amazing bug

wip: going to incremnetally test this.
mark current crashing path.
2026-08-14 19:48:13 +00:00 · 2026-08-11 01:25:40 -04:00 · 2026-08-11 01:25:09 -04:00 · 2026-08-10 23:29:38 -04:00 · 2026-08-10 23:04:43 -04:00 · 2026-08-10 19:53:34 -04:00
16 changed files with 551 additions and 1171 deletions
@@ -175,68 +175,6 @@ WORD_COUNT(mac_gte_sqr_v3, 8)
 ,	shift_aright_var(r_dz, r_dz, r_shift)
 WORD_COUNT(mac_gte_gpf_scale, 13)

-#define mac_normalize_v3s4(...) \
-	load_word(r_src, R_TapePtr, O_(Binds_NormalizeV3S4,src))     /* pop src ptr (scratch addr) */ \
-,	load_word(r_dst, R_TapePtr, O_(Binds_NormalizeV3S4,dst))     /* pop dst ptr (scratch addr) */ \
-,	add_ui_self(   R_TapePtr, S_(Binds_NormalizeV3S4)) \
-,	load_word(r_sx, r_src, O_(V3_S4,x)) \
-,	load_word(r_sy, r_src, O_(V3_S4,y)) \
-,	load_word(r_sz, r_src, O_(V3_S4,z)) \
-,	nop /* load-delay */	/* ── 48-word normalize body (preserved verbatim from ac_normalize_v3s4) ─────── */	/* Stage 1: mtc2 src → IR1/2/3, SQR fires (MAC1/2/3 = IR², IR ← MAC saturated) */ \
-,	gte_mv_to_data_r(r_sx, C2_IR1) \
-,	gte_mv_to_data_r(r_sy, C2_IR2) \
-,	gte_mv_to_data_r(r_sz, C2_IR3) \
-,	nop \
-,	gte_cmdw_sqr	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS */ \
-,	gte_mv_from_data_r(r_sq_y,      C2_MAC1) /* r_sq_y = MAC1 = sx² */ \
-,	gte_mv_from_data_r(r_sq_z,      C2_MAC2) /* r_sq_z = MAC2 = sy² */ \
-,	gte_mv_from_data_r(r_recip_est, C2_MAC3) /* r_recip_est = MAC3 = sz² */ \
-,	nop                                      /* MFC2→GPR load delay (1 slot) */ \
-,	add_u(r_recip_est, r_recip_est, r_sq_z)  /* r_recip_est += sy² */ \
-,	add_u(r_recip_est, r_recip_est, r_sq_y)  /* r_recip_est += sx² (sum = |v|²) */ \
-,	gte_mv_to_data_r(  r_recip_est, C2_LZCS) /* LZCS = |v|² */ \
-,	nop2 \
-,	gte_mv_from_data_r(r_lzcr, C2_LZCR)      /* r_lzcr = LZCR (count of leading bits) */ \
-,	nop                                      /* MFC2→GPR load delay (1 slot) */	/* Stage 3: compute shift amount, align |v|² to bit 24, lookup 1/|v| */ \
-,	and_i(         r_lzcr, r_lzcr, -2)       /* r_lzcr &= ~1 (force even for halving) */ \
-,	li_s(          r_shift, 31)              /* r_shift = 31 */ \
-,	sub_s(         r_shift, r_shift, r_lzcr) /* r_shift = 31 - LZCR */ \
-,	shift_aright(  r_shift, r_shift,  1)     /* r_shift = (31 - LZCR) / 2 */ \
-,	add_si(        r_tmp,   r_lzcr, -24)     /* r_tmp = LZCR - 24 (signed, for branch) */ \
-,	branch_lt_zero(r_tmp, atom_offset(srav_path, aligned_done)) \
-,	nop \
-,	jump_rel(             atom_offset(aligned_done, srav_path)) \
-,	shift_lleft_var(r_recip_est, r_recip_est, r_tmp)     /* BD-slot of branch_equal: r_recip_est = |v|² << (LZCR - 24) */ \
-,	atom_label(srav_path)                                 /* SRAV path: |v|² is small (top bit < bit 24) */ \
-,	li_s(          r_tmp, 24) \
-,	sub_s(         r_tmp, r_tmp, r_lzcr)                 /* r_tmp = 24 - LZCR */ \
-,	shift_aright_var(r_recip_est, r_recip_est, r_tmp)    /* r_recip_est = |v|² >> (24 - LZCR) */ \
-,	atom_label(aligned_done)                                /* Both paths converge here with |v|² aligned to bit 24 */	/* r_recip_est now holds |v|² aligned to bit 24 — convert to byte offset, -64 to skip zero pad. */ \
-,	add_si(        r_recip_est, r_recip_est, -64) \
-,	shift_lleft(   r_recip_est, r_recip_est, 1)           /* r_recip_est *= 2 (half-word index) */	/* Reference OUR local sqrtbl via &-address split. Compiler/linker resolves both halves. */ \
-,	load_upper_i(  r_tmp, u4_hi(& gte_normalize_sqr_tbl)) /* lui */ \
-,	or_i_self(     r_tmp, u4_lo(& gte_normalize_sqr_tbl)) /* ori */ \
-,	add_u(         r_tmp, r_tmp, r_recip_est)             /* r_tmp = sqrtbl base + byte offset (matches libgte 0x80016118: addu t5,t5,t4) */ \
-,	load_half(     r_recip_est, r_tmp, 0)                 /* r_recip_est = sqrtbl[r_recip_est] = 1/|v| estimate */ \
-,	nop                                                   /* retire load_half before MTC2 (matches libgte 0x80016120: nop) */	/* Stage 4: mtc2 IR0..3, GPF (MAC = IR0*IR), mfc2 MAC, srav finalize */ \
-,	gte_mv_to_data_r(r_recip_est, C2_IR0)  /* IR0 = 1/|v| estimate */ \
-,	gte_mv_to_data_r(r_sx,        C2_IR1)  /* IR1 = src.x */ \
-,	gte_mv_to_data_r(r_sy,        C2_IR2)  /* IR2 = src.y */ \
-,	gte_mv_to_data_r(r_sz,        C2_IR3)  /* IR3 = src.z */ \
-,	nop2                                   /* COP2 transfer latency (2 slots) */ \
-,	gte_cmdw_gpf \
-,	gte_mv_from_data_r(r_sx, C2_MAC1)      /* MAC1 → r_sx (overwrites src.x with raw reciprocal-scaled) */ \
-,	gte_mv_from_data_r(r_sy, C2_MAC2) \
-,	gte_mv_from_data_r(r_sz, C2_MAC3) \
-,	shift_aright_var(r_sx, r_sx, r_shift) \
-,	shift_aright_var(r_sy, r_sy, r_shift) \
-,	shift_aright_var(r_sz, r_sz, r_shift)	/* ── I/O wrapper tail (~3 words) ───────────────────────────────────────────── */ \
-,	store_word(r_sx, r_dst, O_(V3_S4,x)) \
-,	store_word(r_sy, r_dst, O_(V3_S4,y)) \
-,	store_word(r_sz, r_dst, O_(V3_S4,z))	/* ── atom_reads(R_TapePtr) atom_writes(R_TapePtr) ────────────────────────── */ \
-,	mac_yield()
-WORD_COUNT(mac_normalize_v3s4, 62)
-
 #define mac_gcmd_push(cmd, reg_transfer, reg_base, port) \
 	load_upper_i(reg_transfer, cmd >> 16) \
 ,	or_i_self(   reg_transfer, cmd & 0xFFFF) \
@@ -83,6 +83,10 @@ FI_ Slice_MipsCode ac_gte_gpf_scale(MipsAtomBuilder_R ab, U4 r_sx, U4 r_sy, U4 r
 	shift_aright_var(r_dz, r_dz, r_shift),
 })

+#pragma endregion MACs (Mips Atom Components)
+
+#pragma region Atom Procs
+
 /* ─── Local copy of PSYQ's sqrtbl (1/sqrt lookup table for VectorNormal). ───
 * Source: PSYQ 4.7 libgte sqrtbl at 0x800185B4 in hello_camera.elf.
 *   objdump -s --start-address=0x800185B4 --stop-address=0x800185F4 hello_camera.elf
@@ -97,7 +101,8 @@ FI_ Slice_MipsCode ac_gte_gpf_scale(MipsAtomBuilder_R ab, U4 r_sx, U4 r_sy, U4 r
 *   Octave 1 (entries 48- 95): mantissa in [0x10000, 0x20000)  output ~[0.707, 0.500]
 *   Octave 2 (entries 96-143): mantissa in [0x20000, 0x40000)  output ~[0.500, 0.354]
 *   Octave 3 (entries144-191): mantissa in [0x40000, 0x80000)  output ~[0.354, 0.251]
- * Within each octave, 8 sub-entries interpolate over the 8 fractional bits of the mantissa (the byte `(0x80 | (i mod 8))` for the lower-byte of the aligned value).
+ * Within each octave, 8 sub-entries interpolate over the 8 fractional bits of the mantissa
+ * (the byte `(0x80 | (i mod 8))` for the lower-byte of the aligned value).
 * Sampling the first value of each octave:
 *   [0]   0x1000 = 1.0000                 ; 1 / sqrt(1.0000)
 *   [48]  0x0e4f = 0.8940                 ; 1 / sqrt(1.2500)
@@ -149,131 +154,121 @@ internal S2 const gte_normalize_sqr_tbl[192] align_(2) = {
 };

 /* ─── Full normalize (all 4 stages inline) ───
+ * Generic 4-stage GTE normalize (SQR → sum+LZCR → align+sqrtbl → GPF+srav).
+ *
+ * Parameterized by caller-provided scratch base + src/dst offsets.
+ * The caller passes r_src_offset and r_dst_offset as compile-time constants
+ * (typically derived from O_ macros in the caller's struct schema, e.g., `O_(CallerBundleScratch, fwd)`).
+ *
+ * This design lets any caller (with a scratch base + struct schema) use `normalize_v3s4_proc` 
+ * without putting magic offsets in the C-side bundle helper — the offsets come from O_ macros at the call site.
+ *
+ * Body uses 9 GPRs (r_src_ptr..r_branch_tmp):
+ *   r_src_ptr, r_dst_ptr : src/dst pointers (computed from r_scratch + caller offsets)
+ *   r_tmp                : scratch (reserved for misc use)
+ *   r_mac1_scratch       : MAC1 result scratch (before sum into r_recip_est)
+ *   r_mac2_scratch       : MAC2 result scratch (clobbered to IR1 in stage 4)
+ *   r_recip_est          : |v|² sum + shift-input + sqrtbl[index] (the main chain)
+ *   r_lzcr               : LZCR value (consumed by stage 3 alignment calc)
+ *   r_shift              : final srav amount (consumed by stage 4 shift_aright_var)
+ *   r_branch_tmp         : scratch (shift count, branch target, sqrtbl base addr)
+ *
+ * Atom_labels are srav_path / aligned_done 
+ * (NOT namespaced — they're internal to this proc; 
+ * the metaprogram's per-atom-name enum emission handles any collision across different atoms/files that share the same labels).
+ *
+ * Pool cost: 11 GPRs (well within the 9-10 caller-trash GPR budget when r_scratch is a wave-context carrier).
+ *
 * Direct port of PSYQ libgte msc02.rel.text VectorNormal disassembly (0x800160a0..0x8001615c).
- *
- * Component variants that could apply:
- *   - `ac_gte_sqr_v3` (line ~56) covers stage 1's `mtc2 IR1/2/3 + nop + gte_cmdw_sqr`.
- *     We do NOT call it because the inlined version of stage 1 is followed immediately by stage 2's `mfc2 MAC1/2/3` chain 
- *     (the operands of `ac_gte_sqr_v3`'s r_sq_x/r_sq_y/r_sq_z would each require an explicit GPR to receive the MAC result,
- *      then a move to land in r_recip_est for the partial-sum chain).
- *     Inlining saves ~3 cycles of `or`-merge + register pressure
- *     (squared MAC3 lands DIRECTLY in r_recip_est which doubles as the partial-sum accumulator and the LZCS input — see r_recip_est row below).
- *   - `ac_gte_gpf_scale` (line ~71) covers stage 4's `mtc2 IR0..3 + nop2 + gte_cmdw_gpf + mfc2 MAC1/2/3 + sra`.
- *     We do NOT call it for the symmetric reason: the normalize in-place semantics overwrite the input regs (r_sx/r_sy/r_sz) with the normalized output,
- *     which `ac_gte_gpf_scale`'s r_dx/r_dy/r_dz output GPRs would not match.
- *   `gte_cmdw_sqr` and `gte_cmdw_gpf` primitive macros ARE used in the inlined body, so changes to those primitives
- *   (e.g., the libgte `fake_cmd` signature bits) propagate automatically. The components remain available for callers that want the explicit GPR-shape variants.
- *
- * Argument aliasing (9 unique physical regs needed, can drop to 8 with r_sq_y ≡ r_lzcr):
- *   r_sx, r_sy, r_sz        : src components in regs (clobbered by mtc2 → IR1/2/3 in stage 1, then by mfc2 MAC1/2/3 in stage 4 — in-place semantics)
- *   r_sq_y, r_sq_z          : MAC2, MAC3 → DIE after stage 2 accumulate (r_sq_y can alias r_lzcr after stage 2 to save one reg)
- *   r_recip_est             : ≡ r_sqmag — multi-purpose (holds |v|² in stage 2, shift-input in stage 3, sqrtbl[index] in stage 4)
- *   r_lzcr                  : LZCR value, alive across stage 3 (srav path needs `24 - LZCR`)
- *   r_shift                 : (31 - LZCR & ~1) >> 1 — final srav amount (stages 3-4)
- *   r_tmp                   : scratch (shift count, branch target, lookup addr, table base)
- *
- * GPR ccount peak: 9.
+ * Words: ~59 (matches libgte 0x800160a0..0x8001615c at +/- 0-2 words for BD-slot reshuffling).
+ * Sqrtbl: hardcoded to 0x800185B4 (libgte msc02.rel.data). Note: swapped to local.
 * Pipeline: clobbers IR0..3, MAC1..3, LZCS, LZCR.
- * Words: ~35 (pending re-gen; matches libgte 0x800160a0..0x8001615c at +/- 0-2 words for BD-slot reshuffling).
- * Sqrtbl: hardcoded to 0x800185B4 (libgte msc02.rel.data). Note: swapped to local. */
-
-/* ─── Binds_NormalizeV3S4 — declared here so the MipsAtom_Proc_ body can reference
- * O_(Binds_NormalizeV3S4,*). Inlined at the proc-call site; not exposed in gen/macs.h. */
-typedef Struct_(Binds_NormalizeV3S4) {
-	U4 src;  /* V3_S4* (scratch address — read from tape) */
-	U4 dst;  /* V3_S4* (scratch address — write to tape) */
-};
-
-/* NOTE: The bundle-specific scratchpad offset schema was intentionally kept out of this file.
- * gte.atom.c is the GENERIC GTE primitives file — it exposes only the parameter-style normalize_v3s4_proc for any future caller. */
-I_ void normalize_v3s4_proc(
-		MipsAtomBuilder_R ab
-	,	U4 r_src              /* GPR code: scratch base carrier (wave-context, e.g., R_T4) */
-	,	U4 r_dst              /* GPR code: scratch dst pointer carrier (wave-context, e.g., R_T5) */
-	,	U4 r_sx, U4 r_sy, U4 r_sz              /* GPR codes: src.x/y/z scratch (atom-local) */
-	,	U4 r_sq_y, U4 r_sq_z                   /* GPR codes: MAC1/2 scratch (atom-local) */
-	,	U4 r_recip_est                         /* GPR code: |v|² sum + shift-input + sqrtbl[index] (atom-local) */
-	,	U4 r_lzcr                              /* GPR code: LZCR value (atom-local) */
-	,	U4 r_shift                             /* GPR code: final srav amount (atom-local) */
-	,	U4 r_tmp                               /* GPR code: scratch (shift count, branch target, lookup addr, table base) */
-)
+ */
 /* MipsAtom_Proc_ wrapper: declares the static MipsCode[] body, then calls atombuilder_unroll(ab, ...) to copy the encoded instructions into the caller's MipsAtomBuilder arena.  */
+I_ void normalize_v3s4_proc(MipsAtomBuilder_R ab,	U4 r_scratch /* GPR code: scratch base carrier (e.g., R_T4 = R_ResolveScratch) */
+	,	U4 r_src_offset, U4 r_dst_offset     /* GPR codes: PARAMETERIZED offsets (caller passes O_ macros) */
+	,	U4 r_src_ptr, U4 r_dst_ptr, U4 r_tmp /* GPR codes: 3 scratch regs (src/dst computed + tmp) */
+	,	U4 r_mac1_scratch, U4 r_mac2_scratch /* GPR codes: 2 more: MAC1/MAC2 scratch */
+	,	U4 r_recip_est                       /* GPR code: |v|² sum + shift-input + sqrtbl[index] */
+	,	U4 r_lzcr, U4 r_shift                /* GPR codes: lzcr + final srav amount */
+	,	U4 r_branch_tmp                      /* GPR code: scratch (shift count, branch target, lookup addr) */
+)
 MipsAtom_Proc_(normalize_v3s4, ab, {
-	/* ── I/O wrapper (~10 words: 3 bind-pop + 3 src-load + 1 nop + 3 dst-store) ─── */
-	load_word(r_src, R_TapePtr, O_(Binds_NormalizeV3S4,src)),     /* pop src ptr (scratch addr) */
-	load_word(r_dst, R_TapePtr, O_(Binds_NormalizeV3S4,dst)),     /* pop dst ptr (scratch addr) */
-	add_ui_self(   R_TapePtr, S_(Binds_NormalizeV3S4)),
-	load_word(r_sx, r_src, O_(V3_S4,x)),
-	load_word(r_sy, r_src, O_(V3_S4,y)),
-	load_word(r_sz, r_src, O_(V3_S4,z)),
+	add_si(r_src_ptr, r_scratch, r_src_offset),     /* r_src_ptr = &src */
+	add_si(r_dst_ptr, r_scratch, r_dst_offset),     /* r_dst_ptr = &dst */
+	nop,
+
+	/* Load src.x/y/z from r_src_ptr (caller-determined address) into r_mac2_scratch/r_recip_est/r_branch_tmp. */
+	load_word(r_mac2_scratch, r_src_ptr, O_(V3_S4,x)),
+	load_word(r_recip_est,    r_src_ptr, O_(V3_S4,y)),
+	load_word(r_branch_tmp,   r_src_ptr, O_(V3_S4,z)),
 	nop, /* load-delay */

