mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-26 21:36:06 -04:00
Drew T
e79d030499
End-to-end proof of the §8b lazy bank composition on a real cracked jr core: lazy isolate -> jtbl_carve into the isolated subseg -> C body -> whole-binary gate -> d19c9580 BYTE-IDENTICAL; R22 clean-fleet 136/136. - func_80182268 (31-ins jr, ov_SC01_077_after) MATCHED first try: shared-tail fallthrough (jtbl cases 3+7 enter case 4's tail) + the u16-shift sign-extend idiom ((s8)(*(u16*)(p+0x70) >> 8) -> lhu/sll16/sra24). Carve collided with the committed func_801734BC carve -> lazy isolation fired exactly as designed. - R14 FINDING (cookbook §41d): the Phase-17 canonical convention "void->s32 return is byte-neutral (§3a-1)" is FALSE for a void body with no `return` — it costs ONE extra instruction. canon_sig_reconcile applies it unconditionally, so it turned a perfect 31-ins MATCH into 32 ins. That extra word made the isolated object's .text 4 B long, shifting EVERY data symbol +4 -> ~271k differing bytes, image +5 B. match_one said MATCH; only the whole-binary gate caught it (G3/P9). - FIX (generalizes the §19 sig_unify lesson): every recovery pass is a FALLBACK, never unconditional. jtbl_family_bank now gates RAW first, reconciled only on failure. - 136/136 byte-identical from a clean tree (R22); 0 NON_MATCHING (G4).