mirror/BFM-decomp
1
0
Fork 0
mirror of https://github.com/Druthulu/BFM-decomp synced 2026-09-26 21:36:06 -04:00
Code Issues Packages Projects Releases Wiki Activity
417 Commits 1 Branch 0 Tags
bce95a13cda63c12970c0fed1383ed19876cfe71
Commit Graph

417 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Drew T bce95a13cd docs(phase-23): T10.7 GLM5.2 A/B result — 10x codegen edge, but def-side wall caps banks (3/18) 
- GLM5.2 vs v3 on 18 hard-band fns: 10/18 vs 1/18 match_one; 3/18 vs ~1/18 whole-binary bank
- def-side loose-typing wall (Phase 16/20) caps banking for ANY drafter (Fable5 review §3c: HOLDS)
- 6 hardest beyond GLM too (0/6 at MAXTOK=16000); ~$1.02 of $25 spent
- 3 strategic options handed to Drew (direct-drafter / flywheel / wall-reconciliation)
 2026-07-01 20:01:14 -06:00
Drew T b1e0391c52 feat(phase-23): T10.7a — recover 1 more GLM bank via fix_arity_callers (func_801577C8) 
- def-side conflict (engine_core.h forward-declared func_801577C8(void) vs GLM's byte-correct
  (s32) def) resolved by relaxing the caller decl to no-proto; strip GLM externs + gate. 136/136.
- FINDING: only +1 of 8 stranded recovers mechanically; the other 7 are the intrinsic Phase-16/20
  DEF-side loose-typing wall (5 have non-(void) conflicting forward-decls, 1 narrow-param) — the
  wall caps ANY drafter, not a v3-tuning artifact (Fable5 review §3c re-test: wall HOLDS)
 2026-07-01 19:27:45 -06:00
Drew T 42719160e7 feat(phase-23): T10.7 — first GLM5.2 cloud-model banks (2 hard-band fns, byte-gated) 
- GLM5.2 via OpenRouter banked func_8013373C + func_8012F8C8 (ov_SC01_077_a.c), 16-22 ins
  hard-band fns v3 could not match; whole-binary byte-gate verified, check-all 136/136
 2026-07-01 19:17:39 -06:00
Drew T 1caff3dc4c feat(phase-23): api_draft MAXTOK env + OpenRouter cost capture (T10.7 reasoning-model support) 
- MAXTOK env (default 512 = local v3 unchanged); reasoning models (GLM5.2) need a high cap or
  they spend the budget on reasoning tokens and return empty content
- accumulate usage.cost from the response -> per-run $ + $/fn readout (OpenRouter reports it)
 2026-07-01 18:29:27 -06:00
Drew T 2d0d63953e docs(phase-23): Fable5 strategy review — scope-vs-reality scorecard, wall re-tests, resource map 
- scorecard: original 2026-06-10 scope vs 22 phases of byte-verified reality (what held,
  what emerged beyond scope, what deviated and should be revisited)
- adversarial pass: P21 no-shortcut + giant scheduler walls HOLD; P16 loose-typing wall has
  a TIMESTAMP GAP (declared 06-19, pre-dating cast_call_sites/block-scope-externs/v3) -> re-test
- July-2026 resources: frontier-on-hard-band (T10.7 re-aim), continuous architect-tier judgment,
  RE-ELEVATE THE PUBLIC FLIP (community labor = the only lever that scales into the proven tail)
- strategic fork: posture A/B/C on the byte-match goal; recommends dual-metric (B), Drew's call
- ranked recs 1-6 + explicit endorsements of what not to change
 2026-07-01 17:25:47 -06:00
Drew T 14fbc0039d docs(phase-23): T10.6 campaign result — 1297 banked, fleet 64.6%, ~92% ≤15 saturated (crash-graceful) 2026-07-01 14:23:09 -06:00
Drew T 70aaad4efb chore(phase-23): regenerate backlog + fleet digests (≤15 campaign end, fleet 64.6%) 2026-07-01 14:20:12 -06:00
Drew T 6069eaa727 feat(decomp): bulk_harvest — +5 fns across 5 binaries, 0 propagated (fleet 64.6%) 2026-07-01 14:17:42 -06:00
