- remap_hseq.gather_externs: carry file-scope externs for body-referenced symbols (extract_unit only
grabbed adjacent ones) — the decl class that blocked per-location bodies indexing a global. func_8015F118
gate-fail -> BYTE-IDENTICAL; the 3 tracker-miss PURE families then bank 133/133 each.
- ran the real whole-binary byte-gate on the 29 substantial matched-exemplar families:
532 members BANKED (byte-gated). Per-family: 3 tracker-miss PURE (0x8015d5e8/0x8015f118/0x801407f4)
bank 100% x133 = 399 byte-perfect (the tracker-fix free win); 1 cross-addr family 50%; 9 zero-bank
families are type-using (Work8016/Prim/...) -> the existing --reconcile/type-lift follow-on (Task 8);
16 families pinned -> Task 7 pin-free re-crack.
- VERDICT: the h_seq machinery (tracker + imm + cross-address + extern-carry) is byte-proven 100% correct
on clean families. GO to scale.
- R22 clean-fleet: make clean + extract-all-136 + check-all = 136 passed, 0 failed. 0 NON_MATCHING (G4).
Metrics: distinct-code 30.3->30.9% (+375 fns), instr-weighted 58.2->58.5%.
- decision-log 2026-07-11 (R31: stratify a mechanical-harvest rate by family/class before judging it).
- Fable5 subagent cracked func_80166994 (trail/afterimage ring recorder, 369 ins) — FULLY
STRUCTURAL, zero register pins -> swept ×134 CLEAN (exemplar + 133 siblings byte-identical).
R22 clean-fleet 136/136; instr-weighted 56.8% -> 57.2%; distinct-code 27.3% -> 28.2%
- NEW IDIOM cookbook §43: a K&R s16-param DEFINITION dissolves the §17/§29 "narrow-param wall".
On MIPS K&R promotes s16->int (ABI-identical to the canon-sig s32), body keeps the in-place
sll aN,16 narrow/extend the (s16)cast form can't reproduce. void->s32 return-flip pair:
split //@EDIT (self-fn, ov077-specific) + engine_core.h ec_edit ×5 (byte-neutral, callers discard)
- family_sweep --edit-remap: split-edits now OPTIONAL (apply where present, never skip; the
whole-binary byte-gate is the sole arbiter, G3/P9) — a sibling lacking the ov077 canon-sig decl
still banks via ec_edit + body. edit-absent tracked, not skipped
- R14: the prior wave's "@stuck: none — MATCH" note on func_80166994 was STALE/FALSE (re-ran DIFF
366/369). Verify a MATCH claim vs the bytes, never a stale note
- structural cracks are the ×134-SAFE ones (contrast §42e pin-heavy families that cc1-SIGABRT in
sibling TUs). Other 6 giants -> cheap-Opus applying §43+§31, Fable5 only on new-class evidence
- weighted_metrics() from .run/sig.*.jsonl + src stubs (executable code only, resident + 134
overlays; main EXE excluded). Two framings: fleet instr-weighted (per-overlay, the decomp.dev
-display number) + dedup distinct-code (each unique h_exact once, the distinct-RE number)
- --fleet now emits THREE labeled metrics into docs/progress.fleet.md: fn-count 74.48% (×134-
inflated), instr-weighted 56.8% (shipped .text), distinct-code 27.3% (of 84,996 unique fns)
- --weighted prints the two weighted numbers standalone; degrades gracefully if sigs absent
- corrects the stale "~30-35% byte-weighted" estimate: the giant campaign since Phase 19 raised
the fleet instr-weighted number to 56.8%; the distinct-code 27.3% is the unique-monster-tail truth
- SETUP §tooling row updated (R21)
- family_sweep.py: new --edit-remap MANIFEST mode (§42e) — per family, symbol-remap the
split-scope //@EDIT old||new per sibling + apply once-global engine_core.h ec_edits
(byte-neutral), stage the family_remap body, gate via harvest_verify (the sole arbiter)
- BANKED 266/266 (0 failed): func_80136824 + func_80136334 (array-decay ptr-flip) ×133
siblings each — full ×134. R22 clean-fleet 136/136, fleet 74.40% -> 74.48%, dedup 1813/0
- R14 FINDING (cookbook §42e addendum + decision-log): the other 4 byte-drift families
(func_80133AB0 zero-reg pin, func_8016DF5C/8013D9B0 GTE-pin, func_80156044 trampoline)
cc1-SIGABRT (Error 134) in the SIBLING TU — hand pins are ov077-TU-context-specific,
NOT mechanically ×134-recoverable; backlogged as ×1/permuter fuel. rtu_match/match_one
are blind here (neutralized/isolation compiles crash too); only make build is truth
- 0 NON_MATCHING in any default build (G4)
- R14 CORRECTION of the prior "family_remap limitation" call: it was a MISDIAGNOSIS. family_remap
succeeds on all droppers; the "remap-fail" family_sweep reports was a mislabeled canon_sig_reconcile
throw ("no definition of func_X found in draft") — the def-finder regex required a leading \n, so a
//@EDIT-stripped raw draft with the fn definition on line 1 was not found.