-	/* ── 48-word normalize body (preserved verbatim from ac_normalize_v3s4) ─────── */
-	// Stage 1: mtc2 src → IR1/2/3, SQR fires (MAC1/2/3 = IR², IR ← MAC saturated)
-	gte_mv_to_data_r(r_sx, C2_IR1),
-	gte_mv_to_data_r(r_sy, C2_IR2),
-	gte_mv_to_data_r(r_sz, C2_IR3),
+	/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */
+	gte_mv_to_data_r(r_mac2_scratch, C2_IR1),
+	gte_mv_to_data_r(r_recip_est,    C2_IR2),
+	gte_mv_to_data_r(r_branch_tmp,   C2_IR3),
 	nop, gte_cmdw_sqr,
-	// Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS
-	gte_mv_from_data_r(r_sq_y,      C2_MAC1), /* r_sq_y = MAC1 = sx² */
-	gte_mv_from_data_r(r_sq_z,      C2_MAC2), /* r_sq_z = MAC2 = sy² */
-	gte_mv_from_data_r(r_recip_est, C2_MAC3), /* r_recip_est = MAC3 = sz² */
-	nop,                                      /* MFC2→GPR load delay (1 slot) */
-	add_u(r_recip_est, r_recip_est, r_sq_z),  /* r_recip_est += sy² */
-	add_u(r_recip_est, r_recip_est, r_sq_y),  /* r_recip_est += sx² (sum = |v|²) */
-	gte_mv_to_data_r(  r_recip_est, C2_LZCS), /* LZCS = |v|² */
+
+	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. */
+	gte_mv_from_data_r(r_mac1_scratch, C2_MAC1),
+	gte_mv_from_data_r(r_mac2_scratch, C2_MAC2),
+	gte_mv_from_data_r(r_lzcr,         C2_MAC3),
+	nop,
+	add_u(r_lzcr, r_lzcr, r_mac2_scratch),
+	add_u(r_lzcr, r_lzcr, r_mac1_scratch),
+	gte_mv_to_data_r(r_lzcr, C2_LZCS),
 	nop2,
-	gte_mv_from_data_r(r_lzcr, C2_LZCR),      /* r_lzcr = LZCR (count of leading bits) */
-	nop,                                      /* MFC2→GPR load delay (1 slot) */
-	// Stage 3: compute shift amount, align |v|² to bit 24, lookup 1/|v|
-	and_i(         r_lzcr, r_lzcr, -2),       /* r_lzcr &= ~1 (force even for halving) */
-	li_s(          r_shift, 31),              /* r_shift = 31 */
-	sub_s(         r_shift, r_shift, r_lzcr), /* r_shift = 31 - LZCR */
-	shift_aright(  r_shift, r_shift,  1),     /* r_shift = (31 - LZCR) / 2 */
-	add_si(        r_tmp,   r_lzcr, -24),     /* r_tmp = LZCR - 24 (signed, for branch) */
-	branch_lt_zero(r_tmp, atom_offset(srav_path, aligned_done)), nop,
-	jump_rel(             atom_offset(aligned_done, srav_path)),
-	shift_lleft_var(r_recip_est, r_recip_est, r_tmp),     /* BD-slot of branch_equal: r_recip_est = |v|² << (LZCR - 24) */
-atom_label(srav_path)                                 /* SRAV path: |v|² is small (top bit < bit 24) */
-	li_s(          r_tmp, 24),
-	sub_s(         r_tmp, r_tmp, r_lzcr),                 /* r_tmp = 24 - LZCR */
-	shift_aright_var(r_recip_est, r_recip_est, r_tmp),    /* r_recip_est = |v|² >> (24 - LZCR) */
-atom_label(aligned_done)                                /* Both paths converge here with |v|² aligned to bit 24 */
-	/* r_recip_est now holds |v|² aligned to bit 24 — convert to byte offset, -64 to skip zero pad. */
-	add_si(        r_recip_est, r_recip_est, -64),
-	shift_lleft(   r_recip_est, r_recip_est, 1),           /* r_recip_est *= 2 (half-word index) */
-	/* Reference OUR local sqrtbl via &-address split. Compiler/linker resolves both halves. */
-	load_upper_i(  r_tmp, u4_hi(& gte_normalize_sqr_tbl)), /* lui */
-	or_i_self(     r_tmp, u4_lo(& gte_normalize_sqr_tbl)), /* ori */
-	add_u(         r_tmp, r_tmp, r_recip_est),             /* r_tmp = sqrtbl base + byte offset (matches libgte 0x80016118: addu t5,t5,t4) */
-	load_half(     r_recip_est, r_tmp, 0),                 /* r_recip_est = sqrtbl[r_recip_est] = 1/|v| estimate */
-	nop,                                                   /* retire load_half before MTC2 (matches libgte 0x80016120: nop) */
-	// Stage 4: mtc2 IR0..3, GPF (MAC = IR0*IR), mfc2 MAC, srav finalize
-	gte_mv_to_data_r(r_recip_est, C2_IR0),  /* IR0 = 1/|v| estimate */
-	gte_mv_to_data_r(r_sx,        C2_IR1),  /* IR1 = src.x */
-	gte_mv_to_data_r(r_sy,        C2_IR2),  /* IR2 = src.y */
-	gte_mv_to_data_r(r_sz,        C2_IR3),  /* IR3 = src.z */
-	nop2,                                   /* COP2 transfer latency (2 slots) */
-	gte_cmdw_gpf,
-	gte_mv_from_data_r(r_sx, C2_MAC1),      /* MAC1 → r_sx (overwrites src.x with raw reciprocal-scaled) */
-	gte_mv_from_data_r(r_sy, C2_MAC2),
-	gte_mv_from_data_r(r_sz, C2_MAC3),
-	shift_aright_var(r_sx, r_sx, r_shift),
-	shift_aright_var(r_sy, r_sy, r_shift),
-	shift_aright_var(r_sz, r_sz, r_shift),
+	gte_mv_from_data_r(r_shift, C2_LZCR),
+	nop,

-	/* ── I/O wrapper tail (~3 words) ───────────────────────────────────────────── */
-	store_word(r_sx, r_dst, O_(V3_S4,x)),
-	store_word(r_sy, r_dst, O_(V3_S4,y)),
-	store_word(r_sz, r_dst, O_(V3_S4,z)),
+	/* Stage 3: compute srav amount (r_lzcr) + align |v|² to bit 24. */
+	and_i(   r_shift, r_shift, -2),
+	li_s(    r_lzcr,           31),
+	sub_s(   r_lzcr, r_lzcr, r_shift),
+	shift_aright(r_lzcr, r_lzcr, 1),
+	/* r_branch_tmp = LZCR - 24 (overwrites r_branch_tmp; src.z no longer needed after SQR) */
+	add_si(  r_branch_tmp, r_shift, -24),
+	branch_lt_zero(r_branch_tmp, atom_offset(srav_path, aligned_done)), nop,
+	jump_rel(atom_offset(aligned_done, srav_path)),
+	shift_lleft_var(r_lzcr, r_lzcr, r_branch_tmp),     /* when r_branch_tmp < 0 (LZCR < 24): shift r_lzcr left by (24-LZCR) */
+	atom_label(srav_path)
+	li_s(    r_branch_tmp, 24),
+	sub_s(   r_branch_tmp, r_branch_tmp, r_shift),
+	shift_aright_var(r_lzcr, r_lzcr, r_branch_tmp),    /* when r_branch_tmp >= 0 (LZCR >= 24): shift r_lzcr right by (LZCR-24) */
+	atom_label(aligned_done)
+	/* r_lzcr holds |v|² aligned to bit 24. */
+	add_si(     r_lzcr, r_lzcr, -64),
+	shift_lleft(r_lzcr, r_lzcr, 1),
+	load_upper_i(r_branch_tmp, u4_hi(& gte_normalize_sqr_tbl)),
+	or_i_self(   r_branch_tmp, u4_lo(& gte_normalize_sqr_tbl)),
+	add_u(r_branch_tmp, r_branch_tmp, r_lzcr),
+	load_half(r_lzcr,   r_branch_tmp, 0), nop,
+
+	/* Stage 4: GPF + srav finalize (r_lzcr = srav_amount carried from stage 3). */
+	gte_mv_to_data_r(r_lzcr,         C2_IR0),
+	gte_mv_to_data_r(r_mac2_scratch, C2_IR1),
+	gte_mv_to_data_r(r_recip_est,    C2_IR2),
+	gte_mv_to_data_r(r_branch_tmp,   C2_IR3),
+	nop2, gte_cmdw_gpf,
+	gte_mv_from_data_r(r_mac2_scratch, C2_MAC1),
+	gte_mv_from_data_r(r_recip_est,    C2_MAC2),
+	gte_mv_from_data_r(r_branch_tmp,   C2_MAC3),
+	shift_aright_var(r_mac2_scratch, r_mac2_scratch, r_lzcr),
+	shift_aright_var(r_recip_est,    r_recip_est,    r_lzcr),
+	shift_aright_var(r_branch_tmp,   r_branch_tmp,   r_lzcr),
+
+	/* Store result.x/y/z to r_dst_ptr (caller-determined dst address). */
+	store_word(r_mac2_scratch, r_dst_ptr, O_(V3_S4,x)),
+	store_word(r_recip_est,    r_dst_ptr, O_(V3_S4,y)),
+	store_word(r_branch_tmp,   r_dst_ptr, O_(V3_S4,z)),

-	/* ── atom_reads(R_TapePtr) atom_writes(R_TapePtr) ────────────────────────── */
 	mac_yield()
 })

-#pragma endregion MACs (Mips Atom Components)
+#pragma endregion Atom Procs

-#pragma region Bsked Atoms
+#pragma region Baked Atoms

 typedef Struct_(Binds_SetGteMT3S2S4) {
 	MT3_S2S4* transform;
@@ -27,9 +27,9 @@
 * to author and compose programs with. From here various conventions can be further applied. 
 * To make things easier  to understand it may be better to focus on what this ABI does not have. 
 * It does not have have any branching within the tape but relative branches within atoms or between atoms.
- * Branching nearly is always downstream. Stack usage is non-existent.
+ * Branching nearly is always downstream. Atuomatic stack usage is non-existent.
 * Push/Pop, FIFO, or Arena/Bump data structures are used by atoms explicitly.
- * In it's current form with the C11 macro dsl, the user also has fullfill manual register allocation per atom.
+ * In it's current form with the C11 macro DSL, the user also has fullfill manual register allocation per atom.
 * 
 * One of the remarkable things about utilizing this ABI is its essentially interopable with CPUs, GPUs, FPGA, 
 * or, basically anything from the 5th generation consoles and onward.
@@ -186,14 +186,14 @@ FI_ void tape_run_a02_s07(Tape tape) { register U4* tape_ptr rgcc(R_TapePtr) = u
 typedef Relative_(FArena) Struct_(TapeBuilder) { U4 ptr; U4 capacity; U4 used; };
 FI_ void        tb_init(TapeBuilder* tb, FArena* arena) { tb->ptr = arena->start; tb->used = 0; }
 FI_ TapeBuilder tb_make_old(             FArena* arena) { return (TapeBuilder){ arena->start, 0 }; }
-FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ mem.ptr, mem.len, 0 }; }
+FI_ TapeBuilder tb_make(Slice mem) { return (TapeBuilder){ u4_(mem.ptr), mem.len, 0 }; }  /* capacity in elements (matches used units) */

 FI_ void tb_emit(TapeBuilder* tb, MipsAtom* atom) { u4_r(tb->ptr)[tb->used] = u4_(atom); ++ tb->used; }
 FI_ void tb_data(TapeBuilder* tb, U4        data) { u4_r(tb->ptr)[tb->used] = u4_(data); ++ tb->used; }
 #define tb_emit_(atom)        tb_emit(& tb, atom)
 #define tb_data_(field, data) tb_data(& tb, u4_(data))

-FI_ void tb_emit_bundle(TapeBuilder_R tb, Slice_MipsAtom atoms) { mem_copy(u4_(tb->ptr), u4_(atoms.ptr), tb->used); tb->used += atoms.len; }
+FI_ void tb_emit_bundle(TapeBuilder_R tb, Slice_MipsAtom atoms) { mem_copy(u4_(tb->ptr), u4_(atoms.ptr), S_slice(atoms)); tb->used += atoms.len; }

 FI_ Tape tb_end  (TapeBuilder* tb) { tb_emit(tb,tape_exit); return (Tape){ C_(U4*,tb->ptr), tb->used }; }
 FI_ Tape tb_slice(TapeBuilder  tb) {                        return (Tape){ C_(U4*,tb.ptr),  tb.used }; }
@@ -242,18 +242,23 @@ typedef Relative_(FArena) Struct_(MipsAtomBuilder) { U4 start; U4 capacity; U4 u
 // to something that can fit within instruction cache?

 FI_ void atombuilder_unroll(MipsAtomBuilder_R ab, Slice_MipsCode code) {
+	/* code.len is in ELEMENTS (per slice_from_array convention); ab->used is also in elements
+	 * (the init uses `ab->used * sizeof(U4)` for byte offset arithmetic — sizeof(U4)==4==sizeof(MipsCode)).
+	 * mem_copy needs BYTES, so we use S_slice(code) for the length. */
 	assert(ab->capacity - ab->used - code.len);
 	U4* dest = (U4*)ab->start + ab->used;  /* write at next-available slot (arena accumulation) */
-	mem_copy(u4_(dest), u4_(code.ptr), code.len);
+	mem_copy(u4_(dest), u4_(code.ptr), S_slice(code));
 	mem_bump(ab->start, ab->capacity, & ab->used, code.len);
 }
 #define atombuilder_unroll_mac(ab, mac) atombuilder_unroll(ab, slice_arg_from_array(Slice_MipsCode, mac))

 // When done authoring, utilize this to cap-off the atom (if not utilizing a MipsAtom_Proc).
 FI_ void atombuilder_end(MipsAtomBuilder_R ab) {
+	/* ac_yield is a MipsCode[] of 4 elements; S_(ac_yield)=bytes, array_len(ac_yield)=elements.
+	 * ab->used is in elements, so mem_bump needs element count. */
 	U4* dest = (U4*)ab->start + ab->used;  /* write at next-available slot */
 	mem_copy(u4_(dest), u4_(ac_yield), S_(ac_yield));
-	mem_bump(ab->start, ab->capacity, & ab->used, S_(ac_yield));
+	mem_bump(ab->start, ab->capacity, & ab->used, array_len(ac_yield));
 }

 #define mipsatom_from_builder(ab) C_(MipsAtom*, (ab).start)
@@ -58,13 +58,13 @@ typedef Struct_(Str8)        { UTF8* ptr; U4 len; };
 typedef Struct_(Slice_Str8)  { Str8* ptr; U4 len; };
 #define slit(string_literal) (Str8){ (UTF8*) string_literal, S_(string_literal) - 1 }

-typedef Struct_(Slice) { U4 ptr, len; }; // Untyped Slice
-FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){ptr, len}; }
+typedef Struct_(Slice) { B1* ptr; U4 len; }; // Untyped Slice (byte-addressable; .len in elements)
+FI_ Slice slice_ut_(U4 ptr, U4 len) { return (Slice){(B1*)ptr, len}; }

 #define Slice_(type)       Struct_(tmpl(Slice,type)) { type* ptr; U4 len; }
 typedef Slice_(B1);
 #define slice_assert(s)    do { assert((s).ptr != 0); assert((s).len > 0); } while(0)
-#define slice_end(slice)   ((slice).ptr + (slice).len)
+#define slice_end(slice)   ((slice).ptr + S_slice(slice) / S_(B1))  /* byte-ptr arithmetic; .len is in elements per slice convention */
 #define S_slice(s)         ((s).len * S_((s).ptr[0]))

 #define slice_ut(ptr,len)  slice_ut_(u4_(ptr),     u4_(len))
@@ -73,16 +73,16 @@ typedef Slice_(B1);

 #define slice_iter(container, iter)     (T_((container).ptr) iter = (container).ptr; iter != slice_end(container); ++ iter)
 #define slice_arg_from_array(type, ...) & (tmpl(Slice,type)) { .ptr = array_decl(type,__VA_ARGS__), .len = array_len( array_decl(type,__VA_ARGS__)) }
-#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = S_(array) }
+#define slice_from_array(type, array)     (tmpl(Slice,type)) { .ptr = array, .len = S_(array) / S_(type) }  /* .len in elements (matches S_slice/slice_arg_from_array convention) */

-FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(s.ptr, s.len); }
+FI_ void slice_zero_(Slice s) { slice_assert(s); mem_zero(u4_(s.ptr), S_slice(s)); }
 #define  slice_zero(s)        slice_zero_(slice_to_ut(s))

 FI_ void slice_copy_(Slice dest, Slice src) {
-	assert(dest.len >= src.len);
+	assert(S_slice(dest) >= S_slice(src));
 	slice_assert(dest);
 	slice_assert(src);
-	mem_copy(dest.ptr, src.ptr, src.len);
+	mem_copy(u4_(dest.ptr), u4_(src.ptr), S_slice(src));
 }
 #define slice_copy(dest, src) do {  \
 	static_assert(T_same(dest, src)); \
@@ -98,8 +98,8 @@ typedef Slice_(U4);
 typedef Opt_(farena)    { U4 alignment, type_width; };
 typedef Struct_(FArena) { U4 start, capacity, used; };
 FI_ void farena_init(FArena_R arena, Slice mem) {  assert(arena != nullptr);
-	arena->start    = mem.ptr;
-	arena->capacity = mem.len;
+	arena->start    = u4_(mem.ptr);
+	arena->capacity = S_slice(mem);  /* FArena.used is in BYTES; capacity must be bytes too */
 	arena->used     = 0;
 }
 FI_ FArena farena_make(Slice mem) { FArena a; farena_init(& a, mem); return a; }
@@ -109,7 +109,7 @@ I_  Slice  farena_push(FArena_R arena, U4 amount, Opt_farena o) {
 	U4 to_commit = align_pow2(desired, o.alignment ?  o.alignment : MEM_ALIGNMENT_DEFAULT);
 	U4 ptr       = arena->start + arena->used;
 	mem_bump(arena->start, arena->capacity, & arena->used, to_commit);
-	return (Slice){ ptr, to_commit };
+	return (Slice){ (B1*)ptr, to_commit };
 }
 FI_ void farena_reset (FArena_R arena) { arena->used = 0; }
 FI_ void farena_rewind(FArena_R arena, U4 save_point) {
@@ -0,0 +1,13 @@
+#ifdef INTELLISENSE_DIRECTIVES
+#pragma once
+#endif
+// Auto-generated by ps1_meta.lua (passes/auto_reg.lua) — DO NOT EDIT
+// Directory: C:\projects\Pikuma\ps1\code\hello_camera
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.c
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.h
+//   source: C:/projects/Pikuma/ps1/code/hello_camera/hello_camera.atom.c
+// Per-phase register allocations resolved by the lua pass.
+// R_<Sym>_Code = <chosen GPR's _Code constant> for every marker in this directory.
+
+#define R_GpTmp_Code R_V0_Code
+
@@ -39,350 +39,3 @@ WORD_COUNT(mac_put_disp_env, 5)
 ,	mac_gcmd_push(gp0_word_nop(), reg_transfer, reg_base, port)
 WORD_COUNT(mac_put_draw_env, 16)

