocess,
+  ~2ms vs. ~56ms) and by `duffle.lua :: ensure_dir` + `to_absolute_path` (avoids
+  `cmd.exe mkdir` + `cd` shell spawns, ~50ms each).
+
 [pscx-redux](https://github.com/grumpycoders/pcsx-redux/): A collection of tools, research, hardware design, and libraries aiming at development and reverse engineering on the PlayStation 1.
 
 * Used as the runtime sandbox emulated the ps1
@@ -66,3 +73,9 @@ scoop install luajit
 ![hello_psyq!](./docs/assets/pcsx-redux_2025-08-05_23-01-19.png)
 ![cube!](./docs/assets/pcsx-redux_2025-10-11_03-04-01.png)
 ![cube and floor!](./docs/assets/pcsx-redux_2026-07-10_22-47-02.png)
+
+Win 11 machine:
+
+![system_info](./docs/assets/system_info.png)
+
+Still haven't gotten around to trying this on linux...

scratch/gte_debug.lua

@@ -0,0 +1,30 @@
+-- gte_debug.lua — defensive version + prints error context.
+local ok, err = pcall(function()
+    print("[debug] PCSX exists:", PCSX ~= nil)
+    print("[debug] PCSX.WebServer exists:", PCSX and PCSX.WebServer ~= nil)
+    print("[debug] PCSX.WebServer.Handlers exists:", PCSX and PCSX.WebServer and PCSX.WebServer.Handlers ~= nil)
+    if not PCSX.WebServer then
+        print("[debug] creating PCSX.WebServer...")
+        PCSX.WebServer = {}
+    end
+    if not PCSX.WebServer.Handlers then
+        print("[debug] creating PCSX.WebServer.Handlers...")
+        PCSX.WebServer.Handlers = {}
+    end
+    print("[debug] type of Handlers:", type(PCSX.WebServer.Handlers))
+
+    PCSX.WebServer.Handlers.gte = function(req)
+        local r = PCSX.getRegisters()
+        local out = { "pc=0x" .. string.format("%x", r.pc) }
+        for i = 0, 31 do
+            out[#out + 1] = string.format("D[%d]=0x%08x C[%d]=0x%08x",
+                i, r.CP2D.r[i], i, r.CP2C.r[i])
+        end
+        return table.concat(out, "\n")
+    end
+    print("[debug] handler registered")
+end)
+
+if not ok then
+    print("[debug] ERROR: " .. tostring(err))
+end

scripts/audit_lua_nesting.lua

@@ -168,28 +168,18 @@ if arg and arg[1] then
 	end
 
 	-- Accept either a directory or a file path. Directory args are
-	-- expanded via `dir /b *.lua` (Windows) or `ls *.lua` (Unix).
+	-- expanded via lfs.dir (native, no subprocess).
+	local lfs = require("lfs")
 	local function is_dir(p)
-		local f = io.open(p, "r")
-		if    f then f:close() return false end
-		return true
+		return lfs.attributes(p, "mode") == "directory"
 	end
 	local function list_lua(dir)
 		local out = {}
-		local cmd
-		if package.config:sub(1, 1) == "\\" then
-			cmd = 'dir /b "' .. dir .. '\\*.lua" 2>nul'
-		else
-			cmd = 'ls -1 "' .. dir .. '"/*.lua 2>/dev/null'
-		end
-		local p = io.popen(cmd)
-		if p then
-			for line in p:lines() do
-				if line:match("%.lua$") then
-					out[#out + 1] = dir .. "/" .. line
-				end
+		if not is_dir(dir) then return out end
+		for entry in lfs.dir(dir) do
+			if entry:match("%.lua$") then
+				out[#out + 1] = dir .. "/" .. entry
 			end
-			p:close()
 		end
 		return out
 	end

scripts/build_psyq.ps1

@@ -81,19 +81,6 @@ $path_psyq          = join-path $path_toolchain  'psyq-4_7'
 $path_psyq_iwyu     = join-path $path_toolchain  'psyq_iwyu'
 $path_psyq_imyu_inc = join-path $path_psyq_iwyu  'include'
 
-function Get-SourceFiles { param([Parameter(Mandatory=$true)] [string[]]$paths, [Parameter(Mandatory=$true)] [string[]]$extensions)
-	$files = @()
-	foreach ($p in $paths) {
-		if (-not (test-path $p)) { continue }
-		foreach ($ext in $extensions) {
-			Get-ChildItem -Path $p -File -Recurse -Filter "*$ext" -ErrorAction SilentlyContinue | ForEach-Object {
-				$files += $_.FullName
-			}
-		}
-	}
-	return ($files | Sort-Object -Unique)
-}
-
 function assemble-unit { param(
 	[string]  $unit,
 	[string]  $link_module,
@@ -153,7 +140,7 @@ function compile-unit { param(
 		$f_arch_no_shared,
 		$f_arch_no_stack_prot
 	)
-	# $compile_args    += $f_std_c23
+	$compile_args    += $f_std_c11
 	$compile_args    += ($f_include + $path_psyq_imyu_inc)
 	$compile_args    += ($f_include + $path_nugget)
 
@@ -193,29 +180,18 @@ 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.
+	# The 16 removed entries (c2, card, cd, comb, ds, gs, gun, hmd, math,
+	# mcrd, mcx, press, sio, snd, spu, tap) had LOAD lines in the map but
+	# ZERO .o files pulled in — they were unused. 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.).
 	$libraries = @(
 		"api",
 		"c",
-		"c2",
-		"card",
-		"cd",
-		"comb",
-		"ds",
 		"etc",
 		"gpu",
-		"gs",
-		"gte",
-		"gun",
-		"hmd",
-		"math",
-		"mcrd",
-		"mcx",
-		"pad",
-		"press",
-		"sio",
-		"snd",
-		"spu",
-		"tap"
+		"gte"
 	)
 	foreach ($lib in $libraries) {
 		$link_args += ($f_link_lib + $lib)
@@ -243,6 +219,52 @@ function make-binary { param([string]$elf, [string]$exe)
 	if ($LASTEXITCODE -ne 0) { Write-Error "Objcopy failed. Aborting."; exit 1 }
 }
 
+function ps1-meta { param(
+		[string]$unity_root,
+		[string[]]$sources,
+		[Parameter(Mandatory=$true)][string]$metadata,
+		[string]$out_root = (join-path $path_build 'gen'),
+		[string[]]$passes = @('--pre-link'),
+		[string[]]$extra_args = @()
+	)
+	# `--unity-root` and `--source` are
+	# mutually exclusive. Exactly one of `$unity_root` / `$sources` must
+	# be supplied; the other must be absent.
+	if ($null -ne $unity_root -and $unity_root -ne '') 
+	{
+		if ($null -ne $sources -and $sources.Count -gt 0) {
+			write-error 'ps1-meta: -unity_root and -sources are mutually exclusive'
+			exit 2
+		}
+	} 
+	elseif ($null -eq $sources -or $sources.Count -eq 0) {
+		write-error 'ps1-meta: either -unity_root <file> or -sources <file...> is required'
+		exit 2
+	}
+
+	$script = join-path $path_scripts 'ps1_meta.lua'
+	$input_summary = if ($null -ne $unity_root -and $unity_root -ne '') {
+		"unity=$unity_root"
+	}
+	else {
+		"$($sources.Count) source(s)"
+	}
+	write-host "ps1-meta  $input_summary, passes=$($passes -join ',')" ` -ForegroundColor Magenta
+
+	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root) + @($extra_args)
+	if ($null -ne $unity_root -and $unity_root -ne '') {
+		$arg_list += @('--unity-root', $unity_root)
+	}
+	else {
+		foreach ($s in $sources) { $arg_list += @('--source', $s) }
+	}
+	& luajit $script @arg_list
+	if ($LASTEXITCODE -ne 0) {
+		write-error "ps1-meta failed (exit $LASTEXITCODE). Aborting."
+		exit $LASTEXITCODE
+	}
+}
+
 function build-hello_psyqo {
 	$includes += @()
 
@@ -317,33 +339,16 @@ function build-graphis_hello {
 }
 # build-graphis_hello
 
-function ps1-meta { param(
-		[Parameter(Mandatory=$true)][string[]]$sources,
-		[Parameter(Mandatory=$true)][string]$metadata,
-		[string]$out_root = (join-path $path_build 'gen'),
-		[string[]]$passes = @('--all')
-	)
-	$script = join-path $path_scripts 'ps1_meta.lua'
-	write-host "ps1-meta  $($sources.Count) source(s), passes=$($passes -join ',')" ` -ForegroundColor Magenta
-	$arg_list = @($passes) + @('--metadata', $metadata) + @('--out-root', $out_root)
-	foreach ($s in $sources) { $arg_list += @('--source', $s) }
-	& luajit $script @arg_list
-	if ($LASTEXITCODE -ne 0) {
-		write-error "ps1-meta failed (exit $LASTEXITCODE). Aborting."
-		exit $LASTEXITCODE
-	}
-}
-
-function build-gte_hello {
+function build-hello_gte {
 	$includes += @()
 
-	$path_module        = join-path $path_code   'gte_hello'
+	$path_module        = join-path $path_code   'hello_gte'
 	$path_duffle        = join-path $path_code   'duffle'
 	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build   'gen'
 
-	$source_dirs  = @($path_duffle, $path_module)
-	$atom_sources = Get-SourceFiles -paths $source_dirs -extensions @('.h', '.c')
-	ps1-meta -sources $atom_sources -metadata $path_atom_metadata -out_root (join-path $path_build 'gen')
+	$src_c = join-path $path_module 'hello_gte.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
 
 	$assemble_args = @()
 	$assemble_args += $f_debug
@@ -359,7 +364,6 @@ function build-gte_hello {
 
 	# assemble-unit $src_asm $module_asm $includes $assemble_args
 
-	$src_c    = join-path $path_module 'hello_gte.c'
 	$module_c = join-path $path_build  'hello_gte_c.o'
 
 	$compile_args = @()
@@ -383,9 +387,172 @@ function build-gte_hello {
 	)
 	link-modules $link_modules $elf $link_args
 	make-binary $elf $exe
-}
-build-gte_hello
 
+	# Post-link: gdb-runtime + dwarf-injection in a single Lua invocation (one luajit cold start).
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen -passes @('--post-link') ` -extra_args @('--elf', $elf)
+
+	$dwarfLineBin     = join-path $path_build_gen 'hello_gte.dwarf_line.bin'
+	$dwarfArangesBin  = join-path $path_build_gen 'hello_gte.dwarf_aranges.bin'
+	$dwarfRnglistsBin = join-path $path_build_gen 'hello_gte.dwarf_rnglists.bin'
+	$injectElf        = join-path $path_build 'hello_gte.dwarf-injected.elf'
+	if ((Test-Path $dwarfLineBin) -and (Test-Path $dwarfArangesBin) -and (Test-Path $dwarfRnglistsBin))
+	{
+		Write-Host "[build] DWARF-injecting $elf -> $injectElf"
+		Copy-Item -LiteralPath $elf -Destination $injectElf -Force
+		# Objcopy call: 3x --update-section for (line, aranges, rnglists).
+		$f_args = @(
+			"--update-section=.debug_line=$dwarfLineBin",
+			"--update-section=.debug_aranges=$dwarfArangesBin",
+			"--update-section=.debug_rnglists=$dwarfRnglistsBin"
+		)
+		& $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
+
+function build-hello_joypad {
+	$includes += @()
+
+	$path_module        = join-path $path_code   'hello_joypad'
+	$path_duffle        = join-path $path_code   'duffle'
+	$path_atom_metadata = join-path $path_duffle 'word_count.metadata.h'
+	$path_build_gen     = join-path $path_build  'gen'
+
+	$src_c = join-path $path_module 'hello_joypad.c'
+	ps1-meta -unity_root $src_c -metadata $path_atom_metadata -out_root $path_build_gen
+
+	$assemble_args = @()
+	$assemble_args += $f_debug
+	$assemble_args += $f_optimize_none
+	$assemble_args += ($f_include + $path_code)
+
+	$src_asm_crt    = join-path $path_nugget_common 'crt0/crt0.s'
+	$module_asm_crt = join-path $path_build         'crt0.o'
+	assemble-unit $src_asm_crt $module_asm_crt $includes $assemble_args
+
+	$module_c = join-path $path_build  'hello_joypad_c.o'
+
+	$compile_args = @()
+	$compile_args += $f_debug
+	$compile_args += $f_optimize_none
+	# $compile_args += $f_optimize_intrinsics
+	# $compile_args += $f_optimize_size
+	# $compile_args += $f_optimize_debug
+	$compile_args += ($f_include + $path_code)
+	compile-unit $src_c $module_c $includes $compile_args
+
+	$elf = join-path $path_build 'hello_joypad.elf'
+	$exe = join-path $path_build 'hello_joypad.ps-exe'
+
+	$link_args = @()
+	$link_args += $f_debug
+	# $link_args += $f_optimize_size
+	$link_modules = @(
+		$module_asm_crt,
+		$module_c
+	)
+	link-modules $link_modules $elf $link_args
+	make-binary $elf $exe
+
+	# 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'
+	$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
 
 # NO idea if this works yet...
 function Send-ToEmulator { param( [string]$exePath )

scripts/duffle_paths.lua

@@ -11,11 +11,10 @@
 ---   ```
 ---
 --- That small bootstrap: (a) locates this helper via `arg[0]` / `debug.getinfo`,
---- (b) loads it (which sets `package.path` + `package.cpath` via cached `git rev-parse`),
+--- (b) loads it (which sets `package.path` + `package.cpath`),
 --- (c) at the bottom calls `require("duffle")` (now resolvable since `package.path` was just set) and returns the duffle M.
 --- Net effect: the caller gets the duffle module in one statement; no separate `dofile(...)` + `require("duffle")` dance.
 ---
---- Replaces the prior 2-line (entry) or 4-line (pass) pattern that had the call site do its own path resolution + duplicated setup.
 
 local M = {}
 
@@ -27,50 +26,25 @@ local CACHE_KEY = "__duffle_repo_root__"
 --- parent of the directory containing this script. We derive it directly from `debug.getinfo(1, "S").source`
 --- (returns `@<path>` for the currently-running chunk).
 ---
---- Replaces the prior `io.popen("git rev-parse --show-toplevel")` approach, which cost ~100-180ms per
---- LuaJIT process on Windows due to git's CLI startup. The path-derive approach costs <1ms.
----
---- If this script's path can't be parsed (shouldn't happen — dofile/debug.getinfo always populates source),
---- fall back to a defensive walk: starting from this script's directory, walk UP until we find a parent that
---- contains a `scripts/` directory. The first match is the repo root.
+--- If `debug.getinfo` can't parse this script's path (shouldn't happen — dofile always populates source),
+--- return nil and let `M.setup()` fail loud.
 --- @return string|nil
 local function find_repo_root()
 	if package.loaded[CACHE_KEY] then return package.loaded[CACHE_KEY] end
 
 	local source = debug.getinfo(1, "S").source
 	-- Strip the leading `@` (Lua's dofile marker) and the trailing `/duffle_paths.lua` filename.
-	-- What remains is the directory containing this script, i.e. `<repo>/scripts/` (with trailing slash or not).
+	-- What remains is the directory containing this script, i.e. `<repo>/scripts/`.
 	local scripts_dir = source and source:match("^@?(.*)[/\\]duffle_paths%.lua$")
-	if scripts_dir then
-		-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
-		local root = scripts_dir:gsub("scripts[\\/]?$", "")
-		root = root:gsub("\\", "/")
-		if root == "" then root = "./" end
-		if not root:match("/$") then root = root .. "/" end
-		package.loaded[CACHE_KEY] = root
-		return root
-	end
+	if not scripts_dir then return nil end
 
-	-- Defensive fallback: walk UP from this script's directory until we find a parent that contains `scripts/`.
-	-- In practice this branch never fires — debug.getinfo always returns a source for dofile()'d chunks.
-	local lfs = pcall(require, "lfs") and require("lfs") or nil
-	if lfs then
-		local dir = source and source:match("^@?(.*[/\\])") or "./"
-		dir = dir:gsub("\\", "/")
-		while dir and dir ~= "" do
-			local candidate_scripts = dir .. "scripts"
-			if lfs.attributes(candidate_scripts, "mode") == "directory" then
-				dir = dir:gsub("/$", "")
-				package.loaded[CACHE_KEY] = dir .. "/"
-				return dir .. "/"
-			end
-			local parent = dir:match("^(.*)/[^/]+/$")
-			if not parent then break end
-			dir = parent .. "/"
-		end
-	end
-
-	return nil
+	-- The repo root is the parent of `scripts/`. Strip the trailing `scripts/` (with or without trailing slash).
+	local root = scripts_dir:gsub("scripts[\\/]?$", "")
+	root = root:gsub("\\", "/")
+	if root == "" then root = "./" end
+	if not root:match("/$") then root = root .. "/" end
+	package.loaded[CACHE_KEY] = root
+	return root
 end
 
 --- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
@@ -83,7 +57,12 @@ end
 function M.setup()
 	local repo_root = find_repo_root()
 	if not repo_root then
-		io.stderr:write("[duffle_paths] git rev-parse failed -- not in a git repo?\n")
+		-- Unreachable in practice: find_repo_root() derives the repo root from this script's
+		-- own source path via debug.getinfo(1, "S").source (no subprocess, no git CLI, <1ms).
+		-- A nil return means the source path did not match the expected
+		-- <repo>/scripts/duffle_paths.lua layout — a packaging bug, not a "missing git repo"
+		-- condition. os.exit(2) is retained so a real failure surfaces loud rather than
+		-- silently producing an unconfigured module table.
 		os.exit(2)
 	end
 

scripts/gdb/gdb_tape_atoms.gdb

@@ -0,0 +1,105 @@
+# scripts/gdb/gdb_tape_atoms.gdb
+#
+# Wrapper for the tape-atom step-debug helpers.
+# The 9 user commands are defined here as STUBS (degraded-state messages). 
+# The real implementations + the per-atom data tables are emitted by `passes/atoms_source_map.lua` 
+# (post-link invocation: `ps1_meta.lua --atoms-source-map --gdb-runtime --elf <elf>`) into `build/gdb_tape_atoms_runtime.gdb`.
+# Sourcing that file RE-DEFINES the commands with real implementations.
+#
+# If `build/gdb_tape_atoms_runtime.gdb` is missing or stale, the stubs remain (E1: no source map).
+# The user just needs to re-run `build_psyq.ps1` to regenerate.
+
+# ?? Stub commands (defined here so they're always present, even if the runtime file is missing). The runtime file overrides these if sourced. ??
+
+define tape_atoms
+  echo "[gdb_tape_atoms] STUB: runtime file build/gdb_tape_atoms_runtime.gdb not found."
+  echo "[gdb_tape_atoms] STUB: run .\\build_psyq.ps1 to regenerate, then re-source this file."
+end
+document tape_atoms
+  List every tape atom symbol in the loaded ELF (code_<name>) with its .rodata address and word count.
+  STUB state: runtime file not sourced. Run build_psyq.ps1 to regenerate.
+end
+
+define break_atom
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document break_atom
+  Set a breakpoint at the start of tape atom <name>. STUB state.
+end
+
+define step_atom
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document step_atom
+  Resume execution until the next atom boundary. STUB state.
+end
+
+define next_atom
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document next_atom
+  Alias for step_atom. STUB state.
+end
+
+define where_in_atom
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document where_in_atom
+  Report current atom name, .rodata addr, word offset, and source line (if known). STUB state.
+end
+
+define stepi_inside_atom
+  echo "[gdb_tape_atoms] STUB: build/gdb_tape_atoms_runtime.gdb not sourced. Run build_psyq.ps1."
+end
+document stepi_inside_atom
+  One MIPS-instruction step, then where_in_atom. STUB state.
+end
+
+define show_c2
+  printf "C2[ 0]   0x%08x\n", $c2_data[0]
+  printf "C2[ 7]   0x%08x  [otz]\n", $c2_data[7]
+  printf "C2[12]   0x%08x  [sxy0]\n", $c2_data[12]
+  printf "C2[13]   0x%08x  [sxy1]\n", $c2_data[13]
+  printf "C2[14]   0x%08x  [sxy2]\n", $c2_data[14]
+  printf "C2[24]   0x%08x  [mac0]\n", $c2_data[24]
+  printf "...\n"
+   echo "(STUB state: only 7 representative regs shown. Run build_psyq.ps1 for full dump.)"
+end
+document show_c2
+  Pretty-print all 32 C2 data registers as hex + named alias. STUB state (7 reg subset).
+end
+
+define show_c2ctl
+  printf "C2CTL[ 0]  0x%08x\n", $c2_control[0]
+  printf "...\n"
+  echo "(STUB state: only 1 reg shown. Run build_psyq.ps1 for full dump.)"
+end
+document show_c2ctl
+  Pretty-print all 32 C2 control registers. STUB state (1 reg subset).
+end
+
+define wave_ctx
+  printf "$t4 = R_FaceCursor     0x%08x\n", $t4
+  printf "$t5 = R_VertBase       0x%08x\n", $t5
+  printf "$t6 = R_OtBase         0x%08x\n", $t6
+  printf "$t7 = R_PrimCursor     0x%08x\n", $t7
+end
+document wave_ctx
+  Pretty-print the 4 wave-context GPRs ($t4..$t7). (wave_ctx works in stub state too.)
+end
+
+
+# ?? Source the runtime file (re-defines commands with real impls + data). ??
+
+# Try to source from project-root-relative path first (the typical case).
+# If the user is in a different CWD, the source will fail and stubs remain.
+# The runtime file path is computed relative to the ELF's source map convention (build/gdb_tape_atoms_runtime.gdb).
+echo [gdb_tape_atoms] Wrapper loaded. Sourcing runtime file...
+# Suppress the "Redefine command" prompts that would otherwise appear when the runtime file overrides the 9 stub commands defined above.
+# The runtime's `define` blocks are intended to overwrite ? there's no ambiguity to confirm.
+set confirm off
+
+# Source the runtime file (re-defines commands with real impls + data).
+source build/gdb_tape_atoms_runtime.gdb
+set confirm on
+echo [gdb_tape_atoms] Runtime sourced successfully (9 commands now have real implementations).

