Found 50% of bare-def corpus completions fail standalone compile (undeclared globals/structs — externs
lived in TU headers). Filtering to compile-verified pairs so the pilot LoRA learns to emit COMPILABLE
matching C, not uncompilable bodies. (Drops harder struct/global fns; corpus v2 = add externs to recover.)
Qwen3.6-35B-A3B stock drafter: fair harness fixed compiles but model stuck at fixed near-misses
(can't refine from diff); full cookbook worse+2.3x slower than curated (dilution). Gets structure,
misses gcc-2.7.2 precision — exactly what src-pair LoRA targets. Forward: fine-tune or permuter-seed.
Gives the no-tool local model the context the agents read: common.h, live matching cookbook
(COOKBOOK_FULL toggle), 2 byte-matched corpus examples. Qwen3.6-35B-A3B result: harness fixes
compile-fails but model stays stuck at fixed near-misses; FULL cookbook worse+2.3x slower than
curated (dilution, gate-confirmed). Bottleneck is model refinement, not context.
docs/gen2-mips-matching-model.md: the BFM/gcc-2.7.2 matching-specialist idea (LoRA on our own
gate-verified pairs — the corpus off-the-shelf RE LLMs lack). export_pairs.py mines 1307 banked
(asm<->C) pairs from build objects (asm/ is gitignored, so disasm the ROM-identical build, splat-like
format) + src defs -> datasets/match_pairs/{pairs,train,test}.jsonl (gitignored, 1174/133 split).
api_draft.py: TEMP env-tunable. .gitignore: datasets/ models/ weights.
Script equivalent of one worker_wave drafter for any OpenAI-compatible endpoint (LM Studio /
llama.cpp / vLLM / OpenRouter). Inlines asm+ghidra_c+toolkit, calls /chat/completions, extracts C,
iterates against match_one (diff fed back, keep best). Output scores as a 'local'/'glm' arm via
ab_score.py. Env: API_BASE/API_KEY/MODEL. Logic self-tested; HTTP is standard OpenAI format.
- from-scratch giant func_801372B0 (207): structurally cracked at xHigh (206/207 ins,
logic 100%%) but byte-blocked by the gcc list-scheduler/regalloc last-mile (§27/§17);
banked the structural draft to backlog as a future permuter head-start
- xHigh finding (Drew's test): giants reach structural match fast — the wall is gcc
determinism, not reasoning depth; xHigh sufficient for this work
- session total: 16 functions ×134 (2 giants type-lift + 14 close=0), fleet 63.22%->63.66%
- docs/{worklist,backlog,progress.fleet}.md regenerated after the func_8015126C x134 bank
- CURRENT_PHASE.md: RESUME HERE block (T2 giant queue + levers + op gotchas) for a fresh session
- grinder stopped (STOP sentinel set); tree clean except R23 db churn
- cookbook §28: the 4-way triage of a close=0 gate-rejected giant (pure-extern plumbing /
masked-residual / struct-walled / regressed-draft) + the canonical-extern recovery (the NEW
lever that banked func_8015126C, which Phase-21 left as a close=0 wall) + the coalescing pin
(register __asm__ + lazy in-&& assign). §28a: PS1-applicable decomp.wiki GCC patterns
(negative-offset loops, branch-duplication, load-coalescing, div-magic table, slti-0)
- tools/recover_giant.py: canonicalize a draft's engine_core.h-callee externs to def-sigs +
move all externs block-scope (validated func_8015126C -> MATCH)
- R14: close=0-giant fan-out NOT uniformly near-free — only func_8015126C was pure-extern;
the rest are struct-walled / masked-residual / regressed -> genuine per-giant hand-work
Drew clarified the neverending loop's PURPOSE: learn idioms. Out of fuel -> analyze ALL
remaining, group by decomp issue, 1 idiom/group, cookbook, wave all applicable, bank, repeat.
ANSWER: this IS the idiom_loop flywheel (--assess groups by gcc-quirk class -> idiom -> distill
-> gen-fuel -> worker_wave -> gate -> repeat); no plan-mode build needed. ONE gap: --assess
clusters the BACKLOG (drafted near-misses), not ALL remaining stubs -> close it with TRIAGE waves
on un-drafted pools (reach-1 next, 281 fns). Effort already right (Max orchestration + Workflow
per-agent xHigh). Recommend a FRESH session (context hygiene); optional clean milestone = close
Phase 21 (reach-134 harvest byte-proven complete) -> Phase 22 idiom-loop at scale. Drew's call.
Drew chose deep re-RE of func_8014EA4C's real frame. Finding: the frame is a DEAD aggregate
copy (8 bytes unaligned-copied to sp+0x20, never read; gcc-2.7.2 keeps it, no DSE). The agent's
buf[16]+memcpy(buf+16) PRECISELY reproduces the exact gcc stack layout (local[0x38] -> frame
0x88, wrong). So the frame is already correct at close=6; the 6 residuals are NOT frame-caused
-- they're loop-body scheduler (global-store vs call-arg-load order) + regalloc (abs vs
in-place ), the irreducible §20/§25 wall. Deep frame-RE is a dead end for schedule/regalloc
residuals -- only the permuter explores that space. 5 levers (cheap-recovery, permuter,
giant-wave, hand-finish, deep-RE) byte-proven exhausted at fleet 63.17% for the reach-134 tail.
Hand-finish attempt on func_8014EA4C (close=6, Drew's choice): 2 byte-proven attempts both
cascaded — q0_arg reorder swapped p/q regalloc (6->123); abs-ternary dropped the load-bearing
dead buf[16] frame hack -> frame 0x60->0x58 -> offset cascade (6->23). Confirms the agent: the
residuals (store/load schedule + abs regalloc) are pressure-locked by the frame-forcing hack;
C-steering breaks the frame. The fix is the decomp-permuter (grinder), not C — it explores the
frame-preserving regalloc/schedule space. Grinder is now auto-permuting func_8014EA4C +
func_80144090 (re-logged close=6/25, token-free). cookbook §27.
