mirror/BFM-decomp
1
0
Fork 0
mirror of https://github.com/Druthulu/BFM-decomp synced 2026-09-27 05:56:00 -04:00
Code Issues Packages Projects Releases Wiki Activity

docs(phase-20): log consolidated state — cached ≤90-ins tail byte-exhausted; T5 is the lever

Browse Source
This commit is contained in:
Drew T
2026-06-21 09:31:20 -06:00
parent a978081b2d
commit d5ddbc587d
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
phase-ends/CURRENT_PHASE.md
+1
View File
@@ -48,6 +48,7 @@ None.
## Progress log
- 2026-06-21 (fresh session, Max): Built `tools/cast_call_sites.py` (the cap tool, §17a-1 callee-cast). Validated on T6 batch-1's 33 gate-fails → 6 byte-identical recoveries (5 propagated ×134 + func_8015F948 local), fleet 58.63%→58.82%, check-all 136/136, dedup 1501→1508. Diagnosed the residual exhaustively (R14): the §20 "~33" projection was wrong — the batch is dominated by the DEF-side loose-typing wall (caller-side blocked by INCLUDE_ASM-declares-nothing), not callee-conflicts. Captured in cookbook §20 RESOLUTION + SETUP inventory (R16/R21/R30). Local commit-checkpoint (Drew pushes, R6).
- 2026-06-21 (Drew order ① -O0 rollout): INVALID PREMISE (R14, byte-proven). The 6 matched -O0 cluster fns are NOT reach-134 byte-identical — only 1/134 overlays shares the bytes; they reference per-overlay data addresses → overlay-LOCAL, not ×134-shareable. The Phase-18/19 "reach-134" was a mislabel. Built+byte-validated the §18 split infra (rollout tool + Makefile %_o0.o pattern + PROVIDE .ld), gate exposed the wall, reverted. -O0 rollout = ×1 per-overlay-unique work (~0.2-0.3%), SKIP. Cookbook §18 corrected. Fleet unchanged 58.82%.
- 2026-06-21 (Drew order ② data-cast): MOOT (R14, byte-proven). 0 drafts have an extern struct/union data conflict; the lone "DATA-conflict" was a typedef-redef (Blk20) masking the def-sig wall. data-cast = code for 0 cases → NOT built. Cookbook §20 corrected. **CONSOLIDATED STATE:** the reach-134 ≤90-ins CACHED tail is byte-EXHAUSTED of cheap recoveries — cap tool banked the callee-conflict fraction (6, +0.19%); the rest are the def-side loose-typing wall + overlay-local + genuine codegen quirks (all byte-proven non-mechanical). Remaining levers: **T5 fresh waves** (uncached reach-134 via Ghidra-C regen — needs MCP cycle + Ultracode, highest potential) · T3a %lo (value dropped — targets are overlay-local; reusable idiom only) · the giants (byte-weighted lever, deep/hard). Fleet 58.82%, 136/136, 0 NON_MATCHING. Reported to Drew for the steer.
- 2026-06-20: Session start; read full load order (PROJECT_CONTEXT + 19 PhaseEnds). Phase 19 closed at 58.00%. Brainstormed strategy with Drew (ML/Wine/debug-symbols/gcc-papermario threads → honest verdicts; the wall is propagation-cap + ~3 uncracked gcc classes, not comprehension). Max set, plan mode. 3 Explore agents mapped the propagation/wave/quirk subsystems; verified the type-lift crux firsthand. Reframed (R14/P9): class-cracking = ~3 bounded spikes. Drew chose **class-crack first** + **skip permuter grind**. Plan approved (gate 1). CURRENT_PHASE + task list created (R28).
## Notes / decisions