scripts/launch_pcsx_debug.ps1

@@ -0,0 +1,94 @@
+# scripts/launch_pcsx_debug.ps1
+#
+# One-shot launcher for debug sessions: starts pcsx-redux with the .ps-exe
+# loaded, the gdb stub enabled, AND the pcsx_debug_helper Lua plugin loaded
+# so external CLI tools (gdb's `shell` command, etc.)
+# can read GTE state via http://localhost:8080/api/v1/lua/gte
+# (the gdb stub doesn't expose COP2 at all).
+#
+# usage:
+#   .\scripts\launch_pcsx_debug.ps1
+#   .\scripts\launch_pcsx_debug.ps1 -ExePath build\hello_gte.ps-exe
+#   .\scripts\launch_pcsx_debug.ps1 -HelperZip scripts\pcsx_debug_helper.zip
+#
+# After launch:
+#   - gdb: target remote localhost:3333
+#   - web: curl http://localhost:8080/api/v1/lua/gte
+#
+# Companion: scripts/debug_psyq.ps1 (bare launch — no .ps-exe, no helper).
+
+[CmdletBinding()]
+param(
+    [string]$PcsxPath = (Join-Path $PSScriptRoot '..\toolchain\pcsx-redux\vsprojects\x64\Release\pcsx-redux.exe'),
+    [string]$ExePath  = (Join-Path $PSScriptRoot '..\build\hello_gte.ps-exe'),
+    [string]$HelperZip = (Join-Path $PSScriptRoot 'pcsx_debug_helper.zip'),
+    [int]   $GdbPort  = 3333,
+    [int]   $WebPort  = 8080
+)
+
+$ErrorActionPreference = 'Stop'
+
+# ── Pre-checks ──
+foreach ($p in @($PcsxPath, $ExePath, $HelperZip)) {
+    if (-not (Test-Path $p)) {
+        Write-Error "Missing: $p"
+        exit 1
+    }
+}
+
+# Kill any existing pcsx-redux so the archive file isn't locked.
+Get-Process pcsx-redux -ErrorAction SilentlyContinue | Stop-Process -Force
+Start-Sleep -Seconds 2
+
+# ── Launch ──
+$absExe  = [System.IO.Path]::GetFullPath($ExePath)
+$absZip  = [System.IO.Path]::GetFullPath($HelperZip)
+
+$args = @(
+    '-gdb', '-run'
+    '-loadexe', "`"$absExe`""
+    '-archive', "`"$absZip`""
+)
+
+Write-Host "Launching pcsx-redux..." -ForegroundColor Cyan
+Write-Host "  ps-exe    : $absExe"
+Write-Host "  helper zip: $absZip"
+Write-Host "  gdb       : localhost:$GdbPort"
+Write-Host "  web       : localhost:$WebPort/api/v1/lua/gte"
+Write-Host ""
+
+Start-Process -FilePath $PcsxPath -ArgumentList $args | Out-Null
+
+# ── Wait for both endpoints to come up ──
+$deadline = (Get-Date).AddSeconds(15)
+while ((Get-Date) -lt $deadline) {
+    $gdbUp = $false
+    $webUp = $false
+    try {
+        $tcp = New-Object System.Net.Sockets.TcpClient
+        $tcp.BeginConnect('localhost', $GdbPort, $null, $null) | Out-Null
+        Start-Sleep -Milliseconds 100
+        $gdbUp = $tcp.Connected
+        $tcp.Close()
+    } catch { $gdbUp = $false }
+    try {
+        $r = Invoke-WebRequest -Uri "http://localhost:$WebPort/" -UseBasicParsing -TimeoutSec 1 -ErrorAction SilentlyContinue
+        $webUp = $r.StatusCode -ne 0
+    } catch { $webUp = $false }
+    if ($gdbUp -and $webUp) { break }
+    Start-Sleep -Milliseconds 500
+}
+
+# ── Smoke-test the gte handler ──
+try {
+    $r = Invoke-WebRequest -Uri "http://localhost:$WebPort/api/v1/lua/gte" -UseBasicParsing -TimeoutSec 5
+    $firstLine = ([System.Text.Encoding]::UTF8.GetString($r.Content) -split "`n")[0]
+    Write-Host "GTE handler OK: $firstLine" -ForegroundColor Green
+} catch {
+    Write-Warning "GTE handler NOT responding: $_"
+    Write-Host "Check the pcsx-redux Lua Console for debug cli messages." -ForegroundColor Yellow
+}
+
+Write-Host ""
+Write-Host "pcsx-redux running. PIDs:" -ForegroundColor Cyan
+Get-Process pcsx-redux | Select-Object Id, ProcessName | Format-Table

scripts/passes/annotation.lua

@@ -1,35 +1,21 @@
 --- passes/annotation.lua — Atom-annotation DSL validator.
 ---
 --- Validates `MipsAtom_(name) atom_info(atom_bind(Binds_X), atom_reads(...), atom_writes(...)) { ... }` declarations in source files.
---- Also reads: `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`)
+--- Also reads `Binds_*` struct declarations (`typedef Struct_(Binds_X) { ... };`).
 ---
---- Source scanning: done ONCE upstream by `duffle.scan_source()` (ps1_meta.lua pre-scans each
---- source and stashes the result in `src.scan`). This pass is pure: read from the scan, run
---- checks, emit findings. No source re-walking.
+--- `duffle.scan_source()` scans each source once upstream; `ps1_meta.lua` stores that result in `src.scan`.
 ---
---- Writes:
----   - `<ctx.out_root>/<dir_basename>.errors.h` — one per module, with `#error` directives on findings (the C compile will surface the error)
----   - The annotations.txt report is rendered by `passes/report.lua` from the per-module results stashed in `ctx.flags._annot_results`
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible
+--- Ownership: the canonical `ctx.shared.corpus` supplies cross-source registries, while each `src.scan` supplies its source's declarations and bodies.
+--- A context without `ctx.shared.corpus` is rejected with an explicit canonical-corpus message.
 
--- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
--- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
--- both standalone and when require'd from the orchestrator.
--- 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.
+-- Bootstrap follows the entry scripts; `scripts/duffle_paths.lua` sets package.path and package.cpath. See `ps1_meta.lua` for the rationale.
+-- `debug.getinfo(1, "S").source` locates this file for standalone and orchestrated runs, then `duffle_paths.lua` returns the loaded `duffle` module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local write_file          = duffle.write_file
-local ensure_dir          = duffle.ensure_dir
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
--- Domain tables (single source of truth in duffle.lua).
-local WAVE_CONTEXT_REGS = duffle.WAVE_CONTEXT_REGS
-local TAPE_ATOM_MACROS  = duffle.TAPE_ATOM_MACROS
-
-local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
+-- The annotation pass reads the source-derived registries from scan_source:
+--   * pipe_ctx.register_alias_registry — for atom_dbg_reg_default(R_X, ...) and atom_reg_types(R_X, ...) member-identity checks
+--   * pipe_ctx.type_name_registry      — for atom_dbg_reg_default(<T>, ...) and atom_reg_types(<T>, ...) type-identity checks
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
@@ -51,8 +37,6 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @field project_root       string
 --- @field upstream           table<string, table>
 --- @field flags              table
---- @field flags._annot_results table[]   -- stashed by annotation pass; consumed by report.lua
---- @field dry_run            boolean
 --- @field verbose            boolean
 
 --- @class PassResult
@@ -68,10 +52,36 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @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)
+
+--- @class DebugSkipMarker  -- sub-shape of scan_source.lua's @class DebugSkipMarker
+--- @field marker_kind    string                 -- exact marker ident read from source. Only "atom_dbg_skip" (bare) is positive.
+--- @field marker_line    integer
+--- @field args           string|nil             -- trimmed text inside the parens (nil when has_parens is false)
+--- @field has_parens     boolean
+--- @field is_bare        boolean                -- true iff marker_kind == "atom_dbg_skip" AND has_parens == false (the only positive form)
+--- @field pending        boolean                -- true while awaiting the following declaration
+--- @field superseded_by_marker_line integer|nil -- set on a marker that was bumped out of the pending slot
+--- @field target_kind    string|nil             -- "atom" | "comp_bare" | "comp_proc" | "unrelated" once observed
 
 --- @class Finding
---- @field line integer  -- source line (or 0 for pass-level)
---- @field msg  string   -- finding message
+--- @field line integer -- source line (or 0 for pass-level)
+--- @field msg  string  -- finding message
+
+--- @class Findings
+--- @field errors   Finding[]
+--- @field warnings Finding[]
+--- @field info     Finding[]
+
+--- @class PipeCtx
+--- @field atom_index     table<string, AtomAnnotation> -- name -> AtomAnnotation (only kind=="atom")
+--- @field binds_index    table<string, BindsStruct>    -- name -> BindsStruct
+--- @field annot_counts   table<string, integer>        -- name -> annotation count (for unique_annotation check)
+--- @field types          table<string, RegTypeDefault> -- from scan_source
+--- @field atom_views     table<string, AtomViewEntry>  -- from scan_source
+--- @field seen_defaults  table<string, integer>        -- duplicate atom_dbg_reg_default detection
+--- @field 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) 
 
 --- @class AnnotatedResult
 --- @field atoms    AtomEntry[]
@@ -82,18 +92,391 @@ local function is_wave_context_reg(n) return WAVE_CONTEXT_REGS[n] ~= nil  end
 --- @field warnings Finding[]
 --- @field info     Finding[]
 
+-- ════════════════════════════════════════════════════════════════════════════
+-- 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.
+--- @param a        AtomAnnotation
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_decl_exists(a, pipe_ctx, findings)
+	if not pipe_ctx.atom_index[a.name] then
+		findings.errors[#findings.errors + 1] = {
+			line = a.line,
+			msg  = string.format("annotation for '%s' has no matching MipsAtom_(%s) { ... }", a.name, a.name),
+		}
+	end
+end
+
+--- Check: every atom may have AT MOST ONE annotation.
+--- Post-loop: needs full-corpus `annot_counts` from pipe_ctx.
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_unique_annotation(pipe_ctx, findings)
+	for name, n in pairs(pipe_ctx.annot_counts) do
+		if n > 1 then
+			findings.errors[#findings.errors + 1] = {
+				line = pipe_ctx.atom_index[name] and pipe_ctx.atom_index[name].line or 0,
+				msg  = string.format("MipsAtom_(%s) has %d annotations (expected at most 1)", name, n),
+			}
+		end
+	end
+end
+
+--- Check: BIND atoms must reference a real Binds_* struct.
+--- I keep this as a warning so the annotation pass can report the common test-fixture case; `check_abi_handoff` in static analysis supplies the build-stopping error.
+--- @param a        AtomAnnotation
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_binds_struct_exists(a, pipe_ctx, findings)
+	if not a.binds                   then return end
+	if pipe_ctx.binds_index[a.binds] then return end
+	findings.warnings[#findings.warnings + 1] = {
+		line = a.line,
+		msg  = string.format("'%s' binds '%s' but no Struct_(%s) { ... } " 
+			.. "declaration found (also flagged as an error by check_abi_handoff in the static-analysis pass)"
+			, a.name, a.binds, a.binds),
+	}
+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 findings Findings
+local function check_macro_word_drift(m, wc, findings)
+	local  declared = wc[m.name]
+	if not declared then
+		findings.errors[#findings.errors + 1] = {
+			line = m.line,
+			msg  = string.format("TAPE_WORDS(%s, %d) but '%s' is not in metadata.h", m.name, m.words, m.name),
+		}
+		return
+	end
+	if declared ~= m.words then
+		findings.errors[#findings.errors + 1] = {
+			line = m.line,
+			msg  = string.format("DRIFT: TAPE_WORDS(%s, %d) but metadata.h declares WORD_COUNT(%s, %d)", m.name, m.words, m.name, declared),
+		}
+		return
+	end
+	findings.info[#findings.info + 1] = {
+		line = m.line,
+		msg  = string.format("OK: %s = %d words", m.name, m.words),
+	}
+end
+
+--- Check: atom_dbg_reg_default(R_X, <type>) targets an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
+--- Pointer depth remains bounded to 0 or 1, and duplicate defaults remain errors.
+--- @param _src     SourceFile -- unused (kept for the per_source shape)
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_semantic_reg_defaults(_src, pipe_ctx, findings)
+	-- Detect duplicate defaults using the ordered occurrence list (the out.types hash only retains the last declaration).
+	local seen_first_line = {}
+	for _, occ in ipairs(pipe_ctx.type_occurrences or {}) do
+		if seen_first_line[occ.reg] == nil then
+			seen_first_line[occ.reg] = occ.source_line
+		else
+			findings.errors[#findings.errors + 1] = {
+				line = occ.source_line,
+				msg  = string.format(
+					"duplicate atom_dbg_reg_default for %q at line %d (first declared at line %d); one default per register",
+					occ.reg, occ.source_line, seen_first_line[occ.reg]),
+			}
+		end
+	end
+	local reg_registry  = pipe_ctx.register_alias_registry or {}
+	local type_registry = pipe_ctx.type_name_registry      or {}
+	for reg, def in pairs(pipe_ctx.types or {}) do
+		if not reg_registry[reg] then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d references unknown register %q (not in register_alias_registry)",
+					def.source_line, reg),
+			}
+		end
+		if def.pointer_depth == nil or def.pointer_depth < 0 or def.pointer_depth > 1 then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d for %q has unsupported pointer depth %d (expected 0 or 1)",
+					def.source_line, reg, def.pointer_depth or -1),
+			}
+		end
+		if not def.type_name or not type_registry[def.type_name] then
+			findings.errors[#findings.errors + 1] = {
+				line = def.source_line,
+				msg  = string.format(
+					"atom_dbg_reg_default at line %d for %q uses unknown type %q (not in type_name_registry)",
+					def.source_line, reg, tostring(def.type_name)),
+			}
+		end
+	end
+end
+
+--- Check: atom_reg_types(R_X, <type>) entries target an alias in `pipe_ctx.register_alias_registry` and a type in `pipe_ctx.type_name_registry`.
+--- A bare `atom_reg` marker opts the `R_<n>` alias into GPR identity; references to R_T0..R_T3 require the same explicit marker.
+--- @param _src     SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_reg_types(_src, pipe_ctx, findings)
+	local reg_registry  = pipe_ctx.register_alias_registry or {}
+	local type_registry = pipe_ctx.type_name_registry      or {}
+	for _, ai in ipairs(pipe_ctx.atom_infos_list or {}) do
+		if ai.reg_type_overrides then
+			for reg, ov in pairs(ai.reg_type_overrides) do
+				if not reg_registry[reg] then
+					findings.errors[#findings.errors + 1] = {
+						line = ai.info_line,
+						msg  = string.format(
+							"atom '%s' has atom_reg_types for %q; compute-register types are restricted to opt-in aliases (%q not in register_alias_registry)",
+							ai.atom_name, reg, reg),
+					}
+				end
+				if not ov.type_name or not type_registry[ov.type_name] then
+					findings.errors[#findings.errors + 1] = {
+						line = ai.info_line,
+						msg  = string.format(
+							"atom '%s' atom_reg_types for %q uses unknown compute type %q (not in type_name_registry)",
+							ai.atom_name, reg, tostring(ov.type_name)),
+					}
+				end
+			end
+		end
+	end
+end
+
+--- Check: atom_view(Binds_X) entries reference a Binds_* struct with at least one field.
+--- @param _src     SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_atom_view_layout(_src, pipe_ctx, findings)
+	for atom_name, view in pairs(pipe_ctx.atom_views or {}) do
+		if not view.binds_name then
+			-- The atom had atom_reg_types but no atom_view; no layout check needed.
+		else
+			local bs = pipe_ctx.binds_index[view.binds_name]
+			if not bs then
+				findings.errors[#findings.errors + 1] = {
+					line = view.info_line,
+					msg  = string.format(
+						"atom '%s' has atom_view(%s) but no Struct_(%s) { ... } declaration was found",
+						atom_name, view.binds_name, view.binds_name),
+				}
+			elseif not bs.fields or #bs.fields == 0 then
+				findings.errors[#findings.errors + 1] = {
+					line = bs.line,
+					msg  = string.format(
+						"atom '%s' has atom_view(%s) but that struct declares zero typed fields",
+						atom_name, view.binds_name),
+				}
+			end
+		end
+	end
+end
+
+--- Check: Binds_* structs require unique field names because atom_view uses those names for typed-field lookup in gdb.
+--- @param _src SourceFile
+--- @param pipe_ctx PipeCtx
+--- @param findings Findings
+local function check_binds_no_duplicate_fields(_src, pipe_ctx, findings)
+	for _, bs in ipairs(pipe_ctx.binds_list or {}) do
+		local seen = {}
+		for _, f in ipairs(bs.fields or {}) do
+			seen[f.name] = (seen[f.name] or 0) + 1
+		end
+		for name, count in pairs(seen) do
+			if count > 1 then
+				findings.errors[#findings.errors + 1] = {
+					line = bs.line,
+					msg  = string.format(
+						"%s has duplicate field name %q (count %d); the typed-view contract requires unique field names",
+						bs.name, name, count),
+				}
+			end
+		end
+	end
+end
+
+-- 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`)
+---   2. marker_kind == "atom_dbg_skip" AND has_parens -> parenthesized form (the marker is bare-only)
+---   3. args  ~= ""                 -> takes no arguments
+---   4. superseded_by_marker_line   -> duplicate marker (cite superseding line)
+---   5. pending + no target_kind    -> dangling (no following declaration)
+---   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 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
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d must be bare; the parenthesized form is no longer accepted (use `atom_dbg_skip MipsAtom_(name) { ... }`)",
+				kind, line),
+		}
+		return
+	end
+
+	if marker.args ~= nil and marker.args ~= "" then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d takes no arguments; found %q", kind, line, marker.args),
+		}
+		return
+	end
+
+	if marker.superseded_by_marker_line then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("duplicate %s marker at line %d; superseded by another %s marker at line %d"
+				, kind, line, kind, marker.superseded_by_marker_line),
+		}
+		return
+	end
+
+	if marker.pending and not marker.target_kind then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("dangling %s marker at line %d: no following MipsAtom_/MipsAtomComp_/MipsAtomComp_Proc_ declaration"
+				, kind, line),
+		}
+		return
+	end
+
+	if marker.target_kind
+		and marker.target_kind ~= "atom"
+		and marker.target_kind ~= "comp_bare"
+		and marker.target_kind ~= "comp_proc" then
+		findings.errors[#findings.errors + 1] = {
+			line = line,
+			msg  = string.format("%s marker at line %d must precede MipsAtom_, MipsAtomComp_, or MipsAtomComp_Proc_; found an unrelated declaration"
+				, kind, line),
+		}
+	end
+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
+--- @param findings Findings
+local function check_wave_context_migration(_src, pipe_ctx, findings)
+	if not (pipe_ctx.types and next(pipe_ctx.types)) then return end
+	if not (pipe_ctx.atom_infos_list) then return end
+	local reg_registry = pipe_ctx.register_alias_registry or {}
+	for _, ai in ipairs(pipe_ctx.atom_infos_list) do
+		if ai.reg_type_overrides then
+			for reg, _ in pairs(ai.reg_type_overrides) do
+				if not reg_registry[reg] then
+					findings.warnings[#findings.warnings + 1] = {
+						line = 0,
+						msg  = "wave-context removed; opt in via #define atom_reg in mips.h "
+							.. "(every R_<alias> that should be visible to the annotation pass "
+							.. "must be enum-declared with the bare atom_reg marker)",
+					}
+					return
+				end
+			end
+		end
+	end
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- CHECK_RULES — data-driven check dispatch (the plex pattern)
+-- ════════════════════════════════════════════════════════════════════════════
+--
+-- Each rule entry picks one of four "shapes" of dispatch:
+--   per_annot(annot, pipe_ctx, findings)        -- runs once per AtomAnnotation
+--   post(pipe_ctx, findings)                    -- runs once after all per_annot calls complete (full-corpus aggregation)
+--   per_macro(macro, wc, findings)              -- runs once per TAPE_WORDS / _Pragma macro declaration
+--   per_skip_marker(marker, pipe_ctx, findings) -- runs once per src.scan.debug_skip_markers entry
+--
+-- Adding a new check = 1 row here + 1 function above. The `validate()` dispatch loop never needs editing.
+
+local CHECK_RULES = {
+	{ name = "atom_decl_exists",          per_annot       = check_atom_decl_exists          },
+	{ name = "binds_struct_exists",       per_annot       = check_binds_struct_exists       },
+	{ name = "unique_annotation",         post            = check_unique_annotation         },
+	{ name = "macro_word_drift",          per_macro       = check_macro_word_drift          },
+	{ name = "skip_marker_validation",    per_skip_marker = check_skip_marker               },
+	{ name = "semantic_reg_defaults",     per_source      = check_semantic_reg_defaults     },
+	{ name = "atom_reg_types",            per_source      = check_atom_reg_types            },
+	{ name = "atom_view_layout",          per_source      = check_atom_view_layout          },
+	{ name = "binds_no_duplicate_fields", per_source      = check_binds_no_duplicate_fields },
+	{ name = "wave_context_migration",    per_source      = check_wave_context_migration    },
+}
+
 -- ════════════════════════════════════════════════════════════════════════════
 -- Validation
 -- ════════════════════════════════════════════════════════════════════════════
 --
--- Pure check: read from src.scan, run validations, emit findings.
--- No source walking; no parsing. The scan was done once upstream.
+-- Pure check: read from src.scan, run validations, emit findings. The scan was done once upstream.
 
