Commit Graph

7 Commits

Author SHA1 Message Date
Drew T 897939eb46 feat(phase-25): T5b batch-1 (partial) — +1877 fns (fleet 71.47->72.02%)
- T5b measure-wave over the <=149-ins cheap tier (83 targets, generated §12-robust
  t5_scaleup.js, Opus xHigh). Drew stopped at ~50 on HTTP-529 overload; processed the
  54 landed drafts (0 truncated) as T5b.
- 46/54 match_one-MATCH (85%); gate banked 16 clean (main 1, _a 10, _after 5) +
  family_sweep +1861 siblings/1 fail across 133 overlays (14 exemplars, ~0 tokens).
- R22 clean-fleet 136/136 byte-identical; dedup-check 1813/0; 0 NON_MATCHING (G4).
- frontier map now 64 exemplars measured (.run/t5_frontier.jsonl). Def-side loose-typing
  wall confirmed dominant at scale (~30/46 match_one-MATCH gate-rejected -> backlog);
  fleet-wide caller-decl reconcile = the #1 T7 lever. 8 near-miss -> permuter fuel (T5c).
- more T5b batches (29 cheap + 34 giants + 5 _o0) + T5c/T5d -> T6 remain (phase OPEN)
2026-07-08 22:56:01 -06:00
Drew T f154822f7b feat(phase-22): cheap-tier A/B experiment harness + findings
ab_match.js (per-arm-model fan-out fork of worker_wave) + ab_score.py (disk-truth
match_one scorer, --measured for real $). Haiku drafters vs Opus on 20 reach1 fns:
proxy 10 vs 15 matches, Haiku ~4.8x more matches/$, parity on <=52-ins band. Whole-binary
gate banked 4 (proxy over-counts; TU-plumbing wall). backlog/fleet refreshed.
2026-06-29 12:52:39 -06:00
Drew T 62fa026f7e feat(phase-21): worker_wave — GIANT-guarded §27 hint (arg-arity decls, sibling templates, coalescing residual) 2026-06-25 16:46:35 -06:00
Drew T e30697a081 docs(phase-21): _a vein is a x134 lever (10/11) — worker_wave region-fix + cont.4 findings
- worker_wave.js: derive match_one self-check --asm-subdir per-target from t.asm
  (was hardcoded main subdir -> every _a agent self-check crashed on missing .s ->
  blind drafting). Enabled the 16/24 close rate on the _a measurement wave.
- CURRENT_PHASE cont.4: the _a vein is x134 (10/11 propagate byte-identical), not the
  cont.3/§24 x1 wall. The gate's propagated:0 was 2 bugs: dedup_propagate all-or-nothing
  batch-revert poisoned by 1 stale -O0 overlay-local straggler (func_8013C360), and
  gate_stage sh() swallows dedup_propagate's non-zero exit. Realized +0.39% (commit:0267).
2026-06-24 20:37:57 -06:00
Drew T f137658171 fix(phase-21): drafters write ONLY their draft .c — distill owns cookbook (prevent unverified near-miss idiom pollution) 2026-06-22 09:36:47 -06:00
Drew T 5fc5d64445 feat(phase-21): R2/R3 — the learning flywheel (class-grouped waves + distill)
- residual class travels with the draft: drafter stamps // @class / // @stuck; gate_stage logs
  the worker's gcc-quirk class to the backlog (not the coarse manifest class).
- wave_targets.py --class <C> / --list-classes: class-grouped re-attempt waves over the backlog's
  classified near-misses (REGALLOC/SCHEDULE/REMAT/STRUCT/IV/LOOPGUARD/LOOSE/PLUMBING).
- orchestrator prep: auto-picks a class-focused wave when a class accrues >= threshold near-misses,
  else a fresh pool harvest; clears the draft dir per wave.
- tools/workflows/distill.js: post-wave agent extracts NEW byte-verified gcc idioms -> cookbook §21
  (R16, conservative); drafters now read the LIVE cookbook so distilled idioms feed forward.
- runbook: the full prep->worker->gate->distill cycle (the Phase-18 close-rate loop, automated).
2026-06-21 13:56:59 -06:00
Drew T 2d7926e9f6 feat(phase-21): T4/T5 tooling — worker Workflow + wave selector + permuter grinder
- tools/workflows/worker_wave.js: parallel toolkit-aware drafter agents (xHigh) per target,
  each writes its best matching C to the drafts dir (gate run by the orchestrator after).
  args parsed robustly (harness may serialize to a JSON string — §20 gotcha).
- tools/wave_targets.py: select a wave batch from the fuel manifest (pool/region/N, skips
  backlog walls). tools/grinder.py: token-free permuter grinder — permute backlog near-misses
  -> gate_stage -> bank; STOP/heartbeat, idles for new worker near-misses, supervisor-friendly.
2026-06-21 12:41:14 -06:00