mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-26 21:36:06 -04:00
Drew T
eb2e72328b
The dedup_extend SC07 lane banked 0/36. Captured every failure's own compiler error rather than guessing: 5 PLUMBING (declaration conflicts) + 4 DIFF per binary, and the PLUMBING set reduces to 4 distinct symbols repeated across all four binaries. Fixed at the lowest available blast radius, each byte-neutral BY CONSTRUCTION and proven by R22: - func_80146C3C (8 of 36): T2, but 1 token. DEFINE_func_8016A08C / DEFINE_func_8016A1CC declared it `(void)` while the SC07 TUs declare `(u8 *a0)`. Both macros already CAST at the call site, so the prototype is codegen-irrelevant -> relax to `()` (no-prototype). Measured first: all 4,020 fleet decls are (void)/()/(u8*)/(u8 *a0) — no default-promotion param anywhere, so `()` is compatible with every one of them (the gcc-2.7.2 `()` dead-end needs a promoting param; there is none). - func_8014F4C0 (4 of 36, self-axis), D_80126CC4 (3), func_8012E5CC (1): T1 binary-local. Conformed the SC07 decls to the fleet-canonical form. Every one has ZERO uncast uses in its TU (verified per file), so the decls were pure splat boilerplate with no codegen effect. Deliberately NOT touched: engine_core.h beyond those 2 lines. The §29 law is about BULK header edits (the 2,046-file sed that still didn't build); 2 targeted macro lines with a measured compatibility argument is not that. R22 clean-fleet: make clean && extract-all && check-all -> 140 passed, 0 failed of 140. Enabling change only; no function banked in this commit. Method note (mine, recorded): I read `corpus.stubs` while `make extract-all` was mid-flight and got garbage — R32's coverage assertion refused to answer instead of returning a wrong stub set. A measurement taken during a rebuild is not a measurement (S27's law, re-earned). Also: `nohup CMD &` inside a backgrounded Bash call makes the harness signal completion of the WRAPPER, not the campaign — the R22 "finished" at 63/140. Waited on `pgrep -x make` instead.