- 6-giant worker_wave (1.16M tokens) BANKED 0. Agents got the SHAPE (§27 recipe works) but the
whole-binary byte-gate (G3/P9) rejected all 6. func_8014F74C close=0 match_one MATCH yet its RAW
draft FAILS the byte-gate (no plumbing/data conflict) -> match_one OVER-predicted (§25 masked
residual). The rest are pure regalloc/scheduling 'not C-steerable -> permuter' (close 6/25/53/91/164).
- Confirms cont.2/Phase-20/Phase-16: giants are HAND-FINISH/permuter fuel, NOT auto-bankable.
cookbook §27 updated. Do NOT scale giant auto-waves for % (1.16M tokens -> 0 banks).
- GATE BUG (R16, noted): canon/cast/sig_unify regressed near-misses (func_8014EA4C 6->128) and
poisoned the backlog. Re-logged the 6 with TRUE closeness + raw drafts (source=giant-raw);
func_8014EA4C(6)+func_80144090(25) now grinder-eligible (token-free).
- INFLECTION (P9): the automated reach-134 harvest is COMPLETE at 63.17% — cheap recovery (0 fresh),
permuter (~0), giant wave (0/6) all exhausted. Forward % = hand-finishing or close Phase 21.
- R14: sig_image h_exact = SHA1(raw IMAGE bytes), UNMASKED (sig_image.py:170-171) -> reach is
the ACCURATE shipped-byte reach; a resolved-reach probe just reproduces it. The real gap is
reach (shipped identical) >= realizable-x134 (shared-C reproduces all N), measurable ONLY by
dedup_propagate's byte-gate on a MATCHED fn. No cheap static probe -> cont.7 'build it' WITHDRAWN.
- R14: func_801775E0 backlog close=1 but saved draft compiles 31-off from source (prologue
save-scheduling/branch-sense/block-order, the S20 wall) -> close=1 was permuter-achieved.
RULE (S26): match_one a draft for SOURCE closeness before assuming a pin crack. The close-1..4
codegen tail is permuter-class (grinder territory), not cheap pin targets.
- INFLECTION (P9): the cached reach-134 automated harvest is COMPLETE; forward %-progress needs
a token-heavy lever (8 giants / per-fn pin-cracking) -> Drew's ROI greenlight. cookbook S26.
- THE FINDING (R14/P9): the _a close=0 recovery banks 0/20 (same def-side multi-way
loose-typing wall as MAIN 0/40, cont.6). cont.6's func_8012F568 was the lone simple-arity
win; the rest need a C type that conflicts caller-decl-vs-def with no single fitting sig.
The cheap close=0 ×134 recovery lever is EXHAUSTED.
- TOOLING (R16, idiom_loop.py --assess was DOUBLY inflated, now fixed):
(1) load_backlog drops banked-since-logged fns (intersect ledger w/ live INCLUDE_ASM stub
set, mirrors backlog._matched_now) -> killed the bogus 'unknown 14' class + ~10/class
stale-matched inflation (schedule 49->39 r134, regalloc 44->39).
(2) DETERMINISTIC-RECOVERY splits FRESH (never recovery-gated = genuine fuel) vs WALLED
(recovery-failed = def-side wall, do NOT re-run). Post-fix: 0 FRESH / 46 WALLED.
- schedule wave NOT fired (ROI-gate, S20/S26): ~50% x134 fraction on n=2, median 15 ins off,
over an h_exact-inflated count. No tokens spent speculatively.
- cont.6 option-3 CONFIRMED: cached reach-134 cheap fuel is dry; the genuine levers are
token-heavy fresh-session work (the 8 giants / per-fn pin-cracking / build a resolved-reach
probe). cookbook S26. 0 banks; fleet 63.17% unchanged; 136/136 byte-identical; 0 NON_MATCHING.
The 2nd straggler class (cont.4 wave-2): cross-overlay loose-typing COMPILE errors (not just
the -O0 byte-mismatch class). The -O0 exclusion alone was insufficient. Refactor the propagation
into a drop-straggler retry loop: extract apply_plan(subplan, restrict=); on a byte-gate failure,
isolate the culprit(s) for the failing overlay (per-fn trial), drop them (kept x1), retry the
survivors. Handles BOTH straggler classes failing ANY overlay; plan strictly shrinks (terminates);
byte-gate stays sole arbiter (a dropped fn never banks where it isn't byte-identical).
Validated end-to-end: wave-2 --auto-from auto-dropped 2 loose-typing stragglers + propagated 3
clean x134 (commit:0271); fleet 62.70 -> 62.82%. Future _a waves now auto-realize x134, no manual probe.
The cont.4 root-cause fixes so future _a waves auto-realize ×134:
- dedup_propagate --auto-from now excludes *_o0.c (-O0) defs: -O0 codegen embeds per-overlay
%lo data, so masked h_exact falsely reports reach-134 (§18/§20). One such straggler
(func_8013C360) reverted 10 clean ×134 matches under the all-or-nothing batch gate (cont.4).
Detected via find_site on the -O0 split file; --addr still forces them. (6 fns excluded.)
- gate_stage: capture dedup_propagate's exit; on a real byte-gate revert (not the benign
'nothing to propagate' no-op) write .run/auto/last_propagate_error.log + add prop_error to
the summary + warn. A swallowed revert previously hid the gain silently.
- 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).
Drew T