- FIX: def-finder regex \n -> (?:^|\n) (also match a def at draft start; strictly additive, low-risk).
- Recovered func_8014FE60 fully: 133/133 siblings banked (fix + engine_core.h DEFINE_func_8014FDF4
extern void->s32 global flip, byte-neutral fleet-wide; caller discards return).
- Residual (the genuine, small --edit-remap): func_8016DF5C/80136334/8013D9B0/80156044 reconcile but
byte-drift per sibling (out-of-body fixes: pointer //@EDIT, no-proto, return-flip not carried per sibling).
- cookbook §42e (the two-layer diagnosis + the forward ×134-leverage-realism rule); decision-log corrected.
- R22 clean-fleet 136/136 BYTE-IDENTICAL from a fully clean tree; NON_MATCHING 7 (0 in default build, G4).
- tools/t7_bank.py: the M1 driver (reconcile-at-bank-time against the CURRENT TU + harvest_verify
whole-binary gate; chunk-bet + per-round re-reconcile for cross-fn ambient mutation; giants first)
- canon_sig_reconcile v3.2: uniquify ALL draft-defined type names AND struct/union TAGS to <name>_<addr>
(byte-neutral) -> collision-proof when many exemplars bank into one TU (the 'redefinition of struct Fr'
class); banked func_8016A290 (284-ins giant) + fixed the inter-draft collisions
- 37/40 non-jumptable M1 exemplars banked BYTE-IDENTICAL (ov_SC01_077 d19c9580, clean rebuild); each is
currently ×1 (M2 sweeps them ×134 next)
- R14 CORRECTION (cookbook §41b): the T6 object-only probe OVER-counted BANKABLE by 7 -- it is blind to
rodata + link. (a) 4 jump-table fns (3 _o0 giants + func_8012ACE0): .text byte-perfect but a switch
jump table in rodata diverges -> REFUTES the T6 'Q3 -O0 REFUTED' claim; F-band jump-table workflow.
(b) 3 last-referencer link-walls (func_8016D688/D1D8/165240): C-ifying the only asm referencer of a
scratch data symbol drops splat's auto-symbol -> ld undefined reference; M-linkwall tier, deferred
- ov_SC01_077 REAL 32 stub-exemplars -> 37 more defined; whole binary byte-identical throughout (G3/P9)
- batch-3: the 3 _o0 giants 3/3 isolation-MATCH at -O0 (new tools/match_one.py --o0 flag);
0 banked — an -O0 in-context byte-diff (canon_sig_reconcile's void->s32 is NOT byte-neutral
at -O0) -> deferred to T7. Frontier 125/127 draftable exemplars measured.