---- Validate one source against its pre-scanned SourceScan payload.
+--- Builds one pass-wide pipe_ctx from the merged `corpus.*` registries and source-ordered `corpus.atom_infos`; per-source declarations and bodies remain in `src.scan`.
+--- The module ownership contract above requires callers to construct `ctx.shared.corpus` through `build_ctx`; the error message below enforces that gate.
 --- @param ctx PassCtx
---- @param src SourceFile
+--- @return PipeCtx
+local function build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared and ctx.shared.corpus
+	if not corpus then
+		error("annotation requires ctx.shared.corpus "
+			.. "(the canonical corpus is the source of truth; "
+			.. "no per-source fallback is supported)", 0)
+	end
+
+	-- `corpus.atom_infos` preserves source order and duplicates; I precompute counts here for `check_unique_annotation` and the per-source checks.
+	local annot_counts = {}
+	for _, info in ipairs(corpus.atom_infos or {}) do
+		if info and info.atom_name then
+			annot_counts[info.atom_name] = (annot_counts[info.atom_name] or 0) + 1
+		end
+	end
+
+	-- Every consumer of these fields observes mutations via the canonical corpus without independently mutable registry construction.
+	return {
+		-- Cross-source lookup tables from corpus.
+		register_alias_registry  = corpus.register_alias_registry or {},
+		type_name_registry       = corpus.type_name_registry      or {},
+		atom_views               = corpus.atom_views              or {},
+		atom_ctxs                = corpus.atom_ctxs               or {},
+		atom_phases              = corpus.atom_phases             or {},
+		binds_by_name            = corpus.binds_by_name           or {},
+		atoms_by_name            = corpus.atoms_by_name           or {},
+		-- Corpus-wide ordered list of atom_info records (source-order + duplicates).
+		atom_infos_list          = corpus.atom_infos              or {},
+		-- Corpus-wide annotation count aggregation (post-rule consumes this).
+		annot_counts             = annot_counts,
+		-- Corpus-wide collisions (recorded by scan_source.merge_corpus_registries).
+		collisions               = corpus.collisions              or {},
+		-- `check_macro_word_drift` reads `corpus.word_counts`, populated by word_count_eval.run.
+		word_counts              = corpus.word_counts or {},
+	}
+end
+
+--- Validate one source against its pre-scanned SourceScan payload + the corpus-wide pipe_ctx.
+--- @param ctx             PassCtx
+--- @param src             SourceFile
+--- @param corpus_pipe_ctx PipeCtx|nil  -- built once per pass from corpus registries; nil builds the same projection here.
 --- @return AnnotatedResult
-local function validate(ctx, src)
+local function validate(ctx, src, corpus_pipe_ctx)
+	corpus_pipe_ctx = corpus_pipe_ctx or build_corpus_pipe_ctx(ctx)
 	local scan = src.scan
 
 	-- Project the pre-scanned atoms to the AtomEntry shape this pass needs.
@@ -115,141 +498,93 @@ local function validate(ctx, src)
 			binds  = info.binds,
 			reads  = info.reads or {},
 			writes = info.writes or {},
-			errors = {},
+			errors = info.errors,
 		}
 	end
 
-	-- Index atoms by name for lookup.
-	local atom_index = {}
-	for _, a in ipairs(atoms) do atom_index[a.name] = a end
+	-- Build a per-source pipe_ctx: shared lookups come from `corpus_pipe_ctx`, while declarations, bodies, types, views, defaults, and occurrences come from `src.scan`.
+	local seen_defaults = {}
+	for reg, _ in pairs(scan.types or {}) do
+		seen_defaults[reg] = (seen_defaults[reg] or 0) + 1
+	end
+	local atom_infos_list = {}
+	for _, ai in ipairs(scan.atom_infos or {}) do
+		atom_infos_list[#atom_infos_list + 1] = ai
+	end
 
-	-- Index binds by name for lookup.
-	local binds_index = {}
-	for _, b in ipairs(scan.binds) do binds_index[b.name] = b end
+	local pipe_ctx = {
+		atom_index               = {},
+		binds_index              = {},
+		annot_counts             = corpus_pipe_ctx.annot_counts,
+		types                    = scan.types or {},
+		type_occurrences         = scan.type_occurrences or {},
+		atom_views               = scan.atom_views or {},
+		seen_defaults            = seen_defaults,
+		atom_infos_list          = atom_infos_list,
+		binds_list               = scan.binds or {},
+		-- See the module ownership contract; these shared lookup tables come from corpus_pipe_ctx.
+		register_alias_registry  = corpus_pipe_ctx.register_alias_registry,
+		type_name_registry       = corpus_pipe_ctx.type_name_registry,
+	}
+	for _, a in ipairs(atoms)      do pipe_ctx.atom_index [a.name] = a end
+	for _, b in ipairs(scan.binds) do pipe_ctx.binds_index[b.name] = b end
 
-	local errors   = {}
-	local warnings = {}
-	local info     = {}
+	-- Findings live in a single struct with three lists (errors / warnings / info).
+	-- Each check writes to the list appropriate for its severity.
+	local findings = { errors = {}, warnings = {}, info = {} }
 
-	-- 1. Every annotated atom must exist as a real MipsAtom_ declaration.
+	-- Lift parse-time errors already recorded in scan_source's atom_info payload into this pass's findings list.
 	for _, a in ipairs(annots) do
-		if not atom_index[a.name] then
-			errors[#errors + 1] = {
-				line = a.line,
-				msg  = string.format("annotation for '%s' has no matching MipsAtom_(%s) { ... }", a.name, a.name),
-			}
-		end
 		if a.errors then
 			for _, msg in ipairs(a.errors) do
-				errors[#errors + 1] = {line = a.line, msg = string.format("'%s': %s", a.name, msg)}
-			end
-		end
-	end
-
-	-- 2. Every atom may have AT MOST ONE annotation (no duplicates).
-	-- (Atoms with ZERO annotations are valid in the new minimal shape.)
-	local count_per_atom = {}
-	for _, a in ipairs(annots) do
-		if a.name then
-			count_per_atom[a.name] = (count_per_atom[a.name] or 0) + 1
-		end
-	end
-	for name, n in pairs(count_per_atom) do
-		if n > 1 then
-			errors[#errors + 1] = {
-				line = atom_index[name] and atom_index[name].line or 0,
-				msg  = string.format("MipsAtom_(%s) has %d annotations (expected at most 1)", name, n),
-			}
-		end
-	end
-
-	-- 3. (Phase validity check DROPPED. Phases were removed from the annotation DSL.)
-
-	-- 4. BIND atoms must reference a real Binds_* struct.
-	for _, a in ipairs(annots) do
-		if a.binds then
-			if not binds_index[a.binds] then
-				-- Demoted from error to warning (2026-07-10): the same condition is now caught by passes/static_analysis.lua's
-				-- check_abi_handoff() as an error. Emitting a warning here keeps the annotation pass from being stop-on-error
-				-- for the common test-fixture case, while still surfacing the issue in the report.
-				-- The static-analysis report remains the source of truth for build-stopping errors.
-				warnings[#warnings + 1] = {
+				findings.errors[#findings.errors + 1] = {
 					line = a.line,
-					msg  = string.format("'%s' binds '%s' but no Struct_(%s) { ... } declaration found (also flagged as an error by check_abi_handoff in the static-analysis pass)", a.name, a.binds, a.binds),
+					msg  = string.format("'%s': %s", a.name, msg),
 				}
 			end
 		end
 	end
 
-	-- 5. BIND writes must be wave-context registers that match Binds_ fields.
+	-- THE per-annotation pipeline. ONE loop. CHECK_RULES dispatches per_annot rules.
 	for _, a in ipairs(annots) do
-		if a.binds and binds_index[a.binds] then
-			local bs = binds_index[a.binds]
-
-			for _, f in ipairs(bs.fields) do
-				local candidate = "R_" .. f.name
-				if not is_wave_context_reg(candidate) then
-					warnings[#warnings + 1] = {
-						line = bs.line,
-						msg  = string.format("%s field '%s' doesn't match a known wave-context register (candidate '%s')", a.binds, f.name, candidate),
-					}
-				end
-			end
-
-			for _, w in ipairs(a.writes) do
-				if not is_wave_context_reg(w) then
-					warnings[#warnings + 1] = {
-						line = a.line,
-						msg  = string.format("%s writes '%s' which is not a known wave-context register", a.name, w),
-					}
-				end
-			end
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_annot then rule.per_annot(a, pipe_ctx, findings) end
 		end
 	end
 
-	-- 6. INFO reads should be wave-context registers (or R_TapePtr for rbind).
-	for _, a in ipairs(annots) do
-		for _, r in ipairs(a.reads) do
-			if not is_wave_context_reg(r) and r ~= "R_TapePtr" then
-				warnings[#warnings + 1] = {
-					line = a.line,
-					msg  = string.format("atom '%s' reads '%s' which is not a known wave-context register", a.name, r),
-				}
-			end
+	-- Post-loop rules (one-shot checks that need full-corpus aggregation in pipe_ctx).
+	for _, rule in ipairs(CHECK_RULES) do
+		if rule.post then rule.post(pipe_ctx, findings) end
+	end
+
+	-- scan_source records each marker in scan.debug_skip_markers; this loop validates each record independently and emits at most one error per marker.
+	-- Valid markers stamp `debug_skip = true` on the following atom or component declaration, which downstream consumers read directly.
+	local skip_markers = scan.debug_skip_markers or {}
+	for _, marker in ipairs(skip_markers) do
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_skip_marker then rule.per_skip_marker(marker, pipe_ctx, findings) end
 		end
 	end
 
-	-- 7. TAPE_WORDS(mac_X, N) ↔ WORD_COUNT(mac_X, N) drift.
-	--    Three outcomes: missing (error), mismatch (error), match (info).
-	local function check_macro_drift(m, declared)
-		if not declared then
-			errors[#errors + 1] = {
-				line = m.line,
-				msg  = string.format("TAPE_WORDS(%s, %d) but '%s' is not in metadata.h", m.name, m.words, m.name),
-			}
-			return
-		end
-		if declared ~= m.words then
-			errors[#errors + 1] = {
-				line = m.line,
-				msg  = string.format("DRIFT: TAPE_WORDS(%s, %d) but metadata.h declares WORD_COUNT(%s, %d)", m.name, m.words, m.name, declared),
-			}
-			return
-		end
-		info[#info + 1] = {
-			line = m.line,
-			msg  = string.format("OK: %s = %d words", m.name, m.words),
-		}
-	end
+	-- Per-macro rules (TAPE_WORDS vs WORD_COUNT drift).
+	local wc = corpus_pipe_ctx.word_counts
 	for _, m in ipairs(scan.macros) do
-		check_macro_drift(m, ctx.shared.word_counts[m.name])
+		for _, rule in ipairs(CHECK_RULES) do
+			if rule.per_macro then rule.per_macro(m, wc, findings) end
+		end
 	end
 
-	-- 8. Information summary.
-	info[#info + 1] = {
+	-- Per-source rules (reg defaults, atom_view layout, compute-register type overrides, Binds_* field uniqueness).
+	-- Each per_source rule sees the full scan payload via pipe_ctx.
+	for _, rule in ipairs(CHECK_RULES) do
+		if rule.per_source then rule.per_source(src, pipe_ctx, findings) end
+	end
+
+	-- Information summary (always emitted).
+	findings.info[#findings.info + 1] = {
 		line = 0,
-		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)",
-			#atoms, #annots, #scan.macros, #scan.binds),
+		msg  = string.format("scanned: %d atom(s), %d annotation(s), %d macro-word-decl(s), %d binds struct(s)"
+			, #atoms, #annots, #scan.macros, #scan.binds),
 	}
 
 	return {
@@ -257,55 +592,9 @@ local function validate(ctx, src)
 		annots   = annots,
 		macros   = scan.macros,
 		binds    = scan.binds,
-		errors   = errors,
-		warnings = warnings,
-		info     = info,
-	}
-end
-
--- ════════════════════════════════════════════════════════════════════════════
--- Per-DIRECTORY (per-module) output: errors.h + annotations.txt
--- ════════════════════════════════════════════════════════════════════════════
-
---- Render `<dir_basename>.errors.h` with `#error` directives for every error found across all sources in the directory.
---- Empty directories (no errors, no atoms) produce no file.
-local function emit_module_errors_h(ctx, dir_basename, atoms_count, errors, sources)
-	if ctx.dry_run then return nil end
-	if atoms_count == 0 and #errors == 0 then
-		return nil
-	end
-	local out_path = ctx.out_root .. "/" .. dir_basename .. ".errors.h"
-	local lines    = {
-		"// Auto-generated by ps1_meta.lua (passes/annotation.lua) — DO NOT EDIT",
-		string.format("// Module: %s   Sources: %d", dir_basename, #sources),
-		"#pragma once",
-		"",
-	}
-	if #errors == 0 then
-		lines[#lines + 1] = "// annotation pass OK"
-	else
-		for _, e in ipairs(errors) do
-			local src_tag = ""
-			if e.source then
-				local src_name = e.source:match("([^/\\]+)$") or e.source
-				src_tag = src_name .. ": "
-			end
-			lines[#lines + 1] = string.format('#error "%s%s (line %d)"', src_tag, e.msg, e.line)
-		end
-	end
-	ensure_dir(ctx.out_root)
-	write_file(out_path, table.concat(lines, "\n") .. "\n")
-	return out_path
-end
-
---- Stash aggregated per-module results for the report pass to consume.
-local function emit_module_annotations_stub(ctx, dir, dir_basename, atoms_count)
-	ctx.flags         = ctx.flags or {}
-	ctx.flags._annot_results = ctx.flags._annot_results or {}
-	ctx.flags._annot_results[#ctx.flags._annot_results + 1] = {
-		dir          = dir,
-		dir_basename = dir_basename,
-		atoms_count  = atoms_count,
+		errors   = findings.errors,
+		warnings = findings.warnings,
+		info     = findings.info,
 	}
 end
 
@@ -327,41 +616,32 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	-- Per-DIRECTORY (per-module) aggregation. Group sources by `src.dir`,
-	-- validate every source in the dir, then emit ONE errors.h per dir.
-	-- `ctx.by_dir` is pre-computed in build_ctx (shared across all passes).
-	local by_dir = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	-- Build the shared pipe_ctx once for this run; every validate() call sees the same cross-source registries.
+	-- The corpus owns the canonical cross-source registries; per-source scans retain body / declaration ownership.
+	local corpus_pipe_ctx = build_corpus_pipe_ctx(ctx)
+	local corpus = ctx.shared.corpus
+
+	-- Group `corpus.sources_by_dir` by module, validate every source in each bucket, and emit one errors.h per directory.
+	local by_dir = (corpus and corpus.sources_by_dir) or {}
 
 	for dir, dir_sources in pairs(by_dir) do
 		local dir_basename = dir:match("([^/\\]+)$") or dir
-
 		local dir_atoms    = 0
 		local dir_errors   = {}
 		local dir_warnings = {}
-		-- Per-source validate() results, cached for the report pass (it reads from this instead of re-validating each source).
-		ctx.flags                       = ctx.flags or {}
-		ctx.flags._annot_source_results = ctx.flags._annot_source_results or {}
 		for _, src in ipairs(dir_sources) do
-			local result = validate(ctx, src)
-			result.source = src.path                                  -- tag for downstream rendering
-			ctx.flags._annot_source_results[src.path] = result          -- stash so report.lua reads from cache instead of re-running validate()
+			local result  = validate(ctx, src, corpus_pipe_ctx)
+			result.source = src.path                           -- tag for downstream rendering
 			dir_atoms = dir_atoms + #result.atoms
 			for _, e in ipairs(result.errors) do
 				dir_errors[#dir_errors + 1] = { line = e.line, msg = e.msg, source = src.path }
-				errors[#errors + 1] = { line = e.line, msg = e.msg }
+				errors    [#errors     + 1] = { line = e.line, msg = e.msg }
 			end
 			for _, w in ipairs(result.warnings) do
 				dir_warnings[#dir_warnings + 1] = { line = w.line, msg = w.msg }
-				warnings[#warnings + 1] = { line = w.line, msg = w.msg }
+				warnings    [#warnings     + 1] = { line = w.line, msg = w.msg }
 			end
 		end
-
-		local err_path = emit_module_errors_h(ctx, dir_basename, dir_atoms, dir_errors, dir_sources)
-		if err_path then
-			table.insert(outputs, { errors_h = err_path })
-		end
-
-		emit_module_annotations_stub(ctx, dir, dir_basename, dir_atoms)
 	end
 
