mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-27 22:45:39 -04:00
Drew T
bff03dbdb5
The 500-fn calibration banked 0/222 across the binary rotation. Root-caused (R14, by reading the code + the run's own backlog — resolving a flat contradiction between two scout agents) to TWO independent bugs in lora_grind's use of gate_stage.run_gate, NOT model quality: - Bug A: good_sha() passed the sha1sum line "<sha> <name>" vs harvest_verify's bare sha1() -> 0 banks for EVERY binary incl. 077 (so the "0/12" was a bug artifact, not an exhausted tail) - Bug B: the gate call left src/asm/out at the hardcoded ov_SC01_077 defaults -> non-077 drafts dropped at the 077 stub-filter, silently (and the asm mis-resolution contaminated the backlog near-miss classification) Fix (tools/gate_stage.py): run_gate resolves src/asm/out/good_sha from `binary` when unset (binary-agnostic, no silent ov_SC01_077 default an overlay inherits; good_sha bare-hash normalized) + a loud negative-control guard (0-overlap binary/src mismatch warns, so a 0 can never again masquerade as 'nothing matched'). tools/lora_grind.good_sha fixed at source. Byte-neutral: make check-all 136/136. Proof: ov_SC01_000 spot-run banked 7/15 (47%) byte-identical (@commit:0322); reach-2 func_8017CE24 propagated x2. ROI finding: 6/7 banks are reach-1 (overlay-unique) -> broad rotation is high bank-RATE / low fleet-% ROI; the fleet lever is reach>=2 targeting (T9) + corpus-v3 (T8). Backlog now correctly classified (4x close=1 = grinder fuel). - docs/gen2-mips-matching-model.md: T7 RESULT section - phase-ends/CURRENT_PHASE.md: T7 done; next = T8 corpus-v3 / T9 reach>=2 selection
