mirror/BFM-decomp
1
0
Fork 0
mirror of https://github.com/Druthulu/BFM-decomp synced 2026-09-28 06:49:47 -04:00
Code Issues Packages Projects Releases Wiki Activity
368 Commits 1 Branch 0 Tags
39ccbee46c5bd2d056dbfe76c8cb7cfba7e3d997
Commit Graph

105 Commits

Author SHA1 Message Date
Drew T fdd3bf3c28 feat(phase-23): inject-extern lever on _a.c — +3 reach-134 fns (fleet 64.09→64.16%) 
- inject_capped_externs --src-file ov_SC01_077_a.c freed 3 more reach-134 fns
  (0x8012A018/0x8012BA10/0x8012FDA8); 077 byte-identical; dedup_propagate
  registered 2 clean ×134 groups + 2 ×133 (ov_SC01_000 straggler kept ×1)
- 1685->1687 shared groups, check-all 136/136, dedup-check 0 failed
- injection lever total this session: 63.82% -> 64.16% (+0.34%, ~0 LLM tokens)
 2026-06-30 11:20:13 -06:00
Drew T 205fc202e9 feat(phase-23): macro-extern-injection lever — 8 reach-134 fns freed ×134 (fleet 63.82→64.09%) 
- tools/inject_capped_externs.py (new): source each propagation-capped reach≥2
  inline match's externs from ov_SC01_077's own file-scope decls, inject them
  block-scope so the lifted macro is self-contained (the Phase-20-backlog
  macro-extern-injection lever; byte-neutral, fail-safe via compiles_standalone)
- 8 reach-134 fns freed from ov_SC01_077.c, 077 byte-identical; dedup_propagate
  --auto-from registered 7 new shared groups ×134 (1678->1685), check-all 136/136
- the LLM reach≥2 measurement run already banked +7 (v3 banks reach≥2: 6/15 in 005
  vs v2's 0/15) — committed earlier (commit:0359, commit:0360)
- cookbook §28d + SETUP tooling inventory (R16/R21/R30)
- fleet 63.82% -> 64.09%, 0 NON_MATCHING, dedup-check 1685/0 failed
 2026-06-30 11:15:11 -06:00
Drew T ac968645f2 feat(phase-23): lora-grind gate — +14 fns x0 propagated (fleet 63.82%) 2026-06-30 07:31:49 -06:00
Drew T 3d4f7778fa feat(phase-23): lora-grind gate — +18 fns x0 propagated (fleet 63.79%) 2026-06-30 06:46:02 -06:00
Drew T 714760a1a0 feat(phase-23): lora-grind gate — +15 fns x0 propagated (fleet 63.77%) 2026-06-30 06:27:19 -06:00
Drew T 37cc7d8dd6 feat(phase-23): lora-grind gate — +12 fns x0 propagated (fleet 63.76%) 2026-06-30 06:08:18 -06:00
Drew T c7775eb8b5 feat(phase-23): lora-grind gate — +15 fns x0 propagated (fleet 63.74%) 2026-06-30 05:36:16 -06:00
Drew T 238a1c00d8 feat(phase-23): lora-grind gate — +17 fns x0 propagated (fleet 63.73%) 2026-06-30 05:18:19 -06:00
Drew T 758ab7de65 feat(phase-23): lora-grind gate — +14 fns x0 propagated (fleet 63.72%) 2026-06-30 05:00:13 -06:00
Drew T bff03dbdb5 fix(phase-23): T7 — lora_grind broad-rotation banking (two gate bugs); ov_SC01_000 7/15 
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
 2026-06-29 23:44:12 -06:00
Drew T c1230fa6ea feat(phase-22): T2 close=0 recovery batch — 14 fns ×134 (fleet 63.29%->63.66%) 
- R14 corrects cookbook §26 ('close=0 recovery exhausted'): recover_giant + the
  whole-binary gate still banks a ~15-20%% tail. Banked 7 close=0 reach-134 fns