- R14 correction: the mechanical reconcile-sweep is NOT the clean 51-target x134 win first
hoped. The frontier "match" status carried un-verified agent claims (5-sample spot-check
= 3/5 genuine); ~19/51 have clean engine_core.h canonicals (the reliable canon_sig_reconcile
tier), the rest hit VARIED walls (callee-sig conflicts, non-identical types Vec3/SVEC,
macro-local data) -> this is the genuine Fable5/T6 residual, not a mechanical sweep.
- docs/phase25-t6-fable-brief.md: the grounded Step-B input package for the fresh Fable5Max
session to author the crack curriculum (95-stub worklist by wall-class + tools + 5 swing
questions). CURRENT_PHASE: T6 hand-off (fresh session, Fable5Max, read the brief). Phase OPEN.
The family-sweep driver validated end-to-end: remap ov_SC01_077 exemplar -> 133 h_norm-siblings,
plain harvest_verify banked 133/133 byte-identical (0 failed). func_80141100 now matched in all 134
overlays. Added a match_one pre-classify so the full sweep gates only clean drafts (type-using
families -> .run/sweep_deferred.txt for the decl-reconcile pass, avoiding bisection blowup).
Two-phase: for each matched ov_SC01_077 fn with unmatched same-address h_norm-siblings, remap
(family_remap) to each sibling + byte-gate into that overlay. Stage all drafts per (overlay,split),
gate each group once via PLAIN harvest_verify (remapped drafts need no transforms — T3). Sizing:
169 exemplars, 22,117 member-remaps, ~3.2 MB potential. Type-using families fail-compile -> logged
for the decl-reconcile pass (not remap failures).
Swing question resolved. h_norm families are byte-shattered by PER-OVERLAY symbols, so free
--tier h_norm dedup fails (0/133 on 0x80141100). BUT the mechanical remap works: disassemble
two h_norm-identical members, positionally pair their reloc targets, substitute the exemplar
per-overlay symbol names -> the sibling C. Verified: reloc decoder 22/22 vs .s; 0x80141100->
ov_SC01_000 whole-binary byte-identical + harvest_verify banked; 12/12 siblings; 7/9 families
x3 = 21/27 match_one. The 2 fails = decl/type plumbing (T7 refinement), not remap fails.
Reshapes the endgame: crack ONE exemplar/family -> remap+gate members (~0 tokens/member); the
11.1 MB is cheaply mechanical. Members no longer need drafting; only exemplars do.
--auto silently skipped the 263-stub ov_SC01_077_after cohort: the drift-check hardcoded
the main asm subdir and gate_stage was never passed src/asm/src_file. Now stub_map() reads
each stub asm subdir from its INCLUDE_ASM line, the drift-check uses the per-fn subdir, and
reconcile_and_gate groups targets by split file, gating each via run_gate(src=,asm=,src_file=).
fix_arity_callers/dedup_propagate already split-aware; harvest_verify self-filters per split.
- housekeeping: deleted stale root gccdump.lreg (gcc default -da RTL dump, dumpbase
"gccdump"; a one-off RTL-inspection run at root CWD, NOT a committed tool). Practice:
RTL dumps go to a .run/ CWD or -dumpbase .run/gccdump (R12)
- make sig-overlays (134 sigs); worklist.py --refresh -> build_fuel_manifest + backlog
render: docs/worklist.md + docs/backlog.md regenerated (stale Jun-26 ranking killed;
the banked whale func_80144B9C dropped from backlog, P9)
- fix tools/sig_image.py docstring: h_norm is the live norm_stream normalizer, not a
"T5 placeholder" (R21/R30)
- CURRENT_PHASE.md: Phase-25 plan-of-record + T0 frontier survey (R14): h_norm 2,764
multi-member families / 11.1 MB; 46% of large solo monsters (3,989/8,670) are family
members; reach-134 h_exact warm-up = 130 classes / 4.11 MB
- func_801770E0 (152 ins, reach-134) propagated ×134 via dedup_propagate --recover.
Chose the NATIVE DEFINE-macro path over a hand-rolled shared header: func_8014E048
(pins+asm) is already ×134 via a DEFINE macro, proving that path handles pin/asm -O2
giants (the whale needed a shared header only because it is -O0 -> separate object).