-#define mac_resolve_look_at__input_and_sub(...) \
-	load_word(r_target_ptr, R_TapePtr, O_(Binds_ResolveLookAtSub,target)) \
-,	load_word(r_eye_ptr,    R_TapePtr, O_(Binds_ResolveLookAtSub,eye)) \
-,	load_word(r_up_in_ptr,  R_TapePtr, O_(Binds_ResolveLookAtSub,up_in)) \
-,	add_ui_self(            R_TapePtr, S_(Binds_ResolveLookAtSub)) \
-,	load_word(r_scratch, R_TapePtr, O_(Binds_ResolveLookAtScratch,scratch_base)) \
-,	add_ui_self(         R_TapePtr, S_(Binds_ResolveLookAtScratch))	/* Stage eye.x/y/z into the scratchpad (atom 6 reads these for the translation
-	 * column). Reuse r_tmp0/r_tmp1/r_tmp2. Offsets via O_(ResolveLookAtScratch,*). */ \
-,	load_word(r_tmp0, r_eye_ptr, O_(P3_S4,x)) \
-,	load_word(r_tmp1, r_eye_ptr, O_(P3_S4,y)) \
-,	load_word(r_tmp2, r_eye_ptr, O_(P3_S4,z)) \
-,	nop /* load-delay */ \
-,	store_word(r_tmp0, r_scratch, O_(ResolveLookAtScratch,eye.x)) \
-,	store_word(r_tmp1, r_scratch, O_(ResolveLookAtScratch,eye.y)) \
-,	store_word(r_tmp2, r_scratch, O_(ResolveLookAtScratch,eye.z))	/* Stage up_in.x/y/z into the scratchpad (atom 2 reads these for the outer
-	 * product with uz). Reuse r_tmp0/r_tmp1/r_tmp2. */ \
-,	load_word(r_tmp0, r_up_in_ptr, O_(V3_S4,x)) \
-,	load_word(r_tmp1, r_up_in_ptr, O_(V3_S4,y)) \
-,	load_word(r_tmp2, r_up_in_ptr, O_(V3_S4,z)) \
-,	nop /* load-delay */ \
-,	store_word(r_tmp0, r_scratch, O_(ResolveLookAtScratch,up_in.x)) \
-,	store_word(r_tmp1, r_scratch, O_(ResolveLookAtScratch,up_in.y)) \
-,	store_word(r_tmp2, r_scratch, O_(ResolveLookAtScratch,up_in.z))	/* Compute fwd = target - eye. */ \
-,	load_word(r_tmp0, r_target_ptr, O_(P3_S4,x)) \
-,	load_word(r_tmp1, r_target_ptr, O_(P3_S4,y)) \
-,	load_word(r_tmp2, r_target_ptr, O_(P3_S4,z)) \
-,	load_word(r_tmp3, r_eye_ptr,    O_(P3_S4,x)) \
-,	load_word(R_AT,   r_eye_ptr,    O_(P3_S4,y)) \
-,	load_word(R_V0,   r_eye_ptr,    O_(P3_S4,z)) \
-,	nop /* load-delay */ \
-,	sub_u(r_tmp0, r_tmp0, r_tmp3) \
-,	sub_u(r_tmp1, r_tmp1, R_AT) \
-,	sub_u(r_tmp2, r_tmp2, R_V0)	/* Store fwd.x/y/z (atom 1 reads these as the normalize src). */ \
-,	store_word(r_tmp0, r_scratch, O_(ResolveLookAtScratch,fwd.x)) \
-,	store_word(r_tmp1, r_scratch, O_(ResolveLookAtScratch,fwd.y)) \
-,	store_word(r_tmp2, r_scratch, O_(ResolveLookAtScratch,fwd.z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__input_and_sub, 34)
-
-#define mac_resolve_look_at__cross_uz_up_in_to_right(...) \
-	add_si(r_g, r_scratch, O_(ResolveLookAtScratch,uz))       /* r_g = &uz */ \
-,	add_si(r_h, r_scratch, O_(ResolveLookAtScratch,up_in))    /* r_h = &up_in */ \
-,	add_si(r_f, r_scratch, O_(ResolveLookAtScratch,right))    /* r_f = &right (out) */ \
-,	nop	/* Load a (uz).x/y/z into r_a/r_b/r_c. */ \
-,	load_word(r_a, r_g, O_(V3_S4,x)) \
-,	load_word(r_b, r_g, O_(V3_S4,y)) \
-,	load_word(r_c, r_g, O_(V3_S4,z)) \
-,	nop	/* Load b (up_in).x/y/z into r_d + R_AT/R_V0 (hardcoded; reusing the
-	 * body's last two loads is fine because the load-delay slot is the nop
-	 * after the third load, and mtc2 below doesn't read these regs). */ \
-,	load_word(r_d,  r_h, O_(V3_S4,x)) \
-,	load_word(R_AT, r_h, O_(V3_S4,y)) \
-,	load_word(R_V0, r_h, O_(V3_S4,z)) \
-,	nop	/* mtc2 a → IR1/2/3, b → D1/2/3 (VXY0/VZ0/VXY1). */ \
-,	gte_mv_to_data_r(r_a,  C2_IR1) \
-,	gte_mv_to_data_r(r_b,  C2_IR2) \
-,	gte_mv_to_data_r(r_c,  C2_IR3) \
-,	gte_mv_to_data_r(r_d,  C2_VXY0)   /* D1 = b.x */ \
-,	gte_mv_to_data_r(R_AT, C2_VZ0)    /* D2 = b.y */ \
-,	gte_mv_to_data_r(R_V0, C2_VXY1)   /* D3 = b.z */ \
-,	nop2  /* MTC2 retirement (CPU→COP2 2-slot delay) */ \
-,	gte_cmdw_outer_product  /* OP fires; MAC1/2/3 = a × b */	/* mfc2 MAC1/2/3 → r_a/r_b/r_c (out.x/y/z). */ \
-,	gte_mv_from_data_r(r_a, C2_MAC1) \
-,	gte_mv_from_data_r(r_b, C2_MAC2) \
-,	gte_mv_from_data_r(r_c, C2_MAC3) \
-,	nop  /* MFC2 retirement */	/* Store out.x/y/z to r_f (out ptr = scratch+32). */ \
-,	store_word(r_a, r_f, O_(V3_S4,x)) \
-,	store_word(r_b, r_f, O_(V3_S4,y)) \
-,	store_word(r_c, r_f, O_(V3_S4,z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__cross_uz_up_in_to_right, 29)
-
-#define mac_resolve_look_at__cross_uz_ux_to_up(...) \
-	add_si(r_g, r_scratch, O_(ResolveLookAtScratch,uz))       /* r_g = &uz */ \
-,	add_si(r_h, r_scratch, O_(ResolveLookAtScratch,ux))       /* r_h = &ux */ \
-,	add_si(r_f, r_scratch, O_(ResolveLookAtScratch,up))       /* r_f = &up (out) */ \
-,	nop	/* Load a (uz).x/y/z into r_a/r_b/r_c. */ \
-,	load_word(r_a, r_g, O_(V3_S4,x)) \
-,	load_word(r_b, r_g, O_(V3_S4,y)) \
-,	load_word(r_c, r_g, O_(V3_S4,z)) \
-,	nop	/* Load b (ux).x/y/z into r_d + R_AT/R_V0. */ \
-,	load_word(r_d,  r_h, O_(V3_S4,x)) \
-,	load_word(R_AT, r_h, O_(V3_S4,y)) \
-,	load_word(R_V0, r_h, O_(V3_S4,z)) \
-,	nop	/* mtc2 a → IR1/2/3, b → D1/2/3 (VXY0/VZ0/VXY1). */ \
-,	gte_mv_to_data_r(r_a,  C2_IR1) \
-,	gte_mv_to_data_r(r_b,  C2_IR2) \
-,	gte_mv_to_data_r(r_c,  C2_IR3) \
-,	gte_mv_to_data_r(r_d,  C2_VXY0) \
-,	gte_mv_to_data_r(R_AT, C2_VZ0) \
-,	gte_mv_to_data_r(R_V0, C2_VXY1) \
-,	nop2 \
-,	gte_cmdw_outer_product \
-,	gte_mv_from_data_r(r_a, C2_MAC1) \
-,	gte_mv_from_data_r(r_b, C2_MAC2) \
-,	gte_mv_from_data_r(r_c, C2_MAC3) \
-,	nop \
-,	store_word(r_a, r_f, O_(V3_S4,x)) \
-,	store_word(r_b, r_f, O_(V3_S4,y)) \
-,	store_word(r_c, r_f, O_(V3_S4,z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__cross_uz_ux_to_up, 29)
-
-#define mac_resolve_look_at__normalize_fwd_to_uz(...) \
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,fwd))     /* r_a = &fwd */ \
-,	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,uz))      /* r_b = &uz  */ \
-,	nop	/* Load src.x/y/z from r_a into r_e/r_f/r_i. */ \
-,	load_word(r_e, r_a, O_(V3_S4,x)) \
-,	load_word(r_f, r_a, O_(V3_S4,y)) \
-,	load_word(r_i, r_a, O_(V3_S4,z)) \
-,	nop /* load-delay */	/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */ \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop \
-,	gte_cmdw_sqr	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. */ \
-,	gte_mv_from_data_r(r_d,         C2_MAC1) \
-,	gte_mv_from_data_r(r_g,         C2_MAC2) \
-,	gte_mv_from_data_r(r_recip_est, C2_MAC3) \
-,	nop \
-,	add_u(r_recip_est, r_recip_est, r_g) \
-,	add_u(r_recip_est, r_recip_est, r_d) \
-,	gte_mv_to_data_r(r_recip_est, C2_LZCS) \
-,	nop2 \
-,	gte_mv_from_data_r(r_h, C2_LZCR) \
-,	nop	/* Stage 3: compute shift amount, align |v|² to bit 24. */ \
-,	and_i(   r_h, r_h, -2) \
-,	li_s(    r_shift, 31) \
-,	sub_s(   r_shift, r_shift, r_h) \
-,	shift_aright(r_shift, r_shift, 1) \
-,	add_si(  r_a, r_h, -24)                 /* r_a = LZCR - 24 (overlapping with r_recip_est; src ptr no longer needed) */ \
-,	branch_lt_zero(r_a, atom_offset(srav_path_fwd_to_uz, aligned_done_fwd_to_uz)) \
-,	nop \
-,	jump_rel(atom_offset(aligned_done_fwd_to_uz, srav_path_fwd_to_uz)) \
-,	shift_lleft_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(srav_path_fwd_to_uz) \
-,	li_s(    r_a, 24) \
-,	sub_s(   r_a, r_a, r_h) \
-,	shift_aright_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(aligned_done_fwd_to_uz)	/* r_recip_est holds |v|² aligned to bit 24. */ \
-,	add_si(  r_recip_est, r_recip_est, -64) \
-,	shift_lleft(r_recip_est, r_recip_est, 1) \
-,	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)) \
-,	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)) \
-,	add_u(r_a, r_a, r_recip_est) \
-,	load_half(r_recip_est, r_a, 0) \
-,	nop	/* Stage 4: GPF + srav finalize. */ \
-,	gte_mv_to_data_r(r_recip_est, C2_IR0) \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop2 \
-,	gte_cmdw_gpf \
-,	gte_mv_from_data_r(r_e, C2_MAC1) \
-,	gte_mv_from_data_r(r_f, C2_MAC2) \
-,	gte_mv_from_data_r(r_i, C2_MAC3) \
-,	shift_aright_var(r_e, r_e, r_shift) \
-,	shift_aright_var(r_f, r_f, r_shift) \
-,	shift_aright_var(r_i, r_i, r_shift)	/* Store result.x/y/z to r_b (dst ptr = scratch+16). */ \
-,	store_word(r_e, r_b, O_(V3_S4,x)) \
-,	store_word(r_f, r_b, O_(V3_S4,y)) \
-,	store_word(r_i, r_b, O_(V3_S4,z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__normalize_fwd_to_uz, 59)
-
-#define mac_resolve_look_at__normalize_right_to_ux(...) \
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,right))    /* r_a = &right */ \
-,	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,ux))       /* r_b = &ux    */ \
-,	nop \
-,	load_word(r_e, r_a, O_(V3_S4,x)) \
-,	load_word(r_f, r_a, O_(V3_S4,y)) \
-,	load_word(r_i, r_a, O_(V3_S4,z)) \
-,	nop \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop \
-,	gte_cmdw_sqr \
-,	gte_mv_from_data_r(r_d,         C2_MAC1) \
-,	gte_mv_from_data_r(r_g,         C2_MAC2) \
-,	gte_mv_from_data_r(r_recip_est, C2_MAC3) \
-,	nop \
-,	add_u(r_recip_est, r_recip_est, r_g) \
-,	add_u(r_recip_est, r_recip_est, r_d) \
-,	gte_mv_to_data_r(r_recip_est, C2_LZCS) \
-,	nop2 \
-,	gte_mv_from_data_r(r_h, C2_LZCR) \
-,	nop \
-,	and_i(   r_h, r_h, -2) \
-,	li_s(    r_shift, 31) \
-,	sub_s(   r_shift, r_shift, r_h) \
-,	shift_aright(r_shift, r_shift, 1) \
-,	add_si(  r_a, r_h, -24) \
-,	branch_lt_zero(r_a, atom_offset(srav_path_right_to_ux, aligned_done_right_to_ux)) \
-,	nop \
-,	jump_rel(atom_offset(aligned_done_right_to_ux, srav_path_right_to_ux)) \
-,	shift_lleft_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(srav_path_right_to_ux) \
-,	li_s(    r_a, 24) \
-,	sub_s(   r_a, r_a, r_h) \
-,	shift_aright_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(aligned_done_right_to_ux) \
-,	add_si(  r_recip_est, r_recip_est, -64) \
-,	shift_lleft(r_recip_est, r_recip_est, 1) \
-,	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)) \
-,	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)) \
-,	add_u(r_a, r_a, r_recip_est) \
-,	load_half(r_recip_est, r_a, 0) \
-,	nop \
-,	gte_mv_to_data_r(r_recip_est, C2_IR0) \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop2 \
-,	gte_cmdw_gpf \
-,	gte_mv_from_data_r(r_e, C2_MAC1) \
-,	gte_mv_from_data_r(r_f, C2_MAC2) \
-,	gte_mv_from_data_r(r_i, C2_MAC3) \
-,	shift_aright_var(r_e, r_e, r_shift) \
-,	shift_aright_var(r_f, r_f, r_shift) \
-,	shift_aright_var(r_i, r_i, r_shift) \
-,	store_word(r_e, r_b, O_(V3_S4,x)) \
-,	store_word(r_f, r_b, O_(V3_S4,y)) \
-,	store_word(r_i, r_b, O_(V3_S4,z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__normalize_right_to_ux, 59)
-
-#define mac_resolve_look_at__normalize_up_to_uy(...) \
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,up))       /* r_a = &up */ \
-,	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,uy))       /* r_b = &uy */ \
-,	nop \
-,	load_word(r_e, r_a, O_(V3_S4,x)) \
-,	load_word(r_f, r_a, O_(V3_S4,y)) \
-,	load_word(r_i, r_a, O_(V3_S4,z)) \
-,	nop \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop \
-,	gte_cmdw_sqr \
-,	gte_mv_from_data_r(r_d,         C2_MAC1) \
-,	gte_mv_from_data_r(r_g,         C2_MAC2) \
-,	gte_mv_from_data_r(r_recip_est, C2_MAC3) \
-,	nop \
-,	add_u(r_recip_est, r_recip_est, r_g) \
-,	add_u(r_recip_est, r_recip_est, r_d) \
-,	gte_mv_to_data_r(r_recip_est, C2_LZCS) \
-,	nop2 \
-,	gte_mv_from_data_r(r_h, C2_LZCR) \
-,	nop \
-,	and_i(   r_h, r_h, -2) \
-,	li_s(    r_shift, 31) \
-,	sub_s(   r_shift, r_shift, r_h) \
-,	shift_aright(r_shift, r_shift, 1) \
-,	add_si(  r_a, r_h, -24) \
-,	branch_lt_zero(r_a, atom_offset(srav_path_up_to_uy, aligned_done_up_to_uy)) \
-,	nop \
-,	jump_rel(atom_offset(aligned_done_up_to_uy, srav_path_up_to_uy)) \
-,	shift_lleft_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(srav_path_up_to_uy) \
-,	li_s(    r_a, 24) \
-,	sub_s(   r_a, r_a, r_h) \
-,	shift_aright_var(r_recip_est, r_recip_est, r_a) \
-,	atom_label(aligned_done_up_to_uy) \
-,	add_si(  r_recip_est, r_recip_est, -64) \
-,	shift_lleft(r_recip_est, r_recip_est, 1) \
-,	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)) \
-,	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)) \
-,	add_u(r_a, r_a, r_recip_est) \
-,	load_half(r_recip_est, r_a, 0) \
-,	nop \
-,	gte_mv_to_data_r(r_recip_est, C2_IR0) \
-,	gte_mv_to_data_r(r_e, C2_IR1) \
-,	gte_mv_to_data_r(r_f, C2_IR2) \
-,	gte_mv_to_data_r(r_i, C2_IR3) \
-,	nop2 \
-,	gte_cmdw_gpf \
-,	gte_mv_from_data_r(r_e, C2_MAC1) \
-,	gte_mv_from_data_r(r_f, C2_MAC2) \
-,	gte_mv_from_data_r(r_i, C2_MAC3) \
-,	shift_aright_var(r_e, r_e, r_shift) \
-,	shift_aright_var(r_f, r_f, r_shift) \
-,	shift_aright_var(r_i, r_i, r_shift) \
-,	store_word(r_e, r_b, O_(V3_S4,x)) \
-,	store_word(r_f, r_b, O_(V3_S4,y)) \
-,	store_word(r_i, r_b, O_(V3_S4,z)) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__normalize_up_to_uy, 59)
-
-#define mac_resolve_look_at__populate_and_translate(...) \
-	load_word(r_look_at, R_TapePtr, O_(Binds_ResolveLookAtPopAndTrans,look_at)) \
-,	add_ui_self(         R_TapePtr, S_(Binds_ResolveLookAtPopAndTrans))	/* Compute the 4 scratch pointers in their dedicated GPRs. */ \
-,	add_si(r_pux,  r_scratch, O_(ResolveLookAtScratch,ux))       /* r_pux  = &ux  */ \
-,	add_si(r_puy,  r_scratch, O_(ResolveLookAtScratch,uy))       /* r_puy  = &uy  */ \
-,	add_si(r_puz,  r_scratch, O_(ResolveLookAtScratch,uz))       /* r_puz  = &uz  */ \
-,	add_si(r_peye, r_scratch, O_(ResolveLookAtScratch,eye))      /* r_peye = &eye */ \
-,	nop	/* ── m[0] = (S2)ux ── */ \
-,	load_word(r_tmp0, r_pux, O_(V3_S4,x)) \
-,	load_word(r_tmp1, r_pux, O_(V3_S4,y)) \
-,	load_word(r_tmp2, r_pux, O_(V3_S4,z)) \
-,	nop \
-,	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[0][0])) \
-,	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[0][1])) \
-,	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[0][2]))	/* ── m[1] = (S2)uy ── */ \
-,	load_word(r_tmp0, r_puy, O_(V3_S4,x)) \
-,	load_word(r_tmp1, r_puy, O_(V3_S4,y)) \
-,	load_word(r_tmp2, r_puy, O_(V3_S4,z)) \
-,	nop \
-,	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[1][0])) \
-,	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[1][1])) \
-,	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[1][2]))	/* ── m[2] = (S2)uz ── */ \
-,	load_word(r_tmp0, r_puz, O_(V3_S4,x)) \
-,	load_word(r_tmp1, r_puz, O_(V3_S4,y)) \
-,	load_word(r_tmp2, r_puz, O_(V3_S4,z)) \
-,	nop \
-,	store_half(r_tmp0, r_look_at, O_(MT3_S2S4,m[2][0])) \
-,	store_half(r_tmp1, r_look_at, O_(MT3_S2S4,m[2][1])) \
-,	store_half(r_tmp2, r_look_at, O_(MT3_S2S4,m[2][2]))	/* ── Translation column t[i] = R * (-eye) ─────────────────────────────
-	 * pos = -eye: load eye.x/y/z from r_peye, negate via sub_u from R_0. */ \
-,	load_word(r_tmp0, r_peye, O_(P3_S4,x)) \
-,	load_word(r_tmp1, r_peye, O_(P3_S4,y)) \
-,	load_word(r_tmp2, r_peye, O_(P3_S4,z)) \
-,	nop \
-,	sub_u(r_tmp0, R_0, r_tmp0)  /* pos.x = -eye.x */ \
-,	sub_u(r_tmp1, R_0, r_tmp1) \
-,	sub_u(r_tmp2, R_0, r_tmp2)	/* mtc2 IR1/2/3 = pos (for MVMVA — input vector registers). */ \
-,	gte_mv_to_data_r(r_tmp0, C2_IR1) \
-,	gte_mv_to_data_r(r_tmp1, C2_IR2) \
-,	gte_mv_to_data_r(r_tmp2, C2_IR3) \
-,	nop2	/* MVMVA: MAC1/2/3 = R * IR with cv=0 (no TR vector), mx=0 (rotation matrix),
-	 * sf=0 (no shift), v=0 (V0 = IR1/2/3, no far-plane clipping). The pre-set
-	 * rotation matrix is the one set by the preceding set_gte_world atom.
-	 * gte_cmdw_mvmva is parameterless and defaults to cv=0/mx=0/sf=0/v=0. */ \
-,	gte_cmdw_mvmva \
-,	nop  /* GTE interlock */	/* mfc2 MAC1/2/3 → r_tmp0/r_tmp1/r_tmp2 (sign-extended into 32-bit GPRs).
-	 * MAC1/2/3 hold R*v with no TR add and no perspective divide — exactly the
-	 * 3 distinct world-space translation values we need for t[0..2]. */ \
-,	gte_mv_from_data_r(r_tmp0, C2_MAC1) \
-,	gte_mv_from_data_r(r_tmp1, C2_MAC2) \
-,	gte_mv_from_data_r(r_tmp2, C2_MAC3) \
-,	nop \
-,	store_word(r_tmp0, r_look_at, O_(MT3_S2S4,t[0])) \
-,	store_word(r_tmp1, r_look_at, O_(MT3_S2S4,t[1])) \
-,	store_word(r_tmp2, r_look_at, O_(MT3_S2S4,t[2])) \
-,	mac_yield()
-WORD_COUNT(mac_resolve_look_at__populate_and_translate, 50)
-
@@ -8,36 +8,6 @@
 #pragma region hello_camera