 	return { outputs = outputs, errors = errors, warnings = warnings }

scripts/passes/atoms_source_map.lua

@@ -0,0 +1,572 @@
+--- passes/atoms_source_map.lua — Per-.word source-line map emitter for tape atoms.
+---
+--- Writer: this pass, given `atom.paths` (the per-atom mutable surface owned by `emission_model`). Readers:
+--- `passes/dwarf_injection.lua` (synthesizes DW_TAG_inlined_subroutine + per-word line program rows) and the gdb-runtime
+--- wrapper at `scripts/gdb/gdb_tape_atoms.gdb` (loads the source map via `source <path>`).
+---
+--- Inputs from `atom.paths`: the ordered `items` stream, dense `word_events`, `invocations` views. Outputs: 
+---  one `WORD N LINE L TEXT T` line per emitted `.word`, plus the per-word provenance form that DWARF synthesis consumes.
+---
+--- Two output forms:
+---   1. Markdown form: Handled by `passes/report.lua` (writes `<module>.atoms.md`).
+---      The render functions `render_source_map` + `render_provenance` are exported for `report.lua` to call directly.
+---      Compile artifacts (`*.macs.h`, `*.offsets.h`) stay in `<source_dir>/gen/`.
+---   2. `gdb_tape_atoms_runtime.gdb`: Post-link opt-in (`ctx.flags.gdb_runtime`),
+---      so the gdb wrapper script + the generated runtime script share the same canonical location.
+---      Triggered by `--post-link` or `--gdb-runtime`.
+---
+--- Output forma (sourcemap.txt form):
+---   ```
+---   # FORMAT_VERSION 1
+---   # auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT
+---   ATOM <name> "<abs-source-path>" <total_words>
+---   WORD 0  LINE 49  TEXT load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+---   WORD 1  LINE 49  TEXT load_half_u(R_T0, R_FaceCursor, 0 * S_(S2)),
+---   ... (one WORD line per .word emitted by the atom body) ...
+---   ENDATOM
+---   ATOM <next-name> "<abs-source-path>" <total_words>
+---   ...
+---   ENDATOM
+---   ```
+---
+--- Marker records are zero-width in `atom.paths.items`, so they emit no WORD rows in the dense word view.
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Module-scope requires + package.path setup
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- 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 elf_dwarf = require("elf_dwarf")
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Constants
+-- ════════════════════════════════════════════════════════════════════════════
+
+-- Format version emitted as the first line. Bump + add a migration test if the format changes; 
+-- the gdb runtime loader rejects mismatches (E2).
+local FORMAT_VERSION = 1
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Type declarations
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- @class AtomSourceMapCtx
+--- @field shared             table   -- `ctx.shared`
+--- @field shared.corpus      table   -- source-order registry; single writer is build_ctx
+--- @field shared.word_counts table
+--- @field out_root           string  -- output root (e.g. "build/gen")
+--- @field flags              table   -- `ctx.flags`; reads `flags.gdb_runtime` + `flags.elf_path`
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- Atom-path renderers
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Join word boundaries (from `items`) to per-word call text + source lines (from `word_events`).
+--- @param atom table
+--- @return table[], integer
+local function canonical_word_entries(atom)
+	local paths = atom.paths 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
+	end
+
+	local entries = {}
+	for index, event in ipairs(events) do
+		local item = word_items[index] or {}
+		entries[#entries + 1] = {
+			pos        = event.i or (index - 1),
+			line       = event.call_line or item.line or 0,
+			text       = event.call_text or item.call_text or "",
+			body_line  = event.body_line or item.body_line or item.line or 0,
+			invocation = (event.outermost_invocation_id
+				and paths.invocations
+				and paths.invocations[event.outermost_invocation_id]) or nil,
+		}
+	end
+	return entries, #events
+end
+
+--- Render one atom's provenance stanza. Format 1 line shapes:
+---   `WORD N  CALL <src-path>:<src-line>  MACRO <name> "<def-path>:<def-line>" BODY <line>` (component invocation)
+---   `WORD N  CALL <src-path>:<src-line>  RAW` (raw `.word` outside any mac_* component)
+--- Component identity comes from the outermost invocation record; the count-table lookup confirms the component was
+--- declared in `corpus.word_counts` (populated by word_count_eval + components passes).
+--- @param src table
+--- @param atom table
+--- @param wc table -- identity alias of corpus.word_counts
+--- @return string[], integer
+local function emit_provenance_stanza(src, atom, wc)
+	local lines    = {}
+	local rel_path = src.path:gsub("\\\\", "/")
+	local entries, total = canonical_word_entries(atom)
+
+	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
+
+	for _, entry in ipairs(entries) do
+		local inv = entry.invocation
+		local macro_count = inv and wc["mac_" .. inv.component_name]
+		if inv and macro_count ~= nil then
+			lines[#lines + 1] = string.format(
+				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				entry.pos, rel_path, entry.line, inv.component_name,
+				inv.def_path or "", inv.def_line or 0, entry.body_line)
+		else
+			lines[#lines + 1] = string.format(
+				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+		end
+	end
+
+	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
+--- @return string
+local function render_provenance(src, wc)
+	local lines = {}
+	lines[#lines + 1] = "# FORMAT_VERSION 1"
+	lines[#lines + 1] = "# auto-generated by ps1_meta.lua (passes/atoms_source_map.lua) — DO NOT EDIT"
+	lines[#lines + 1] = "# Per-.word provenance: maps each emitted .word to its call site (atom body"
+	lines[#lines + 1] = "# file:line) and, when the word was emitted by a `mac_X(...)` component invocation,"
+	lines[#lines + 1] = "# the component's definition file:line + the per-word BODY line. Used by"
+	lines[#lines + 1] = "# dwarf_injection to synthesize DW_TAG_inlined_subroutine instances + per-word"
+	lines[#lines + 1] = "# line program rows for native source-level step into component bodies."
+
+	local function append(atom)
+		local stanza = emit_provenance_stanza(src, atom, wc)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render one atom's stanza for the sourcemap.txt form (ATOM header line, N WORD lines, ENDATOM marker).
+--- Returns (lines, total_words).
+--- @param src    table
+--- @param atom   table
+--- @param wc     table
+--- @return string[], integer
+local function emit_atom_stanza(src, atom)
+	local lines    = {}
+	local rel_path = src.path:gsub("\\\\", "/")
+	local entries, total = canonical_word_entries(atom)
+
+	lines[#lines + 1] = string.format('ATOM %s "%s" 0', atom.raw_name or atom.name, rel_path)
+	for _, entry in ipairs(entries) do
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+	end
+	lines[1] = lines[1]:gsub(" 0$", " " .. tostring(total))
+	lines[#lines + 1] = "ENDATOM"
+	return lines, total
+end
+
+--- Render the full source map file content for one source (one .atoms.sourcemap.txt per source). Mirrors offsets.lua's
+--- `project_atoms` shape: scan.atoms + scan.raw_atoms, no kind filter.
+--- @param src table
+--- @param wc table
+--- @return string
+local function render_source_map(src)
+	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"
+
+	local function append(atom)
+		local stanza = emit_atom_stanza(src, atom)
+		for _, line in ipairs(stanza) do lines[#lines + 1] = line end
+	end
+	for _, atom in ipairs(src.scan.atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+	for _, atom in ipairs(src.scan.raw_atoms or {}) do
+		if atom.paths then append(atom) end
+	end
+
+	return table.concat(lines, "\n") .. "\n"
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- gdb-runtime emission (post-link, addresses via nm)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Escape a string for embedding in a gdb `set $var = "..."` literal.
+--- gdb uses C-style escaping; we escape `\` and `"` (newlines were flattened earlier).
+--- @param s string
+--- @return string
+local function gdb_escape(s)
+	return (s:gsub("\\", "\\\\"):gsub('"', '\\"'))
+end
+
+--- Build the list of atoms with addresses + word entries. Shared helper for the gdb-runtime file emission.
+--- @param ctx PassCtx
+--- @return table[] -- list of {idx, name, src_path, file_base, addr, size_bytes, words, entries}
+local function build_atom_table(ctx)
+	local addrs = elf_dwarf.read_nm(ctx.flags.elf_path)
+	local corpus = ctx.shared and ctx.shared.corpus
+	local matched = {}
+
+	for _, src in ipairs(corpus.source_order or {}) do
+		local file_base = src.path:match("([^/\\\\]+)$") or src.path
+		local function append(atom)
+			if not atom.paths then return end
+			local name = atom.raw_name or atom.name
+			local info = addrs[name]
+			if not info then return end
+			local entries, total = canonical_word_entries(atom)
+			matched[#matched + 1] = {
+				name       = name,
+				src_path   = src.path,
+				file_base  = file_base,
+				addr       = info[1],
+				size_bytes = info[2],
+				words      = total,
+				entries    = entries,
+			}
+		end
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do append(atom) end
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do append(atom) end
+	end
+
+	-- Deterministic order: sort by address (matches `nm` output ordering).
+	table.sort(matched, function(a, b) return a.addr < b.addr end)
+	for i, a in ipairs(matched) do a.idx = i - 1 end
+	return matched
+end
+
+--- Append the 9 gdb command definitions to `lines`. Pure gdb scripting — addresses come from `nm`, the convenience
+--- vars set in `emit_gdb_runtime` provide printf args, and each command is a static sequence of `printf` / `tbreak` /
+--- `if ... end` blocks. The Lua pass emits N atoms' worth of lines; runtime iteration is gdb's job.
+---
+--- 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.
+--- @param lines   table  -- output line buffer (mutated in place)
+--- @param matched table  -- list of atom records from `build_atom_table`
+local function append_gdb_commands(lines, matched)
+	-- ── tape_atoms ──
+	-- Hardcoded one printf per atom. No loop.
+	lines[#lines + 1] = "define tape_atoms"
+	for _, a in ipairs(matched) do
+		-- gdb 12.1 quirk: literals in printf args require an attached target.
+		-- Use the per-atom convenience vars set above as printf args.
+		lines[#lines + 1] = string.format('  printf "  code_%%-32s @ 0x%%08x  %%4d words\\n", $__atom_name_%d, $__atom_addr_%d, $__atom_words_%d',
+			a.idx, a.idx, a.idx)
+	end
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document tape_atoms"
+	lines[#lines + 1] = "  List every tape atom symbol in the loaded ELF (code_<name>) with .rodata addr + word count."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── break_atom (generic) + per-atom break_atom_X ──
+	lines[#lines + 1] = "define break_atom"
+	lines[#lines + 1] = '  echo "Usage: break_atom_<exact_name>  (pick from the list below)"'
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format('  printf "  break_atom_%%-32s\\n", $__atom_name_%d', a.idx)
+	end
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document break_atom"
+	lines[#lines + 1] = "  Generic help: lists the per-atom break_atom_<name> commands."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format("define break_atom_%s", a.name)
+		lines[#lines + 1] = string.format("  break *$__atom_addr_%d", a.idx)
+		lines[#lines + 1] = string.format('  printf "  Breakpoint set at code_%s (0x%%08x)\\n", $__atom_addr_%d', a.name, a.idx)
+		lines[#lines + 1] = "end"
+		lines[#lines + 1] = string.format("document break_atom_%s", a.name)
+		lines[#lines + 1] = string.format("  Set a breakpoint at code_%s.", a.name)
+		lines[#lines + 1] = "end"
+		lines[#lines + 1] = ""
+	end
+
+	-- ── step_atom / next_atom ──
+	-- Hardcoded one tbreak per atom. No loop.
+	lines[#lines + 1] = "define step_atom"
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format("  tbreak *$__atom_addr_%d", a.idx)
+	end
+	lines[#lines + 1] = "  continue"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document step_atom"
+	lines[#lines + 1] = "  Set one-shot BPs at every atom + continue. Stops at the next atom boundary."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	lines[#lines + 1] = "define next_atom"
+	lines[#lines + 1] = "  step_atom"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document next_atom"
+	lines[#lines + 1] = "  Alias for step_atom."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── where_in_atom ──
+	-- Hardcoded one outer-if per atom; inside, one inner-if per WORD entry.
+	lines[#lines + 1] = "define where_in_atom"
+	lines[#lines + 1] = "  set $__pc = (unsigned int)$pc"
+	lines[#lines + 1] = "  set $__matched = 0"
+	for _, a in ipairs(matched) do
+		-- Precompute end_addr (gdb 12.1's expression evaluator chokes on `addr + words*4`).
+		lines[#lines + 1] = string.format("  set $__end_%d = $__atom_addr_%d + $__atom_words_%d * 4", a.idx, a.idx, a.idx)
+		lines[#lines + 1] = string.format("  if $__pc >= $__atom_addr_%d && $__pc < $__end_%d", a.idx, a.idx)
+		lines[#lines + 1] = string.format('    printf "atom:    code_%%s\\n", $__atom_name_%d', a.idx)
+		lines[#lines + 1] = '    printf "addr:    0x%08x\\n", $__pc'
+		lines[#lines + 1] = string.format("    set $__word = ($__pc - $__atom_addr_%d) / 4", a.idx)
+		lines[#lines + 1] = string.format('    printf "word:    %%d/%%d\\n", $__word, $__atom_words_%d', a.idx)
+		-- One inner-if per WORD entry. Each word's line + text hardcoded.
+		for _, we in ipairs(a.entries) do
+			lines[#lines + 1] = string.format("    if $__word == %d", we.pos)
+			-- Escape TEXT for printf format string.
+			local escaped_text = we.text:gsub("%%", "%%%%"):gsub('"', '\\"')
+			lines[#lines + 1] = string.format('      printf "source:  %%s:%%d  %%s\\n", $__atom_file_%d, %d, "%s"', a.idx, we.line, escaped_text)
+			lines[#lines + 1] = "    end"
+		end
+		-- Fallback for words beyond the source map (shouldn't happen if nm matches).
+		local max_word = 0
+		if #a.entries > 0 then max_word = a.entries[#a.entries].pos end
+		lines[#lines + 1] = string.format('    if $__word > %d', max_word)
+		lines[#lines + 1] = '      printf "source:  (no source-map entry for word %%d; map may be stale)\\n", $__word'
+		lines[#lines + 1] = "    end"
+		lines[#lines + 1] = "    set $__matched = 1"
+		lines[#lines + 1] = "  end"
+	end
+	lines[#lines + 1] = "  if !$__matched"
+	lines[#lines + 1] = '    echo PC is not inside any known atom (in .text or unmapped region).'
+	lines[#lines + 1] = "  end"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document where_in_atom"
+	lines[#lines + 1] = "  Report current atom name, .rodata addr, word offset, and source line."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── stepi_inside_atom ──
+	-- Hardcoded one if-containment-check per atom (no loop).
+	-- Precompute end_addr in Lua so we don't ask gdb to evaluate `addr + words*4` inside the if condition
+	-- (gdb 12.1's expression evaluator chokes on the `*` and emits a misleading 'function malloc' error in some gdb builds).
+	lines[#lines + 1] = "define stepi_inside_atom"
+	lines[#lines + 1] = "  set $__in_atom = 0"
+	lines[#lines + 1] = "  set $__did_step = 0"
+	lines[#lines + 1] = "  set $__pc = (unsigned int)$pc"
+	for _, a in ipairs(matched) do
+		-- Precompute end_addr in the convenience var (single expression gdb handles).
+		lines[#lines + 1] = string.format("  set $__end_%d = $__atom_addr_%d + $__atom_words_%d * 4", a.idx, a.idx, a.idx)
+		lines[#lines + 1] = string.format("  if $__pc >= $__atom_addr_%d && $__pc < $__end_%d", a.idx, a.idx)
+		lines[#lines + 1] = "    set $__in_atom = 1"
+		lines[#lines + 1] = "    stepi"
+		lines[#lines + 1] = "    set $__did_step = 1"
+		lines[#lines + 1] = "  end"
+	end
+	lines[#lines + 1] = "  if !$__did_step"
+	lines[#lines + 1] = '    echo [gdb_tape_atoms] stepi_inside_atom: PC is not inside any atom; refusing to step.'
+	lines[#lines + 1] = "  end"
+	lines[#lines + 1] = "  where_in_atom"
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document stepi_inside_atom"
+	lines[#lines + 1] = "  One MIPS-instruction step, then where_in_atom. The step-and-see-source-line workflow."
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = ""
+
+	-- ── wave_ctx ──
+	lines[#lines + 1] = "define wave_ctx"
+	lines[#lines + 1] = '  printf "$t4 = R_FaceCursor     0x%08x\\n", $t4'
+	lines[#lines + 1] = '  printf "$t5 = R_VertBase       0x%08x\\n", $t5'
+	lines[#lines + 1] = '  printf "$t6 = R_OtBase         0x%08x\\n", $t6'
+	lines[#lines + 1] = '  printf "$t7 = R_PrimCursor     0x%08x\\n", $t7'
+	lines[#lines + 1] = "end"
+	lines[#lines + 1] = "document wave_ctx"
+	lines[#lines + 1] = "  Pretty-print the 4 wave-context GPRs ($t4=R_FaceCursor, $t5=R_VertBase, $t6=R_OtBase, $t7=R_PrimCursor). Requires target attached."
+	lines[#lines + 1] = "end"
+end
+
+--- Emit the gdb-runtime file (post-link). Pure gdb scripting — addresses come from `mipsel-none-elf-nm -S`, get embedded
+--- in `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`, and load via `set $var = ...` + `define ... end` blocks at gdb
+--- source-time.
+--- @param ctx PassCtx
+local function emit_gdb_runtime(ctx)
+	if not (ctx.flags and ctx.flags.gdb_runtime) then return end
+	local  elf_path = ctx.flags.elf_path
+	if not elf_path or elf_path == "" then
+		io.stderr:write("[atoms_source_map] --gdb-runtime requires --elf <elf>\n")
+		return
+	end
+	if lfs.attributes(elf_path, "mode") ~= "file" then
+		io.stderr:write(string.format(
+			"[atoms_source_map] --gdb-runtime: ELF not found at %s\n", elf_path))
+		return
+	end
+
+	local matched = build_atom_table(ctx)
+	if #matched == 0 then
+		io.stderr:write("[atoms_source_map] --gdb-runtime: no atoms matched against nm symbols (stale scan?).\n")
+		return
+	end
+
+	local lines = {}
+	lines[#lines + 1] = "# Auto-generated by ps1_meta.lua (passes/atoms_source_map.lua)"
+	lines[#lines + 1] = "# DO NOT EDIT — re-run ps1_meta.lua --atoms-source-map --gdb-runtime to regenerate"
+	lines[#lines + 1] = "# Sourced by scripts/gdb/gdb_tape_atoms.gdb (the wrapper)."
+	lines[#lines + 1] = "# Pure gdb scripting — no Python, no Tcl, no Guile required."
+	lines[#lines + 1] = "# Commands are FULLY HARDCODED per-atom because gdb doesn't do nested"
+	lines[#lines + 1] = "# `$` substitution in var names (`$foo_$i` is one literal identifier)."
+	lines[#lines + 1] = "# Per-atom convenience vars ($__atom_name_<i> etc.) are set so gdb's"
+	lines[#lines + 1] = "# `printf` has valid expression args (gdb 12.1 quirks: literals in"
+	lines[#lines + 1] = "# printf args require an attached target; convenience-var args do not)."
+	lines[#lines + 1] = string.format("# %d atoms from ELF: %s", #matched, elf_path)
+	lines[#lines + 1] = ""
+
+	-- Format version + count + ELF path (the latter is referenced by the load-line).
+	lines[#lines + 1] = "set $__atom_format_version = " .. FORMAT_VERSION
+	lines[#lines + 1] = string.format("set $__atom_count = %d", #matched)
+	lines[#lines + 1] = string.format('set $__elf_path = "%s"', gdb_escape(elf_path))
+	lines[#lines + 1] = ""
+
+	-- Per-atom convenience vars (used as printf args; literals aren't accepted
+	-- without an attached target on gdb 12.1).
+	for _, a in ipairs(matched) do
+		lines[#lines + 1] = string.format('set $__atom_name_%d = "%s"', a.idx, gdb_escape(a.name))
+		lines[#lines + 1] = string.format("set $__atom_addr_%d = 0x%x", a.idx, a.addr)
+		lines[#lines + 1] = string.format("set $__atom_words_%d = %d", a.idx, a.words)
+		lines[#lines + 1] = string.format('set $__atom_file_%d = "%s"', a.idx, gdb_escape(a.file_base))
+	end
+	lines[#lines + 1] = ""
+
+	-- The 9 commands (each `define ... end` overrides the wrapper's stub).
+	lines[#lines + 1] = "# ── 9 user commands (overrides wrapper stubs) ──"
+	append_gdb_commands(lines, matched)
+	lines[#lines + 1] = ""
+
+	-- Confirmation line for the source operator.
+	lines[#lines + 1] = 'printf "[gdb_tape_atoms] runtime loaded %d atoms from %s\\n", $__atom_count, $__elf_path'
+
+	local out_path
+	-- Move out of `<out_root>/gdb_tape_atoms_runtime.gdb` to `<out_root>/../gdb_tape_atoms_runtime.gdb` when the conventional `<out_root>` is `<build>/gen`
+	-- (any equivalent spelling — relative, absolute backslash, absolute forward-slash, trailing-separator variants).
+	-- This puts the gdb runtime alongside the ELF at `build/` rather than under the report subdir.
+	local function ends_with_gen_dir(p)
+		if type(p) ~= "string" then return false end
+		return p:match("[/\\]gen[/\\]?$") ~= nil or p == "build/gen" or p == "build\\gen"
+	end
+	if ends_with_gen_dir(ctx.out_root) then
+		-- Strip the trailing `/gen` segment, then write the runtime script under `build/`.
+		-- e.g. "C:/projects/Pikuma/ps1/build/gen" -> "C:/projects/Pikuma/ps1/build".
+		local parent = ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		out_path = parent .. "/gdb_tape_atoms_runtime.gdb"
+	else
+		out_path = ctx.out_root .. "/gdb_tape_atoms_runtime.gdb"
+	end
+	duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file_lf(out_path, table.concat(lines, "\n") .. "\n")
+	-- io.stderr:write(string.format("[atoms_source_map] wrote %s (%d atoms)\n", out_path, #matched))
+end
+
+-- ════════════════════════════════════════════════════════════════════════════
+-- M — module exports
+-- ════════════════════════════════════════════════════════════════════════════
+
+local M = {}
+
+-- Expose the pure render functions so `report.lua` and the focused tests can call them directly without triggering the file-emit path.
+M.render_source_map = render_source_map
+M.render_provenance = render_provenance
+
+--- Render ONE atom's sourcemap stanza.
+--- @param atom table   -- atom record (must have `atom.paths` populated)
+--- @return string
+function M.render_atom_source_map(atom)
+	assert(type(atom) == "table", "render_atom_source_map: atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_source_map: atom.paths must be a table")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		lines[#lines + 1] = string.format("WORD %d  LINE %d  TEXT %s",
+			entry.pos, entry.line, entry.text)
+	end
+	lines[#lines + 1] = "ENDATOM"
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render ONE atom's provenance stanza — no per-file format header, no enumeration of other atoms.
+---
+--- `rel_path` is the source path (forward-slashes) embedded in every `CALL` line.
+--- The .md caller (report.lua) is expected to derive this once per `## <source>` heading and pass it down for each atom in that source.
+--- @param atom      table   -- atom record (must have `atom.paths` populated)
+--- @param wc        table   -- identity alias of `corpus.word_counts`
+--- @param rel_path  string  -- source path (forward-slashes) for `CALL` fields
+--- @return string
+function M.render_atom_provenance(atom, wc, rel_path)
+	assert(type(atom)       == "table", "render_atom_provenance:  atom must be a table")
+	assert(type(atom.paths) == "table", "render_atom_provenance:  atom.paths must be a table")
+	assert(type(rel_path)   == "string", "render_atom_provenance: rel_path must be a string")
+	local entries, total = canonical_word_entries(atom)
+	local lines = {}
+	lines[#lines + 1] = string.format("ATOM %s  %d", (atom.raw_name or atom.name), total)
+	for _, entry in ipairs(entries) do
+		local inv = entry.invocation
+		local macro_count = inv and wc and wc["mac_" .. inv.component_name]
+		if inv and macro_count ~= nil then
+			lines[#lines + 1] = string.format(
+				'WORD %d  CALL %s:%d  MACRO %s "%s:%d" BODY %d',
+				entry.pos, rel_path, entry.line, inv.component_name,
+				inv.def_path or "", inv.def_line or 0, entry.body_line)
+		else
+			lines[#lines + 1] = string.format(
+				"WORD %d  CALL %s:%d  RAW", entry.pos, rel_path, entry.line)
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Pass entry. For each source that declares at least one `MipsAtom_(name)` / `MipsCode code_<name>`,
+--- emit two files in `<out_root>/`: `<basename>.atoms.sourcemap.txt` (per-word call-site map) and `<basename>.atoms.provenance.txt`
+--- (per-word definition + body line, resolved via the outermost `mac_X(...)` invocation).
+--- When `ctx.flags.gdb_runtime` is true and `ctx.flags.elf_path` exists, also emit the post-link gdb script `<ctx.out_root>/gdb_tape_atoms_runtime.gdb`.
+--- @param ctx PassCtx
+--- @return PassResult
+function M.run(ctx)
+	local outputs  = {}
+	local errors   = {}
+	local warnings = {}
+
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+		error("atoms_source_map.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	end
+
+	-- Word counts come from `corpus.word_counts` (populated by word_count_eval + components passes).
+	local wc = corpus.word_counts or {}
+	if not next(wc) then
+		warnings[#warnings + 1] = {
+			line = 0,
+			msg  = "atoms_source_map: corpus.word_counts is empty; the word-counts + components passes may not have populated it. Check the PASSES dep edges.",
+		}
+	end
+
+	-- atoms.sourcemap.txt + atoms.provenance.txt content moved to report.lua via `<module>.atoms.md` markdown file.
+	-- This pass emits only the post-link gdb_runtime artifact (see emit_gdb_runtime below).
+
+	-- Optionally emit the gdb-runtime form (post-link, one file per build).
+	if ctx.flags and ctx.flags.gdb_runtime then
+		emit_gdb_runtime(ctx)
+	end
+
+	return { outputs = outputs, errors = errors, warnings = warnings }
+end
+
+return M

scripts/passes/components.lua

@@ -1,38 +1,27 @@
 --- passes/components.lua — Component-macro header generator.
 ---
---- Reads the pre-scanned SourceScan payload (produced once upstream by `duffle.scan_source`)
---- for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations, then does
---- per-source backward lookups for the function-args string (from the preceding `FI_ MipsAtom ac_X(...)`
---- function declaration) and the preceding comment block (for LSP/IntelliSense signature docs).
+--- 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.
 ---
---- Emits a per-directory `<dir_basename>.macs.h` containing one `#define mac_X(sig) \` macro per component
---- + `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
+--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations,
+--- then resolves the function-args string from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration via a backward walk.
+---
+--- Emits one `<dir_basename>.macs.h` per source with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
 
---- @class Component
---- @field name    string
---- @field body    string
---- @field args    string|nil
---- @field line    integer
---- @field comment string|nil
-
---- @class M
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Bootstrap: same as entry scripts. See `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
--- both standalone and when require'd from the orchestrator.
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
-local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
+local _bootstrap_dir  = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
+local duffle          = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -40,12 +29,12 @@ local word_count_eval = require("word_count_eval")
 
 -- Atom component declaration identifiers.
 local ATOM_COMP_PROC  = "MipsAtomComp_Proc_"
-local MIPS_ATOM       = "MipsAtom"  -- prefix on the function declaration that wraps an AtomComp_Proc_
+local MIPS_ATOM       = "Slice_MipsCode"  -- prefix on the function declaration that wraps an AtomComp_Proc_
 
 -- Component-name prefixes.
-local AC_PREFIX       = "ac_"        -- arg to MipsAtomComp_(ac_X); the X is the atom name
+local AC_PREFIX       = "ac_"  -- arg to MipsAtomComp_(ac_X); the X is the atom name
 local AC_PREFIX_LEN   = 3
-local MAC_PREFIX      = "mac_"       -- prefix on generated macros; the rest is the atom name
+local MAC_PREFIX      = "mac_" -- prefix on generated macros; the rest is the atom name
 local MAC_PREFIX_LEN  = 4
 