Clean fleet check-all 136/136 (R22), dedup-check 1811->1812/0.
- R14: func_8014E048 was ALREADY ×134 (T6 §A) — the whale-session handoff was stale;
only func_801770E0 + func_801372B0 actually remained ×1.
- tooling (reusable): dedup_propagate.overlay_files now also scans the whale-rollout
_o0b/_after splits — post-whale-region fns (func_801770E0 in _after.c) were invisible
for both source-def-find and stub-replacement. func_801770E0 extern block made
contiguous (a comment between externs made find_site drop 5 externs; byte-neutral).
The single biggest byte-weight lever (770 ins ×134 ~ +1.6% byte-weight), banked across all 134
overlays as -O0 compiled C (was ×1 in ov_SC01_077, commit:0463).
- ROLLOUT (tools/rollout_whale_o0.py): per single-file overlay, LINE-split <ov>.c at the whale
(splat emits in vram order), carve the yaml code subseg into before/<ov>_o0b(-O0)/<ov>_after,
thin <ov>_o0b.c #includes the shared src/shared/func_80144B9C.h (the whale C moved there — DRY,
TU-local typedefs, not a fragile 200-line DEFINE_ macro). Makefile WHALE_O0B_OBJS wildcard -O0.
- DEDUP: group E_func_80144B9C (134 members, h_exact 74186b97, source=the header; group_members
keys on binary+vram so a header-share validates like a macro-share).
- MILESTONE: clean fleet check-all 136/136 (R22); dedup-check 1811/0; FLEET REAL 225816->225949,
byte-identical 65.91->65.95% function-count (the gain is byte-weighted). Validated on ov_SC01_000
+ spot-checks SC02/SC03/SC07 before the full fleet.
- cookbook §38 + SETUP inventory.
- tools/reconcile_decls.py: the DATA-symbol analog of cast_call_sites.py. A fleet-majority
canonical-type ORACLE (engine_core.h-authoritative + plurality across all overlays+resident —
the picker gen_harvest_targets lacked) + a byte-neutral access-cast TRANSFORM (array/struct/
scalar/ptr taxonomy, single-pass so no double-wrap). Mechanizes the manual giant decl-reconcile
that func_80129CF8 got by hand.
- wired into gate_stage after cast_call_sites (canon -> cast -> reconcile_decls -> byte-gate);
idempotent / no-op on drafts without a data conflict -> can't regress the wave; the whole-binary
byte-gate stays the sole arbiter (G3/P9).
- VALIDATED: oracle 8/8 correct picks for func_80129CF8's symbols; full loose giant ->
reconcile_decls -> swapped into the DEFINE_ macro -> make build BINARY=ov_SC01_077 = BYTE-IDENTICAL
d19c9580; idempotent (0 reconciled) on the already-canonical committed draft.
- find_site extern-collection: allow a trailing `/* comment */` after the `;`. The
comment-blind regex `^\s*extern\b.*;\s*$` stopped the backward scan at
`extern u8 D_801152A8[]; /* canonical TU type */`, dropping every EARLIER extern →
compiles_standalone failed on the now-undeclared callees/data (func_80135A4C,
func_80133784, D_801152A8). R14: THIS — not "pin/asm" as the backlog framed it — was
func_8014E048's real self-containment blocker.
- propagated func_8014E048 (the T5b S11 pins+barrier crack) + func_80157580 ×134;
dedup 1797→1799 groups (0 failed). CLEAN fleet check-all 136/136 BYTE-IDENTICAL (R22).
Fleet byte-identical 65.40% → 65.48%.
find_site + dedup_propagate --recover + build_engine_types fixes, then re-bank
the 13 recover_integration leaf-MATCHes and propagate each across all 134 overlays.
- find_site: match INDENTED inline defs (was column-0 only, silently dropping every
recover_integration-banked def from propagation — T6 blocker 1). Unit-tested:
indented defs match; indented call-exprs (if/assign/bare/return) correctly rejected.