-// --- atom: resolve_look_at__normalize_fwd_to_uz (62 words) ---
-
-#define _atom_offset_srav_path_fwd_to_uz_aligned_done_fwd_to_uz  6
-#define _atom_offset_aligned_done_fwd_to_uz_srav_path_fwd_to_uz  1
-
-enum {
-    atom_offset_srav_path_fwd_to_uz_aligned_done_fwd_to_uz = _atom_offset_srav_path_fwd_to_uz_aligned_done_fwd_to_uz,
-    atom_offset_aligned_done_fwd_to_uz_srav_path_fwd_to_uz = _atom_offset_aligned_done_fwd_to_uz_srav_path_fwd_to_uz,
-};
-
-// --- atom: resolve_look_at__normalize_right_to_ux (62 words) ---
-
-#define _atom_offset_srav_path_right_to_ux_aligned_done_right_to_ux  6
-#define _atom_offset_aligned_done_right_to_ux_srav_path_right_to_ux  1
-
-enum {
-    atom_offset_srav_path_right_to_ux_aligned_done_right_to_ux = _atom_offset_srav_path_right_to_ux_aligned_done_right_to_ux,
-    atom_offset_aligned_done_right_to_ux_srav_path_right_to_ux = _atom_offset_aligned_done_right_to_ux_srav_path_right_to_ux,
-};
-
-// --- atom: resolve_look_at__normalize_up_to_uy (62 words) ---
-
-#define _atom_offset_srav_path_up_to_uy_aligned_done_up_to_uy  6
-#define _atom_offset_aligned_done_up_to_uy_srav_path_up_to_uy  1
-
-enum {
-    atom_offset_srav_path_up_to_uy_aligned_done_up_to_uy = _atom_offset_srav_path_up_to_uy_aligned_done_up_to_uy,
-    atom_offset_aligned_done_up_to_uy_srav_path_up_to_uy = _atom_offset_aligned_done_up_to_uy_srav_path_up_to_uy,
-};
-
 // --- atom: pad_input_cube_rotation (60 words) ---

 #define _atom_offset_dpad_left_exit_dpad_left         6
@@ -17,6 +17,7 @@
 #	include "duffle/psyq.atom.c"
 #		include "gen/offsets.h"
 #		include "gen/macs.h"
+#		include "gen/auto_reg.h"
 #	include "hello_camera.h"
 #endif

@@ -93,16 +94,35 @@ MipsAtomComp_Proc_(ac_put_draw_env, ab, {
 #pragma region Atom Procs
 // Modular Atoms

+/* Scratchpad layout for the resolve_look_at bundle.
+ * The chain atoms communicate entirely via the wave-context GPR carrier R_ResolveScratch (R_T4) + hardcoded offsets into smem.scratchpad
+ * (PS1 hardware scratchpad at 0x1F800000).
+ *
+ * Atom 0 (input_and_sub) STAGES the C-side inputs (eye, up_in) into the scratchpad;
+ * AT THE SAME TIME it computes fwd = target - eye and stores it at scratch+0.
+ * Atoms 1-6 then read/write specific scratchpad offsets internally using
+ * `r_scratch + hardcoded_offset` — no tape-data pointers are passed between atoms.
+ *   +0   fwd   (atom 0 writes; atom  1 reads)
+ *   +16  uz    (atom 1 writes; atoms 2 + 4 read)
+ *   +32  right (atom 2 writes; atom  3 reads)
+ *   +48  ux    (atom 3 writes; atoms 4 + 6 read)
+ *   +64  up    (atom 4 writes; atom  5 reads)
+ *   +80  uy    (atom 5 writes; atom  6 reads)
+ *   +96  eye   (atom 0 stages from C-side pointer; atom 6 reads)
+ *   +128 up_in (atom 0 stages from C-side pointer; atom 2 reads)
+ */
+
+// enum {
+// 	R_LookAt    = R_T0 atom_reg atom_type(MT3_S2S4*),
+// 	R_CamEye    = R_T1 atom_reg atom_type(P3_S4*),
+// 	R_CamTarget = R_T2 atom_reg atom_type(P3_S4*),
+// 	R_WorldUp   = R_T3 atom_reg atom_type(V3_S4*),
+// };
+
 enum {
 	/* Wave-context GPR carrier for the resolve_look_at bundle: the scratch base.
-	* Set by atom 0 (popped from tape), read by atoms 1-6 (used as pointer base).
-	* Type is U4* — this holds the scratch base address (smem.scratchpad value).
-	*
-	* Other wave-context carriers (R_ResolveUzPtr / UxPtr / UyPtr) used in the
-	* prior design were dropped: the new chain atoms compute their src/dst
-	* addresses internally from R_ResolveScratch + hardcoded_offset. */
+	* Set by atom 0 (popped from tape), read by atoms 1-6 (used as pointer base). */
 	R_ResolveScratch = R_T4 atom_reg atom_type(U4*),
-#define R_ResolveScratch_Code  R_T4_Code
 };
 typedef Struct_(Binds_ResolveLookAt) {
 	MT3_S2S4* look_at;
@@ -111,11 +131,7 @@ typedef Struct_(Binds_ResolveLookAt) {
 	V3_S4*    up_in;
 };

-/* Per-atom bind-pop structs for the resolve_look_at bundle.
- * Atom 0 (input_and_sub) is the ONLY atom that touches the C-side pointers +
- * scratch base. Atoms 1-6 use scratch + hardcoded offsets internally.
- * Field types are U4 (raw pointer value) because the structs are populated
- * by the frame-time bundle helper with the literal C-side pointer values. */
+/* Per-atom bind-pop structs for the resolve_look_at bundle. */
 typedef Struct_(Binds_ResolveLookAtScratch) {
 	U4 scratch_base;  /* U4 (scratch base address — populated by helper with u4_(smem.scratchpad)) */
 };
@@ -125,30 +141,25 @@ typedef Struct_(Binds_ResolveLookAtScratch) {
 *
 * Each slot is 16 bytes: V3_S4 is already 16 bytes (4 × S4 = x/y/z/pad).
 * The struct fields are contiguous — slot i starts at offset i*16.
- * Used by the assembly via O_(ResolveLookAtScratch, fld.x/y/z) which resolves
- * to a compile-time byte offset. NOT a runtime struct — the struct is purely
- * a schema for offsets; the assembly uses `r_scratch + O_(...)` to compute
- * slot addresses at runtime.
+ * Used by the assembly via O_(ResolveLookAtScratch, fld.x/y/z) which resolves to a compile-time byte offset.
+ * NOT a runtime struct — the struct is purely a schema for offsets; the assembly uses `r_scratch + O_(...)` to compute slot addresses at runtime.
 *
 * Slot producers/consumers (referenced by the resolve_look_at chain atoms):
- *
- *   +0    fwd       atom 0 writes (target - eye); atom 1 (normalize) reads
- *   +16   uz        atom 1 writes (normalize fwd); atoms 2 + 4 read (cross operands)
- *   +32   right     atom 2 writes (cross uz x up_in); atom 3 (normalize) reads
- *   +48   ux        atom 3 writes (normalize right); atoms 4 + 6 read
- *   +64   up        atom 4 writes (cross uz x ux); atom 5 (normalize) reads
- *   +80   uy        atom 5 writes (normalize up); atom 6 reads
- *   +96   eye       atom 0 stages (C-side input); atom 6 reads (translation column)
+ *   +0    fwd    0 writes (target - eye);     atom  1 (normalize) reads
+ *   +16   uz     1 writes (normalize fwd);    atoms 2 + 4 read (cross operands)
+ *   +32   right  2 writes (cross uz x up_in); atom  3 (normalize) reads
+ *   +48   ux     3 writes (normalize right);  atoms 4 + 6 read
+ *   +64   up     4 writes (cross uz x ux);    atom  5 (normalize) reads
+ *   +80   uy     5 writes (normalize up);     atom  6 reads
+ *   +96   eye    0 stages (C-side input);     atom  6 reads (translation column)
 *   +112  target reserved (currently written nowhere — kept for symmetry w/ eye)
- *   +128  up_in     atom 0 stages (C-side input); atom 2 reads (cross operand)
+ *   +128  up_in  0 stages (C-side input); atom 2 reads (cross operand)
 *
 * Fields use P3_S4 (point) for eye/target (RGA: affine point, implicit weight 1);
- * V3_S4 (vector) for fwd/uz/right/ux/up/uy/up_in (RGA: Euclidean vector). P3_S4
- * is a storage alias of V3_S4 (see math.h comment: "Storage alias of V3_S4.
+ * V3_S4 (vector) for fwd/uz/right/ux/up/uy/up_in (RGA: Euclidean vector).
+ * P3_S4 is a storage alias of V3_S4 (see math.h comment: "Storage alias of V3_S4.
 * Use P3_S4 when the value is a point.") — both are 16 bytes.
- *
- * Moved from gte.atom.c (Task 12.11): gte.atom.c is the GENERIC GTE primitives
- * file and must not know about any specific atom bundle's scratch layout. */
+ */
 typedef Struct_(ResolveLookAtScratch) {
 	V3_S4 fwd;       /* offset  +0  (16 bytes — 4 S4 fields incl. internal pad) */
 	V3_S4 uz;        /* offset +16 (16 bytes) */
@@ -161,38 +172,23 @@ typedef Struct_(ResolveLookAtScratch) {
 	V3_S4 up_in;     /* offset +128 (16 bytes) */
 };

-/* ─── resolve_look_at bundle chain atoms (Task 5) ────────────────────────────
- * 7 unique atom procs in the resolve_look_at bundle (4 chain atoms + 3 normalize
- * variants). All 7 are runtime-built MipsAtom_Proc_ atoms: each function declares
- * a static MipsCode[] body, then calls atombuilder_unroll() to append it to the
- * caller's MipsAtomBuilder arena. Task 6's resolve_look_at_init() uses this pattern
- * to pre-build the bundle into the static arena (smem.resolve_look_at_arena).
+/* ─── resolve_look_at bundle chain atoms ────────────────────────────
+ * 4 unique atom procs in the resolve_look_at bundle (4 chain atoms + 3 calls to generic normalize_v3s4_proc).
+ * All 4 chain atoms are runtime-built MipsAtom_Proc_ atoms: each function declares a static MipsCode[] body, 
+ * then calls atombuilder_unroll() to append it to the caller's MipsAtomBuilder arena. resolve_look_at_init() 
+ * uses this pattern to pre-build the bundle into the static arena (smem.resolve_look_at_arena).
 *
- * Atom roster (positions 0-6 in the bundle):
- *   Atom 0: resolve_look_at__input_and_sub            (chain atom)
- *   Atom 1: resolve_look_at__normalize_fwd_to_uz      (normalize wrapper)
- *   Atom 2: resolve_look_at__cross_uz_up_in_to_right  (chain atom)
- *   Atom 3: resolve_look_at__normalize_right_to_ux    (normalize wrapper)
- *   Atom 4: resolve_look_at__cross_uz_ux_to_up        (chain atom)
- *   Atom 5: resolve_look_at__normalize_up_to_uy       (normalize wrapper)
- *   Atom 6: resolve_look_at__populate_and_translate   (chain atom)
+ * Atom roster:
+ *	0: resolve_look_at__input_and_sub              (chain atom)
+ *	1: normalize_v3s4_proc (gte.atom.c)            (generic normalize; called for fwd→uz)
+ *	2: resolve_look_at__cross_uz_up_in_to_right    (chain atom)
+ *	3: normalize_v3s4_proc (gte.atom.c)            (generic normalize; called for right→ux)
+ *	4: resolve_look_at__cross_uz_ux_to_up          (chain atom)
+ *	5: normalize_v3s4_proc (gte.atom.c)            (generic normalize; called for up→uy)
+ *	6: resolve_look_at__populate_and_translate     (chain atom)
 *
- * The 3 normalize wrappers are CHAIN-SPECIFIC — they hardcode src/dst scratch
- * offsets in the body (computed via r_scratch + O_(ResolveLookAtScratch, fld)).
- * The generic normalize_v3s4_proc (in gte.atom.c) takes src/dst as GPR parameters
- * and is NOT used by this bundle. (Layering rule: gte.atom.c contains only
- * generic GTE primitives; bundle-specific code lives in this file.)
- *
- * The 3 normalize procs were moved from gte.atom.c to this file in Task 12.11
- * (user feedback: "normalize is not supposed to be aware of a specific scratch
- * for one atom bundle"). The procs were renamed to resolve_look_at__normalize_*_proc
- * to make their bundle-specific nature clear.
- *
- * Lua metaprogram support (Task 12.10): the metaprogram auto-emits
- * `atom_offset__X__Y` defs in gen/offsets.h for each atom_label/atom_offset pair
- * in the body. The 3 normalize procs each have internal branches (srav_path /
- * aligned_done variants) and get their per-proc-instance defs (e.g.,
- * `atom_offset_srav_path_fwd_to_uz_aligned_done_fwd_to_uz`).
+ * The generic normalize_v3s4_proc is a parameterized 4-stage GTE normalize (SQR → mfc2 → LZCS → GPF → srav);
+ * it accepts scratch base + offset args so any caller (with a scratch base + struct schema) can use it.
 */

 typedef Struct_(Binds_ResolveLookAtSub) {
@@ -201,21 +197,17 @@ typedef Struct_(Binds_ResolveLookAtSub) {
 	U4 up_in;   /* U4 (C-side V3_S4* — read by atom 0 directly; staged into scratchpad by atom 0) */
 };

-/* Atom 0 in the bundle: input_and_sub. Stages C-side inputs into the scratchpad
- * and computes fwd = target - eye.
- *
- * Inputs (C-side pointers popped from the tape; NOT scratchpad addresses):
+/* Atom 0 in the bundle: input_and_sub. Stages C-side inputs into the scratchpad and computes fwd = target - eye.
+ * Inputs (C-side pointers popped from the tape):
 *   r_target_ptr : P3_S4* (C-side struct; atom 0 reads target.x/y/z directly)
 *   r_eye_ptr    : P3_S4* (C-side struct; staged into scratchpad at +96/+100/+104)
 *   r_up_in_ptr  : V3_S4* (C-side struct; staged into scratchpad at +128/+132/+136)
- *
 * Wave-context output:
 *   r_scratch    : R_ResolveScratch (R_T4) — scratch base, read by atoms 1-6
 * 
 * Bind-pop layout:
 *   Binds_ResolveLookAtSub     = 12 bytes (target + eye + up_in ptrs)
 *   Binds_ResolveLookAtScratch =  4 bytes (scratch_base)
- *
 * Staging work:
 *   * Stage eye.x/y/z   → scratch+96/+100/+104 (for atom 6's translation column)
 *   * Stage up_in.x/y/z → scratch+128/+132/+136 (for atom 2's outer-product operand)
@@ -235,11 +227,8 @@ typedef Struct_(Binds_ResolveLookAtSub) {
 *
 * Pool cost: 8 GPRs + R_T4 (carrier) + R_AT + R_V0 (hardcoded) = 11 GPRs.
 */
-I_ void resolve_look_at__input_and_sub_proc(MipsAtomBuilder_R ab
-	,	U4 r_target_ptr
-	,	U4 r_eye_ptr
-	,	U4 r_up_in_ptr
-	,	U4 r_scratch
+I_ void resolve_look_at__input_and_sub_proc(MipsAtomBuilder_R ab,	U4 r_scratch
+	,	U4 r_target_ptr,U4 r_eye_ptr, U4 r_up_in_ptr
 	,	U4 r_tmp0,      U4 r_tmp1,    U4 r_tmp2, U4 r_tmp3
 ) MipsAtom_Proc_(resolve_look_at__input_and_sub, ab, {
 	/* Pop the 3 C-side pointers + scratch_base from the tape. */
@@ -291,9 +280,8 @@ I_ void resolve_look_at__input_and_sub_proc(MipsAtomBuilder_R ab
 })

 /* Atoms 2 + 4 in the bundle: out = a × b (GTE outer product on IR/D vectors).
- * No bind pop — the three operand pointers (a, b, out) are derived in-body
- * from r_scratch + hardcoded_offset. Each atom has its own variant because
- * the offsets are baked into the body and each atom uses unique GPRs.
+ * No bind pop — the three operand pointers (a, b, out) are derived in-body from r_scratch + hardcoded_offset.
+ * Each atom has its own variant because the offsets are baked into the body and each atom uses unique GPRs.
 *
 * GTE register layout (per PSX-SPX + duffle gte.h):
 *   IR1/2/3  = a.x/y/z    (mtc2)
@@ -307,8 +295,7 @@ I_ void resolve_look_at__input_and_sub_proc(MipsAtomBuilder_R ab
 */

 /* Atom 2: cross uz × up_in → right. */
-I_ void resolve_look_at__cross_uz_up_in_to_right_proc(MipsAtomBuilder_R ab
-	,	U4 r_scratch
+I_ void resolve_look_at__cross_uz_up_in_to_right_proc(MipsAtomBuilder_R ab, U4 r_scratch
 	,	U4 r_a, U4 r_b, U4 r_c /* load a.x/y/z; result out.x/y/z */
 	,	U4 r_d                 /* load b.x */
 	,	U4 r_f, U4 r_g, U4 r_h /* r_f = &right (out ptr), r_g = &uz, r_h = &up_in */
@@ -325,9 +312,9 @@ I_ void resolve_look_at__cross_uz_up_in_to_right_proc(MipsAtomBuilder_R ab
 	load_word(r_c, r_g, O_(V3_S4,z)),
 	nop,

-	/* Load b (up_in).x/y/z into r_d + R_AT/R_V0 (hardcoded; reusing the
-	 * body's last two loads is fine because the load-delay slot is the nop
-	 * after the third load, and mtc2 below doesn't read these regs). */
+	/* Load b (up_in).x/y/z into r_d + R_AT/R_V0 
+	(hardcoded; reusing the body's last two loads is fine because the load-delay slot is the nop after the third load, 
+	and mtc2 below doesn't read these regs). */
 	load_word(r_d,  r_h, O_(V3_S4,x)),
 	load_word(R_AT, r_h, O_(V3_S4,y)),
 	load_word(R_V0, r_h, O_(V3_S4,z)),
@@ -359,8 +346,7 @@ I_ void resolve_look_at__cross_uz_up_in_to_right_proc(MipsAtomBuilder_R ab
 })