Drew T 8f15e2bad9 feat(decomp): bulk_harvest — +38 fns across 23 binaries, 1 propagated (fleet 64.59%) 2026-07-01 14:10:47 -06:00
Drew T 34ef8ad960 feat(decomp): bulk_harvest — +35 fns across 21 binaries, 0 propagated (fleet 64.58%) 2026-07-01 13:28:40 -06:00
Drew T d53b6ae56a feat(decomp): bulk_harvest — +45 fns across 28 binaries, 2 propagated (fleet 64.57%) 2026-07-01 12:45:44 -06:00
Drew T 8ad248fa3d feat(decomp): bulk_harvest — +48 fns across 35 binaries, 1 propagated (fleet 64.56%) 2026-07-01 12:10:38 -06:00
Drew T f2b9c24735 feat(decomp): bulk_harvest — +56 fns across 38 binaries, 2 propagated (fleet 64.54%) 2026-07-01 11:32:03 -06:00
Drew T a95c19447d feat(decomp): bulk_harvest — +58 fns across 39 binaries, 1 propagated (fleet 64.53%) 2026-07-01 10:44:57 -06:00
Drew T ecf01e32a3 feat(decomp): bulk_harvest — +58 fns across 38 binaries, 0 propagated (fleet 64.51%) 2026-07-01 09:59:28 -06:00
Drew T ef27dedb83 feat(decomp): bulk_harvest — +51 fns across 36 binaries, 2 propagated (fleet 64.49%) 2026-07-01 09:21:07 -06:00
Drew T eb696bf35d feat(decomp): bulk_harvest — +49 fns across 41 binaries, 0 propagated (fleet 64.48%) 2026-07-01 08:45:39 -06:00
Drew T 416b3f37a6 feat(decomp): bulk_harvest — +56 fns across 46 binaries, 2 propagated (fleet 64.46%) 2026-07-01 08:05:11 -06:00
Drew T 67a0718148 feat(decomp): bulk_harvest — +54 fns across 48 binaries, 3 propagated (fleet 64.45%) 2026-07-01 07:23:56 -06:00
Drew T 00064af879 feat(decomp): bulk_harvest — +47 fns across 43 binaries, 1 propagated (fleet 64.43%) 2026-07-01 06:46:07 -06:00
Drew T 7e9dd8f03a feat(decomp): bulk_harvest — +53 fns across 48 binaries, 2 propagated (fleet 64.42%) 2026-07-01 06:05:00 -06:00
Drew T c1afac9f76 feat(decomp): bulk_harvest — +70 fns across 58 binaries, 2 propagated (fleet 64.4%) 2026-07-01 05:17:55 -06:00
Drew T f55052e2b5 feat(decomp): bulk_harvest — +64 fns across 55 binaries, 1 propagated (fleet 64.38%) 2026-07-01 04:39:21 -06:00
Drew T b4693430a0 feat(decomp): bulk_harvest — +66 fns across 55 binaries, 1 propagated (fleet 64.36%) 2026-07-01 04:00:29 -06:00
Drew T 3d90707d41 feat(decomp): bulk_harvest — +69 fns across 62 binaries, 1 propagated (fleet 64.34%) 2026-07-01 03:22:46 -06:00
Drew T 7fa2134326 feat(decomp): bulk_harvest — +75 fns across 64 binaries, 0 propagated (fleet 64.32%) 2026-07-01 02:41:58 -06:00
Drew T d2d15c9e41 feat(decomp): bulk_harvest — +78 fns across 69 binaries, 1 propagated (fleet 64.3%) 2026-07-01 02:03:59 -06:00
Drew T 7efe5ce10d feat(decomp): bulk_harvest — +72 fns across 63 binaries, 5 propagated (fleet 64.28%) 2026-07-01 01:29:20 -06:00
Drew T 95586a6751 feat(decomp): bulk_harvest — +69 fns across 67 binaries, 6 propagated (fleet 64.26%) 2026-07-01 00:54:27 -06:00
Drew T 8dc42f12b5 docs(gen3): parking-lot for asset export/Unity + verified mesh-state survey 
- docs/gen3-parking-lot.md (NEW): town mesh/texture + Musashi->Unity export ideas, PARKED (Gen3).
  Verified survey (read-only, 2026-07-01): BFM uses the libgs TMD model path — real jal
  GsMapModelingData + GsLinkObject5 call sites (EXE 0x8001C2xx-0x8001C9xx model subsystem +
  ov_SC02_005); render = stock libgpu; no libgs sort/anim calls (custom anim -> Musashi harder);
  textures via VRAM rip + type-0 'graphics' blobs; no exporter tooling yet. Decisive next probes listed.
