Files
BFM-decomp/config
Drew T d593ee20fe feat(phase-18): Step-3b calibration — §17-toolkit wave banks 9 reach-134 matches (55.58->55.94%)
- upgraded harvest wave (.run/harvest_wave_3b.js) teaches the §17 TOOLKIT: triage the
  match_one diff -> reconstruct / array-decay / register-pins+barrier (instead of stubbing
  the regalloc tail). Calibration: 16 fresh tractable reach-134 targets.
- whole-binary close-rate 9/16 = 56% (vs Phase-17 baseline 33%): 2 via register pins
  (func_8012CB64, func_801468C8), 2 via array-decay (func_80149544, func_80149584),
  5 via clean reconstruction. Each reach-134 -> propagated x134 (config/dedup.us.yaml +9).
- FINDING (R16 flywheel): the match_one->gate gap (12->6 then ->9) is the §3a call-site
  CASTS the agents skipped + sig_unify narrowing the extern to canonical -> conflict.
  Recovered 3 by call-site arity-cast / void-return cast (codegen-neutral). Step-1 prompt
  must teach call-site casts + re-validate after sig_unify.
- make check-all 136/136 byte-identical (R22); dedup-check 1432 validated, 0 failed;
  0 NON_MATCHING in any default build (G4).
2026-06-20 13:40:11 -06:00
..