- dedup_propagate --recover: on a straggler byte-gate failure, FIRST no-proto that
overlay's conflicting caller extern + re-gate (Part B, byte-neutral, same lever as
fix_arity_callers --any-proto); else EXCLUDE only that overlay (Part A, ×N-1) rather
than the historical all-or-nothing drop. Wired into gate_stage.
- build_engine_types: comment-aware find_defs/find_typedefs (blank_comments). The
generated header's own "...typedef lift" comment was captured as a bogus `typedef
vec`, self-colliding and blocking every --strip. NOTE: full --strip still conflicts
with the _a/_o0 split files, so split-file overlays need a TARGETED lift.
- banked + propagated 13 fns ×134: func_8014F74C 801542A4 8015BE94 8015F380 80160F00
801653B8 80166244 8016E778 801732C4 8017331C 80173374 80174554 801745AC.
func_8014F74C needed PosT/MoveT lifted to engine_types.h (targeted, byte-neutral).
- dedup 1784→1797 groups (0 failed). CLEAN fleet check-all 136/136 BYTE-IDENTICAL (R22:
make clean && extract-all && check-all). Fleet REAL 224073, byte-identical 65.40%.
R14: the confounding stale-asm/ tree (13 missing .s) that masked the first --recover
test is fixed by re-extract; the 3 pre-existing --auto-from stragglers (0x80174650/
8012A018/80165CA0) are pin/asm + uncaptured-local-macro (SHB) bodies, correctly dropped.
The integration-recovery tool for leaf-MATCH-but-whole-binary-gate-rejected fns ("declaration/TU
plumbing" — the dominant residual gate_stage's canon/cast/sig_unify pipeline doesn't reach).
- tools/recover_integration.py (NEW): batch recovery — gather leaf-MATCH candidates (--auto from
the backlog, drift-checked R14; or --funcs/--from-file) → no-proto their conflicting caller decls
→ gate_stage (byte-gate + log). 2-PASS snapshot/restore: pass 1 finds the bankable set, pass 2
re-banks ONLY winners from the clean snapshot (so non-banks are never corrupted).
- tools/fix_arity_callers.py: extended with --binary — scan+rewrite the overlay's OWN inline caller
decls (src/<bin>/<bin>*.c), not just engine_core.h. That was THE gap: a conflicting caller extern
is often inline in the overlay src (e.g. func_8016E778's `extern void f(void)` vs def `f(int)`),
which fix_arity_callers never saw -> the fn stayed unbanked.
- VALIDATED: banked 13 leaf-MATCH fns (func_8014F74C/801542A4/8015BE94/8015F380/80160F00/801653B8/
80166244/8016E778/801732C4/8017331C/80173374/80174554/801745AC), CLEAN-verified together
(ov_SC01_077 d19c9580). The banks themselves are reverted here (they re-bank via the tool and
will land ×134 once propagation-recovery lands — cleaner than committing ×1).
- R14 lesson (clean-verify caught it, R22): fix_arity_callers --revert is LOSSY for --any-proto
(()->(void), not back to the original args) -> corrupted non-banks; fixed with the 2-pass snapshot.
- REMAINING T6 (×134 propagation-recovery, 3 diagnosed blockers): (1) dedup_propagate find_site
misses INDENTED inline defs (Phase-15 class); (2) overlay-local-type lift; (3) auto-reconcile the
straggler's conflicting caller externs (the flagship func_80132784 / ov_SC02_005 class, done by
hand — needs automating). See CURRENT_PHASE.
Pre-existing latent breakage (Phase-21 close commit:0292, "+62 PsyQ names"), found during T5b's
R22 fleet verify. A genuinely-clean `make clean` + re-extract + build of `main` failed with
dozens of `can't open asm/nonmatchings/800c3/func_*.s` — clean-rebuild broken since Phase 21,
masked all along by incremental builds reusing stale pre-rename .s/.o (the exact R22 failure mode).