115 lines
5.1 KiB
Plaintext
115 lines
5.1 KiB
Plaintext
# ============================================================
|
|
# ROM-content policy — see PROJECT_CONTEXT.md rule H1
|
|
#
|
|
# While the repo is PRIVATE, ROM-derived content MAY be committed
|
|
# (asm/, assets/, extracted/, decompressed .CD output). Compliance is
|
|
# handled by a pre-public history scrub + a rom->decoder regeneration
|
|
# tool. Two things stay ignored regardless:
|
|
# 1. the raw multi-GB disc dump (GitHub rejects >100 MB files; it is
|
|
# trivially reproducible from the disc), and
|
|
# 2. pure regenerated build churn / large binary caches.
|
|
# Re-add asm/ assets/ extracted/ to this list before going public.
|
|
# ============================================================
|
|
|
|
# Raw disc dump — never committed (size + reproducible)
|
|
/Brave Fencer Musashi (USA)/
|
|
/disks/
|
|
|
|
# RAM-dump corpus (dumps/*.bin) is COMMITTED while private (H1 relaxed; static, not regenerable;
|
|
# off-machine backup). Joins the pre-public ROM-content scrub. INDEX.md documents each.
|
|
|
|
# Project-local runtime scratch (MCP server log + stop sentinel) — replaces /tmp use
|
|
/.run/
|
|
|
|
# Regenerated build output (deterministic churn)
|
|
/build/
|
|
/expected/
|
|
|
|
# Regenerated splat outputs — recreated by `make extract` from the committed
|
|
# config/splat.us.exe.yaml + config/symbols.us.txt + the EXE. The durable RE record
|
|
# is the symbol file, not these (build scaffolding, not committed work).
|
|
/asm/
|
|
/assets/
|
|
/include/include_asm.h
|
|
/include/macro.inc
|
|
/include/labels.inc
|
|
/include/gte_macros.inc
|
|
/undefined_syms_auto.txt
|
|
/undefined_funcs_auto.txt
|
|
|
|
# Ghidra project — COMMITTED as the irreplaceable RE-work backup (private repo; rule R20,
|
|
# Drew 2026-06-15). NOT regenerable by hand (annotations/types/structs beyond symbols.us.txt).
|
|
# Track the project; exclude ONLY live-session transients so a running/restarted MCP server
|
|
# doesn't dirty the tree. Re-commit after significant RE + a Ghidra save. (Ghidra pinned at
|
|
# 12.1 -> the committed .rep restores directly.) Joins the pre-public ROM-content scrub list.
|
|
/ghidra/*.lock
|
|
/ghidra/**/*.lock
|
|
/ghidra/*.lock~
|
|
/ghidra/**/*.lock~
|
|
/ghidra/**/tmp*.ps
|
|
|
|
# Toolchain downloads & caches (not source)
|
|
/.venv/
|
|
__pycache__/
|
|
# Compiler binaries are ignored (regenerable downloads); the committed *.sha256
|
|
# checksum record is re-included so a fresh clone can verify its download (sotn
|
|
# pattern). Contents-exclude form (not /tools/bin/) so the !re-include can apply.
|
|
/tools/bin/*
|
|
!/tools/bin/*.sha256
|
|
# PsyQ SDK — Sony-copyrighted (stays PRIVATE; excluded from the public mirror, like ROM content).
|
|
# Back up the SMALL gathered working artifacts (lib40/ *.LIB, lib40_elf/ *.a, the PsyQ *.EXE tools,
|
|
# psyq-obj-parser, conv47/) — hard to re-source if the upstream archives go down (rule R20). Ignore
|
|
# ONLY the >100MB raw source archives (GitHub rejects them; the working libs are already extracted
|
|
# from them — the disc-dump-equivalent, re-sourceable from redump if ever truly needed).
|
|
/tools/psyq/psyq40usa.zip
|
|
/tools/psyq/*.bin
|
|
/tools/psyq/*.cue
|
|
/tools/pcsx-redux/
|
|
*.tar.gz
|
|
# ...EXCEPT the old-gcc cc1 compiler tarballs — back them up (gathered compilers; small; verified by
|
|
# tools/bin/CHECKSUMS.sha256). Extracted cc1 binaries stay ignored (regenerated by `tar xzf`, SETUP §4.7). R20.
|
|
!/tools/bin/*.tar.gz
|
|
|
|
# Cross-validation scratch: compiled brave oracle + its ROM-derived .dir output
|
|
/tools/brave-CUE/_build/
|
|
|
|
# Phase-18 compiler-quirk research reference repos (gitignored; re-clonable from GitHub, exact SHA
|
|
# recorded in docs/SETUP.md). pmret/gcc-papermario = the gcc-2.7.2 source (read the reload/CSE/
|
|
# scheduling/regalloc passes) + ladysilverberg/xenogears-decomp = our exact compiler's decomp (mine
|
|
# transferable idioms). Research reference, not a build dependency — not committed (avoids repo bloat).
|
|
/tools/reference/
|
|
|
|
# Editor / OS noise
|
|
.vscode/
|
|
Thumbs.db
|
|
|
|
# Asset/EXE extraction: commit the tooling + manifest + the retail EXE, NOT the
|
|
# bulk payloads (deterministically regenerable from the dump; ~hundreds of MB
|
|
# incl. FMV) and NOT the prototype EXEs. Per-ROM subfolders under extracted/:
|
|
# extracted/retail/ — retail SLUS-00726 disc extraction (3 files committed)
|
|
# extracted/proto/ — prototype main EXEs (ROM-derived; fully ignored)
|
|
# Nested allowlist: ignore everything under extracted/, re-include only the 3
|
|
# retail committed files (EXE + manifest + manifest hash — always safe; sotn
|
|
# precedent). This keeps the "zero ROM-derived bulk staged" guarantee.
|
|
/extracted/*
|
|
!/extracted/retail/
|
|
/extracted/retail/*
|
|
!/extracted/retail/SLUS_007.26
|
|
!/extracted/retail/manifest.jsonl
|
|
!/extracted/retail/manifest.sha1
|
|
|
|
# NOTE (H1 relaxed, private phase): assets/ and other ROM-derived content stay
|
|
# committable while the repo is private. asm/ + the splat-generated include macros +
|
|
# undefined_*_auto.txt are IGNORED above: `make extract` regenerates them deterministically
|
|
# from the committed yaml + symbols, so they are build scaffolding, not durable RE work.
|
|
# (Phase 5 refined the earlier "commit asm/" stance once the regeneration flow existed;
|
|
# the durable record is config/symbols.us.txt.)
|
|
|
|
# Gen2 matching-model corpus + weights (R20: large artifacts out of git)
|
|
datasets/
|
|
models/
|
|
*.gguf
|
|
*.safetensors
|
|
.venv-train/
|
|
unsloth_compiled_cache/
|