mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-27 05:56:00 -04:00
Drew T
0acd4cc5d4
feat(phase-29): the family campaign — +18.6pp instr in 25 sessions; engine spent on evidence; roadmap v2 (v1.28.0) - T1 diff_regions.py RESOLVED the swing number: (a) TOOLING — the "~3% h_seq ceiling" was an -O0 compile-flag artifact (REGALLOC 0/106); the third structural wall to resolve to our own tooling. - The campaign: crack waves + propagate-behind-each-crack + the §20 broad type-lift (154 types) + the jtbl carve waves + the SC07-region sweeps. SESSION-25 alone: 2,713 members, +100,572 ins, +1,620 unique fns, every named blocker closed — NOT ONE a compiler wall. - Doctrine shifts (decision-log): fresh cracks are the only lever that moves distinct-code; the propagation cap gates de-duplication, not coverage; the wave bottleneck is INTEGRATION (~92% byte-correct, ~27% bank) -> the tiered T0/T1 blast-radius recovery driver (14/36 = 39% measured); gate_stage call-site-casts are the integration spine (bulk header edits BREAK builds). - T98 characterised the residue: 80 families / 960 members / 173 distinct (29 all-STRUCT by-design + 51 gate-failing, 3 sweeps at 0) -> T99 burn-down (derived from digest git history, R33): final-four-day decay +2.7 -> +0.3pp/day. Closed on the measurement, gate-2 confirmed. - roadmap-to-100 v2: P30 Recovery & Concentration -> P31 Scope-Complete + Main & Resident -> P32 Behemoths+Walls -> P33 Verify+Flip. member_adapt + family-adapt fine-tune VACATED. - Honesty ledger: 7 recorded errors + the month-old shared byte-gate default defect (found+fixed); the -O0 Arm-A splat wall named and carried as P30 T2 instrument work, never force-banked. - cookbook §54-§121 (68 sections); ~15 decision-log entries; calibration re-measured; R37 proposed.