 /* Atom 4: cross uz × ux → up. */
-I_ void resolve_look_at__cross_uz_ux_to_up_proc(MipsAtomBuilder_R ab
-	,	U4 r_scratch
+I_ void resolve_look_at__cross_uz_ux_to_up_proc(MipsAtomBuilder_R ab,	U4 r_scratch
 	,	U4 r_a, U4 r_b, U4 r_c                  /* load a.x/y/z; result out.x/y/z */
 	,	U4 r_d                                  /* load b.x */
 	,	U4 r_f, U4 r_g, U4 r_h                  /* r_f = &up (out ptr), r_g = &uz, r_h = &ux */
@@ -404,259 +390,10 @@ I_ void resolve_look_at__cross_uz_ux_to_up_proc(MipsAtomBuilder_R ab
 	mac_yield()
 })

-/* Atoms 1, 3, 5 in the bundle: chain-specific normalize wrappers around the
- * generic normalize_v3s4_proc (gte.atom.c). The generic proc takes src/dst as
- * GPR parameters; these wrappers HARDCODE src/dst via r_scratch + O_(ResolveLookAtScratch, fld)
- * so the C-side bundle helper doesn't need to push scratchpad addresses via
- * tb_data between atoms. (Task 12.8 fix: eliminate magic offsets.)
- *
- * The 4-stage normalize body (SQR → mfc2 → LZCS → GPF → srav) is identical to
- * the generic version (GPR-renamed); cycle counts match. The only per-atom
- * difference is the (src, dst) scratch offsets and the per-proc atom_label
- * suffixes (srav_path_fwd_to_uz, srav_path_right_to_ux, srav_path_up_to_uy) so
- * the per-proc-instance offsets are emitted disjointly in gen/offsets.h.
- *
- * GPR pool (10 free regs: R_T0..R_T3, R_T5..R_T7, R_V0, R_V1, R_AT; R_T4 reserved for R_ResolveScratch):
- *   r_a       : src ptr (overlaps with r_recip_est carrier after the 3 src-loads)
- *   r_b       : dst ptr (saved throughout)
- *   r_e/r_f/r_i : src.x/y/z → result.x/y/z (preserved across stages 1-2 via r_d/r_g/r_recip_est scratch)
- *   r_d/r_g   : MAC1/2 scratch (dead after stage 2)
- *   r_h       : LZCR (saved across stages 3-4)
- *   r_recip_est : |v|² accumulator + sqrtbl[index] + 1/|v| (saved throughout)
- *   r_shift   : final srav amount (saved across stages 3-4)
- *
- * The Lua metaprogram (Task 12.10) auto-emits:
- *   - `mac_resolve_look_at__normalize_<from>_to_<to>` alias in gen/macs.h
- *   - `atom_offset__srav_path_<from>_to_<to>__aligned_done_<from>_to_<to>` defs in gen/offsets.h
- */
-
-/* Atom 1: normalize fwd (scratch+0) → uz (scratch+16). */
-I_ void resolve_look_at__normalize_fwd_to_uz_proc(MipsAtomBuilder_R ab
-	,	U4 r_scratch
-	,	U4 r_a, U4 r_b                    /* src/dst scratch pointers */
-	,	U4 r_e, U4 r_f, U4 r_i             /* src.x/y/z → result.x/y/z */
-	,	U4 r_d, U4 r_g                     /* MAC1/2 scratch (dead after stage 2) */
-	,	U4 r_h                             /* LZCR */
-	,	U4 r_recip_est
-	,	U4 r_shift
-) MipsAtom_Proc_(resolve_look_at__normalize_fwd_to_uz, ab, {
-	/* Compute src/dst pointers from r_scratch. */
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,fwd)),     /* r_a = &fwd */
-	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,uz)),      /* r_b = &uz  */
-	nop,
-
-	/* Load src.x/y/z from r_a into r_e/r_f/r_i. */
-	load_word(r_e, r_a, O_(V3_S4,x)),
-	load_word(r_f, r_a, O_(V3_S4,y)),
-	load_word(r_i, r_a, O_(V3_S4,z)),
-	nop, /* load-delay */
-
-	/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop, gte_cmdw_sqr,
-
-	/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. */
-	gte_mv_from_data_r(r_d,         C2_MAC1),
-	gte_mv_from_data_r(r_g,         C2_MAC2),
-	gte_mv_from_data_r(r_recip_est, C2_MAC3),
-	nop,
-	add_u(r_recip_est, r_recip_est, r_g),
-	add_u(r_recip_est, r_recip_est, r_d),
-	gte_mv_to_data_r(r_recip_est, C2_LZCS),
-	nop2,
-	gte_mv_from_data_r(r_h, C2_LZCR),
-	nop,
-
-	/* Stage 3: compute shift amount, align |v|² to bit 24. */
-	and_i(   r_h, r_h, -2),
-	li_s(    r_shift, 31),
-	sub_s(   r_shift, r_shift, r_h),
-	shift_aright(r_shift, r_shift, 1),
-	add_si(  r_a, r_h, -24),                 /* r_a = LZCR - 24 (overlapping with r_recip_est; src ptr no longer needed) */
-	branch_lt_zero(r_a, atom_offset(srav_path_fwd_to_uz, aligned_done_fwd_to_uz)), nop,
-	jump_rel(atom_offset(aligned_done_fwd_to_uz, srav_path_fwd_to_uz)),
-	shift_lleft_var(r_recip_est, r_recip_est, r_a),
-	atom_label(srav_path_fwd_to_uz)
-	li_s(    r_a, 24),
-	sub_s(   r_a, r_a, r_h),
-	shift_aright_var(r_recip_est, r_recip_est, r_a),
-	atom_label(aligned_done_fwd_to_uz)
-	/* r_recip_est holds |v|² aligned to bit 24. */
-	add_si(  r_recip_est, r_recip_est, -64),
-	shift_lleft(r_recip_est, r_recip_est, 1),
-	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)),
-	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)),
-	add_u(r_a, r_a, r_recip_est),
-	load_half(r_recip_est, r_a, 0),
-	nop,
-
-	/* Stage 4: GPF + srav finalize. */
-	gte_mv_to_data_r(r_recip_est, C2_IR0),
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop2,
-	gte_cmdw_gpf,
-	gte_mv_from_data_r(r_e, C2_MAC1),
-	gte_mv_from_data_r(r_f, C2_MAC2),
-	gte_mv_from_data_r(r_i, C2_MAC3),
-	shift_aright_var(r_e, r_e, r_shift),
-	shift_aright_var(r_f, r_f, r_shift),
-	shift_aright_var(r_i, r_i, r_shift),
-
-	/* Store result.x/y/z to r_b (dst ptr = scratch+16). */
-	store_word(r_e, r_b, O_(V3_S4,x)),
-	store_word(r_f, r_b, O_(V3_S4,y)),
-	store_word(r_i, r_b, O_(V3_S4,z)),
-
-	mac_yield()
-})
-
-/* Atom 3: normalize right (scratch+32) → ux (scratch+48). */
-I_ void resolve_look_at__normalize_right_to_ux_proc(MipsAtomBuilder_R ab
-	,	U4 r_scratch
-	,	U4 r_a, U4 r_b
-	,	U4 r_e, U4 r_f, U4 r_i
-	,	U4 r_d, U4 r_g
-	,	U4 r_h
-	,	U4 r_recip_est
-	,	U4 r_shift
-) MipsAtom_Proc_(resolve_look_at__normalize_right_to_ux, ab, {
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,right)),    /* r_a = &right */
-	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,ux)),       /* r_b = &ux    */
-	nop,
-	load_word(r_e, r_a, O_(V3_S4,x)),
-	load_word(r_f, r_a, O_(V3_S4,y)),
-	load_word(r_i, r_a, O_(V3_S4,z)),
-	nop,
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop, gte_cmdw_sqr,
-	gte_mv_from_data_r(r_d,         C2_MAC1),
-	gte_mv_from_data_r(r_g,         C2_MAC2),
-	gte_mv_from_data_r(r_recip_est, C2_MAC3),
-	nop,
-	add_u(r_recip_est, r_recip_est, r_g),
-	add_u(r_recip_est, r_recip_est, r_d),
-	gte_mv_to_data_r(r_recip_est, C2_LZCS),
-	nop2,
-	gte_mv_from_data_r(r_h, C2_LZCR),
-	nop,
-	and_i(   r_h, r_h, -2),
-	li_s(    r_shift, 31),
-	sub_s(   r_shift, r_shift, r_h),
-	shift_aright(r_shift, r_shift, 1),
-	add_si(  r_a, r_h, -24),
-	branch_lt_zero(r_a, atom_offset(srav_path_right_to_ux, aligned_done_right_to_ux)), nop,
-	jump_rel(atom_offset(aligned_done_right_to_ux, srav_path_right_to_ux)),
-	shift_lleft_var(r_recip_est, r_recip_est, r_a),
-	atom_label(srav_path_right_to_ux)
-	li_s(    r_a, 24),
-	sub_s(   r_a, r_a, r_h),
-	shift_aright_var(r_recip_est, r_recip_est, r_a),
-	atom_label(aligned_done_right_to_ux)
-	add_si(  r_recip_est, r_recip_est, -64),
-	shift_lleft(r_recip_est, r_recip_est, 1),
-	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)),
-	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)),
-	add_u(r_a, r_a, r_recip_est),
-	load_half(r_recip_est, r_a, 0),
-	nop,
-	gte_mv_to_data_r(r_recip_est, C2_IR0),
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop2,
-	gte_cmdw_gpf,
-	gte_mv_from_data_r(r_e, C2_MAC1),
-	gte_mv_from_data_r(r_f, C2_MAC2),
-	gte_mv_from_data_r(r_i, C2_MAC3),
-	shift_aright_var(r_e, r_e, r_shift),
-	shift_aright_var(r_f, r_f, r_shift),
-	shift_aright_var(r_i, r_i, r_shift),
-	store_word(r_e, r_b, O_(V3_S4,x)),
-	store_word(r_f, r_b, O_(V3_S4,y)),
-	store_word(r_i, r_b, O_(V3_S4,z)),
-	mac_yield()
-})
-
-/* Atom 5: normalize up (scratch+64) → uy (scratch+80). */
-I_ void resolve_look_at__normalize_up_to_uy_proc(MipsAtomBuilder_R ab
-	,	U4 r_scratch
-	,	U4 r_a, U4 r_b
-	,	U4 r_e, U4 r_f, U4 r_i
-	,	U4 r_d, U4 r_g
-	,	U4 r_h
-	,	U4 r_recip_est
-	,	U4 r_shift
-) MipsAtom_Proc_(resolve_look_at__normalize_up_to_uy, ab, {
-	add_si(r_a, r_scratch, O_(ResolveLookAtScratch,up)),       /* r_a = &up */
-	add_si(r_b, r_scratch, O_(ResolveLookAtScratch,uy)),       /* r_b = &uy */
-	nop,
-	load_word(r_e, r_a, O_(V3_S4,x)),
-	load_word(r_f, r_a, O_(V3_S4,y)),
-	load_word(r_i, r_a, O_(V3_S4,z)),
-	nop,
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop, gte_cmdw_sqr,
-	gte_mv_from_data_r(r_d,         C2_MAC1),
-	gte_mv_from_data_r(r_g,         C2_MAC2),
-	gte_mv_from_data_r(r_recip_est, C2_MAC3),
-	nop,
-	add_u(r_recip_est, r_recip_est, r_g),
-	add_u(r_recip_est, r_recip_est, r_d),
-	gte_mv_to_data_r(r_recip_est, C2_LZCS),
-	nop2,
-	gte_mv_from_data_r(r_h, C2_LZCR),
-	nop,
-	and_i(   r_h, r_h, -2),
-	li_s(    r_shift, 31),
-	sub_s(   r_shift, r_shift, r_h),
-	shift_aright(r_shift, r_shift, 1),
-	add_si(  r_a, r_h, -24),
-	branch_lt_zero(r_a, atom_offset(srav_path_up_to_uy, aligned_done_up_to_uy)), nop,
-	jump_rel(atom_offset(aligned_done_up_to_uy, srav_path_up_to_uy)),
-	shift_lleft_var(r_recip_est, r_recip_est, r_a),
-	atom_label(srav_path_up_to_uy)
-	li_s(    r_a, 24),
-	sub_s(   r_a, r_a, r_h),
-	shift_aright_var(r_recip_est, r_recip_est, r_a),
-	atom_label(aligned_done_up_to_uy)
-	add_si(  r_recip_est, r_recip_est, -64),
-	shift_lleft(r_recip_est, r_recip_est, 1),
-	load_upper_i(r_a, u4_hi(& gte_normalize_sqr_tbl)),
-	or_i_self(r_a, u4_lo(& gte_normalize_sqr_tbl)),
-	add_u(r_a, r_a, r_recip_est),
-	load_half(r_recip_est, r_a, 0),
-	nop,
-	gte_mv_to_data_r(r_recip_est, C2_IR0),
-	gte_mv_to_data_r(r_e, C2_IR1),
-	gte_mv_to_data_r(r_f, C2_IR2),
-	gte_mv_to_data_r(r_i, C2_IR3),
-	nop2,
-	gte_cmdw_gpf,
-	gte_mv_from_data_r(r_e, C2_MAC1),
-	gte_mv_from_data_r(r_f, C2_MAC2),
-	gte_mv_from_data_r(r_i, C2_MAC3),
-	shift_aright_var(r_e, r_e, r_shift),
-	shift_aright_var(r_f, r_f, r_shift),
-	shift_aright_var(r_i, r_i, r_shift),
-	store_word(r_e, r_b, O_(V3_S4,x)),
-	store_word(r_f, r_b, O_(V3_S4,y)),
-	store_word(r_i, r_b, O_(V3_S4,z)),
-	mac_yield()
-})
-
 typedef Struct_(Binds_ResolveLookAtPopAndTrans) {
 	U4 look_at;  /* U4 (MT3_S2S4* — destination matrix address) */
 };
-/* Atom 6 in the bundle: write look_at->m[][] from ux/uy/uz, then compute
- * the translation column t[] = R * (-eye).
+/* Atom 6 in the bundle: write look_at->m[][] from ux/uy/uz, then compute the translation column t[] = R * (-eye).
 *
 * GPR codes (assigned by resolve_look_at_init):
 *   r_look_at   : MT3_S2S4* (popped from tape; output matrix destination)
@@ -666,14 +403,15 @@ typedef Struct_(Binds_ResolveLookAtPopAndTrans) {
 *   r_peye      : pointer to eye  (offset O_(ResolveLookAtScratch,eye))
 *   r_tmp0/1/2  : atom-local scratch (load + MVMVA + store temps)
 *
- * The 4 pointer regs (r_pux/r_puy/r_puz/r_peye) are DEDICATED — they hold the scratch addresses for the entire body.
+ * 4 pointer regs (r_pux/r_puy/r_puz/r_peye) are DEDICATED — they hold the scratch addresses for the entire body.
 * They are computed in-body via `add_si(r_px, r_scratch, O_(ResolveLookAtScratch, field))` so no tape-data pointer is needed.
 *
 * Struct layout (per duffle/math.h):
 *   MT3_S2S4 { A3x3_S2 m; A3_S4 t; }  →  m[][] is S2 packed (9 × 2 = 18 bytes at offset 0) 
 *                                        t[0/1/2] is S4     (3 × 4 = 12 bytes at offset 18)
 *
- * Translation column: GTE MVMVA with the world rotation matrix pre-set (helper emits set_gte_world before the bundle, per the bundle design).
+ * Translation column: GTE MVMVA with the world rotation matrix pre-set
+ * (helper emits set_gte_world before the bundle, per the bundle design).
 * MVMVA computes R * pos (with cv=0/mx=0/sf=0/v=0); MAC1/2/3 = R * (-eye).
 * Pool cost: r_look_at (1) + r_scratch (R_T4 carrier) + 4 ptr regs + 3 tmp regs = 9 GPRs.
 */
@@ -737,16 +475,14 @@ I_ void resolve_look_at__populate_and_translate_proc(MipsAtomBuilder_R ab
 	gte_mv_to_data_r(r_tmp2, C2_IR3),
 	nop2,

-	/* MVMVA: MAC1/2/3 = R * IR with cv=0 (no TR vector), mx=0 (rotation matrix),
-	 * sf=0 (no shift), v=0 (V0 = IR1/2/3, no far-plane clipping). The pre-set
-	 * rotation matrix is the one set by the preceding set_gte_world atom.
+	/* MVMVA: MAC1/2/3 = R * IR with cv=0 (no TR vector), mx=0 (rotation matrix), sf=0 (no shift), v=0 (V0 = IR1/2/3, no far-plane clipping).
+	 * The pre-set rotation matrix is the one set by the preceding set_gte_world atom.
 	 * gte_cmdw_mvmva is parameterless and defaults to cv=0/mx=0/sf=0/v=0. */
 	gte_cmdw_mvmva,
 	nop,  /* GTE interlock */

 	/* mfc2 MAC1/2/3 → r_tmp0/r_tmp1/r_tmp2 (sign-extended into 32-bit GPRs).
-	 * MAC1/2/3 hold R*v with no TR add and no perspective divide — exactly the
-	 * 3 distinct world-space translation values we need for t[0..2]. */
+	 * MAC1/2/3 hold R*v with no TR add and no perspective divide — exactly the 3 distinct world-space translation values we need for t[0..2]. */
 	gte_mv_from_data_r(r_tmp0, C2_MAC1),
 	gte_mv_from_data_r(r_tmp1, C2_MAC2),
 	gte_mv_from_data_r(r_tmp2, C2_MAC3),
@@ -821,32 +557,52 @@ internal MipsAtom_(screen_env_init) atom_info(atom_phase(screen_init)
 	mac_yield(),
 };

+/* gp_screen_init's GPR setup. Tests the mixed user-pinning + auto-reg pattern:
+ *   - R_IO_BaseAddr = R_T4 (user-pinned via atom_reg; pre-existing)
+ *   - R_GP1_Offset  = R_T2 (user-pinned via atom_reg; NEW -- for GPIO_PORT1_OFFSET)
+ *   - R_ScreenX     = R_T5 (user-pinned via atom_reg; used as a transfer and GTE setup reg)
+ *   - R_GpTmp       = auto-allocated by the lua pass and used for several GPU transfers;
+ *                    the C preprocessor resolves it to the chosen free pool GPR.
+ *
+ * For gp_screen_init, the auto-reg pool exclusions are:
+ *   user_pinned (from the corpus register_alias_registry) : R_T0..R_T7 (all 8 user-pinned across hello_camera.atom.c)
+ *   body-parsed physical registers                         : aliases resolve through the registry;
+ *                                                           the body uses R_ScreenX, not raw R_T5
+ *   source_pool after both subtractions                   : {R_V0, R_V1} only
+ *   R_GpTmp gets R_V0 (the first-fit choice). Its repeated GPU-transfer use proves that the
+ *   auto-reg allocation is active while the R_ScreenX references prove the pinned alias is used.
+ *   R_TapePtr (R_T9), R_AtomJmp (R_T8), R_AT are excluded from the POOL by construction in
+ *   passes/auto_reg.lua -- see the "obvious exclusions" comment block at the top of that file.
+ */
 enum {
 	R_IO_BaseAddr = R_T4 atom_reg, /* Caller-pinned: IO_BASE_ADDR = 0x1F800000 */
+	R_GP1_Offset  = R_T2 atom_reg, /* Caller-pinned: GPIO_PORT1_OFFSET = 0x10 */
+	atom_auto_reg(gp_screen_init, R_GpTmp),  /* Auto-allocated scratch; resolved to a free pool GPR by the lua pass. C-preprocessor expands to R_GpTmp = R_GpTmp_Code with an atom_auto_reg trailing comment. */
 #define R_IO_BaseAddr_Code R_T4_Code
+#define R_GP1_Offset_Code  R_T2_Code
 };
 internal MipsAtom_(gp_screen_init) atom_info(atom_phase(screen_init), atom_reads(R_IO_BaseAddr)) {
 	store_word(R_0, R_IO_BaseAddr, GPIO_PORT1_OFFSET),                                   /* GP1(00h) Reset */
-	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_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) */
+	mac_gcmd_push(gp1_word_ResetCmdBuffer(),   R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(01h) ClearFIFO; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_AcknowledgeIRQ(),   R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(02h) AckIRQ; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_DisplayOn(),        R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(03h) Display ON; uses pinned R_ScreenX as the transfer reg. */
+	mac_gcmd_push(gp1_word_dma_to_gpu(),       R_GpTmp,   R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(04h) DMADirection=2 (CPU->GPU). libpsyx's per-frame PutDrawEnv/DrawOTag use DMA2; without this the DMA queue never drains. Uses auto-allocated R_GpTmp. */
+	mac_gcmd_push(gp1_word_StartDisplayArea(), R_GpTmp,   R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* GP1(05h) StartDisplayArea (X=0, Y=0); uses auto-allocated R_GpTmp. */