  (func_80156ECC/80147E44/8015ADB0/801661CC/80166054/8012CFA8/8012A62C); dedup_propagate
  --auto-from swept those + 7 pre-existing inline-matched bonus fns ×134 (14 total)
- the other ~33 close=0 are the genuine DEF-side loose-typing / masked-residual wall
- func_80132784 HARD-DEFER (hoist-vs-remat + register-lifetime-reuse not C-expressible)
- §28c distilled: close=0-not-exhausted recipe + the dedup_propagate registry-skip
  recovery (register via append_groups; bytes were already check-all-correct)
- make check-all 136/136 byte-identical (R22); dedup 1619->1633 validated/0 failed;
  0 NON_MATCHING (G4)
 2026-06-26 04:14:06 -06:00
Drew T d789d9c242 feat(phase-22): T2 — giant func_80163C2C banked ×134 via Blk16/Buf32 type-lift (fleet 63.25%->63.29%) 
- second struct-walled close=0 giant (167 ins): lift Blk16(u32x4)+Buf32(2xBlk16) to
  src/shared/engine_types.h, remove the Blk16 dupes from ov_SC01_077.c + the -O0 split
  ov_SC01_077_a.c (byte-neutral), recover_giant, harvest_verify MATCH, dedup_propagate
  ×134 (+22378 ins byte-weighted)
- struct-walled trio done: func_80156B74 + func_80163C2C banked; func_8014F74C is a
  masked residual (grinder fuel, not type-liftable)
- make check-all 136/136 byte-identical (R22); dedup 1618->1619 validated/0 failed;
  0 NON_MATCHING (G4)
 2026-06-26 03:51:11 -06:00
Drew T b5a1fe6dab feat(phase-22): T2 — giant func_80156B74 banked ×134 via engine_types.h type-lift (fleet 63.22%->63.25%) 
- struct-walled close=0 giant (cookbook §28 case #3) is bankable: lift the S8/B8
  typedefs from ov_SC01_077.c's prelude to src/shared/engine_types.h (shared via
  engine_core.h; byte-neutral), recover_giant block-scopes externs, whole-binary
  harvest_verify MATCH, dedup_propagate ×134 (214 ins, +28676 ins byte-weighted)
- cookbook §28b: the type-lift recipe + the serialize-make-jobs gotcha (concurrent
  make corrupted an .o and dropped the registry write; re-ran dedup_propagate idempotent)
- make check-all 136/136 byte-identical (R22); dedup 1617->1618 validated/0 failed;
  0 NON_MATCHING (G4)
 2026-06-26 03:45:35 -06:00
Drew T 75dc869099 feat(phase-22): T1 — calibrate hand-loop; func_8015126C banked x134 (fleet 63.18%->63.22%) 
- func_8015126C (254-ins top-5 giant): cracked the §25 coalescing residual with a
  register-pin idiom (cmp79 __asm__("$4") + lazy in-&& assign -> $a0, not hoisted)
  + the canonical-extern recovery (rewrite each engine_core.h-callee extern to its
  canonical def-sig, move ALL externs block-scope) so it banks whole-binary, not just
  match_one. dedup_propagate x134; make check-all 136/136 byte-identical; dedup-check
  1617 validated / 0 failed
- proves the Phase-22 loop end-to-end (hand-crack -> bank -> x134 propagate -> verify)
- CURRENT_PHASE.md: T0/T1 closed; the close=0-giant wall = draft self-contained externs
  conflicting with engine_core.h canonical sigs, deterministically fixable (T4 lever)
 2026-06-26 02:54:39 -06:00