- CURRENT_PHASE.md: T10 done + T10.6 campaign in-flight + T10.7 OpenRouter-next notes (crash-recovery)
 2026-07-01 00:52:23 -06:00
Drew T 6eee685acc feat(decomp): bulk_harvest — +81 fns across 74 binaries, 18 propagated (fleet 64.23%) 2026-07-01 00:14:55 -06:00
Drew T 337e120c2a feat(decomp): bulk_harvest — +4 fns across 4 binaries, 0 propagated (fleet 64.21%) 2026-06-30 23:39:18 -06:00
Drew T 5ba5500f82 feat(decomp): bulk_harvest — +4 fns across 4 binaries, 1 propagated (fleet 64.21%) 2026-06-30 23:37:38 -06:00
Drew T 35bc7c3873 feat(phase-23): bulk_harvest --cycles campaign mode (resumable, per-cycle commit, dry/STOP-aware) 2026-06-30 23:34:31 -06:00
Drew T 265e930f65 chore(phase-23): regenerate backlog + fleet digests (T10 banks) 2026-06-30 23:23:04 -06:00
Drew T e9f199530e docs(phase-23): T10 harvester built + measured — 65% bank-rate, parallel gate ~75x, vLLM justified 
- gen2-mips-matching-model.md: T10 RESULT (bulk_harvest phase-separation; 52/80=65% on fresh
  SC03 <=15-ins; gate 0.4s/fn/8w vs ~30s/fn serial; drafting now the sole bottleneck -> vLLM)
- SETUP.md: bulk_harvest.py inventory row + gate_stage/harvest_verify parallel-safe params (R21)
- CURRENT_PHASE.md: T10 done (built+measured), T10.5 vLLM split out as Drew-gated; progress log
 2026-06-30 23:22:46 -06:00
Drew T aa2af72f0c feat(decomp): bulk_harvest — +52 fns across 40 binaries, 12 propagated (fleet 64.2%) 2026-06-30 23:17:02 -06:00
Drew T a36c5ded0a feat(decomp): bulk_harvest — +9 fns across 9 binaries, 12 propagated (fleet 64.19%) 2026-06-30 22:52:53 -06:00
Drew T c5cca1c6d4 feat(phase-23): T10 tooling — parallel-safe gate spine + bulk_harvest 
- gate_stage.run_gate: optional lock_path (per-binary lock), verified_out/failed_out
  (per-worker harvest scratch), compute_fleet toggle — all default to today's behavior,
  so grinder/orchestrator/lora_grind are byte-unaffected
- harvest_verify: --verified-out/--failed-out (per-worker result paths for parallel gating)
- bulk_harvest.py (NEW): phase-separated saturation factory — bulk-draft (GPU) ->
  ProcessPool parallel-gate over distinct binaries (build/<bin>/** isolated) ->
  dedupe-once + one commit; round-robin fuel spread; whole-binary byte-gate stays the
  sole arbiter (G3/P9); on-demand/bounded (STOP sentinel); lora_grind kept as fallback
 2026-06-30 22:48:26 -06:00
Drew T 69189570da docs(phase-23): session close — handoff CURRENT_PHASE for the harvester-factory build (T10) 
- fleet 64.19%, 136/136 byte-clean; injection lever +0.34% banked this session
- ★ NEXT TASK rewritten: build phase-separated + vLLM batched harvester (Drew option c)
  with the full fuel map (~4,597 unique ≤15 funcs, SC03=12,497), the throughput finding
  (~48s/func + GPU-idle-during-gate), the applied fixes, and the parked items
- supersedes the stale reach≥2/14B plan; 14B-for-giants now AFTER ≤15 saturation
 2026-06-30 15:59:47 -06:00
Drew T 7dbdd1ce47 perf(phase-23): cap api_draft max_tokens 4096->512 (no-stop ramble) + capture saturation/throughput findings 
- the model rambled to 4096 tokens on no-stop-token drafts (~80-130s each); a
  ≤15-ins fn's C is ~100-300 toks, so 512 truncates the ramblers (gate-rejected anyway)
- CURRENT_PHASE: ≤15 band NOT saturated (~4,597 unique / 4,087 untried, SC03=12,497);
  drafting ~15-18s/func serial is the throughput wall; architecture = phase-separate +
  parallel-gate + vLLM batched serving (the real GPU lever); fresh bank-rate still TBD
 2026-06-30 15:52:04 -06:00
Drew T f72aa74bd4 feat(phase-23): lora-grind gate — +22 fns x0 propagated (fleet 64.19%) 2026-06-30 15:49:45 -06:00
Drew T 5a984886af perf(phase-23): lora_grind computes progress.py --fleet only on propagate sweeps (was every batch, ~14s overhead dominating small batches) 2026-06-30 15:32:56 -06:00
Drew T eb3b18efcc feat(phase-23): lora_grind --min-nins (skip the saturated small band for a fresh-band pass) 2026-06-30 15:05:25 -06:00
Drew T 72dabb470c feat(phase-23): broad v3 harvest (48 banks) + final propagate sweep; docs 
- bounded broad v3 harvest banked 48 small fns across SC01/SC02 overlays (mostly
  reach-1 ×1; fleet 64.16->64.18%), final propagate sweep +4 groups (1687->1691)
- check-all 136/136 byte-identical, dedup-check 0 failed
- docs/community-matching-model-plan.md (publishing the drafter — parked)
- docs/sister-decomps.md (VS cross-project plan + idiom-discovery rationale — parked)
- CURRENT_PHASE.md: 2026-07-01 session arc (injection lever +0.34%, the type-walled-14
  deferred, the saturation-campaign plan)
 2026-06-30 13:54:53 -06:00
Drew T 71e686ad2e feat(phase-23): lora-grind gate — +4 fns x0 propagated (fleet 64.18%) 2026-06-30 13:50:51 -06:00
Drew T 7b1b68cdad feat(phase-23): lora-grind gate — +6 fns x0 propagated (fleet 64.18%) 2026-06-30 13:37:06 -06:00
Drew T 613df42ed0 feat(phase-23): lora-grind gate — +7 fns x0 propagated (fleet 64.18%) 2026-06-30 13:24:13 -06:00
Drew T 7792566bfd feat(phase-23): lora-grind gate — +7 fns x0 propagated (fleet 64.17%) 2026-06-30 13:11:19 -06:00
Drew T 39ccbee46c feat(phase-23): lora-grind gate — +3 fns x0 propagated (fleet 64.17%) 2026-06-30 12:58:40 -06:00