-	/* GP1: DisplayMode + Display Ranges */
-	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_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),
+	/* GP1: DisplayMode + Display Ranges. */
+	mac_gcmd_push(gp1_word_display_mode_320x240_15bit_ntsc, R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_horizontal_range_ntsc,           R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_vertical_range_ntsc,             R_ScreenX, R_IO_BaseAddr, GPIO_PORT1_OFFSET),

 	/* GTE: SetGeomOffset (OFX, OFY) — ScreenRes_CenterX, ScreenRes_CenterY. */
-	load_upper_i(R_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),
+	load_upper_i(R_ScreenX, ScreenRes_CenterX), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFX_Code),
+	load_upper_i(R_ScreenX, ScreenRes_CenterY), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_OFY_Code),

 	/* GTE: SetGeomScreen (H) — CR26 (per PSX-SPX / libpsyx), value is the raw projection-plane distance, NOT shifted. */
-	add_ui(R_T5, R_0, ScreenZ), gte_mv_to_ctrl_r(R_T5, gte_cr_H_Code),
+	add_ui(R_ScreenX, R_0, ScreenZ), gte_mv_to_ctrl_r(R_ScreenX, gte_cr_H_Code),

 	/* GP1: DisplayEnable — bit 0 = 0 (Display ON). */
-	mac_gcmd_push(gp1_word_DisplayOn(), R_T5, R_IO_BaseAddr, GPIO_PORT1_OFFSET),
+	mac_gcmd_push(gp1_word_DisplayOn(), R_GpTmp, R_IO_BaseAddr, GPIO_PORT1_OFFSET), /* Uses auto-allocated R_GpTmp. */
 	mac_yield(),
 };

@@ -1015,43 +771,8 @@ atom_label(exit_circle_z)
 	mac_yield_tail(),
 };

-/* Scratchpad layout for the resolve_look_at bundle.
- * The chain atoms communicate entirely via the wave-context GPR carrier
- * R_ResolveScratch (R_T4) + hardcoded offsets into smem.scratchpad
- * (PS1 hardware scratchpad at 0x1F800000).
- *
- * Atom 0 (input_and_sub) STAGES the C-side inputs (eye, up_in) into the scratchpad;
- * AT THE SAME TIME it computes fwd = target - eye and stores it at scratch+0.
- * Atoms 1-6 then read/write specific scratchpad offsets internally using
- * `r_scratch + hardcoded_offset` — no tape-data pointers are passed between atoms.
- *
- *   +0   fwd   (atom 0 writes; atom 1 reads)
- *   +16  uz    (atom 1 writes; atoms 2 + 4 read)
- *   +32  right (atom 2 writes; atom 3 reads)
- *   +48  ux    (atom 3 writes; atoms 4 + 6 read)
- *   +64  up    (atom 4 writes; atom 5 reads)
- *   +80  uy    (atom 5 writes; atom 6 reads)
- *   +96  eye   (atom 0 stages from C-side pointer; atom 6 reads)
- *   +128 up_in (atom 0 stages from C-side pointer; atom 2 reads)
- *
- * No struct view is required — the C-side bundle helper passes only C-side
- * pointers (target, eye, up_in, look_at) and the scratch base address;
- * the assembly hardcodes all inter-slot offsets. The original Task 12.7 magic
- * offsets `& smem.scratchpad[N]` in the C-side helper were eliminated by this
- * redesign; the user feedback was: "you didn't have to use magic offsets into
- * the scratchpad memory. those are harcoded." */
-
 enum {
-	R_LookAt    = R_T0 atom_reg atom_type(MT3_S2S4*),
-	R_CamEye    = R_T1 atom_reg atom_type(P3_S4*),
-	R_CamTarget = R_T2 atom_reg atom_type(P3_S4*),
-	R_WorldUp   = R_T3 atom_reg atom_type(V3_S4*),
-};
-
-
-
-enum {
-	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* VRAM output cursor (primitive buffer) */
+	R_PrimCursor = R_T7 atom_reg atom_type(U4*),    /* Output cursor (primitive buffer) */
 	R_FaceCursor = R_T4 atom_reg atom_type(V4_S2*), /* Cube face-index cursor (V4_S2*); floor context switches to V3_S2* via atom_phase */
 	R_VertBase   = R_T5 atom_reg atom_type(V3_S2*), /* Base address of the vertex array */
 	R_OtBase     = R_T6 atom_reg atom_type(U4*),    /* Base address of the Ordering Table */
@@ -1060,7 +781,6 @@ enum {
 #define R_VertBase_Code    R_T5_Code
 #define R_OtBase_Code      R_T6_Code
 };
-
 typedef Struct_(Binds_CubeTri) {
 	U4     PrimCursor;
 	V4_S2* FaceCursor;
@@ -1097,7 +817,7 @@ MipsAtom_(cube_g4_face) atom_info(atom_phase(cube_g4),

 	gte_mv_from_data_r(R_T0, C2_MAC0), nop,
 	branch_le_zero(R_T0, atom_offset(cull, cube_g4_face_exit)),
-		/* BD-slot: write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
+		/* BD-slot: Write the prim tag (R_0=0; overwrites the legacy tag word in the prim_buffer).
 		 * If branch IS taken (face culled), the body is skipped and this 0-tag is stranded —
 		 * harmless because the OT entry that points to this prim is created later. */
 		store_word(R_0, R_PrimCursor, O_(Poly_G4, tag)),
@@ -43,6 +43,7 @@
 #pragma region Hello Camera Headers
 #	include "gen/macs.h"
 #	include "gen/offsets.h"
+#	include "gen/auto_reg.h"

 #include "hello_camera.h"
 #pragma endregion Hello Camera Headers
@@ -136,16 +137,23 @@ resolve_look_at_c11(MT3_S2S4* look_at, P3_S4* eye, P3_S4* target, V3_S4* up_in)
 * After this returns, the smem.resolve_look_at_atom_addrs[] array contains valid MIPS atom pointers
 * for the frame-time bundle helper to emit via tb_emit(tb, captured_addr).
 *
- * 7 atoms are within hello_camera.atom.c:
+ * 4 unique procs in hello_camera.atom.c (chain atoms 0, 2, 4, 6); atoms 1, 3, 5
+ * share the GENERIC normalize_v3s4_proc from gte.atom.c (called 3x with different
+ * O_(ResolveLookAtScratch,...) offsets):
 * 	0: resolve_look_at__input_and_sub_proc
- * 	1: resolve_look_at__normalize_fwd_to_uz_proc
+ * 	1: normalize_v3s4_proc                  (fwd → uz; offsets 0, 16)
 * 	2: resolve_look_at__cross_uz_up_in_to_right_proc
- * 	3: resolve_look_at__normalize_right_to_ux_proc
+ * 	3: normalize_v3s4_proc                  (right → ux; offsets 32, 48)
 * 	4: resolve_look_at__cross_uz_ux_to_up_proc
- * 	5: resolve_look_at__normalize_up_to_uy_proc
+ * 	5: normalize_v3s4_proc                  (up → uy; offsets 64, 80)
 * 	6: resolve_look_at__populate_and_translate_proc
 *
- * (gte.atom.c contains only normalize_v3s4_proc — bundle-specific scratch layout is no longer exposed to the GTE primitives file.)
+ * Task 12.16 promotion: the bundle-specific resolve_look_at__chain_normalize_proc
+ * has been promoted to the generic normalize_v3s4_proc (gte.atom.c), which now
+ * takes r_scratch + r_src_offset + r_dst_offset as U4 parameters. The 3 callers
+ * pass O_(ResolveLookAtScratch,...) macros as offset args. The metaprogram emits
+ * one set of `atom_offset__normalize_v3s4__srav_path__aligned_done` defs
+ * (namespaced by atom name) in duffle/gen/offsets.h, shared by all 3 callers.
 *
 * GPR pool per atom: 10 free GPRs (R_T0..R_T3 + R_T5..R_T7 + R_V0 + R_V1 + R_AT).
 * R_T4 is reserved as the wave-context carrier (R_ResolveScratch).
@@ -161,75 +169,82 @@ internal void resolve_look_at_init(void) {
 	 * computes fwd = target - eye; binds R_ResolveScratch (R_T4) as the wave-context carrier for atoms 1-6.
 	 * The body hardcodes R_AT and R_V0 as eye.y/eye.z temps (the existing sub_u(eye.x, eye.y, eye.z) chain from the prior Task 12.7 design). */
 	smem.resolve_look_at_atom_addrs[0] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__input_and_sub_proc(ab,
+	resolve_look_at__input_and_sub_proc(ab, R_ResolveScratch,
 		R_T0,  /* r_target_ptr (popped from tape) */
 		R_T1,  /* r_eye_ptr    (popped from tape) */
 		R_T2,  /* r_up_in_ptr  (popped from tape) */
-		R_ResolveScratch, /* r_scratch (wave-context carrier; popped from tape) */
-		R_T3,             /* r_tmp0 */
-		R_T5,             /* r_tmp1 */
-		R_T6,             /* r_tmp2 */
-		R_T7);            /* r_tmp3 */
+		R_T3, R_T5, R_T6,  R_T7); /* r_tmp<0-3> */

-	/* Atom 1: resolve_look_at__normalize_fwd_to_uz — src=scratch+0, dst=scratch+16 (HARDCODED in body).
-	 * GPR pool: r_scratch (R_T4 carrier) + 10 body GPRs = 11.
-	 *   r_a/r_b (R_T0/R_T1) : src/dst pointers (r_a overlaps r_recip_est after the loads)
-	 *   r_e/r_f/r_i (R_T2/R_T3/R_T5) : src.x/y/z → result.x/y/z
-	 *   r_d/r_g (R_T6/R_T7) : MAC1/2 scratch (dead after stage 2)
-	 *   r_h (R_V0) : LZCR
-	 *   r_recip_est (R_V1), r_shift (R_AT) : saved throughout */
+	/* Atom 1: normalize_v3s4_proc (generic, from gte.atom.c) — src=scratch+0=fwd, dst=scratch+16=uz.
+	 * The proc takes r_src_offset + r_dst_offset as U4 PARAMETERS — we pass the O_(...) macros here (evaluating to numeric literals 0 and 16).
+	 * The 4-stage body is identical across the 3 call sites (atoms 1, 3, 5); only the offset args differ.
+	 * GPR pool: r_scratch (R_T4 carrier) + 9 body GPRs = 10.
+	 *   r_src_ptr      (R_T0) : src ptr
+	 *   r_dst_ptr      (R_T1) : dst ptr
+	 *   r_tmp          (R_T2) : unused (reserved for symmetry)
+	 *   r_mac1_scratch (R_T3) : MAC1 scratch
+	 *   r_mac2_scratch (R_T5) : src.x → result.x (carries through stages 1-2)
+	 *   r_recip_est    (R_T6) : src.y → result.y
+	 *   r_lzcr         (R_T7) : |v|² accumulator + srav amount (single reg)
+	 *   r_shift        (R_V0) : LZCR (saved across stages 3-4)
+	 *   r_branch_tmp   (R_V1) : src.z → result.z (reused after stage 1)
+	 */
 	smem.resolve_look_at_atom_addrs[1] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__normalize_fwd_to_uz_proc(ab,
-		R_ResolveScratch, /* r_scratch (wave-context carrier; src/dst base) */
-		R_T0, R_T1,       /* r_a, r_b (src/dst ptrs) */
-		R_T2, R_T3, R_T5, /* r_e, r_f, r_i (src components) */
-		R_T6, R_T7,       /* r_d, r_g (MAC scratch) */
-		R_V0,             /* r_h (LZCR) */
-		R_V1,             /* r_recip_est */
-		R_AT);            /* r_shift */
+	normalize_v3s4_proc(ab, R_ResolveScratch, /* r_scratch (wave-context carrier) */
+		O_(ResolveLookAtScratch, fwd),          /* r_src_offset = 0 */
+		O_(ResolveLookAtScratch, uz),           /* r_dst_offset = 16 */
+		R_T0, R_T1, R_T2,                       /* r_src_ptr, r_dst_ptr, r_tmp */
+		R_T3,                                   /* r_mac1_scratch */
+		R_T5,                                   /* r_mac2_scratch */
+		R_T6,                                   /* r_recip_est */
+		R_T7,                                   /* r_lzcr */
+		R_V0,                                   /* r_shift */
+		R_V1);                                  /* r_branch_tmp */

 	/* Atom 2: resolve_look_at__cross_uz_up_in_to_right — a=scratch+16, b=scratch+128,
 	 * out=scratch+32 (HARDCODED in body). GPR pool: r_scratch + 7 body + R_AT + R_V0 = 10. */
 	smem.resolve_look_at_atom_addrs[2] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__cross_uz_up_in_to_right_proc(ab,
-		R_ResolveScratch,   /* r_scratch (wave-context carrier; src/dst base) */
+	resolve_look_at__cross_uz_up_in_to_right_proc(ab, R_ResolveScratch, /* r_scratch (wave-context carrier; src/dst base) */
 		R_T0, R_T1, R_T2, /* r_a, r_b, r_c (a.x/y/z → out.x/y/z) */
 		R_T3,             /* r_d (b.x) */
 		R_T5,             /* r_f (out ptr = scratch+32) */
 		R_T6,             /* r_g (a ptr = scratch+16) */
 		R_T7);            /* r_h (b ptr = scratch+128) */

-	/* Atom 3: resolve_look_at__normalize_right_to_ux — src=scratch+32, dst=scratch+48 (HARDCODED). */
+	/* Atom 3: normalize_v3s4_proc (generic, from gte.atom.c) — src=scratch+32=right, dst=scratch+48=ux. */
 	smem.resolve_look_at_atom_addrs[3] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__normalize_right_to_ux_proc(ab,
-		R_ResolveScratch,
-		R_T0, R_T1,
-		R_T2, R_T3, R_T5,
-		R_T6, R_T7,
+	normalize_v3s4_proc(ab, R_ResolveScratch,
+		O_(ResolveLookAtScratch, right), /* r_src_offset = 32 */
+		O_(ResolveLookAtScratch, ux),    /* r_dst_offset = 48 */
+		R_T0, R_T1, R_T2,
+		R_T3,
+		R_T5,
+		R_T6,
+		R_T7,
 		R_V0,
-		R_V1,
-		R_AT);
+		R_V1);

 	/* Atom 4: resolve_look_at__cross_uz_ux_to_up — a=scratch+16, b=scratch+48, out=scratch+64 (HARDCODED). */
 	smem.resolve_look_at_atom_addrs[4] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__cross_uz_ux_to_up_proc(ab,
-		R_ResolveScratch,
+	resolve_look_at__cross_uz_ux_to_up_proc(ab, R_ResolveScratch,
 		R_T0, R_T1, R_T2,
 		R_T3,
 		R_T5,  /* r_f (out ptr = scratch+64) */
 		R_T6,  /* r_g (a ptr = scratch+16) */
 		R_T7); /* r_h (b ptr = scratch+48) */

-	/* Atom 5: resolve_look_at__normalize_up_to_uy — src=scratch+64, dst=scratch+80 (HARDCODED). */
+	/* Atom 5: normalize_v3s4_proc (generic, from gte.atom.c) — src=scratch+64=up, dst=scratch+80=uy. */
 	smem.resolve_look_at_atom_addrs[5] = (MipsAtom*)u4_v(ab->start + ab->used * sizeof(U4));
-	resolve_look_at__normalize_up_to_uy_proc(ab,
-		R_ResolveScratch,
-		R_T0, R_T1,
-		R_T2, R_T3, R_T5,
-		R_T6, R_T7,
+	normalize_v3s4_proc(ab, R_ResolveScratch,
+		O_(ResolveLookAtScratch, up), /* r_src_offset = 64 */
+		O_(ResolveLookAtScratch, uy), /* r_dst_offset = 80 */
+		R_T0, R_T1, R_T2,
+		R_T3,
+		R_T5,
+		R_T6,
+		R_T7,
 		R_V0,
-		R_V1,
-		R_AT);
+		R_V1);

 	/* Atom 6: resolve_look_at__populate_and_translate — write look_at->m[][] from ux/uy/uz (computed from r_scratch+offset internally),
 	then compute translation column t[] = R * (-eye). GPR pool: r_look_at + r_scratch + 4 ptr regs + 3 tmp regs = 9. */
@@ -336,20 +351,48 @@ void update(PrimitiveArena* pa, U4* ordering_buf)
 	A2_S2 p;    //???
 	S4 flag; //????