Drew T 1f4aac62fb docs(phase-21): cont.8 — collaboration sweep (clean verdicts) + 62 PsyQ names + SETUP -G fix 
- reach-1 idiom loop: +16 ×1 matches (commit:0289, commit:0291) + chained-store-fanout idiom
  (commit:0290); giants pivot 0 auto-banks (2× = the byte-proven wall) -> 6 ranked
  hand-finish near-misses in backlog (2 close: func_8015126C close=2, func_8014EE14 close=6)
- collaboration sweep CLOSED with clean verdicts (5 scouts): parity research (we're SOTA;
  no neural/GPU/diffusion shortcut), decomp.dev survey (4 gcc-2.7.2-psx twins; VS=CC0 the
  clean pool partner), census (no Square symbols; Croc=SDK-naming only), VS corpus (76
  byte-verified pairs, low-priority asset), dedup probe (CLEAN NEGATIVE — BFM shares only
  PsyQ with sisters, AKAO refuted, 0 engine code) -> no external shortcut to the gcc wall
- config/symbols.us.txt: +62 PsyQ library/BIOS names from cross-project byte-signature
  dedup (xdedup vs VS/Xenogears); byte-safe (main still 143dbb89, R22-verified)
- docs/SETUP.md §5.3: corrected stale "-G TBD" marker -> RESOLVED Phase 5 (-G0) (R21)
- docs/backlog.md: giant near-misses logged with true worker closeness (cont.7d-safe)
- CURRENT_PHASE.md cont.8: session log + 3 parked R&D items + process lessons
- fleet 63.18%, 136/136 byte-identical, 0 NON_MATCHING
 2026-06-25 23:56:12 -06:00
Drew T e5ae5958d5 feat(phase-21): recover-a gate — +1 fns x1 propagated (fleet 63.17%) 2026-06-25 00:57:32 -06:00
Drew T 1c1b75026a feat(phase-21): worker gate — +7 fns x4 propagated (fleet 63.13%) 2026-06-24 23:16:18 -06:00
Drew T 3943c34ac4 feat(phase-21): worker gate — +6 fns x4 propagated (fleet 62.98%) 2026-06-24 22:48:09 -06:00
Drew T 05e98070ba feat(phase-21): drop-straggler propagation banks 3 _a fns x134 (fleet 216104 / 344010 = 62.82%) 
dedup_propagate now isolates+drops cross-overlay stragglers instead of all-or-nothing
revert. Wave-2: 2 loose-typing stragglers (func_8012E5CC/func_8012C750 compile-fail
ov_SC01_000) auto-dropped + kept x1; 3 clean fns propagated x134.
 2026-06-24 22:09:47 -06:00
Drew T 1e3410ffc1 feat(phase-21): propagate 10 _a-wave fns x134 — _a vein is a real x134 lever (fleet 215699 / 344010 = 62.70%) 2026-06-24 20:35:19 -06:00
Drew T ccd61d6528 feat(phase-21): worker gate — +1 fns x1 propagated (fleet 62.31%) 2026-06-23 19:35:21 -06:00
Drew T b5eb1718af feat(phase-21): grinder gate — +1 fns x1 propagated (fleet 62.27%) 2026-06-23 14:15:31 -06:00
Drew T 3074584982 feat(phase-21): grinder gate — +1 fns x1 propagated (fleet 62.23%) 2026-06-23 12:45:22 -06:00
Drew T c2bad12d6c feat(phase-21): worker gate — +13 fns x12 propagated (fleet 62.19%) 2026-06-23 12:39:39 -06:00
Drew T 296d834467 feat(phase-21): worker gate — +13 fns x13 propagated (fleet 61.73%) 2026-06-23 12:18:12 -06:00
Drew T 3299d7f1a2 feat(phase-21): worker gate — +1 fns x1 propagated (fleet 61.16%) 2026-06-22 10:51:47 -06:00
Drew T 382a559502 feat(phase-21): worker gate — +2 fns x2 propagated (fleet 61.12%) 2026-06-22 09:27:29 -06:00