 -- ASCII byte values used in tokenization.
@@ -70,12 +59,10 @@ local GEN_SUBDIR      = "gen"
 --- @field sources            SourceFile[]           -- all source files in the build
 --- @field metadata_path      string                 -- path to word_count.metadata.h
 --- @field shared             table                  -- cross-pass shared state
---- @field shared.word_counts table<string, integer> -- populated by word-counts + components
 --- @field out_root           string                 -- output root (e.g. "build/gen")
 --- @field project_root       string                 -- project root (e.g. "code/")
 --- @field upstream           table<string, table>   -- per-pass upstream outputs
 --- @field flags              table                  -- CLI flags
---- @field dry_run            boolean                -- if true, compute but don't write
 --- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
@@ -84,11 +71,13 @@ local GEN_SUBDIR      = "gen"
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class Component
---- @field name    string        -- atom name (without `ac_` prefix)
---- @field body    string        -- brace-delimited body (without the braces)
---- @field args    string|nil    -- function-args string (function form only)
---- @field line    integer       -- source line of the declaration
---- @field comment string|nil    -- preceding `/* */` or `//` comment block (signature doc)
+--- @field name       string     -- atom name (without `ac_` prefix)
+--- @field body       string     -- brace-delimited body (without the braces)
+--- @field args       string|nil -- function-args string (function form only)
+--- @field line       integer    -- source line of the declaration
+--- @field comment    string|nil -- scanner-owned `declaration_comment`; the components pass reads it from the scanner record
+--- @field kind       string     -- "comp_bare" | "comp_proc"
+--- @field debug_skip boolean    -- mirror of `a.debug_skip` (scanner-owned); true iff a bare `atom_dbg_skip` marker immediately preceded the declaration
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Local helpers (file I/O + path normalization)
@@ -97,44 +86,39 @@ local GEN_SUBDIR      = "gen"
 local M = {}
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Function-args extraction (precedes MipsAtomComp_Proc_ invocations)
+-- Back-walk helpers (composed into the entry point below: find_function_args_for)
+--
+-- Only the function-args lookup for proc components occurs here.
+-- The preceding-comment walk occur in `scan_source.lua` — `a.declaration_comment` carries the resolved comment,
+-- so this file reads it forward rather than re-walking the source.
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
--- Returns the position of the open paren, or nil if not found.
--- @param source string
--- @param name string
--- @param before_pos integer
--- @return integer|nil
-local function find_last_name_open_paren(source, name, before_pos)
-	local search     = source:sub(1, before_pos)
-	local name_open  = name .. "("
-	local last_idx   = nil
-	local scan_pos   = 1
-	while true do
-		local found = search:find(name_open, scan_pos, true)  -- plain (no regex)
-		if not found then break end
-		last_idx = found
-		scan_pos = found + #name_open
-	end
-	return last_idx
-end
-
 --- Find the args of the function declaration that immediately precedes a `MipsAtomComp_Proc_` invocation of the given name.
 --- Returns the args string (e.g., `"U4 off, U4 code, U1 r, U1 g, U1 b"`) or nil if no function declaration is found.
 ---
 --- Convention: function form is
----   `FI_ MipsAtom ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
+---   `FI_ Slice_MipsCode ac_X(args) MipsAtomComp_Proc_(ac_X, { body })`
 --- We find the LAST occurrence of `"ac_X("` before `before_pos` and extract the args from inside the parens.
---- We then verify the preceding context ends with `MipsAtom` (the function-decl keyword
---- with possible qualifiers between).
+--- We then verify the preceding context ends with `Slice_MipsCode`
+--- (the function-decl keyword with possible qualifiers between).
 ---
---- @param source string
---- @param name string
+--- @param source     string
+--- @param name       string
 --- @param before_pos integer
 --- @return string|nil
 local function find_function_args_for(source, name, before_pos)
-	local last_idx = find_last_name_open_paren(source, name, before_pos)
+	-- Find the LAST occurrence of `name + "("` in `source[1..before_pos]`.
+	local name_open  = name .. "("
+	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).
@@ -153,178 +137,14 @@ local function find_function_args_for(source, name, before_pos)
 	return inner
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Preceding-comment-block extraction
--- ════════════════════════════════════════════════════════════════════════════
-
--- Skip whitespace (space/tab/newline/CR) backward from `pos`, returning the position of the first non-whitespace char.
--- @param source string
--- @param pos integer
--- @return integer
-local function skip_ws_backward(source, pos)
-	local back = pos - 1
-	while back > 0 do
-		local ch = source:sub(back, back)
-		if ch == " " or ch == "\t" or ch == "\n" or ch == "\r" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
--- Find the opening `/*` for a block comment whose `*/` ends at `close_pos`.
--- Returns the position of `/`, or nil if not found.
--- @param source string
--- @param close_pos integer  -- position of the closing `*` of `*/`
--- @return integer|nil
-local function find_block_comment_open(source, close_pos)
-	local prefix  = source:sub(1, close_pos - 1)
-	local open_at = nil
-	for scan = #prefix - 1, 1, -1 do
-		if prefix:sub(scan, scan + 1) == "/*" then
-			open_at = scan
-			break
-		end
-	end
-	return open_at
-end
-
--- Walk back from `open_at` over leading spaces + tabs to include the indentation before the `/*` in the captured comment.
--- @param source string
--- @param open_at integer
--- @return integer
-local function extend_left_over_indent(source, open_at)
-	local start = open_at
-	while start > 1 do
-		local ch = source:sub(start - 1, start - 1)
-		if ch == " " or ch == "\t" then
-			start = start - 1
-		else
-			break
-		end
-	end
-	return start
-end
-
--- Walk back from `line_end` to the start of the source line (the most recent `\n` or position 1).
--- @param source string
--- @param line_end integer
--- @return integer
-local function find_line_start(source, line_end)
-	local start = line_end
-	while start > 1 and source:sub(start - 1, start - 1) ~= "\n" do
-		start = start - 1
-	end
-	return start
-end
-
--- (internal) Capture one `/* ... */` block comment whose closing `*/`
--- ends at `close_end_pos`. Returns (block_text, new_scan_pos) where `new_scan_pos`
--- is where to continue scanning for more comments, or nil if no block comment was found.
-local function capture_block_comment(source, close_end_pos)
-	local open_at = find_block_comment_open(source, close_end_pos)
-	if not open_at then return nil end
-	local block_start = extend_left_over_indent(source, open_at)
-	return source:sub(block_start, close_end_pos), block_start
-end
-
--- (internal) Capture one `// ...` line comment ending at `line_end_pos`.
--- Returns (comment_text, new_scan_pos) or nil if the line is not a `//` comment.
-local function capture_line_comment(source, line_end_pos)
-	local line_start = find_line_start(source, line_end_pos)
-	local line       = source:sub(line_start, line_end_pos)
-	if line:sub(1, 2) == "//" then
-		return line, line_start - 1
-	end
-	return nil
-end
-
---- Find the contiguous comment block immediately preceding `pos` in `source`.
---- Returns the comment text (with the `/* */` or `//` markers preserved) or an empty string if no comment is adjacent.
----
---- Used to copy signature comments from the source declaration (`MipsAtomComp_` / `MipsAtomComp_Proc_` / function decl)
---- over to the generated `mac_X` macro, so LSP/IntelliSense displays the args doc.
----
---- @param source string
---- @param pos integer
---- @return string
-local function preceding_comment_block(source, pos)
-	local scan_pos = pos
-	local pieces   = {}
-	while true do
-		local non_ws = skip_ws_backward(source, scan_pos)
-		if non_ws == 0 then break end
-
-		local is_block_close = non_ws >= 2 and source:sub(non_ws - 1, non_ws) == "*/"
-		local is_line_end    = source:sub(non_ws, non_ws) == "\n" or source:sub(non_ws, non_ws) == "\r"
-
-		if is_block_close then
-			local block_text, new_scan_pos = capture_block_comment(source, non_ws)
-			if not block_text then break end
-			table.insert(pieces, 1, block_text)
-			scan_pos = new_scan_pos
-		elseif is_line_end then
-			local line_text, new_scan_pos = capture_line_comment(source, non_ws)
-			if not line_text then break end
-			table.insert(pieces, 1, line_text)
-			scan_pos = new_scan_pos
-		else
-			break
-		end
-	end
-	if #pieces == 0 then return "" end
-	return table.concat(pieces, "\n")
-end
-
 -- ════════════════════════════════════════════════════════════════════════════
 -- Argument-name extraction
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Walk `trimmed` backward from `pos` over trailing whitespace / asterisks / brackets, returning the position of the first
--- non-trailer character (i.e. the end of the identifier).
--- @param trimmed string
--- @param pos integer
--- @return integer
-local function trim_trailer_back(trimmed, pos)
-	local back = pos
-	while back > 0 do
-		local ch = trimmed:sub(back, back)
-		if ch == " " or ch == "\t" or ch == "*" or ch == "]" or ch == "[" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
--- Walk `trimmed` backward from `pos` over identifier chars (alnum + `_`),
--- returning the position just before the identifier starts.
--- @param trimmed string
--- @param pos integer
--- @return integer
-local function trim_ident_back(trimmed, pos)
-	local back = pos
-	while back > 0 do
-		local ch = trimmed:sub(back, back)
-		if duffle.is_alnum(ch) or ch == "_" then
-			back = back - 1
-		else
-			break
-		end
-	end
-	return back
-end
-
 --- Extract just the parameter NAMES from a function-args string (stripping type annotations). E.g.,
 ---   `"U4 off, U4 code, U1 r, U1 g, U1 b"`  ->  `{"off", "code", "r", "g", "b"}`
 ---   `"U4 *ptr"`                            ->  `{"ptr"}`
 ---   `""`                                   ->  nil
----
---- No regex — uses `duffle.is_alnum` + plain string ops.
----
 --- @param args_str string|nil
 --- @return string[]|nil
 local function extract_arg_names(args_str)
@@ -334,8 +154,27 @@ local function extract_arg_names(args_str)
 	for _, tok in ipairs(tokens) do
 		local trimmed = duffle.trim(tok)
 		if trimmed ~= "" then
-			local ident_end = trim_trailer_back(trimmed, #trimmed)
-			local ident_start = trim_ident_back(trimmed, ident_end) + 1
+			-- Find the identifier at the end: walk back over trailers (whitespace + `*` + `[]`),
+			-- then walk back over the identifier chars (alnum + `_`).
+			local ident_end = #trimmed
+			while ident_end > 0 do
+				local ch = trimmed:sub(ident_end, ident_end)
+				if ch == " " or ch == "\t" or ch == "*" or ch == "]" or ch == "[" then
+					ident_end = ident_end - 1
+				else
+					break
+				end
+			end
+			local ident_start = ident_end
+			while ident_start > 0 do
+				local ch = trimmed:sub(ident_start, ident_start)
+				if duffle.is_alnum_byte(string.byte(ch)) or ch == "_" then
+					ident_start = ident_start - 1
+				else
+					break
+				end
+			end
+			ident_start = ident_start + 1
 			local name = trimmed:sub(ident_start, ident_end)
 			if name ~= "" then names[#names + 1] = name end
 		end
@@ -348,26 +187,34 @@ end
 -- Component projection (read from pre-scanned SourceScan)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
--- Does per-source backward lookups for args (preceding function decl) and comment (preceding comment block).
--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table
--- instead of calling duffle.tokenize_body again.
--- @param source string  -- the full source text (needed for backward lookups)
--- @param scan  table   -- SourceScan from duffle.scan_source
--- @return Component[]
+--- Project pre-scanned MipsAtomComp_ / MipsAtomComp_Proc_ entries into Component shape.
+--- Reads the scanner-owned `declaration_comment` (resolved by scan_source.lua, skipping backward across an associated bare `atom_dbg_skip` marker when present).
+--- Per-source backward lookups remain in place only for the function `args` of proc components.
+--- That lookup is unique to components.lua and stays separate from the declaration-comment walk.
+--- Carries `body_tokens` forward from scan-source so word_count_rec reads from the precomputed table instead of calling duffle.tokenize_body again.
+--- Carries the scanner-owned `debug_skip` flag forward so the generated projection can emit `/* atom_dbg_skip */`
+--- before the authored comment and so `update_canonical_components` can mirror the same field onto `corpus.components[name]`.
+--- @param source string  -- the full source text (needed for backward lookups)
+--- @param scan   table   -- SourceScan from duffle.scan_source
+--- @return Component[]
 local function project_components(source, scan)
 	local out = {}
 	for _, a in ipairs(scan.atoms) do
 		if a.kind == "comp_bare" or a.kind == "comp_proc" then
-			local args    = find_function_args_for(source, a.raw_name, a.ident_pos)
-			local comment = preceding_comment_block(source, a.ident_pos)
+			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.
+			local comment = a.declaration_comment or ""
 			out[#out + 1] = {
 				line        = a.line,
 				name        = a.name,
 				body        = a.body,
+				body_off    = a.body_off,
 				body_tokens = a.body_tokens,
 				args        = args,
 				comment     = comment,
+				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
+				debug_skip  = a.debug_skip == true,
 			}
 		end
 	end
@@ -379,7 +226,6 @@ end
 -- ════════════════════════════════════════════════════════════════════════════
 
 -- Convert `//` line comments to `/* */` block comments in a token.
---
 -- C macros use `\` line-continuations; a `//` comment before `\` would consume the continuation,
 -- breaking the macro. We convert `//` to `/* */` so the multi-line macro structure is preserved.
 --
@@ -403,8 +249,8 @@ local function convert_line_comments_to_block(s)
 			while eol <= len and result:byte(eol) ~= BYTE_NEWLINE do
 				eol = eol + 1
 			end
-			local before   = result:sub(1, pos - 1)
-			local comment  = result:sub(pos + 2, eol - 1)  -- skip the `//`
+			local before  = result:sub(1, pos - 1)
+			local comment = result:sub(pos + 2, eol - 1)  -- skip the `//`
 			local after
 			if eol <= len and result:byte(eol) == BYTE_NEWLINE then
 				after = " */" .. result:sub(eol)  -- keep the newline
@@ -422,10 +268,11 @@ end
 -- Word-count computation (memoized recursive lookup)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. Returns the ident unchanged
--- if it doesn't start with the prefix (so a non-component ident like `mask_upper` falls through to the wc-table branch).
--- @param ident string|nil
--- @return string|nil
+--- Strip the `mac_` prefix from a component-call ident so we can look it up against the components-by-name table. 
+--- Returns the ident unchanged if it doesn't start with the prefix 
+--- (so a non-component ident like `mask_upper` falls through to the wc-table branch).
+--- @param ident string|nil
+--- @return string|nil
 local function strip_mac_prefix(ident)
 	if not ident then return nil end
 	if ident:sub(1, MAC_PREFIX_LEN) == MAC_PREFIX then
@@ -434,13 +281,13 @@ local function strip_mac_prefix(ident)
 	return ident
 end
 
--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
--- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
--- @param name string -- the component name (without `mac_`)
--- @param comp_by_name table<string, Component>
--- @param wc table<string, integer>
--- @param cache table<string, integer>
--- @return integer
+--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
+--- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
+--- @param name         string -- the component name (without `mac_`)
+--- @param comp_by_name table<string, Component>
+--- @param wc           table<string, integer>
+--- @param cache        table<string, integer>
+--- @return integer
 local function word_count_rec(name, comp_by_name, wc, cache)
 	if cache[name] ~= nil then return cache[name] end
 	cache[name] = -1  -- mark in-progress (cycle detection)
@@ -475,13 +322,11 @@ end
 
 --- Compute word counts for every component in `components` in a single pass.
 --- The name-lookup table + memoization cache are built ONCE (per source) instead of per-component,
---- so the cache survives across siblings and a component's recursive `mac_Y(...)` references hit memoized values
---- instead of re-walking the body. Previously each call rebuilt both tables (O(N) tables per call → O(N^2)).
----
+--- so the cache survives across siblings and a component's recursive `mac_Y(...)` 
+--- references hit memoized values instead of re-walking the body.
 --- Cycle detection (A -> B -> A) is preserved via the in-progress `-1` sentinel in `cache`.
----
 --- @param components Component[]
---- @param wc table<string, integer>
+--- @param wc         table<string, integer>
 --- @return table<string, integer>  -- map of component name (without `mac_`) -> word count
 local function count_all_components(components, wc)
 	local comp_by_name = {}
@@ -518,19 +363,12 @@ local function split_comment_lines(s)
 	return out
 end
 
---- Split an atom body by top-level commas; drop empty tokens.
---- @param body string
---- @return string[]
-local function tokens_from_body(body)
-	return duffle.tokenize_body_simple(body)
-end
-
 --- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
 --- @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
+	if    arg_names and #arg_names > 0 then
 		return table.concat(arg_names, ", ")
 	end
 	return "..."
@@ -540,7 +378,7 @@ end
 --- The last 2 chars are always that pair.
 local function strip_trailing_continuation(lines)
 	local last = lines[#lines]
-	if last:sub(-2) == " \\" then
+	if    last:sub(-2) == " \\" then
 		lines[#lines] = last:sub(1, -3)
 	end
 end
@@ -561,20 +399,31 @@ end
 
 --- Build the list of lines for one component
 --- (signature comment, `#define mac_X(...)` line with backslash-continued tokens, then `WORD_COUNT(mac_X, N)` entry).
---- @param c Component
+--- For skipped components, a `/* atom_dbg_skip */` marker comment is emitted immediately before the authored comment block.
+--- The marker is a single line, the comment comes next, and the `#define` line follows. The `debug_skip` stamp is scanner-owned
+--- (`a.debug_skip == true` on the declaration record); the components pass projects it directly.
+--- @param c          Component
 --- @param components Component[]
---- @param wc table<string, integer>
+--- @param wc         table<string, integer>
 --- @return string[]  -- list of lines for this component
 local function build_component_lines(c, counts)
 	local lines = {}
 
+	-- Marker comment: emitted once for every skipped component.
+	-- The marker is scanner-owned (declared by `atom_dbg_skip` immediately before the declaration in the source);
+	-- the components pass projects `c.debug_skip` and emits the marker as a generated comment.
+	if c.debug_skip then
+		lines[#lines + 1] = "/* atom_dbg_skip */"
+	end
+
 	if c.comment and c.comment ~= "" then
 		for _, line in ipairs(split_comment_lines(c.comment)) do
 			lines[#lines + 1] = line
 		end
 	end
 
-	local tokens = tokens_from_body(c.body)
+	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)
 	-- Direct lookup against the per-source precomputed `counts` table (built once by count_all_components).
 	local n      = counts[c.name]
@@ -594,10 +443,10 @@ end
 -- Per-source emit logic
 -- ════════════════════════════════════════════════════════════════════════════
 
--- 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
--- @return string[]
+--- 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
+--- @return string[]
 local function header_boilerplate(src)
 	return {
 		-- #pragma once wrapped in #ifdef INTELLISENSE_DIRECTIVES, matching the convention in lottes_tape.h.
@@ -610,8 +459,8 @@ local function header_boilerplate(src)
 		"// Component atoms (MipsAtomComp_(ac_*)) -> macro variants (mac_*)",
 		"",
 		-- Self-contained: define WORD_COUNT if not already defined.
-		-- We use the same definition here so the auto-generated entries below expand to compile-time constants whether
-		-- the metadata file is included first or not.
+		-- We use the same definition here so the auto-generated entries below expand
+		-- to compile-time constants whether the metadata file is included first or not.
 		"#ifndef WORD_COUNT",
 		"#define WORD_COUNT(name, count)  enum { words_##name = (count) };",
 		"#endif",
@@ -619,32 +468,28 @@ local function header_boilerplate(src)
 	}
 end
 
--- Compute the output path for one source's `.macs.h` file.
--- The pre-rework convention uses the *directory* basename
--- (not the source file basename) — e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
--- This matches what the C codebase #includes.
--- @param src SourceFile
--- @return string  -- the output directory
--- @return string  -- the full output path
+--- Compute the output path for one source's `.macs.h` file.
+--- The pre-rework convention uses the *directory* basename (not the source file basename)
+---  e.g. `code/duffle/lottes_tape.h` produces `code/duffle/gen/duffle.macs.h`.
+--- This matches what the C codebase #includes.
+--- @param src SourceFile
+--- @return string  -- the output directory
+--- @return string  -- the full output path
 local function compute_macs_h_path(src)
-	local out_dir  = src.dir .. "/" .. GEN_SUBDIR
-	local out_path = out_dir .. "/" .. duffle.basename_no_ext(src.dir) .. ".macs.h"
+	local  out_dir  = src.dir .. "/" .. GEN_SUBDIR
+	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. Writes in BINARY mode so LF line endings are
---- preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
----
---- Honors `ctx.dry_run`: prints the intended path but does not write the file.
----
---- @param ctx PassCtx
---- @param src SourceFile
+--- Emit a per-source `.macs.h` header with the `mac_X` macros + `WORD_COUNT` entries. 
+--- Writes in BINARY mode so LF line endings are preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
+--- @param ctx        PassCtx
+--- @param src        SourceFile
 --- @param components Component[]
---- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
 --- @return string|nil  -- path to the written file (nil if no components)
 local function emit_component_macros_h(ctx, src, components, counts)
 	if #components == 0 then return nil end
-
 	local out_dir, out_path = compute_macs_h_path(src)
 	local lines             = header_boilerplate(src)
 
@@ -655,12 +500,6 @@ local function emit_component_macros_h(ctx, src, components, counts)
 	end
 
 	local content = table.concat(lines, "\n") .. "\n"
-
-	if ctx.dry_run then
-		print(string.format("  -> %s (dry-run)", out_path))
-		return out_path
-	end
-
 	duffle.ensure_dir(out_dir)
 	duffle.write_file_lf(out_path, content)
 	print(string.format("  -> %s", out_path))
@@ -671,15 +510,90 @@ end
 -- Pass entry
 -- ════════════════════════════════════════════════════════════════════════════
 