-	// Camera Look at
+	// Camera Look at (Tape) + inline C11 fallback — bundle runs, then C11 inlines the look_at.
+	// Currently: bundle's atom 0 (input_and_sub) runs + C11 does the rest. As bundle atoms
+	// are incrementally fixed, the corresponding C11 lines get commented out.
 	if (1)
 	{
-		camera_look_at_c11(& smem.cam, & smem.cube.pos, & v3s4(0, -fp_one, 0));
-	}
-	// Camera look at (Tape)
-	{
-		MT3_S2S4* look_at = & smem.cam.look_at;
-		P3_S4*    eye     = & smem.cam.pos;
-		V3_S4*    up_in   = & v3s4(0, -fp_one, 0);
-
 		tb.used = 0; tb_scope_run(& tb) {
 			resolve_look_at(& tb, & smem.cam.look_at, & smem.cam.pos, & smem.cube.pos, & v3s4(0, -fp_one, 0));
 		}
+
+		// RGA(Lengyel): Build matrix expansion of a rigid transformation. Corresponding motor is not constructed; we write the LA form for GTE.
+		// Preconditions: eye != target, up_in not collinear with (target - eye).
+		V3_S4 right, up, forward;
+		V3_S4 ux, uy, uz;
+		V3_S4 pos, off;
+
+		// forward = smem.cube.pos; sub_v3s4(& forward, smem.cam.pos); // RGA(Lengyel): Affine point - point = zero-weight direction.  (now done by bundle atom 0)
+		// Read fwd from scratchpad[+0] (atom 0's output)
+		forward.x = u4_v(0x1F800000)[0];
+		forward.y = u4_v(0x1F800000)[1];
+		forward.z = u4_v(0x1F800000)[2];
+		forward.pad = u4_v(0x1F800000)[3];
+		// normalize_v3s4(& forward, & uz);                  // RGA(Lengyel): Normalize the direction bulk. Not finite-point unitization.  (now done by bundle atom 1)
+		// Read uz from scratchpad[+16] (atom 1's output)
+		uz.x = u4_v(0x1F800010)[0];
+		uz.y = u4_v(0x1F800010)[1];
+		uz.z = u4_v(0x1F800010)[2];
+		uz.pad = u4_v(0x1F800010)[3];
+
+		cross_v3s4(& uz,   & v3s4(0, -fp_one, 0), & right); normalize_v3s4(& right, & ux); // RGA(Lengyel): Complement(Wedge(forward, up_in)) -> right axis.
+		cross_v3s4(& uz, & ux,    & up);    normalize_v3s4(& up,    & uy); // RGA(Lengyel): Complement(Wedge(forward, right)) -> up axis.
+
+		// RGA(Lengyel): matrix expansion of the world-to-camera rotation (basis rows).
+		smem.cam.look_at.m[0][0] = ux.x; smem.cam.look_at.m[0][1] = ux.y; smem.cam.look_at.m[0][2] = ux.z;
+		smem.cam.look_at.m[1][0] = uy.x; smem.cam.look_at.m[1][1] = uy.y; smem.cam.look_at.m[1][2] = uy.z;
+		smem.cam.look_at.m[2][0] = uz.x; smem.cam.look_at.m[2][1] = uz.y; smem.cam.look_at.m[2][2] = uz.z;
+
+		pos = smem.cam.pos; mul_v3s4(& pos, v3s4(-1,-1,-1)); // RGA(Lengyel): -eye in world coordinates (spatial bulk only; implicit weight is dropped).
+
+		// RGA(Lengyel): R * (-eye) is the full matrix translation column.
+		// Motor translator would store half this displacement in m.xyz; GTE consumes full column.
+		mul_m3s2_v3s4(& smem.cam.look_at, & pos, & off);
+		trans_m3s2(   & smem.cam.look_at,        & off);
 	}

 	// Draw cube
@@ -515,8 +515,7 @@ local function splice_c_lines(source)
 		local splice_len = nil
 		if byte == BYTE_BACKSLASH and source:byte(pos + 1) == BYTE_NEWLINE then
 			splice_len = 2
-		elseif byte == BYTE_BACKSLASH and source:byte(pos + 1) == BYTE_CR
-			and source:byte(pos + 2) == BYTE_NEWLINE then
+		elseif byte == BYTE_BACKSLASH and source:byte(pos + 1) == BYTE_CR and source:byte(pos + 2) == BYTE_NEWLINE then
 			splice_len = 3
 		end

@@ -2095,7 +2094,8 @@ local E_MAC_PREFIX_LEN = 4
 ---   * Unknown `mac_X` (not in `component_index`): fall back to `word_counts[ident]` if present; otherwise emit one opaque event so the cycle budget accounts for the word.
 ---   * Marker Tokens (`atom_label(...)` / `atom_offset(...)`): Zero events (they are pure metaprogram hints).
 ---
--- Cycle protection: a per-expansion `visiting` set tracks components currently on the expansion stack; a re-entry produces a deterministic `{kind = "cycle", ...}` error and aborts that branch (does NOT hang, does NOT recurse).
+--- Cycle protection: a per-expansion `visiting` set tracks components currently on the expansion stack;
+--- a re-entry produces a deterministic `{kind = "cycle", ...}` error and aborts that branch (does NOT hang, does NOT recurse).
 ---
 --- Pure: reads `body_entry` / `component_index` / `word_counts`. Memoization is the caller's responsibility.
 --- Callers wanting `word_events` / `word_event_errors` precomputed for many atoms should memoize them per atom.
@@ -47,8 +47,7 @@ local function find_repo_root()
 	return root
 end

--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and
--- `package.cpath` (for `lpeg.dll`).
+--- Set `package.path` (for `require("duffle")` + `require("passes.X")`) and `package.cpath` (for `lpeg.dll`).
 ---
 --- This script does NOT touch the OS environment: no `os.setenv`, no `os.putenv`, no `$PATH` mods. 
 --- It just sets `package.path` and `package.cpath` (the standard Lua way to register module search dirs). 
@@ -4,22 +4,21 @@
 --- Runs a deterministic first-fit allocator in the `R_T0..R_T7 + R_V0..R_V1` pool (10 physical GPRs).
 --- Emits one `#define R_<Sym>_Code R_Tn_Code` per marker into per-directory `gen/auto_reg.h`.
 ---
--- User-pinned GPRs (added 2026-08-10): The corpus's `register_alias_registry` is consulted to
--- exclude GPRs the user has pinned via `atom_reg` + `_Code` defs (e.g. wave-context carriers like
--- `R_ResolveScratch = R_T4 atom_reg`). These GPRs are unavailable to EVERY atom's source pool,
--- not just to atoms in the same phase — wave-context carriers are preserved across atoms by the
--- wave-context discipline and must never be reallocated.
+--- User-pinned GPRs : The corpus's `register_alias_registry` is consulted to exclude GPRs the user has pinned via 
+--- `atom_reg` + `_Code` defs (e.g. carriers like `R_ResolveScratch = R_T4 atom_reg`). 
+--- These GPRs are unavailable to EVERY atom's source pool.
+--- Carriers are preserved across atoms by context discipline and must never be reallocated.
 --- Per-atom body parsing also catches alias references (R_<Alias>) and hardcoded R_Tn references,
 --- so the user can write either `R_T4` or `R_ResolveScratch` in an atom body and the pass will
 --- exclude R_T4 from that atom's pool.
 ---
 --- Conflict detection: If the user hardcodes `R_Tn` in an atom body that shares a phase with an auto-reg that picked `R_Tn`,
--- emit `phase_register_clash` as an info finding (no build stop). Should be unreachable after the
--- user-pinning + body-parsing fix above; kept as a defensive safety net.
+--- emit `phase_register_clash` as an info finding (no build stop).
+--- Should be unreachable after the user-pinning + body-parsing fix above; kept as a defensive safety net.
 ---
--- Pool exhaustion: if a phase declares more `R_<Sym>` mappings than the 10-register pool can hold,
+--- Pool exhaustion: If a phase declares more `R_<Sym>` mappings than the 10-register pool can hold,
 --- emit `phase_register_pool_exhausted` as a build-stopping error.
---
+
 --- @class AutoRegResult
 --- @field outputs  table[]  -- {kind=, path=} entries
 --- @field errors   table[]  -- {line=, msg=}  entries (build-stops)
@@ -28,20 +27,57 @@
 local _bootstrap_dir = debug.getinfo(1, "S").source:match("^@?(.*[/\\])") or "./"
 local duffle         = dofile(_bootstrap_dir .. "../duffle_paths.lua")

-- The fixed allocation pool: 10 physical GPRs whose `R_<Sym>_Code` macros exist in mips.h (lines 92-107).
-- Each pool entry is the PHYSICAL GPR ident (R_T0 etc.);
-- `gpr .. "_Code"` resolves to the matching `R_Tn_Code` constant the source code references via `#define R_Load_Code R_T0_Code`.
-- Excluded: R_AT (assembler temp per lottes_tape.h:86), R_T8 (deferred to ac_yield_load pattern),
--           R_T9 (R_TapePtr; owned by the tape runtime).
+--- ════════════════════════════════════════════════════════════════════════════
+--- THE GPR ALLOCATION POOL — what is allocatable, and (more importantly) WHY
+--- ════════════════════════════════════════════════════════════════════════════
+---
+--- The auto-reg pass picks physical GPRs for `atom_auto_reg(...)` / `phase_auto_reg(...)` markers.
+--- It allocates from a FIXED 10-register pool.
+--- This comment block makes the inclusion AND exclusion criteria obvious so a reader doesn't have 
+--- to grep lottes_tape.h + mips.h to understand the design.
+---
+--- ── WHAT'S IN THE POOL (10 GPRs, all caller-trash per the O32 ABI) ────────
+---     R_T0..R_T7  (GPR codes 8..15), R_V0..R_V1 (GPR codes 2..3)  
+---      The workhorse of every atom body. The uesr should be aware of atom allocation across atoms they chain. 
+---      If they have a collision it means either they didn't saturate the register file optimally for a phase,
+---      or the may have made the workload to large for the run.
+---
+--- ── WHAT'S NOT IN THE POOL — and WHY (the "obvious exclusions") ────────────
+---   R_T9  (GPR code 25)  — R_TapePtr, the tape instruction stream pointer.
+---     Owned by the tape runtime (in tape_run / tape_run_a02_s07).
+---     `rgcc(R_TapePtr)` register-variable ties the C compiler's view to $t9 across the whole tape_run.
+---     The auto-reg pass MUST NOT clobber this; doing so would desync the C-side tape pointer from the
+---     hardware pointer and crash on the next tape_run.
+---
+---   R_T8  (GPR code 24)  — R_AtomJmp, the atom-jump register used by the 4-word yield handshake.
+---     Every `mac_yield()` / `mac_yield_tail` does `load_word R_AtomJmp, R_TapePtr, 0` then
+---     `jump_reg R_AtomJmp`. The auto-reg pass MUST NOT clobber this either, or the atom dispatcher breaks.
+---     Owned by the tape runtime, same family as R_TapePtr.
+---
+---   R_AT  (GPR code 1)   — Assembler temporary. Reserved by the MIPS O32 ABI for pseudoinstruction expansion
+---     (lottes_tape.h:86, mips.h:93). The ISA's psuedo instructions use it as a scratch temporary.
+---
+---   R_A0..A3 (codes 4..7)   — Function arguments. Used in tape_run_a02_s07, see below.
+---   R_S0..S7 (codes 16..23) — Callee-saved. Preserved across C-ABI calls by convention.
+---     The `tape_run_a02_s07` variant clobbers them deliberately, but the default `tape_run` does NOT.
+---     Kept out of POOL to preserve the conservative default.
+---     Add them in a separate "big clobber" pool if/when needed.
+---
+---   R_K0/K1       (codes 26..27) — Kernel / interrupt handler reserves. Never touched by user code; OS-internal.
+---   R_GP/SP/FP/RA (codes 28..31) — Stack frame + return-address. Owned by the C compiler; never allocatable.
+---   R_0           (code 0)       — Hardwired zero. Cannot be written.
+---
 local POOL = {
    "R_T0", "R_T1", "R_T2", "R_T3",
    "R_T4", "R_T5", "R_T6", "R_T7",
    "R_V0", "R_V1",
 }

-- Map from integer MIPS GPR code (the `code` field on AliasEntry) to the physical GPR ident
-- in POOL. The standard MIPS O32 ABI register numbering matches mips.h's R_*_Code #defines
-- (mips.h:92-123). Only the POOL entries matter for auto_reg — non-pool aliases are out of scope.
+-- Map from integer MIPS GPR code (the `code` field on AliasEntry) to the physical GPR ident in POOL.
+-- The standard MIPS O32 ABI register numbering matches mips.h's R_*_Code #defines (mips.h).
+-- Only the POOL entries matter for auto_reg — non-pool aliases
+-- (R_AT=1, R_A0..A3=4..7, R_T8=24, R_T9=25, R_K0/K1=26..27, R_GP/SP/FP/RA=28..31) 
+-- are deliberately omitted — see the comment block above for the WHY of each exclusion.
 local INT_CODE_TO_POOL_GPR = {
    [2]  = "R_V0", [3]  = "R_V1",
    [8]  = "R_T0", [9]  = "R_T1", [10] = "R_T2", [11] = "R_T3",
@@ -71,8 +107,9 @@ local function allocate_phase(phase_label, decls)
        if not next_gpr then
            errors[#errors + 1] = {
                line = 0,
-                msg  = string.format(
-                    "phase_register_pool_exhausted: phase '%s' requested symbol '%s' but the pool has no remaining registers (max 10 per phase: R_T0..R_T7 + R_V0..R_V1). Split the phase or use hardcoded GPRs."
+                msg  = string.format("phase_register_pool_exhausted: "
+									.. "phase '%s' requested symbol '%s' but the pool has no remaining registers "
+									.. "(max 10 per phase: R_T0..R_T7 + R_V0..R_V1). Split the phase or use hardcoded GPRs."
                  , phase_label, sym),
            }
            return result, errors
@@ -83,21 +120,17 @@ local function allocate_phase(phase_label, decls)
 end

 -- Build two projections from corpus.register_alias_registry:
--   user_pinned   -- { [physical_gpr_ident] = true }   -- GPRs unavailable to auto_reg globally
--                                                       -- (wave-context carriers, file-scope pinned aliases)
+--   user_pinned   -- { [physical_gpr_ident] = true } -- GPRs unavailable to auto_reg globally (wave-context carriers, file-scope pinned aliases)
 --   alias_to_gpr  -- { [alias_ident]        = physical_gpr_ident } -- for body parsing
-- Both projections are derived from the same set of entries: every AliasEntry in
-- register_alias_registry has `has_atom_reg = true` (only those entries are added to the
-- registry; see passes/scan_source.lua parse_enum_entry). Each entry's `code` is the integer
-- MIPS GPR number (0..31); INT_CODE_TO_POOL_GPR translates it back to the physical GPR ident.
-- Aliases whose `code` points to a non-POOL GPR (e.g. R_S0, R_T8, R_K1) are ignored — they
-- don't affect the auto_reg pool, and they're already excluded from POOL above.
+-- Both projections are derived from the same set of entries: every AliasEntry in register_alias_registry has `has_atom_reg = true` 
+-- (only those entries are added to the registry; see passes/scan_source.lua parse_enum_entry).
+-- Each entry's `code` is the integer MIPS GPR number (0..31); INT_CODE_TO_POOL_GPR translates it back to the physical GPR ident.
+-- Aliases whose `code` points to a non-POOL GPR (e.g. R_S0, R_T8, R_K1) are ignored — 
+-- they don't affect the auto_reg pool, and they're already excluded from POOL above.
 local function build_user_pins(corpus)
    local user_pinned  = {}
    local alias_to_gpr = {}
-    if not corpus.register_alias_registry then
-        return user_pinned, alias_to_gpr
-    end
+    if not corpus.register_alias_registry then return user_pinned, alias_to_gpr end
    for alias_name, alias_entry in pairs(corpus.register_alias_registry) do
        if alias_entry.has_atom_reg and alias_entry.code then
            local gpr = INT_CODE_TO_POOL_GPR[alias_entry.code]
@@ -111,10 +144,10 @@ local function build_user_pins(corpus)
 end

 -- Find every physical GPR referenced in the atom body, via EITHER:
--   (a) a hardcoded physical GPR ident (R_T\d+|R_V\d+|R_A\d+|R_S\d+) — the existing regex;
--   (b) an alias ident (R_<Alias>) resolved via alias_to_gpr back to its physical GPR ident.
-- Returns { [physical_gpr_ident] = count }. The clash-detection and source-pool-exclusion logic
-- only needs the presence of each GPR (boolean test), but keeping the count preserves the
+--   (a) A hardcoded physical GPR ident (R_T\d+|R_V\d+|R_A\d+|R_S\d+) — the existing regex;
+--   (b) An alias ident (R_<Alias>) resolved via alias_to_gpr back to its physical GPR ident.
+-- Returns { [physical_gpr_ident] = count }. Clash-detection and source-pool-exclusion logic
+-- only needs the presence of each GPR (boolean test), but keeping count preserves the
 -- original find_hardcoded_rn shape so callers can switch without churn.
 -- The alias pattern is sorted lexicographically to keep the regex deterministic.
 local function find_used_gprs(body_text, alias_to_gpr)
@@ -190,12 +223,10 @@ function M.run(ctx)
    end

    -- 0. Build the user-pinned GPR exclusion set + alias-to-GPR resolution map.
-    --    Wave-context carriers (e.g. `R_ResolveScratch = R_T4 atom_reg` in
-    --    hello_camera.atom.c) MUST NOT be allocated to any auto-reg marker — they're
-    --    preserved across atoms by the wave-context discipline. The corpus's
-    --    register_alias_registry is the source of truth for these opt-in pins.
-    --    Body references to those aliases (via alias_to_gpr) are also excluded on a
-    --    per-atom basis in step 2 below.
+    --  Wave-context carriers (e.g. `R_ResolveScratch = R_T4 atom_reg` in hello_camera.atom.c)
+    --  MUST NOT be allocated to any auto-reg marker — they're preserved across atoms by the wave-context discipline.
+		--  The corpus's register_alias_registry is the source of truth for these opt-in pins.
+    --  Body references to those aliases (via alias_to_gpr) are also excluded on a per-atom basis in step 2 below.
    local user_pinned, alias_to_gpr = build_user_pins(corpus)

    -- 1. Allocate phase pools first (phase declarations take precedence over per-atom declarations).
@@ -213,8 +244,8 @@ function M.run(ctx)

    -- 2. Allocate per-atom auto-regs. If the atom scope matches a phase, reuse the phase pool.
    -- Otherwise, allocate a private pool for the atom.
-    -- The phase membership is in `corpus.atom_phases[phase_label].atoms` (an array of atom names declared via `atom_phase(<phase>)` in the atom's `atom_info` line). 
-		-- Build a reverse map `atom_name -> phase_label` so the lookup is O(1) per atom scope.
+    -- The phase membership is in `corpus.atom_phases[phase_label].atoms` (an array of atom names declared via `atom_phase(<phase>)`
+		-- in the atom's `atom_info` line). Build a reverse map `atom_name -> phase_label` so the lookup is O(1) per atom scope.
    local atom_name_to_phase = {}
    for phase_label, entry in pairs(corpus.atom_phases or {}) do
        for _, atom_name in ipairs(entry.atoms or {}) do
@@ -229,8 +260,7 @@ function M.run(ctx)
        --   (a) every GPR already committed (phase allocations + prior atom allocations)
        --   (b) every USER-PINNED GPR (wave-context carriers + file-scope pinned aliases)
        --   (c) every GPR referenced in the atom's body — either hardcoded R_X or alias R_Xxx
-        --       (the latter resolved via alias_to_gpr; this catches cases where the user
-        --       wrote R_ResolveScratch instead of R_T4 directly)
+        --       (the latter resolved via alias_to_gpr; this catches cases where the user wrote R_ResolveScratch instead of R_T4 directly)
        -- Atoms whose scope matches a phase share the global pool with the phase allocations;
        -- the original `source_pool = phase_allocations[phase_label]` form used the phase
        -- allocation MAP as a pool, but that map has no array part, so `table.remove(source_pool, 1)`
@@ -259,7 +289,8 @@ function M.run(ctx)
            if not next_gpr then
                errors[#errors + 1] = {
                    line = 0,
-                    msg  = string.format("phase_register_pool_exhausted: atom '%s' requested symbol '%s' but no free registers remain in its scope pool."
+                    msg  = string.format("phase_register_pool_exhausted: atom '%s' requested symbol '%s' "
+											.. "but no free registers remain in its scope pool."
                    	, atom_scope, sym),
                }
            else
@@ -271,10 +302,10 @@ function M.run(ctx)