Drew T 72e71811d3 feat(phase-21): grinder gate — +1 fns x1 propagated (fleet 61.04%) 2026-06-22 08:54:00 -06:00
Drew T b1272109c0 feat(phase-21): worker gate — +3 fns x2 propagated (fleet 61.0%) 2026-06-22 08:51:03 -06:00
Drew T 9e9278d4a8 feat(phase-21): grinder gate — +1 fns x1 propagated (fleet 60.93%) 2026-06-22 05:31:06 -06:00
Drew T c63a41b1a8 feat(phase-21): worker gate — +1 fns x1 propagated (fleet 60.89%) 2026-06-22 05:07:05 -06:00
Drew T 43a356b3fb feat(phase-21): worker gate — +6 fns x5 propagated (fleet 60.85%) 2026-06-22 04:25:30 -06:00
Drew T 06496b5238 feat(phase-21): worker gate — +4 fns x4 propagated (fleet 60.65%) 2026-06-22 03:30:16 -06:00
Drew T 33e21d3050 feat(phase-21): worker gate — +3 fns x3 propagated (fleet 60.5%) 2026-06-22 02:50:42 -06:00
Drew T 77644039b8 feat(phase-21): worker gate — +6 fns x6 propagated (fleet 60.38%) 2026-06-22 02:10:22 -06:00
Drew T 7d46a6056e feat(phase-21): worker gate — +4 fns x4 propagated (fleet 60.15%) 2026-06-22 01:34:03 -06:00
Drew T 8a6db99b85 feat(phase-21): worker gate — +3 fns x3 propagated (fleet 59.99%) 2026-06-22 00:53:33 -06:00
Drew T 1332f43d55 feat(phase-21): worker gate — +4 fns x4 propagated (fleet 59.87%) 2026-06-22 00:05:35 -06:00
Drew T 28767a6a07 feat(phase-21): worker gate — +1 fns x1 propagated (fleet 59.72%) 2026-06-21 23:26:10 -06:00
Drew T a1b29eac8a feat(phase-21): recovery gate — +2 fns x2 propagated (fleet 59.68%) 2026-06-21 22:51:51 -06:00
Drew T 125222d14e feat(phase-21): worker gate — +7 fns x6 propagated (fleet 59.6%) 2026-06-21 22:35:16 -06:00
Drew T 7f272eb6b8 feat(phase-21): worker gate — +5 fns x4 propagated (fleet 59.37%) 2026-06-21 22:02:17 -06:00
Drew T 1d7c31393c feat(phase-21): worker gate — +7 fns x4 propagated (fleet 59.21%) 2026-06-21 15:17:27 -06:00
Drew T 4d00a40609 feat(phase-21): worker gate — +4 fns x2 propagated (fleet 59.05%) 2026-06-21 13:50:39 -06:00
Drew T bee543722d feat(phase-21): worker gate — +1 fns x1 propagated (fleet 58.98%) 2026-06-21 13:04:40 -06:00
Drew T 76c5feb116 feat(phase-21): worker gate — +1 fns x1 propagated (fleet 58.94%) 2026-06-21 12:50:16 -06:00
Drew T c2e93e50cd feat(phase-21): T3 — backlog ledger + shared gate stage (validated, +0.08%) 
- tools/backlog.py: near-miss ledger (.run/backlog.jsonl + docs/backlog.md ranked by
  reach->closeness->size; drop-now-matched filter, P9). tools/gate_stage.py: the shared
  deterministic spine both worker+grinder call (canon_resident_calls -> cast_call_sites ->
  sig_unify -> harvest_verify --chunk 1 -> dedup_propagate --auto-from -> backlog -> summary).
- validated end-to-end on a real 30-draft m2c sample: banked func_801710DC + propagated x134
  (dedup 1506->1508), 29 near-misses logged+ranked. make check-all 136/136 byte-identical
  (R22); fleet 58.82%->58.90%; dedup-check 1508 validated/0 failed (G3/P9).
- capped-fns reframe (R14): they're matched-inline (propagation-blocked), not draftable ->
  deferred to the worker; not claiming the unbanked +0.3% (P9).