--- (internal) Extend `ctx.shared.word_counts` with this source's component macros
--- so offsets sees them without re-reading the file.
--- @param ctx PassCtx
--- @param components Component[]
--- @param counts table<string, integer>  -- precomputed word counts (from count_all_components)
-local function update_shared_word_counts(ctx, components, counts)
-	local wc = ctx.shared.word_counts
+--- (internal) Extend `corpus.word_counts` with this source's component macros so offsets sees them without re-reading the file.
+--- First declaration wins: a later caller's count is dropped (the existing entry from the first source is preserved).
+--- @param corpus     table  -- the corpus
+--- @param components Component[]
+--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
+local function update_canonical_word_counts(corpus, components, counts)
+	local wc = corpus.word_counts
 	for _, c in ipairs(components) do
-		wc["mac_" .. c.name] = counts[c.name]
+		local key = "mac_" .. c.name
+		if wc[key] == nil then
+			wc[key] = counts[c.name]
+		end
+	end
+end
+
+--- @class ComponentDef
+--- @field name       string  -- bare name (without ac_/mac_ prefix)
+--- @field line       integer -- definition source line (line of `MipsAtomComp_(ac_X)` / `MipsAtomComp_Proc_(ac_X, ...)`)
+--- @field path       string  -- absolute source path of the definition
+--- @field kind       string  -- "comp_bare" | "comp_proc"
+--- @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").
+--- The pass does NOT write to `ctx.shared.components` (ownership follows the canonical contract).
+--- The `debug_skip` field mirrors the scanner-owned declaration record (`c.debug_skip`).
+--- No parallel skip map is built here; consumers that need the per-component skip state read `corpus.components[name].debug_skip` directly.
+--- @param corpus     table  -- the corpus
+--- @param src        SourceFile
+--- @param components Component[]
+local function update_canonical_components(corpus, src, components)
+	local rel_path = src.path:gsub("\\", "/")
+	for _, c in ipairs(components) do
+		-- Keyed by bare name (e.g. `yield`, `load_tri_indices`).
+		-- The atoms_source_map pass looks up components by bare name from the corpus;
+		-- `mac_` prefix lives at the call-site identifier and is stripped before lookup.
+		if corpus.components[c.name] == nil then
+			corpus.components[c.name] = {
+				name       = c.name,
+				line       = c.line,
+				path       = rel_path,
+				kind       = c.kind or "comp_bare",
+				debug_skip = c.debug_skip == true,
+			}
+		else
+			-- A second declaration of the same bare name: record a typed collision so static-analysis + the report can surface it.
+			-- Identical-shape declarations (same path + line) reuse the first-wins entry without a collision record.
+			local existing = corpus.components[c.name]
+			if existing.path ~= rel_path or existing.line ~= c.line then
+				local kind     = c.kind or "comp_bare"
+				local first_kind = existing.kind or "comp_bare"
+				corpus.collisions[#corpus.collisions + 1] = {
+					kind              = "component",
+					name              = c.name,
+					first_site        = { path = existing.path, line = existing.line },
+					conflicting_site  = { path = rel_path,       line = c.line },
+					first_shape       = "kind=" .. first_kind,
+					conflicting_shape = "kind=" .. kind,
+				}
+			end
+		end
+	end
+end
+
+--- (internal) Populate `corpus.component_body_index` with this source's body index entries.
+--- First declaration wins; later declarations are dropped (no separate collision record: the components collision is already surfaced by `update_canonical_components`).
+--- The pass writes to `corpus.component_body_index` only (the corpus owns this projection).
+--- @param corpus     table  -- the corpus
+--- @param src        SourceFile
+--- @param components Component[]
+--- @param scan       table  -- the SourceScan payload (for line_of)
+local function update_canonical_component_body_index(corpus, src, components, scan)
+	local line_of = scan and scan.line_of
+	for _, c in ipairs(components) do
+		if corpus.component_body_index[c.name] == nil then
+			corpus.component_body_index[c.name] = {
+				body_tokens = c.body_tokens,
+				body_off    = c.body_off,
+				line_of     = line_of,
+				source      = src.path,
+				declaration = c.line,
+				kind        = c.kind,
+			}
+		end
 	end
 end
 
@@ -690,16 +604,41 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	for _, src in ipairs(ctx.sources) do
+	-- Corpus ownership gate.
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("components.run requires ctx.shared.corpus.", 0)
+	end
+	if type(corpus.source_order) ~= "table" then
+		error("components.run requires ctx.shared.corpus.source_order.", 0)
+	end
+	if type(corpus.word_counts) ~= "table" then
+		error("components.run requires ctx.shared.corpus.word_counts; "
+			.. "word_count_eval.run must run before components.run "
+			.. "(see PASSES deps).", 0)
+	end
+
+	-- Projection ownership:
+	--   * `corpus.word_counts["mac_"..name]` — current component count
+	--   * `corpus.components[name]`         — bare-name component definition
+	--   * `corpus.component_body_index[name]` — body / line_of / source index
+	-- The pass writes to the corpus only; consumers read from the corpus directly.
+
+	for _, src in ipairs(corpus.source_order) do
 		-- project_components reads from src.scan + does backward lookups on src.text
 		local components = project_components(src.text, src.scan)
 		if #components > 0 then
-			-- Compute word counts for ALL components once (was: rebuilt per call inside the helpers).
-			local counts = count_all_components(components, ctx.shared.word_counts)
+			-- Compute all component word counts once per source.
+			-- Use `corpus.word_counts` so the recursive lookup sees both authored-metadata entries
+			-- (loaded by word_count_eval.run) AND same-source component entries (populated earlier in this loop by `update_canonical_word_counts`).
+			local counts = count_all_components(components, corpus.word_counts)
 			local macs_path = emit_component_macros_h(ctx, src, components, counts)
 			if macs_path then
 				outputs[#outputs + 1] = { macs_h = macs_path }
-				update_shared_word_counts(ctx, components, counts)
+				-- Populate the projections AFTER disk emission (so the byte-identical `.macs.h` contract is preserved before any current-count mutation).
+				update_canonical_word_counts(corpus, components, counts)
+				update_canonical_components(corpus, src, components)
+				update_canonical_component_body_index(corpus, src, components, src.scan)
 			end
 		end
 	end

scripts/passes/emission_model.lua

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

scripts/passes/offsets.lua

@@ -21,18 +21,12 @@
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-local word_count_eval = require("word_count_eval")
-local count_token_words = word_count_eval.count_token_words
+local duffle = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Marker-call identifiers inside atom bodies.
-local LABEL_MARKER    = "atom_label"
-local OFFSET_MARKER   = "atom_offset"
-
 -- Offset macro/enum naming prefixes (the emitted header uses these).
 local OFFSET_MACRO_PREFIX = "_atom_offset_"
 local OFFSET_ENUM_PREFIX  = "atom_offset_"
@@ -52,16 +46,10 @@ local OFFSET_MACRO_COL = 44
 --- @field scan     table   -- pre-scanned SourceScan payload (from duffle.scan_source)
 
 --- @class PassCtx
---- @field sources            SourceFile[]           -- all source files in the build
---- @field metadata_path      string                 -- path to word_count.metadata.h
 --- @field shared             table                  -- cross-pass shared state
---- @field shared.word_counts table                  -- macro name -> word count
+--- @field shared.corpus      table                  -- canonical corpus projection
+--- @field shared.word_counts table
 --- @field out_root           string                 -- output root (e.g. "build/gen")
---- @field project_root       string                 -- project root (e.g. "code/")
---- @field upstream           table<string, table>   -- per-pass upstream outputs
---- @field flags              table                  -- CLI flags
---- @field dry_run            boolean                -- if true, compute but don't write
---- @field verbose            boolean                -- log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -69,10 +57,10 @@ local OFFSET_MACRO_COL = 44
 --- @field warnings table[]  -- {line=, msg=} entries; build-succeeds
 
 --- @class BranchOffset
---- @field tag    string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
---- @field target string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
---- @field pos    integer  -- the branch's word position within the atom body
---- @field offset integer  -- computed `target_word - branch_word - 1`
+--- @field tag         string   -- the marker tag (e.g. "F" in `atom_offset(F, T)`)
+--- @field target      string   -- the target label name (e.g. "T" in `atom_offset(F, T)`)
+--- @field branch_word integer  -- branch word position within the atom body
+--- @field offset      integer  -- computed `target_word - branch_word - 1`
 
 --- @class AtomData
 --- @field name        string         -- atom name
@@ -80,188 +68,85 @@ local OFFSET_MACRO_COL = 44
 --- @field offsets     BranchOffset[] -- per-branch offset list
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Per-token marker-call helpers (atom_label / atom_offset inside bodies)
+-- Canonical marker projection
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Extract comma-separated identifier args from a parenthesized group after a function-like macro call.
--- Returns (args, after_paren) where `after_paren` is the position just past the closing `)`, or nil if `token` did not start with `(`.
--- @param token string
--- @param after_ident integer
--- @return string[], integer|nil
-local function extract_ident_args(token, after_ident)
-	local arg_start = duffle.skip_ws_and_cmt(token, after_ident)
-	if token:sub(arg_start, arg_start) ~= "(" then return {}, nil end
-	local inner, after_paren = duffle.read_parens(token, arg_start)
-	-- scan: <marker>(<args>)
+-- MARKER_PROJECTORS is the marker-kind data table.
+-- The emission-model pass already records marker word positions;
+-- this pass only projects those records into the label/branch lookup shape needed by offset computation.
+local MARKER_PROJECTORS = {
+	label = function(state, marker)
+		state.labels[marker.name] = marker.word_index
+	end,
+	offset = function(state, marker)
+		state.branches[#state.branches + 1] = {
+			tag         = marker.name,
+			target      = marker.target,
+			branch_word = marker.word_index,
+		}
+	end,
+}
 
-	local args      = {}
-	local pos       = 1
-	local inner_len = #inner
-	while pos <= inner_len do
-		pos = duffle.skip_ws_and_cmt(inner, pos)
-		if pos > inner_len then break end
-		local ident, after = duffle.read_ident(inner, pos)
-		if ident and ident ~= "" then
-			table.insert(args, ident)
-			pos = after
-		else
-			pos = pos + 1
-		end
-		pos = duffle.skip_ws_and_cmt(inner, pos)
-		if pos <= inner_len and inner:sub(pos, pos) == "," then pos = pos + 1 end
+--- Project canonical marker records into the two lookup tables used by the offset renderer.
+--- No source text, body text, or body token is inspected.
+--- @param markers table[] -- atom.paths.markers
+--- @return table<string, integer>, table[]
+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
 	end
-
-	return args, after_paren
-end
-
--- (internal) Record a `atom_label(name)` marker — `at_pos` is the branch-free word position within the atom body.
--- @param labels table<string, integer>
--- @param args string[]
--- @param at_pos integer
-local function record_label_marker(labels, args, at_pos)
-	if #args >= 1 then labels[args[1]] = at_pos end
-end
-
--- (internal) Record a `atom_offset(tag, target)` marker.
--- @param branches table[]  -- list of {pos=, target=, tag=}
--- @param args string[]
--- @param at_pos integer
-local function record_offset_marker(branches, args, at_pos)
-	if #args >= 2 then
-		table.insert(branches, { pos = at_pos, target = args[2], tag = args[1] })
-	end
-end
-
---- Scan a single token for atom_label/atom_offset markers, walking through balanced groups transparently (so nested calls are found).
---- @param token string
---- @param at_pos integer    -- the branch-free word position of this token in the body
---- @param labels table<string, integer>
---- @param branches table[]
-local function scan_for_atom_markers(token, at_pos, labels, branches)
-	local pos     = 1
-	local tok_len = #token
-	while pos <= tok_len do
-		pos = duffle.skip_ws_and_cmt(token, pos)
-		if pos > tok_len then break end
-		local ch = token:sub(pos, pos)
-		if duffle.is_alpha(ch) then
-			local ident, after = duffle.read_ident(token, pos)
-			if ident == LABEL_MARKER then
-				local args, after_paren = extract_ident_args(token, after)
-				record_label_marker(labels, args, at_pos)
-				pos = after_paren or after
-			elseif ident == OFFSET_MARKER then
-				local args, after_paren = extract_ident_args(token, after)
-				record_offset_marker(branches, args, at_pos)
-				pos = after_paren or after
-			else
-				pos = after
-			end
-		else
-			local nx = duffle.skip_str_or_cmt(token, pos)
-			pos = (nx > pos) and nx or (pos + 1)
-		end
-	end
-end
-
--- (internal) Count words emitted by the rest of `tok` after a marker call
--- (the marker call itself emits 0 words, but the source pattern may bundle the marker with the next instruction on the same line,
--- separated by no top-level comma).
--- Returns the word count contributed by that rest.
--- @param tok string
--- @param word_counts table
--- @return integer
-local function count_marker_rest(tok, word_counts)
-	-- duffle.find_marker_call_end returns the position PAST the closing `)` of the marker call
-	-- (or nil if `tok` isn't a marker call). Canonical impl in duffle.lua is faster than the
-	-- file-local copy that used to live here (byte-indexed, no `tok:sub` per char).
-	local marker_end = duffle.find_marker_call_end(tok)
-	if not marker_end or marker_end >= #tok then return 0 end
-	local rest = duffle.trim(tok:sub(marker_end))
-	if rest == "" then return 0 end
-	return count_token_words(rest, word_counts)
-end
-
--- (internal) Is this token a marker call (`atom_label` or `atom_offset`)?
--- @param tok string
--- @return boolean
-local function is_marker_token(tok)
-	local leading_ident = duffle.read_ident(tok, 1)
-	return leading_ident == LABEL_MARKER or leading_ident == OFFSET_MARKER
-end
-
---- Scan an atom body for labels + branches, count total words.
---- Returns (labels, branches, total_words).
---- @param body string
---- @param word_counts table
---- @return table<string, integer>, table[], integer
--- scan_atom_body: walk pre-tokenized body for atom_label/atom_offset markers + word counts.
--- Uses `atom.body_tokens` from the SourceScan payload (pre-tokenized by scan-source pass).
--- @param body_tokens table[]  -- {{tok=string, rel=integer}, ...} from duffle.tokenize_body
--- @param word_counts table
--- @return table, table, integer  -- labels, branches, total_words
-local function scan_atom_body(body_tokens, word_counts)
-	local pos      = 0
-	local labels   = {}
-	local branches = {}
-	for _, t in ipairs(body_tokens) do
-		local tok = t.tok
-		if is_marker_token(tok) then
-			-- Marker call: record at the current pos, do NOT advance pos.
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + count_marker_rest(tok, word_counts)
-		else
-			local words = count_token_words(tok, word_counts)
-			scan_for_atom_markers(tok, pos, labels, branches)
-			pos = pos + words
-		end
-	end
-	return labels, branches, pos
+	return state.labels, state.branches
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Offset computation + header generation
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Compute branch offsets as `target_word - branch_word - 1`
--- (the standard MIPS branch-immediate encoding).
--- @param labels table<string, integer>
--- @param branches table[]
--- @return BranchOffset[]
+--- Compute branch offsets as `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding).
+--- @param labels table<string, integer>
+--- @param branches table[]
+--- @return BranchOffset[]
 local function compute_offsets(labels, branches)
 	local results = {}
 	for _, br in ipairs(branches) do
 		local target = labels[br.target]
 		if not target then
-			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.pos .. ")")
+			error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")")
 		end
-		results[#results + 1] = { target = br.target, tag = br.tag, offset = target - br.pos - 1 }
+		results[#results + 1] = {
+			target      = br.target,
+			tag         = br.tag,
+			branch_word = br.branch_word,
+			offset      = target - br.branch_word - 1,
+		}
 	end
 	return results
 end
 
--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
--- @param s string
--- @param w integer
--- @return string
+--- Right-pad `s` with spaces to width `w`. If `s` is already `w` or wider, no padding is added.
+--- @param s string
+--- @param w integer
+--- @return string
 local function pad_right(s, w)
 	return s .. string.rep(" ", math.max(0, w - #s))
 end
 
--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
--- @param r BranchOffset
--- @return table
-local function make_offset_const(r)
+--- (internal) Build a constant-table entry `{macro_name, enum_name, value}` from a BranchOffset.
+--- @param bo BranchOffset
+--- @return table
+local function make_offset_const(bo)
 	return {
-		macro_name = OFFSET_MACRO_PREFIX .. r.tag .. "_" .. r.target,
-		enum_name  = OFFSET_ENUM_PREFIX  .. r.tag .. "_" .. r.target,
-		value      = r.offset,
+		macro_name = OFFSET_MACRO_PREFIX .. bo.tag .. "_" .. bo.target,
+		enum_name  = OFFSET_ENUM_PREFIX  .. bo.tag .. "_" .. bo.target,
+		value      = bo.offset,
 	}
 end
 
--- (internal) Emit one atom's offset constants + enum into the lines buffer.
--- @param add fun(s: string)
--- @param atom AtomData
+--- (internal) Emit one atom's offset constants + enum into the lines buffer.
+--- @param add  fun(s: string)
+--- @param atom AtomData
 local function emit_atom_offsets(add, atom)
 	if #atom.offsets == 0 then return end
 	add("// --- atom: " .. atom.name .. " (" .. atom.total_words .. " words) ---")
@@ -282,10 +167,10 @@ local function emit_atom_offsets(add, atom)
 	add("")
 end
 
--- Generate the per-source .offsets.h header.
--- @param source_path string
--- @param atoms_data AtomData[]
--- @return string
+--- Generate the per-source .offsets.h header.
+--- @param source_path string
+--- @param atoms_data  AtomData[]
+--- @return string
 local function generate_header(source_path, atoms_data)
 	local basename = duffle.basename_no_ext(source_path)
 
@@ -307,59 +192,41 @@ local function generate_header(source_path, atoms_data)
 	return table.concat(lines, "\n") .. "\n"
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- M — module exports
--- ════════════════════════════════════════════════════════════════════════════
-
 local M = {}
 
--- Project the pre-scanned SourceScan entries into the {name, body, body_tokens} shape this pass needs.
--- MipsAtom_ entries have kind="atom"; MipsCode code_<name> entries have kind="raw_atom".
--- `body_tokens` is set by scan-source on every `scan.atoms[i]` / `scan.raw_atoms[i]`; we carry it forward
--- so `scan_atom_body` reads from the precomputed table directly (no per-atom tokenize_body fallback).
--- @param scan table  -- SourceScan from duffle.scan_source
--- @return table[]  -- list of {name=, body=, body_tokens=}
-local function project_atoms(scan)
-	local out = {}
-	for _, a in ipairs(scan.atoms) do
-		out[#out + 1] = { name = a.raw_name, body = a.body, body_tokens = a.body_tokens }
-	end
-	for _, a in ipairs(scan.raw_atoms) do
-		out[#out + 1] = { name = a.name, body = a.body, body_tokens = a.body_tokens }
-	end
-	return out
-end
-
--- (internal) Process one source: project atoms from scan, scan bodies, write header.
--- Returns the offsets_h path if a header was written, or nil.
--- @param ctx PassCtx
--- @param src SourceFile
--- @return string|nil  -- the offsets_h path
+--- (internal) Process one source: render offsets from canonical atom paths.
+--- Returns the offsets_h path if a header was written, or nil.
+--- @param ctx PassCtx
+--- @param src SourceFile
+--- @return string|nil  -- the offsets_h path
 local function process_source(ctx, src)
-	local atoms = project_atoms(src.scan)
-	if #atoms == 0 then return nil end
-
 	local atoms_data = {}
-	for _, atom in ipairs(atoms) do
-		local labels, branches, total = scan_atom_body(atom.body_tokens, ctx.shared.word_counts)
+	local scan = src.scan or {}
+
+	local function append_atom(atom)
+		local paths = atom and atom.paths
+		if not paths then return end
+		local labels, branches = project_markers(paths.markers)
 		atoms_data[#atoms_data + 1] = {
-			name        = atom.name,
-			total_words = total,
+			name        = atom.raw_name or atom.name,
+			total_words = #(paths.word_events or {}),
 			offsets     = compute_offsets(labels, branches),
 		}
 	end
 
+	for _, atom in ipairs(scan.atoms or {}) do append_atom(atom) end
+	for _, atom in ipairs(scan.raw_atoms or {}) do append_atom(atom) end
+	if #atoms_data == 0 then return nil end
+
 	local out_path = src.dir .. "/gen/" .. duffle.basename_no_ext(src.dir) .. ".offsets.h"
-	if not ctx.dry_run then
-		duffle.ensure_dir(duffle.dirname(out_path))
-		duffle.write_file(out_path, generate_header(src.path, atoms_data))
-	end
+	duffle.ensure_dir(duffle.dirname(out_path))
+	duffle.write_file(out_path, generate_header(src.path:gsub("/", "\\"), atoms_data))
 	return out_path
 end
 
 --- Run the offsets pass.
---- For each source, emits a per-module `<dir_basename>.offsets.h` containing `#define _atom_offset_F_T = N` constants 
---- for every `atom_offset(F, T)` reference in the source's atoms.
+--- For each canonical source, emits a per-module `<dir_basename>.offsets.h`
+--- containing constants for every marker recorded in atom.paths.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
@@ -367,7 +234,12 @@ function M.run(ctx)
 	local errors   = {}
 	local warnings = {}
 
-	for _, src in ipairs(ctx.sources) do
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" or type(corpus.source_order) ~= "table" then
+		error("offsets.run requires ctx.shared.corpus.source_order (canonical corpus).", 0)
+	end
+
+	for _, src in ipairs(corpus.source_order) do
 		local out_path = process_source(ctx, src)
 		if out_path then
 			outputs[#outputs + 1] = { offsets_h = out_path }

scripts/passes/report.lua

@@ -5,26 +5,31 @@
 ---   - `build/gen/<dir_basename>.annotations.txt` — one per source-directory containing atoms; aggregates across all sources in the directory.
 ---   - `build/gen/annotation_validation.txt` — the project summary.
 ---
---- The annotation pass stashes per-MODULE summary entries in `ctx.flags._annot_results` (set by `passes/annotation.lua`). 
---- This pass re-validates each source via `annotation.validate()` to get the detailed per-source results needed for the report. 
----
---- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
---- Lua 5.3 compatible.
+--- 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.
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- 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.
--- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
 -- Bootstrap: see `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
--- Uses `debug.getinfo` to find this file's own directory, so it works
--- both standalone and when require'd from the orchestrator.
+-- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
 -- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
+
+-- Load the annotation pass so we can re-validate each source against the canonical corpus projection.
+-- The annotation pass exposes `M.validate`, which returns the per-source AnnotationResult (atoms / annots / macros / binds / errors / warnings)
+-- that the report pass renders into the per-module `<dir_basename>.annotations.txt` output.
+local annotation = dofile(_bootstrap_dir .. "annotation.lua")
+
+-- Load atoms_source_map for the `render_source_map` / `render_provenance` module functions (used by `render_module_atoms_md` to produce `<module>.atoms.md` without re-walking source tokens).
+-- The pass itself emits no per-source files anymore; we only consume the two pure renderers here.
+-- Defined BEFORE the renderer functions below so their upvalues resolve to this local (not the global `atoms_source_map`, which is nil).
+local atoms_source_map = dofile(_bootstrap_dir .. "atoms_source_map.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Constants
@@ -40,8 +45,7 @@ local SECTION_HEADER_MACROS    = "── Macro word-count declarations ───
 local SECTION_HEADER_ERRORS    = "── Errors ──────────────────────────────────────────────"
 local SECTION_HEADER_WARNINGS  = "── Warnings ────────────────────────────────────────────"
 
--- Lua pattern that captures the basename (last path segment) of a
--- forward- or back-slash separated path.
+-- Lua pattern that captures the basename (last path segment) of a forward- or back-slash separated path.
 local BASENAME_PATTERN = "([^/\\]+)$"
 
 -- Debug flag name — set to truthy in `_G` to enable verbose logging.
@@ -61,16 +65,14 @@ local PASS_NAME = "report"
 --- @field basename string  -- filename without extension
 