    -- 3. Conflict-with-hardcoded detection (defensive — should be unreachable now).
    -- The source_pool exclusion in step 2 (b) + (c) already accounts for both user-pinned GPRs
-    -- and body-referenced GPRs (hardcoded R_Tn OR alias R_<Alias>). An auto-reg allocation that
-    -- matched an existing body reference would be impossible by construction. This warning is kept
-    -- as a defensive safety net for cases the body scanner might miss (e.g. macros that expand to
-    -- register references the scanner cannot resolve).
+    -- and body-referenced GPRs (hardcoded R_Tn OR alias R_<Alias>).
+		-- An auto-reg allocation that matched an existing body reference would be impossible by construction.
+		-- This warning is kept as a defensive safety net for cases the body scanner might miss
+		-- (e.g. macros that expand to register references the scanner cannot resolve).
    -- For each resolved (scope, sym) -> R_Tn mapping, scan the atom body source for used GPRs.
    for atom_scope, decls in pairs(atom_allocations) do
        local atom = corpus.atoms_by_name and corpus.atoms_by_name[atom_scope]
@@ -284,7 +315,8 @@ function M.run(ctx)
                if used_in_body[allocated_gpr] and used_in_body[allocated_gpr] > 0 then
                    warnings[#warnings + 1] = {
                        line  = atom.line or 0,
-                        msg   = string.format("phase_register_clash: atom '%s' has hardcoded '%s' in its body AND an auto-reg marker '%s' that was allocated to '%s' (same phase). Resolve by removing the hardcoded reference or renaming the auto-reg."
+                        msg   = string.format("phase_register_clash: atom '%s' has hardcoded '%s' in its body AND an auto-reg marker '%s' "
+													.. "that was allocated to '%s' (same phase). Resolve by removing the hardcoded reference or renaming the auto-reg."
                        	, atom_scope, allocated_gpr, sym, allocated_gpr),
                    }
                end
@@ -300,7 +332,8 @@ function M.run(ctx)
        for _, src in ipairs(sources) do
            -- Collect every (sym -> gpr) entry that originated from a source in this directory.
            -- `src.scan.atom_auto_regs` is keyed by ATOM SCOPE NAME; `pairs(t)` iterates KEYS so `scope_name` here is the scope ident (e.g. "cube_g4_face").
-						-- The previous `for _, scan_atom_auto` form silently assigned the VALUE (a `{sym = sym}` table) to the variable, which made `atom_allocations[scan_atom_auto]` a table-indexed lookup that never resolved.
+						-- The previous `for _, scan_atom_auto` form silently assigned the VALUE (a `{sym = sym}` table) to the variable,
+						-- which made `atom_allocations[scan_atom_auto]` a table-indexed lookup that never resolved.
            for scope_name in pairs(src.scan and src.scan.atom_auto_regs or {}) do
                for sym, gpr in pairs(atom_allocations[scope_name] or {}) do
                    per_dir_mappings[sym] = gpr
@@ -3,9 +3,12 @@
 --- Ownership: `corpus.word_counts`, `corpus.components`, and `corpus.component_body_index`.
 --- Scanner owns `declaration_comment` and `debug_skip` on each declaration record; this pass projects both forward.
 ---
--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)`, `MipsAtomComp_Proc_(ac_X, { body })`, and `MipsAtom_Proc_(X, ab, { body })` declarations,
+--- Reads the pre-scanned SourceScan payload from `duffle.scan_source` for `MipsAtomComp_(ac_X)` and `MipsAtomComp_Proc_(ac_X, { body })` declarations (kind="comp_bare" / "comp_proc"),
 --- then resolves the function-args string from the preceding `FI_ Slice_MipsCode ac_X(...)` declaration via a backward walk.
 ---
+--- `MipsAtom_Proc_(X, ab, { body })` declarations (kind="atom_proc") are ATOMS, not components, and are deliberately excluded —
+--- atoms get emitted via `tb_emit(tb, code_<name>)` linker symbols, not inlined as `mac_*` macros.
+---
 --- Emits one `gen/macs.h` per *immediate source directory* with `#define mac_X(sig) \` macros plus `WORD_COUNT(mac_X, N)` entries for downstream offset computation.
 --- All sources inside the same directory contribute to the same file (per-directory aggregation).
 --- The directory itself is the namespace, so the filename does not repeat the module name.
@@ -76,7 +79,7 @@ local MACS_FILENAME   = "macs.h"
 --- @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" | "atom_proc"
+--- @field kind       string     -- "comp_bare" | "comp_proc"  (atom_proc is NOT a component — see `project_components`)
 --- @field debug_skip boolean    -- Mirror of `a.debug_skip` (scanner-owned); true iff a bare `atom_dbg_skip` marker immediately preceded the declaration

 -- ════════════════════════════════════════════════════════════════════════════
@@ -200,16 +203,17 @@ end
 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" or a.kind == "atom_proc" then
-			-- `MipsAtom_Proc_` atoms have no `FI_ Slice_MipsCode ac_X(...)` function-decl prelude
-			-- (the macro sits inside a wrapping `I_ void <proc_name>(...)` body), so the function-args
-			-- lookup is meaningless; signature defaults to `...` (variadic-ignored).
-			-- The `mac_<name>` alias expansion discards the `ab` (atom-builder) arg the same way
-			-- `MipsAtomComp_Proc_` components do.
-			local args = nil
-			if a.kind ~= "atom_proc" then
-				args = find_function_args_for(source, a.raw_name, a.ident_pos)
-			end
+		-- Only `MipsAtomComp_(ac_X)` (kind="comp_bare") and `MipsAtomComp_Proc_(ac_X, ...)` (kind="comp_proc")
+		-- are COMPONENTS — they get inlined via `mac_<name>` aliases inside atom bodies.
+		-- `MipsAtom_Proc_` (kind="atom_proc") is an ATOM (ends with `mac_yield()`); it gets emitted via
+		-- `tb_emit(tb, code_<name>)` (linker symbol), NOT inlined as a macro. Including `atom_proc` here
+		-- would incorrectly emit `mac_<name>` aliases for atoms, polluting `gen/macs.h`.
+		-- See `docs/duffle_dsl_primer.md` §"mac_* aliases" for the contract.
+		if a.kind == "comp_bare" or a.kind == "comp_proc" then
+			-- Function-args lookup is meaningful for `MipsAtomComp_Proc_` components
+			-- (the macro sits inside `FI_ Slice_MipsCode ac_X(...)`); the alias expansion
+			-- discards the `ab` (atom-builder) arg the same way both forms do.
+			local args = find_function_args_for(source, a.raw_name, a.ident_pos)
 			-- Comment ownership: scan_source.lua stamps `declaration_comment` on the record by walking backward past any associated bare marker.
 			-- The pass reads `declaration_comment` directly.
 			local comment = a.declaration_comment or ""
@@ -221,7 +225,7 @@ local function project_components(source, scan)
 				body_tokens = a.body_tokens,
 				args        = args,
 				comment     = comment,
-				kind        = a.kind,  -- "comp_bare" | "comp_proc" | "atom_proc"; provenance emitter reads this.
+				kind        = a.kind,  -- "comp_bare" | "comp_proc"; provenance emitter reads this.
 				debug_skip  = a.debug_skip == true,
 			}
 		end
@@ -400,8 +404,7 @@ local function cycle_cost_rec(name, comp_by_name, latency, cache)
 end

 --- (internal) Recursive GP0 prim-buffer contribution. Count `store_word` / `store_half` / `store_byte` 
--- calls in the component body that target `R_PrimCursor` (these are the
--- RAM-side prim-buffer words the macro contributes), recursing through nested `mac_*` calls.
+--- calls in the component body that target `R_PrimCursor` (these are the RAM-side prim-buffer words the macro contributes), recursing through nested `mac_*` calls.
 --- Only `R_PrimCursor`-targeting stores count. Stores targeting other registers (e.g. `R_OtBase`, heap pointers) are not prim-buffer contributions.
 --- @param name         string
 --- @param comp_by_name table<string, Component>
@@ -484,10 +487,9 @@ end

 --- Determine the macro signature: function-args list (function form) or variadic-ignored (bare form).
 --- For `MipsAtomComp_Proc_` components, the leading `ab` (atom-builder) arg is dropped:
--- the generated `mac_<name>` macros are inline-expansion aliases for baked atoms; their bodies
--- don't reference `ab` (the builder is only consumed by the procedural `atombuilder_unroll` line
--- that `MipsAtomComp_Proc_` appends after the body). Inline callers therefore don't need to thread
--- a builder context.
+--- the generated `mac_<name>` macros are inline-expansion aliases for baked atoms; their bodies don't reference `ab`
+--- (the builder is only consumed by the procedural `atombuilder_unroll` line that `MipsAtomComp_Proc_` appends after the body).
+--- Inline callers therefore don't need to thread a builder context.
 --- @param args_str string|nil
 --- @return string
 local function signature_from_args(args_str)
@@ -544,7 +546,7 @@ local function build_component_lines(c, counts)

 	-- Marker comment: emitted once for every skipped component.
 	-- The marker is scanner-owned (declared by `atom_dbg_skip` immediately before the declaration in the source);
-	-- the components pass projects `c.debug_skip` and emits the marker as a generated comment.
+	-- This pass projects `c.debug_skip` and emits the marker as a generated comment.
 	if c.debug_skip then
 		lines[#lines + 1] = "/* atom_dbg_skip */"
 	end
@@ -578,8 +580,8 @@ end

 --- Build the boilerplate header lines (the `#ifdef INTELLISENSE_DIRECTIVES` block,
 --- the `// Auto-generated` comment, the `// Source:` line, and the self-contained `WORD_COUNT` macro definition).
--- @param dir     string       -- the absolute source directory
--- @param sources SourceFile[] -- sources contributing to this directory (for the header comment)
+--- @param dir     string       -- Absolute source directory
+--- @param sources SourceFile[] -- Sources contributing to this directory (for the header comment)
 --- @return string[]
 local function header_boilerplate(dir, sources)
 	local source_lines = { "// Directory: " .. duffle.to_absolute_path(dir) .. "/" }
@@ -610,9 +612,9 @@ end
 --- Compute the per-directory output path for `.macs.h`.
 ---  e.g. any source in `code/duffle/` produces `code/duffle/gen/macs.h` regardless of source filename.
 --- The directory name is the namespace; the filename does not repeat it.
--- @param dir string  -- the absolute source directory
--- @return string  -- the output directory
--- @return string  -- the full output path
+--- @param dir string  -- Absolute source directory
+--- @return string  -- Output directory
+--- @return string  -- Full output path
 local function compute_macs_h_path(dir)
 	local  out_dir  = dir .. "/" .. GEN_SUBDIR
 	local  out_path = out_dir .. "/" .. MACS_FILENAME
@@ -622,11 +624,11 @@ end
 --- Emit a per-directory `.macs.h` header with the aggregated `mac_X` macros + `WORD_COUNT` entries.
 --- Writes in BINARY mode so LF line endings are preserved (the git blob is LF; Windows text-mode would emit CRLF and break the byte-identical diff).
 --- @param ctx        PassCtx
--- @param dir        string                  -- the absolute source directory
--- @param sources    SourceFile[]            -- sources contributing to this directory (for the header comment)
--- @param components Component[]             -- aggregated components from all sources in this directory
--- @param counts     table<string, integer>  -- precomputed word counts (from count_all_components)
--- @return string|nil  -- path to the written file (nil if no components)
+--- @param dir        string                  -- Absolute source directory
+--- @param sources    SourceFile[]            -- Sources contributing to this directory (for the header comment)
+--- @param components Component[]             -- Aggregated components from all sources in this directory
+--- @param counts     table<string, integer>  -- Precomputed word counts (from count_all_components)
+--- @return string|nil -- Path to the written file (nil if no components)
 local function emit_component_macros_h(ctx, dir, sources, components, counts)
 	if #components == 0 then return nil end
 	local out_dir, out_path = compute_macs_h_path(dir)
@@ -665,11 +667,11 @@ local function update_canonical_word_counts(corpus, components, counts)
 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" | "atom_proc"
--- @field debug_skip boolean -- mirror of the scanner-owned `a.debug_skip`; consumers read this directly
+--- @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"  (atom_proc is NOT a component)
+--- @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").
@@ -703,9 +703,9 @@ end
 ---     `{comp_name, call_file, call_line, comp_file, comp_line, start_pos, end_pos, body_lines, debug_skip}`. `body_lines[k]`
 ---     is the k-th word's source line within the component body.
 ---
--- @param corpus table -- the corpus from `ctx.shared.corpus`
+--- @param corpus table -- From `ctx.shared.corpus`
 --- @param addrs  table -- ELF symbols keyed by atom name from `elf_dwarf.read_nm`
--- @return table[]  -- list of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
+--- @return table[]  -- List of {name, addr, size_bytes, words, entries, invocations, debug_skip?}
 local function build_atom_table(corpus, addrs)
 	-- Cross-ref: keep only atoms present in BOTH the nm symbol table AND `corpus.atoms_by_name`. Output is sorted by ascending addr.
 	local atoms_by_name = corpus.atoms_by_name or {}
@@ -834,10 +834,10 @@ end
 --- This is intentional: silently falling back to a hardcoded GPR would mask the missing opt-in.
 ---
 --- Pre-tokenized: `body_tokens` is the scan-source pass's pre-split list of top-level statements (each entry is a single `load_*` call or other statement).
--- @param body_tokens table[]  -- the atom's pre-tokenized body statements (from atom.body_tokens)
--- @param binds_name  string   -- expected Binds_X name (skip pairs with mismatching binds)
--- @param registries  table    -- merged registries from collect_per_source_registries
--- @return table[]             -- list of {reg = <MIPS index>, field = <field name>}
+--- @param body_tokens table[]  -- The atom's pre-tokenized body statements (from atom.body_tokens)
+--- @param binds_name  string   -- Expected Binds_X name (skip pairs with mismatching binds)
+--- @param registries  table    -- Merged registries from collect_per_source_registries
+--- @return table[]             -- List of {reg = <MIPS index>, field = <field name>}
 local function parse_body_load_pairs(body_tokens, binds_name, registries)
 	local pairs = {}
 	local reg_index_by_name = (registries and registries.register_alias_registry) or {}
@@ -880,9 +880,9 @@ end
 --- The piece chain uses (DW_OP_regN, DW_OP_piece, ULEB128(field_size)).
 ---
 --- Binds fields come from `scan.binds`; the per-source `scan.binds[i].fields` already carries the typed-field record after the scan-source generalization.
--- @param corpus     table   -- the corpus from `ctx.shared.corpus`
--- @param atom_table table[] -- the cross-ref'd atom table from build_atom_table
--- @param registries table   -- merged registries from collect_per_source_registries
+--- @param corpus     table   -- From `ctx.shared.corpus`
+--- @param atom_table table[] -- Cross-ref'd atom table from build_atom_table
+--- @param registries table   -- Merged registries from collect_per_source_registries
 --- @return table, table      -- (rbind_atoms, rbind_structs)
 local function parse_rbind_atoms(corpus, atom_table, registries)
 	registries = registries or {}
@@ -944,7 +944,7 @@ local function parse_rbind_atoms(corpus, atom_table, registries)
 						binds     = ai.binds,
 						fields    = struct.fields, -- {name, offset} from scan.binds
 						bytes     = struct.bytes,
-						regs      = pairs,         -- ordered list of {reg, field}
+						regs      = pairs,         -- Ordered list of {reg, field}
 						info_line = ai.info_line,
 					}
 					table.insert(struct.atom_names, atom_name)
@@ -1780,7 +1780,8 @@ local function build_inserted_children(main_cu_offset, main_cu_end_excl, atom_ta
 	emit(uleb128(ABBREV_TYPED_VIEW_POINTER))  -- DW_TAG_pointer_type (abbrev 110; NOT 9; void chain target)
 	emit(elf_dwarf.write_u32_le(ref4_of(void_chain_offset)))  -- 4-byte ref4: points at the void base_type's tag byte
 	-- type_chain_offsets["void|1"] is what step (f) of the per-RR_<R_Name> chain looks up.
-	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type. The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
+	type_chain_offsets["void|1"] = void_chain_offset  -- both the base_type offset and the pointer_type are emitted consecutively; the OUTERMOST is the pointer_type.
+	-- The variable's DW_AT_type must reference the pointer_type, not the base_type. Patch below.
 	-- Capture the pointer_type's offset (the last-thing-emitted DIE start) and overwrite the lookup.
 	-- The pointer_type was emitted as: uleb(9) (1 byte) + 4-byte ref4 = 5 bytes. Its tag byte is at void_chain_offset + 8 (the base_type's 8 bytes: 1 tag + 5 name + 1 byte_size + 1 encoding).
 	local ptr_void_offset = void_chain_offset + 8
@@ -4,9 +4,17 @@
 --- for `MipsAtom_(name)` and `MipsCode code_<name>` declarations, computes the word offset
 --- from each `atom_offset(F, T)` marker to its target `atom_label(T)` declaration, and emits
 --- `gen/offsets.h` with one `#define _atom_offset_F_T = N` per branch.
+---
 --- Per-directory aggregation: every source in the same directory contributes to the same `gen/offsets.h`.
 --- The directory itself is the namespace; the filename does not repeat the module name.
 ---
+--- (Task 12.16 note: atom-namespaced enum names — e.g., `atom_offset__normalize_v3s4__srav_path__aligned_done` —
+--- were considered to prevent cross-atom label collisions, but the C-side `atom_offset(F, T)` macro in
+--- `code/duffle/dsl.atom.h` doesn't know the current atom_name at expansion time, so any namespacing
+--- on the metaprogram side breaks the C build. Reverted. The C-side would need a per-atom
+--- `CURRENT_ATOM` #define (set by `MipsAtom_`/`MipsAtom_Proc_` macros) plus an updated `atom_offset`
+--- macro that uses it. That's a coordinated refactor — deferred to a future track.)
+---
 --- The offset is `target_word - branch_word - 1` (the standard MIPS branch-immediate encoding: branch_offset = relative_pc_in_words - 1).

 -- ════════════════════════════════════════════════════════════════════════════
@@ -39,8 +39,8 @@
 --- `── Info` section renders finding-level info between `── Warnings` and the per-atom cycle counts.
 ---
 --- The structural handshake checks (`mac_yield_uniformity`, `hazard_nop_use`, `control_transfer_delay_slot_use`) skip atoms/components with `debug_skip == true`.
--- The `atom_dbg_skip` marker designates runtime-helper declarations whose structure is fixed by the tape runtime (e.g. `tape_exit`, `ac_yield`).
--- Flagging them as "missing mac_yield" or "BD slot is redundant" is signal noise, not a logic failure. 
+--- `atom_dbg_skip` marker designates runtime-helper declarations whose structure is fixed by the tape runtime (e.g. `tape_exit`, `ac_yield`).
+--- Flagging them as "missing mac_yield" or "BD slot is redundant".
 --- Other checks (transfer_hazards, gpu_portstore_shape, abi_handoff, enum_alias_membership, …) still apply to debug_skip declarations because real hazards / typos can still surface in them.
 ---
 --- The orchestrator (`ps1_meta.lua`) wires this module in via the PASSES table:
@@ -452,7 +452,7 @@ local function is_cop2_consumer_of(consumer_event, destination, producer_rel)
 end

 -- True iff `consumer_event` reads the GPR operand at any position the destination register occupies.
-- The read-position lookup consults `duffle.OPERAND_READ_POSITIONS` for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
+-- read_pos lookup consults `duffle.OPERAND_READ_POSITIONS` for the consumer's encoder and walks each `args[pos]` to find an operand-equal match.
 local function is_gpr_consumer_of(consumer_event, destination)
 	local consumer_token = consumer_event.encoder or consumer_event.ident
 	local read_pos       = duffle.OPERAND_READ_POSITIONS or {}
Author	SHA1	Message	Date
ed	a2d79d65eb	amazing bug	2026-08-11 01:25:40 -04:00
ed	bebcc6a585	wip: going to incremnetally test this.	2026-08-11 01:25:09 -04:00
ed	ece21ed368	mark current crashing path.	2026-08-10 23:29:38 -04:00
ed	144c605ad8	some more review. not working still.	2026-08-10 23:04:43 -04:00
ed	4afd1af0fd	started to review this...	2026-08-10 19:53:34 -04:00