 2026-06-21 12:32:21 -06:00
Drew T 6c5bd1c4a9 feat(phase-20): cast_call_sites.py — the §17a-1 callee-cast cap tool; 6 recovered, fleet 58.63%->58.82% 
- tools/cast_call_sites.py: per draft, callee whose canonical TU sig != draft's
  intended sig -> decl->canonical (kills in-TU conflicting types, keeps symbol in
  scope) + call-site cast ((ret(*)(args))func_X)(args) (codegen-neutral: gcc folds
  the cast of a known symbol -> direct jal). Pure --in/--out; whole-binary gate is
  the arbiter (G3/P9). Pipeline: canon_resident_calls -> cast_call_sites ->
  sig_unify -> harvest_verify --chunk 1
- recovered 6 of T6 batch-1's 33 gate-fails byte-identical (func_80153C44/8015CF58/
  801711FC/80161BE0/801683D8/8015F948); 5 propagated x134 + 1 local; check-all 136/136
- R14 CORRECTION of cookbook §20's '~33' projection (byte-proven): the batch is NOT
  mostly callee-conflicts — ~6 callee (cast-recovered), ~18 DEF-side loose-typing wall
  (def's matchable sig conflicts with the unchangeable caller-canonical; caller-side
  fix BLOCKED because INCLUDE_ASM declares no symbol -> a shared macro's internal
  extern is load-bearing in the 133 stub overlays), + a small DATA tail. The def-side
  needs re-drafting under the caller-canonical sig (future wave), not a text transform
- cookbook §20 RESOLUTION + SETUP tool inventory (R16/R21/R30); CURRENT_PHASE log (R28)
- fleet 58.63% -> 58.82%, dedup 1501 -> 1508, 0 NON_MATCHING (G4)
 2026-06-21 02:04:39 -06:00
Drew T 9991e4d6be feat(phase-20): T6 batch-1 — harvest wave 41/48 match_one, 8 banked (7 x134), fleet 58.35->58.63% 
- Ultracode wave on 48 cached reach-134 <=90-ins targets: 41/48 match_one MATCH (85%)
- whole-binary byte-gate: 8/41 banked (canon 3 + sig_unify 5); 7 propagated x134 + 1 local;
  dedup 1494->1501, check-all 136/136, +0.28%
- FINDING (R14/P9): match_one->gate gap ~80% this batch = EXTERN-CODEGEN-COUPLING (a callee's
  declared sig changes the call codegen; match_one uses the draft externs, the gate uses the
  overlay decls). The 33 lost are byte-correct vs the wrong externs -> need a callee-extern-
  canonicalize pass (NOT sig_unify, which rewrites the def-sig and regresses byte-correct drafts)
- the 7 DIFFs = gcc-class diagnoses: T3c store-vs-load (8014F2E0/80150528) CONFIRMED unsteerable
  (sched.c); SS10 hoist-vs-remat (80149374/801493D0); IV-combine (80177AD4); hoisted-invariant
  order (80177F84); 80161A90 = -O1 (build-infra)
 2026-06-21 00:57:04 -06:00
Drew T 4f5f73c480 feat(phase-20): T1 — typedef lift closes the type-blocked propagation cap (9 fns ×134) 
- build_engine_types.py: add find_typedefs() (brace-aware) + typedef_name() (anon-struct,
  fn-ptr, alias) + same-name/overlap collision guards; lift 14 typedefs to engine_types.h
  in source order after the named structs (A801593E4 -> S801593E4 dep preserved)
- engine_types.h: 48 named types + 14 typedefs; ov_SC01_077 byte-neutral (still d19c9580)
- dedup_propagate --auto-from: 9 reach-134 fns propagated x134 (8013E4B4 80148C9C 8014A238
  8014C118 80153978 801593E4 80168664 8016B964 8017AD0C); engine_core.h +9 DEFINE macros
- FINDING (R14/P9): cap was 16; only 9 were TYPE-blocked (all freed). The other 7 fail on
  undeclared callees/data (80142B2C 801535F4 80155E30 80157580 801576A8 80168F40 80170B48)
  -> the SS19 declaration-plumbing class, a recovery-tooling lever, NOT the type-lift
- check-all 136/136 byte-identical; fleet 58.00%% -> 58.35%%; dedup 1485 -> 1494 / 0 failed; 0 NON_MATCHING (G4)
 2026-06-20 23:49:58 -06:00