 --- @class PassCtx
---- @field sources            SourceFile[]            -- all source files in the build
---- @field metadata_path      string                  -- path to word_count.metadata.h
---- @field shared             table                   -- cross-pass shared state
---- @field out_root           string                  -- output root (e.g. "build/gen")
---- @field project_root       string                  -- project root (e.g. "code/")
---- @field upstream           table<string, table>    -- per-pass upstream outputs
---- @field flags              table                   -- CLI flags + per-pass stash
---- @field flags._annot_results ModuleEntry[]         -- stashed by annotation pass
---- @field dry_run            boolean                 -- if true, compute but don't write
---- @field verbose            boolean                 -- if true, log diagnostic info
+--- @field sources            SourceFile[]         -- all source files in the build
+--- @field metadata_path      string               -- path to word_count.metadata.h
+--- @field shared             table                -- cross-pass shared state
+--- @field out_root           string               -- output root (e.g. "build/gen")
+--- @field project_root       string               -- project root (e.g. "code/")
+--- @field upstream           table<string, table> -- per-pass upstream outputs
+--- @field flags              table                -- CLI flags + per-pass stash
+--- @field verbose            boolean              -- if true, log diagnostic info
 
 --- @class PassResult
 --- @field outputs  table[]  -- {kind=, path=} entries describing emit files
@@ -139,305 +141,458 @@ local PASS_NAME = "report"
 -- Per-MODULE annotation report (aggregated across all sources in a dir)
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
--- @param path string
--- @return string
+--- Extract the basename (last path segment) of a forward- or back-slash separated path. Returns the input unchanged if no separator is found.
+--- @param path string
+--- @return string
 local function source_basename(path)
 	return path:match(BASENAME_PATTERN) or path
 end
 
--- (internal) Format a single annotation entry as one rendered line.
--- @param a AnnotEntry
--- @param src_name string
--- @return string
-local function format_annot_line(a, src_name)
-	if a.error then
-		return string.format("  ✗ line %d  %s  [ERROR: %s]  [%s]", a.line, a.macro or "?", a.error, src_name)
+-- ════════════════════════════════════════════════════════════════════════════
+-- Markdown renderers (consolidated-report-files refactor, 2026-07-26)
+-- ════════════════════════════════════════════════════════════════════════════
+
+--- Render the thin project-wide summary (`build/atom_meta_report.summary.md`).
+--- @param all_results { module:string, atoms:integer, annots:integer, binds:integer,
+---                       macros:integer, findings:integer, errors:integer,
+---                       warnings:integer, info:integer }[]
+--- @return string
+local function render_project_summary(all_results)
+	local lines = {
+		"# Project summary",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua).",
+		"",
+		"| module | atoms | annots | binds | macros | findings | errors | warnings | info |",
+		"|--------|-------|--------|-------|--------|----------|--------|----------|------|",
+	}
+	local totals = { atoms = 0, annots = 0, binds = 0, macros = 0,
+		findings = 0, errors = 0, warnings = 0, info = 0 }
+	for _, e in ipairs(all_results) do
+		lines[#lines + 1] = string.format(
+			"| %s | %d | %d | %d | %d | %d | %d | %d | %d |",
+			e.module, e.atoms, e.annots, e.binds, e.macros,
+			e.findings, e.errors, e.warnings, e.info)
+		totals.atoms    = totals.atoms    + e.atoms
+		totals.annots   = totals.annots   + e.annots
+		totals.binds    = totals.binds    + e.binds
+		totals.macros   = totals.macros   + e.macros
+		totals.findings = totals.findings + e.findings
+		totals.errors   = totals.errors   + e.errors
+		totals.warnings = totals.warnings + e.warnings
+		totals.info     = totals.info     + e.info
 	end
-	local line = string.format("  ●  line %d  %s  [%s]", a.line, a.name, src_name)
-	if a.binds               then line = line .. "  binds=" .. a.binds end
-	if #a.reads > 0          then line = line .. "  reads={" .. table.concat(a.reads, ",") .. "}" end
-	if #a.writes > 0         then line = line .. "  writes={" .. table.concat(a.writes, ",") .. "}" end
-	return line
-end
-
--- (internal) Tally totals across all results in a module.
--- @param results AnnotationResult[]
--- @return integer, integer, integer, integer, integer, integer
-local function tally_module_totals(results)
-	local total_atoms, total_annots, total_binds, total_macros = 0, 0, 0, 0
-	local total_errors, total_warnings = 0, 0
-	for _, r in ipairs(results) do
-		total_atoms    = total_atoms    + #r.atoms
-		total_annots   = total_annots   + #r.annots
-		total_binds    = total_binds    + #r.binds
-		total_macros   = total_macros   + #r.macros
-		total_errors   = total_errors   + #r.errors
-		total_warnings = total_warnings + #r.warnings
-	end
-	return total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings
-end
-
--- (internal) Section renderer: per-source atom declarations.
-local function render_module_atoms_section(add, results)
-	add(SECTION_HEADER_ATOMS)
-	for _, r in ipairs(results) do
-		local src_name = source_basename(r.source)
-		for _, a in ipairs(r.atoms) do
-			add(string.format("  MipsAtom_(%s)   line %d   [%s]", a.name, a.line, src_name))
-		end
-	end
-	add("")
-end
-
--- (internal) Section renderer: per-source annotation entries.
-local function render_module_annots_section(add, results)
-	add(SECTION_HEADER_ANNOTS)
-	for _, r in ipairs(results) do
-		local src_name = source_basename(r.source)
-		for _, a in ipairs(r.annots) do
-			add(format_annot_line(a, src_name))
-		end
-	end
-	add("")
-end
-
--- (internal) Section renderer: per-source Binds_* struct declarations.
-local function render_module_binds_section(add, results)
-	add(SECTION_HEADER_BINDS)
-	for _, r in ipairs(results) do
-		local src_name = source_basename(r.source)
-		for _, b in ipairs(r.binds) do
-			add(string.format("  %s   line %d   %d bytes   [%s]", b.name, b.line, b.bytes, src_name))
-			for _, f in ipairs(b.fields) do
-				add(string.format("    +%2d: %s", f.offset, f.name))
-			end
-		end
-	end
-	add("")
-end
-
--- (internal) Section renderer: per-source macro word-count declarations.
-local function render_module_macros_section(add, results)
-	add(SECTION_HEADER_MACROS)
-	for _, r in ipairs(results) do
-		local src_name = source_basename(r.source)
-		for _, m in ipairs(r.macros) do
-			add(string.format("  %s  line %d  words=%d  [%s]", m.name, m.line, m.words, src_name))
-		end
-	end
-	add("")
-end
-
--- (internal) Section renderer: per-source errors (one-line + "(none)" if empty).
-local function render_module_errors_section(add, results, total_errors)
-	add(SECTION_HEADER_ERRORS)
-	if total_errors == 0 then
-		add("  (none)")
-	else
-		for _, r in ipairs(results) do
-			local src_name = source_basename(r.source)
-			for _, e in ipairs(r.errors) do
-				add(string.format("  ✗ line %d  %s  [%s]", e.line, e.msg, src_name))
-			end
-		end
-	end
-	add("")
-end
-
--- (internal) Section renderer: per-source warnings (one-line + "(none)" if empty).
-local function render_module_warnings_section(add, results, total_warnings)
-	add(SECTION_HEADER_WARNINGS)
-	if total_warnings == 0 then
-		add("  (none)")
-	else
-		for _, r in ipairs(results) do
-			local src_name = source_basename(r.source)
-			for _, w in ipairs(r.warnings) do
-				add(string.format("  ⚠ line %d  %s  [%s]", w.line, w.msg, src_name))
-			end
-		end
-	end
-	add("")
-end
-
---- Render the per-MODULE annotation report (one `<dir_basename>.annotations.txt`).
---- @param dir string                  -- module directory path
---- @param sources SourceFile[]        -- sources in this module
---- @param results AnnotationResult[]  -- per-source validate() results
---- @return string                     -- the rendered report text
-local function render_module_report(dir, sources, results)
-	local lines = {}
-	local function add(s) lines[#lines + 1] = s end
-
-	add(RULE_THICK)
-	add("ANNOTATION PASS — module " .. source_basename(dir))
-	add(RULE_THICK)
-	add(string.format("Sources: %d", #sources))
-	for _, s in ipairs(sources) do add("  " .. s.path) end
-	add("")
-
-	local total_atoms, total_annots, total_binds, total_macros, total_errors, total_warnings = tally_module_totals(results)
-	add(string.format("Atoms: %d   Annotations: %d   Binds structs: %d   Macro decls: %d",
-		total_atoms, total_annots, total_binds, total_macros))
-	add("")
-
-	render_module_atoms_section(add, results)
-	render_module_annots_section(add, results)
-	render_module_binds_section(add, results)
-	render_module_macros_section(add, results)
-	render_module_errors_section(add, results, total_errors)
-	render_module_warnings_section(add, results, total_warnings)
-
+	lines[#lines + 1] = string.format(
+		"| **TOTAL** | %d | %d | %d | %d | %d | %d | %d | %d |",
+		totals.atoms, totals.annots, totals.binds, totals.macros,
+		totals.findings, totals.errors, totals.warnings, totals.info)
 	return table.concat(lines, "\n") .. "\n"
 end
 
--- ════════════════════════════════════════════════════════════════════════════
--- Per-project summary
--- ════════════════════════════════════════════════════════════════════════════
-
---- Render the per-project summary (`build/gen/annotation_validation.txt`).
---- Aggregates totals across all sources; lists per-source error counts if any source has errors.
---- @param all_results AnnotationResult[]
+--- Render the per-module verbose source-map markdown (`build/<module>.atoms.md`).
+--- Per-source sub-section, per-atom stanza with sourcemap + provenance rows.
+--- Pulls sourcemap + provenance from `atoms_source_map` (no second source walk).
+--- @param dir string
+--- @param dir_sources SourceFile[]
+--- @param wc table<string, integer>
 --- @return string
-local function render_project_report(all_results)
-	local lines = {}
+local function render_module_atoms_md(dir, dir_sources, wc)
+	local dir_basename = source_basename(dir)
+	local lines = {
+		"# " .. dir_basename .. " — atoms (verbose source map)",
+		"> Per-word call-site + provenance. Auto-generated.",
+		"",
+	}
+	for _, src in ipairs(dir_sources) do
+		local src_name = source_basename(src.path)
+		lines[#lines + 1] = "## " .. src_name
+		lines[#lines + 1] = ""
+		-- For each atom with a projection, render its sourcemap + provenance.
+		local atoms_list = {}
+		for _, atom in ipairs((src.scan or {}).atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
+		end
+		for _, atom in ipairs((src.scan or {}).raw_atoms or {}) do
+			if atom.paths then atoms_list[#atoms_list + 1] = atom end
+		end
+		if #atoms_list == 0 then
+			lines[#lines + 1] = "_(no atom projections)_"
+			lines[#lines + 1] = ""
+		else
+			-- Per-source forward-slash path (same one `emit_atom_stanza` / `emit_provenance_stanza` would derive;
+			-- computed once per `## <source>` heading and reused by each atom's `WORD N CALL ...` field).
+			local rel_path = src.path:gsub("\\\\", "/")
+			for _, atom in ipairs(atoms_list) do
+				lines[#lines + 1] = string.format(
+					"### atom: %s (line %d, %d words)",
+					atom.name, atom.line or 0, #(atom.paths.items or {}))
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Sourcemap** — per-word call site:"
+				lines[#lines + 1] = "```"
+				-- Per-atom invariant: call the per-atom renderers, NOT the per-source ones.
+				-- The per-source renderers enumerate every atom in `src`;
+				-- calling them in a per-atom loop would repeat the whole source under every `### atom:` heading.
+				lines[#lines + 1] = atoms_source_map.render_atom_source_map(atom):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+				lines[#lines + 1] = "**Provenance** — per-word definition + body:"
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = atoms_source_map.render_atom_provenance(atom, wc, rel_path):gsub("\n+$", "")
+				lines[#lines + 1] = "```"
+				lines[#lines + 1] = ""
+			end
+		end
+	end
+	return table.concat(lines, "\n") .. "\n"
+end
+
+--- Render the consolidated per-module markdown (`build/<module>.atom_meta_report.md`).
+--- Aggregates annotation + static-analysis content across all sources in `dir`.
+--- Annotations come from re-running `annotation.validate()` per source (the existing pattern);
+--- static-analysis comes from `corpus.static_analysis_results[dir_basename]` (populated by `static_analysis.lua` — no second corpus_pipe_ctx build).
+--- @param dir string
+--- @param dir_sources SourceFile[]
+--- @param annot_results AnnotationResult[]
+--- @param sa_results table                 -- corpus.static_analysis_results[dir_basename]
+--- @return string
+local function render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+	local dir_basename = source_basename(dir)
+	local lines = {
+		"# " .. dir_basename .. " — atom meta report",
+		"> Auto-generated by ps1_meta.lua (passes/report.lua). Do not edit.",
+		"",
+	}
 	local function add(s) lines[#lines + 1] = s end
 
-	local total_atoms, total_annots, total_macros, total_binds = 0, 0, 0, 0
-	local total_errors, total_warnings = 0, 0
-	for _, r in ipairs(all_results) do
-		total_atoms    = total_atoms    + #r.atoms
-		total_annots   = total_annots   + #r.annots
-		total_macros   = total_macros   + #r.macros
-		total_binds    = total_binds    + #r.binds
-		total_errors   = total_errors   + #r.errors
-		total_warnings = total_warnings + #r.warnings
+	-- Module summary table.
+	local n_atoms        = 0
+	local n_annot        = 0
+	local n_binds        = 0
+	local n_macros       = 0
+	local n_bare, n_proc = 0, 0
+	for _, r in ipairs(annot_results) do
+		n_atoms  = n_atoms  + #r.atoms
+		n_annot  = n_annot  + #r.annots
+		n_binds  = n_binds  + #r.binds
+		n_macros = n_macros + #r.macros
+	end
+	for _, a in ipairs(sa_results.atoms or {}) do
+		if     a.kind == "comp_bare" then n_bare = n_bare + 1
+		elseif a.kind == "comp_proc" then n_proc = n_proc + 1
+		end
 	end
 
-	add(RULE_THICK)
-	add("ANNOTATION VALIDATION — project summary")
-	add(RULE_THICK)
-	add("")
-	add(string.format("Atoms:        %d", total_atoms))
-	add(string.format("Annotations:  %d", total_annots))
-	add(string.format("Macros:       %d", total_macros))
-	add(string.format("Binds:        %d", total_binds))
-	add("")
-	add(string.format("Errors:       %d", total_errors))
-	add(string.format("Warnings:     %d", total_warnings))
+	add("## Module summary"); add("")
+	add("| metric | value |"); add("|--------|-------|")
+	add(string.format("| sources | %d |", #dir_sources))
+	add(string.format("| atoms | %d (atoms: %d, comp_bare: %d, comp_proc: %d) |",
+		#(sa_results.atoms or {}),
+		#(sa_results.atoms or {}) - n_bare - n_proc, n_bare, n_proc))
+	add(string.format("| annotations | %d |", n_annot))
+	add(string.format("| binds structs | %d |", n_binds))
+	add(string.format("| macro decls | %d |", n_macros))
+	add(string.format("| findings | %d (errors: %d, warnings: %d, info: %d) |",
+		#(sa_results.findings or {}),
+		#(sa_results.errors   or {}),
+		#(sa_results.warnings or {}),
+		#(sa_results.info     or {})))
 	add("")
 
-	if total_errors > 0 then
-		add("Per-source error counts:")
-		for _, r in ipairs(all_results) do
-			if #r.errors > 0 then
-				local src_name = source_basename(r.source)
-				add(string.format("  %s : %d error(s)", src_name, #r.errors))
+	-- Sources
+	add("## Sources"); add("")
+	for _, s in ipairs(dir_sources) do add("- `" .. s.path .. "`") end
+	add("")
+
+	-- Atoms (annotation)
+	add("## Atoms"); add("")
+	add("| kind | name | source | line |"); add("|------|------|--------|------|")
+	for _, r in ipairs(annot_results) do
+		local src_name = source_basename(r.source)
+		for _, a in ipairs(r.atoms) do
+			add(string.format("| atom | %s | %s | %d |", a.name, src_name, a.line))
+		end
+	end
+	add("")
+
+	-- Annotations
+	add("## Annotations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | name | binds | reads | writes |")
+		add("|--------|------|------|-------|-------|--------|")
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, a in ipairs(r.annots) do
+				local binds  = a.binds  or "—"
+				local reads  = (#a.reads  > 0 and table.concat(a.reads,  ",")) or "—"
+				local writes = (#a.writes > 0 and table.concat(a.writes, ",")) or "—"
+				add(string.format("| %s | %d | %s | %s | %s | %s |",
+					src_name, a.line, a.name, binds, reads, writes))
+			end
+		end
+	end
+	add("")
+
+	-- Binds_* structs
+	add("## Binds_* structs"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, b in ipairs(r.binds) do
+				add(string.format("### %s (%s:%d, %d bytes)",
+					b.name, src_name, b.line, b.bytes))
+				for _, f in ipairs(b.fields) do
+					add(string.format("- `+%d %s`", f.offset, f.name))
+				end
+				add("")
+			end
+		end
+	end
+
+	-- Macro decls
+	add("## Macro word-count declarations"); add("")
+	if #annot_results == 0 then
+		add("_(none)_")
+	else
+		add("| source | line | macro declaration |")
+		add("|--------|------|-------------------|")
+		for _, r in ipairs(annot_results) do
+			local src_name = source_basename(r.source)
+			for _, m in ipairs(r.macros) do
+				add(string.format("| %s | %d | %s |",
+					src_name, m.line, m.name))
+			end
+		end
+	end
+	add("")
+
+	-- Findings by atom (static-analysis)
+	add("## Static analysis — findings by atom"); add("")
+	local by_atom = {}
+	for _, f in ipairs(sa_results.findings or {}) do
+		by_atom[f.atom] = by_atom[f.atom] or {}
+		by_atom[f.atom][#by_atom[f.atom] + 1] = f
+	end
+	if next(by_atom) == nil then
+		add("_(no findings)_")
+	else
+		for _, a in ipairs(sa_results.atoms or {}) do
+			local fs = by_atom[a.name]
+			if fs then
+				add(string.format("### %s", a.name))
+				for _, f in ipairs(fs) do
+					add(string.format("- `[%s] %s`", f.check, f.msg))
+				end
+				add("")
+			end
+		end
+	end
+
+	-- Errors / Warnings / Info
+	local function add_findings(label, entries)
+		add(string.format("## %s", label))
+		if #entries == 0 then
+			add("_(none)_")
+		else
+			for _, e in ipairs(entries) do
+				add(string.format("- line %d  %s", e.line, e.msg))
 			end
 		end
 		add("")
 	end
+	add_findings("Errors",   sa_results.errors   or {})
+	add_findings("Warnings", sa_results.warnings or {})
+	add_findings("Info",     sa_results.info     or {})
+
+	-- Per-atom cycle counts (path-aware)
+	add("## Per-atom cycle counts (path-aware, best case, no stalls)"); add("")
+	add("| atom | source | min | max | branches | paths | notes |")
+	add("|------|--------|-----|-----|----------|-------|-------|")
+	local sorted = {}
+	for _, a in ipairs(sa_results.atoms or {}) do sorted[#sorted + 1] = a end
+	table.sort(sorted, function(x, y)
+		return ((x.paths or {}).cycles_max or 0) > ((y.paths or {}).cycles_max or 0)
+	end)
+	for _, a in ipairs(sorted) do
+		local p        = a.paths or {}
+		local src_name = a.source_path and source_basename(a.source_path) or ""
+		local notes   = ""
+		if p.has_loops then notes = notes .. " [loop!]" end
+		if p.unknown_macros and #p.unknown_macros > 0 then
+			notes = notes .. " [unknown: " .. table.concat(p.unknown_macros, ", ") .. "]"
+		end
+		add(string.format("| %s | %s | %d | %d | %d | %d | %s |",
+			a.name, src_name,
+			p.cycles_min or 0, p.cycles_max or 0,
+			p.branches or 0, p.paths or 0, notes))
+	end
+	add("")
+
+	-- Per-source scan summary
+	add("## Per-source scan summary"); add("")
+	for _, src in ipairs(dir_sources) do
+		local src_atoms = {}
+		for _, a in ipairs(sa_results.atoms or {}) do
+			if a.source_path == src.path then src_atoms[#src_atoms + 1] = a end
+		end
+		if #src_atoms > 0 then
+			local mn, mx = math.huge, -1
+			for _, a in ipairs(src_atoms) do
+				local p = a.paths or {}
+				if (p.cycles_min or 0) < mn then mn = p.cycles_min or 0 end
+				if (p.cycles_max or 0) > mx then mx = p.cycles_max or 0 end
+			end
+			local path_str
+			if mx > 0 then
+				path_str = string.format("  cycles=%d..%d", mn, mx)
+			else
+				path_str = string.format("  %d cycles", mn)
+			end
+			add(string.format("- `%s` — %d atom%s%s",
+				src.basename, #src_atoms,
+				#src_atoms == 1 and "" or "s", path_str))
+		end
+	end
+	add("")
 
 	return table.concat(lines, "\n") .. "\n"
 end
 
 -- ════════════════════════════════════════════════════════════════════════════
--- Orchestration helpers
+-- REPORT_RENDERERS — data-driven report dispatch (one row per file kind)
 -- ════════════════════════════════════════════════════════════════════════════
-
--- (internal) Pull per-source validate() results from the annotation pass's stash.
--- The annotation pass runs first in the dep chain and caches results in `ctx.flags._annot_source_results`; we read from there instead of re-validating each source.
--- Returns the list of module results + the flat list of all results (for the project-wide summary).
--- @param ctx PassCtx
--- @param dir_sources SourceFile[]
--- @return AnnotationResult[], AnnotationResult[]
-local function lookup_module_results(ctx, dir_sources)
-	local src_cache      = (ctx.flags and ctx.flags._annot_source_results) or {}
-	local module_results = {}
-	local all_results    = {}
-	for _, src in ipairs(dir_sources) do
-		local result = src_cache[src.path]
-		if result then
-			result.source = src.path                                    -- defensive (annotation tags it too; this guards against cache misses from earlier iterations)
-			module_results[#module_results + 1] = result
-			all_results[#all_results + 1] = result
-		end
-	end
-	return module_results, all_results
-end
-
--- (internal) Does this module's results contain anything worth emitting?
--- @param module_results AnnotationResult[]
--- @return boolean
-local function module_has_content(module_results)
-	for _, r in ipairs(module_results) do
-		if #r.atoms  > 0 or #r.annots > 0 or #r.binds    > 0
-		or #r.macros > 0 or #r.errors > 0 or #r.warnings > 0 then
-			return true
-		end
-	end
-	return false
-end
-
--- (internal) Log a debug message if `_G[DEBUG_FLAG]` is truthy.
--- @param fmt string
-local function debug_log(fmt, ...)
-	if _G[DEBUG_FLAG] then
-		io.stderr:write(string.format("[%s] " .. fmt, PASS_NAME, ...))
-	end
-end
+-- `once = true` means render once at the project level (not per-module).
+-- `basename(dir_basename)` yields the file's basename for that kind.
+-- `gather(ctx, dir, dir_sources [, all_modules])` returns the rendered string.
+local REPORT_RENDERERS = {
+	{
+		name     = "atom_meta_report",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atom_meta_report" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			-- Annotations: re-run `annotation.validate()` per source (the existing pattern).
+			local annot_results = {}
+			for _, src in ipairs(dir_sources) do
+				if src.scan then
+					local r = annotation.validate(ctx, src, nil)
+					r.source = src.path
+					annot_results[#annot_results + 1] = r
+				end
+			end
+			-- Static-analysis: read stashed projection (no re-validate).
+			local dir_basename = dir:match("([^/\\]+)$") or dir
+			local sa_results   = (ctx.shared.corpus.static_analysis_results or {})[dir_basename] or {}
+			return render_module_meta_report(dir, dir_sources, annot_results, sa_results)
+		end,
+	},
+	{
+		name     = "atoms",
+		ext      = "md",
+		basename = function(dir_basename) return dir_basename .. ".atoms" end,
+		once     = false,
+		gather   = function(ctx, dir, dir_sources)
+			return render_module_atoms_md(dir, dir_sources,
+				ctx.shared.corpus.word_counts or {})
+		end,
+	},
+	{
+		name     = "summary",
+		ext      = "md",
+		basename = function(_dir_basename) return "atom_meta_report.summary" end,
+		once     = true,
+		gather   = function(_ctx, _dir, _dir_sources, all_modules)
+			return render_project_summary(all_modules)
+		end,
+	},
+}
 
 -- ════════════════════════════════════════════════════════════════════════════
--- M — module exports
+-- M — public pass surface
 -- ════════════════════════════════════════════════════════════════════════════
 
 local M = {}
 
---- Run the report pass. 
---- Renders one `<dir_basename>.annotations.txt` per source-directory that has content, plus the project-wide `annotation_validation.txt` summary.
+--- Run the report pass. Emits 1 `atom_meta_report.summary.md` per build + 2 `atom_meta_report.md` + 2 `atoms.md` files per module (duffle + gte_hello).
+--- Reads `corpus.static_analysis_results` (added in Phase 1) to populate per-module findings without re-running validate().
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
-	local outputs  = {}
-	local errors   = {}
-	local warnings = {}
+	local outputs = {}
+	local corpus  = ctx.shared and ctx.shared.corpus
+	local by_dir  = (corpus and corpus.sources_by_dir) or {}
 
-	local module_entries = (ctx.flags and ctx.flags._annot_results) or {}
-	local by_dir         = ctx.by_dir or duffle.group_sources_by_dir(ctx.sources)
+	-- `out_path_root`: when the conventional `out_root` is `build/gen` (any spelling — relative, absolute, separator variants).
+	-- Write the md files to `build/` (parent of `gen/`) instead of nested under `gen/`.
+	-- Mirrors the `gdb_tape_atoms_runtime.gdb` relocation.
+	local function ends_with_gen(p)
+		return type(p) == "string" and (p:match("[/\\]gen[/\\]?$") ~= nil
+			or p == "build/gen" or p == "build\\gen")
+	end
+	local out_root_effective = ends_with_gen(ctx.out_root)
+		and ctx.out_root:gsub("[/\\]gen[/\\]?$", "")
+		or ctx.out_root
 
-	if not ctx.dry_run then duffle.ensure_dir(ctx.out_root) end
+	duffle.ensure_dir(out_root_effective)
 
-	local all_results_for_summary = {}
-	for _, entry in ipairs(module_entries) do
-		debug_log("entry: dir=%s basename=%s atoms_count=%d dir_sources=%d\n", entry.dir, entry.dir_basename, entry.atoms_count, #(by_dir[entry.dir] or {}))
+	-- Aggregator for the project-wide `once = true` summary renderer.
+	local all_modules = {}
 
-		if entry.atoms_count > 0 or #(by_dir[entry.dir] or {}) > 0 then
-			local dir_sources                 = by_dir[entry.dir] or {}
-			local module_results, all_results = lookup_module_results(ctx, dir_sources)
-			for _, r in ipairs(all_results) do
-				all_results_for_summary[#all_results_for_summary + 1] = r
+	for dir, dir_sources in pairs(by_dir) do
+		local dir_basename = dir:match("([^/\\]+)$") or dir
+
+		-- Per-renderer dispatch for the per-module renderers (once = false).
+		for _, renderer in ipairs(REPORT_RENDERERS) do
+			if not renderer.once then
+				local body     = renderer.gather(ctx, dir, dir_sources)
+				local out_path = out_root_effective .. "/" .. renderer.basename(dir_basename) .. "." .. renderer.ext
+				duffle.write_file(out_path, body)
+				outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
 			end
+		end
 
-			if module_has_content(module_results) then
-				local out_path = ctx.out_root .. "/" .. entry.dir_basename .. ".annotations.txt"
-				if not ctx.dry_run then
-					duffle.write_file(out_path, render_module_report(entry.dir, dir_sources, module_results))
-				end
-				outputs[#outputs + 1] = { annotations_txt = out_path }
-			else
-				debug_log("  -> no content; skipping\n")
+		-- For the summary, compute per-module totals once (re-validating annotations per source — same pattern as the meta_report renderer).
+		local annot_results = {}
+		for _, src in ipairs(dir_sources) do
+			if src.scan then
+				local r = annotation.validate(ctx, src, nil)
+				r.source = src.path
+				annot_results[#annot_results + 1] = r
 			end
 		end
+		local n_annot, n_binds, n_macros = 0, 0, 0
+		for _, r in ipairs(annot_results) do
+			n_annot  = n_annot  + #r.annots
+			n_binds  = n_binds  + #r.binds
+			n_macros = n_macros + #r.macros
+		end
+		local sa_results = (corpus.static_analysis_results or {})[dir_basename] or {}
+		all_modules[#all_modules + 1] = {
+			module   = dir_basename,
+			atoms    = #(sa_results.atoms or {}),
+			annots   = n_annot,
+			binds    = n_binds,
+			macros   = n_macros,
+			findings = #(sa_results.findings or {}),
+			errors   = #(sa_results.errors   or {}),
+			warnings = #(sa_results.warnings or {}),
+			info     = #(sa_results.info     or {}),
+		}
+	end
+
+	-- Project-wide renderer (once = true): write the summary file.
+	for _, renderer in ipairs(REPORT_RENDERERS) do
+		if renderer.once then
+			local body     = renderer.gather(ctx, nil, nil, all_modules)
+			local out_path = out_root_effective .. "/" .. renderer.basename("") .. "." .. renderer.ext
+			duffle.write_file(out_path, body)
+			outputs[#outputs + 1] = { kind = renderer.name, path = out_path }
+		end
 	end
 
-	if not ctx.dry_run and #all_results_for_summary > 0 then
-		local summary_path = ctx.out_root .. "/annotation_validation.txt"
-		duffle.write_file(summary_path, render_project_report(all_results_for_summary))
-		outputs[#outputs + 1] = { summary_txt = summary_path }
-	end
-
-	return { outputs = outputs, errors = errors, warnings = warnings }
+	return { outputs = outputs, errors = {}, warnings = {} }
 end
 
 return M

scripts/passes/word_count_eval.lua

@@ -1,11 +1,17 @@
 --- word_count_eval.lua — Word-counting logic for the tape-atom metaprogram pipeline.
 ---
---- Three responsibilities:
----   1. **Public utilities** (used by `passes/components.lua`, `passes/offsets.lua`, `passes/annotation.lua`):
----        - `M.count_token_words(token, wc)` — words emitted by one token
----        - `M.scan_dir(dir, suffix)` — glob walk for *.macs.h
----   2. **Pass entry** `M.run(ctx)` — loads metadata.h + *.macs.h into `ctx.shared.word_counts` for downstream passes.
----   3. **Internal helpers** for the body scanner.
+--- Two responsibilities:
+---   1. **Public utility** `M.count_token_words(token, wc)`: Used by `passes/offsets.lua`, `passes/annotation.lua`, and other passes.
+---   2. **Pass entry** `M.run(ctx)`: Loads the authored `word_count.metadata.h` into `ctx.shared.corpus.word_counts` for downstream passes.
+---      The generated `.macs.h` files are OUTPUT artifacts and are NOT inputs to this pass;
+---      Current component counts are owned by `passes/components.lua` (which populates `corpus.word_counts` and `corpus.component_body_index`
+---      AFTER computing each current count from the just-built body + `corpus.word_counts`).
+---
+--- **Canonical contract**:
+---   * `ctx.shared.corpus.word_counts` is the count table.
+---   * `corpus.word_counts` is the sole count table. Consumers read `corpus.word_counts` directly.
+---   * `ctx.shared.components` and `ctx.shared.component_body_index` are NOT created by this pass (projections only).
+---   * No `.macs.h` recursive discovery (no `scan_dir`, no scan cache, no `_invalidate_scan_cache`).
 ---
 --- **Conventions**: tabs (1/level), EmmyLua annotations, no regex,
 --- Lua 5.3 compatible.
@@ -14,31 +20,11 @@
 -- Module-scope requires + package.path setup
 -- ════════════════════════════════════════════════════════════════════════════
 
--- Resolve `arg[0]` to an absolute-ish script directory so that `require("duffle")` resolves against `scripts/` regardless of CWD.
--- Note: this boilerplate is duplicated in 6 other entry scripts; a Phase-6 extraction target (`duffle.setup_package_path()`).
--- Bootstrap: see `ps1_meta.lua` for the rationale.
 -- Bootstrap: load `scripts/duffle_paths.lua` (sets package.path + package.cpath).
 -- Uses `debug.getinfo` to find this file's own directory, so it works both standalone and when require'd from the orchestrator.
--- Bootstrap: load `duffle_paths.lua` via `debug.getinfo(1, "S").source` (works both standalone + when require'd).
 -- duffle_paths.lua sets package.path then returns `require("duffle")` at the bottom, so the dofile value IS the duffle module.
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
-local duffle        = dofile(_bootstrap_dir .. "../duffle_paths.lua")
-
--- ════════════════════════════════════════════════════════════════════════════
--- Constants
--- ════════════════════════════════════════════════════════════════════════════
-
--- Windows separator chars — used to convert `dir` output (which uses `\`) into POSIX paths (which our scripts expect).
-local PATH_SEP_BACKSLASH = "\\"
-local PATH_SEP_FORWARD   = "/"
-
--- Fallback glob command (subprocess). Used when `lfs` (LuaFileSystem) is not available.
--- Scoped to `code\` to avoid walking `.git/`, `toolchain/`, `build/`, etc.
-local DIR_GLOB_CMD = 'dir /b /s "%s\\code\\%s" 2>nul'
-
--- Try to load lfs (LuaFileSystem). If available, scan_dir uses native directory enumeration (~2ms)
--- instead of spawning `dir /b /s` as a subprocess (~56ms). Built by update_deps.ps1 into toolchain/lfs/lfs.dll.
-local lfs = pcall(require, "lfs") and require("lfs") or nil
+local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")
 
 -- ════════════════════════════════════════════════════════════════════════════
 -- Type declarations
@@ -57,12 +43,12 @@ local lfs = pcall(require, "lfs") and require("lfs") or nil
 --- @field sources            SourceFile[]  -- all source files in the build
 --- @field metadata_path      string        -- path to word_count.metadata.h
 --- @field shared             table         -- cross-pass shared state
---- @field shared.word_counts WordCounts    -- populated by this pass
+--- @field shared.corpus      table         -- canonical corpus (required)
+--- @field shared.corpus.word_counts WordCounts -- canonical count table (populated by this pass)
 --- @field out_root           string        -- output root (e.g. "build/gen")
 --- @field project_root       string        -- project root (e.g. "code/")
 --- @field upstream           table<string, table>  -- per-pass upstream outputs
 --- @field flags              table         -- CLI flags
---- @field dry_run            boolean       -- if true, compute but don't write
 --- @field verbose            boolean       -- if true, log diagnostic info
 
 --- @class PassResult
@@ -84,9 +70,8 @@ local M = {}
 --- For most tokens (regular MIPS instructions) this returns 1.
 --- For `mac_X(...)` calls, this returns the resolved word count from `wc` (recursively if needed). For `nop2` etc., returns wc[name].
 --- For unknown macros, returns 1 and (optionally) warns.
----
---- @param token string        -- a single token from split_top_level_commas
---- @param wc WordCounts       -- the shared word-count table
+--- @param token string     -- a single token from split_top_level_commas
+--- @param wc    WordCounts -- the shared word-count table
 --- @return integer
 function M.count_token_words(token, wc)
 	local s = duffle.trim(token)
@@ -101,112 +86,42 @@ function M.count_token_words(token, wc)
 	return 1
 end
 
--- ┌────────────────────────────────────────────────────────────────────┐
--- │ Shared utility: scan_dir                                           │
--- └────────────────────────────────────────────────────────────────────┘
-
---- Recursively scan a directory for files matching a glob suffix.
---- No regex per the no_regex constraint — uses plain byte matching via `dir /b /s` on Windows.
----
---- The `.macs.h` files produced by the components pass always live at `<project_root>/<module>/gen/`.
---- We can shortcut the `dir /b /s` walk by listing modules first (one `dir /b /ad`), then walking each `<module>/gen/`
---- (one `dir /b` per module, no recursion). 
---- For projects with 2 modules and 0 .macs.h files, this drops the cost from ~52ms 
---- (full recursive walk of the entire project tree) to ~5ms.
----
---- @param dir string        -- directory to scan (absolute or relative)
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
--- Cache the scan_dir result per (dir, suffix) in package.loaded. 
--- Each `io.popen` call on Windows is ~50-100ms of subprocess overhead, so caching the result saves a fixed cost on every build.
--- The cache persists for the lifetime of the Lua process (cleared when ps1_meta.lua exits).
--- If a build removes/creates .macs.h files mid-process, the caller can invalidate by calling `M._invalidate_scan_cache()`.
-local SCAN_CACHE_KEY = "__word_count_eval_scan_cache__"
-
---- Scan `code/` for files matching `suffix` (e.g. `*.macs.h`).
---- Uses `lfs` (LuaFileSystem) when available — native directory enumeration at ~2ms.
---- Falls back to `dir /b /s` subprocess (~56ms) when `lfs` is not compiled.
----
---- @param dir string        -- project root directory
---- @param suffix string     -- file pattern, e.g. "*.macs.h"
---- @return string[]
-function M.scan_dir(dir, suffix)
-	local key = dir .. "\0" .. suffix
-
-	local cache = package.loaded[SCAN_CACHE_KEY]
-	if    cache and cache[key] then return cache[key] end
-
-	local results = {}
-
-	if lfs then
-		-- Native walk: list code/<module>/gen/ for matching files. Zero subprocess spawns.
-		local code_dir = dir .. "/code"
-		if lfs.attributes(code_dir, "mode") == "directory" then
-			for mod_name in lfs.dir(code_dir) do
-				if mod_name ~= "." and mod_name ~= ".." then
-					local gen_path = code_dir .. "/" .. mod_name .. "/gen"
-					if lfs.attributes(gen_path, "mode") == "directory" then
-						for fname in lfs.dir(gen_path) do
-							if fname:match("%.macs%.h$") then
-								results[#results + 1] = gen_path .. "/" .. fname
-							end
-						end
-					end
-				end
-			end
-		end
-	else
-		-- Fallback: single `dir /b /s` subprocess scoped to code\.
-		local pipe = io.popen(DIR_GLOB_CMD:format(dir, suffix))
-		if pipe then
-			for raw_line in pipe:lines() do
-				results[#results + 1] = raw_line:gsub(PATH_SEP_BACKSLASH, PATH_SEP_FORWARD)
-			end
-			pipe:close()
-		end
-	end
-
-	-- Cache the result (including empty results).
-	cache      = cache or {}
-	cache[key] = results
-	package.loaded[SCAN_CACHE_KEY] = cache
-
-	return results
-end
-
---- Invalidate the scan cache (call after creating new .macs.h files in the same Lua process — usually not needed).
-function M._invalidate_scan_cache() package.loaded[SCAN_CACHE_KEY] = nil end
-
 -- ┌────────────────────────────────────────────────────────────────────┐
 -- │ Pass entry: M.run(ctx) — "word-counts" pass                        │
 -- └────────────────────────────────────────────────────────────────────┘
 
---- Load metadata.h + scan for existing *.macs.h files into ctx.shared.word_counts.
---- Loading the .macs.h files is idempotent: entries from later (current-build) .macs.h files override metadata.h entries of the same name.
+--- 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.
 ---
+--- Contract:
+---   * `ctx.shared.corpus` MUST exist (canonical corpus ownership).
+---   * `ctx.metadata_path` MUST be a readable file path to the authored `word_count.metadata.h`.
+---   * The pass assigns exactly one table to `corpus.word_counts`.
+---     Consumers read the corpus-owned table directly.
+---     Consumers must read `corpus.word_counts` directly.
 --- @param ctx PassCtx
 --- @return PassResult
 function M.run(ctx)
-	local wc = {}
-
-	-- 1. Load metadata.h (the encoding-macro source of truth).
-	local meta_counts = duffle.load_word_counts(ctx.metadata_path)
-	for name, count in pairs(meta_counts) do wc[name] = count end
-
-	-- 2. Scan project_root recursively for *.macs.h files (component-macro source).
-	local macs_files = M.scan_dir(ctx.project_root, "*.macs.h")
-	for _, macs_path in ipairs(macs_files) do
-		local ok, mc = pcall(duffle.load_word_counts, macs_path)
-		if not ok then
-			io.stderr:write(string.format("[word_count_eval] parse error in '%s': %s\n", macs_path, tostring(mc)))
-		elseif type(mc) ~= "table" then
-			io.stderr:write(string.format("[word_count_eval] '%s' did not return a table (got %s)\n", macs_path, type(mc)))
-		else
-			for name, count in pairs(mc) do wc[name] = count end
-		end
+	-- 1. Canonical-corpus ownership gate.
+	local corpus = ctx.shared and ctx.shared.corpus
+	if type(corpus) ~= "table" then
+		error("word_count_eval.run requires ctx.shared.corpus (canonical corpus). The fixture must install the corpus before running this pass.", 0)
 	end
 
-	ctx.shared.word_counts = wc
+	-- 2. metadata_path gate.
+	if type(ctx.metadata_path) ~= "string" or ctx.metadata_path == "" then
+		error("word_count_eval.run requires ctx.metadata_path (path to the authored word_count.metadata.h).", 0)
+	end
+
+	-- 3. Load authored metadata. Generated .macs.h files are NOT scanned
+	--    (the pass computes their counts from the just-built bodies after disk emission; see passes/components.lua).
+	local wc = duffle.load_word_counts(ctx.metadata_path)
+
+	-- 4. Assign the count table. ONE assignment, no copy. The assignment creates no secondary alias.
+	corpus.word_counts = wc
+
 	return { outputs = {}, errors = {}, warnings = {} }
 end
 

scripts/pcsx_debug_helper/autoexec.lua

@@ -0,0 +1,51 @@
+-- autoexec.lua - pcsx_debug_helper plugin entry point.
+-- Packaged in scripts/pcsx_debug_helper.zip. Loaded by pcsx-redux via the -archive CLI flag (see scripts/launch_pcsx_debug.ps1).
+--
+-- Registers two web handlers for external CLI tools:
+--   /api/v1/lua/gte - full GTE state (32 data + 32 control regs + PC)
+--   /api/v1/lua/gp  - GP state summary (screenshot endpoint + VRAM endpoint refs)
+--
+-- The GTE handler reads COP2 regs via PCSX.getRegisters().CP2D/CP2C.
+-- The pcsx-redux gdb stub doesn't expose COP2, so this is the only way for external tools to see GTE state.
+--
+-- The GP handler is a thin pointer: 
+-- pcsx-redux's Lua API exposes only PCSX.GPU.takeScreenShot() (no GPUSTAT, no GP0/GP1 command log, no display state). For richer GP state, the existing web endpoints are the practical path:
+--   /api/v1/state/still      - PNG screenshot
+--   /api/v1/gpu/vram/raw     - VRAM raw bytes (1MB)
+--
+-- Companion: scripts/gdb/gdb_tape_atoms.gdb (covers GPRs + atom-aware stepping).
+
+local function register_handlers()
+    if not PCSX.WebServer then PCSX.WebServer = {} end
+    if not PCSX.WebServer.Handlers then PCSX.WebServer.Handlers = {} end
+
+    -- ── GTE state ──
+    PCSX.WebServer.Handlers.gte = function(req)
+        local r = PCSX.getRegisters()
+        local out = { "pc=0x" .. string.format("%x", r.pc) }
+        for i = 0, 31 do
+            out[#out + 1] = string.format("D[%d]=0x%08x C[%d]=0x%08x",
+                i, r.CP2D.r[i], i, r.CP2C.r[i])
+        end
+        return table.concat(out, "\n")
+    end
+
+		-- ── GP state (pointer to existing endpoints) ──
+		-- pcsx-redux's Lua GPU API exposes only takeScreenShot(); no GPUSTAT / GP0 / GP1 command log / display state.
+		-- We point to the existing web endpoints that DO expose those (when the emulator is actually rendering. Paused-at-BP frames won't have a fresh frame).
+    PCSX.WebServer.Handlers.gp = function(req)
+        local out = {
+            "gpu_screenshot_png=http://localhost:8080/api/v1/state/still",
+            "vram_raw=http://localhost:8080/api/v1/gpu/vram/raw (1MB VRAM)",
+            "gpustat=NOT_AVAILABLE_VIA_LUA",
+            "gp_command_log=NOT_AVAILABLE_VIA_LUA (use pcsx-redux Debug > GPU Logger)",
+            "hint_run_emulator_unpaused_for_screenshot",
+        }
+        return table.concat(out, "\n")
+    end
+end
+
+local ok, err = pcall(register_handlers)
+if ok then print("[pcsx_debug_helper] handlers registered: gte, gp")
+else       print("[pcsx_debug_helper] registration failed: " .. tostring(err))
+end

scripts/update_deps.ps1

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