- Root cause (byte-proven): Phase-21 xdedup renamed 62 PsyQ library functions to their proper
names (InitHeap, FlushCache, GetTPage, SysEnqIntRP, SpuWrite, CdMix, __main …) in
symbols.us.txt, but never regenerated/renamed the committed src/*.c stub refs that call them by
the OLD func_<ADDR> name. splat's FRESH regeneration of a stub .c uses the CURATED names (move
src/800c3.c aside -> splat writes INCLUDE_ASM(InitHeap) + emits InitHeap.s); the committed stubs
were simply stale.
- Fix: rename all 62 INCLUDE_ASM(func_<ADDR>) -> the curated name across 12 files (800c.c 1,
800c3.c 22, apicard1/2/4 4/6/2, boot.c 1, libcd1.c 8, libetc.c 6, libgpu.c 3, sgap.c 1,
snd1.c 7, snd2.c 1). Pure rename to match splat's canonical output; byte-neutral.
- VERIFIED: main clean-builds 143dbb89 from a fully clean tree; full clean fleet check-all 136/136.
- Lesson (cookbook): a symbols.us.txt rename must be propagated to (a) shared-macro bodies
(engine_core.h — the T5b GetTPage fix) AND (b) INCLUDE_ASM stub refs (this), AND verified by a
genuinely-clean check-all (make clean + full re-extract), never incremental. Added a lint check.
- p16_permute.hide_asm: b64literal-pragma carrier for register pins + GTE __asm__ blocks.
pycparser parses the pragma; decomp-permuter's process_pragmas decodes it back so cc1 sees
the real pins/asm (regalloc steered, mvmva compiles). No submodule edit (reuses its own carrier).
- drop_preproc_and_scalar_typedefs: keep #define + custom struct/typedefs, drop only #include
(fixes func_801412A8 'OTLINK undeclared'); +f32 typedef gap; make_base_c hides asm.
- compile.sh/compile_o0.sh: prepend .include "macro.inc" so GTE mvmva assembles (the real build
gets it via include_asm.h, which base.c omits + -DPERMUTER disables). Byte-neutral for non-GTE.
compile_o0.sh (-O0) auto-selected for _o0 targets.
- run_masked.py: expr_type->int fallback for hidden-pin vars -> 0 internal-permuter-failures
(perm_split_assignment/perm_temp_for_expr no longer KeyError on pinned drafts).
- ALL 5 seeds parse+compile (base 4/110/36/52/77). FLAGSHIP func_80132784 (4/400) CLOSED to a
masked-0 that match_one confirms MATCH (400 ins) -> .run/wave/func_80132784.win.c for T4 gate.
- no build-input changed (136/136 untouched)
- tools/idiom_hunt.py: group backlog near-misses by residual class -> GLM names the reusable idiom +
emits corrected C -> byte-gate to validate; captures reasoning; HARD --budget cap
- CALIBRATION ($0.51 total, 2 classes): struct + regalloc-order -> 0 banks. GLM re-derives our OWN
idioms (register-pin §17, array-of-struct §18, type-width §25) and confirms walls, but banks nothing
new — the backlog near-misses are the residual our idioms already failed on (irreducible/def-side wall).
- VERDICT: the new-idiom well is DRY (Fable5 review confirmed empirically for $0.51, not $300 overnight).
GLM's value stays: direct drafter for FRESH def-conflict-free hard fns (~22%), not an idiom generator.
- tools/glm_reconcile.py (NEW): aim GLM's reasoning at the DEF-side loose-typing wall (body + conflicting
TU decls + reconciliation toolkit -> consistent buildable byte-identical decls); captures reasoning
(.run/glm_reason/, idiom source R16); relax-in-any-TU-file + crash-robust call
- api_draft: REASON=1 saves the reasoning trace per draft (idiom mining on any GLM run)
- fix_arity_callers: --any-proto (relax any prototype, not just (void))
- RESULT: GLM's reasoning is expert-level (store-width/sh-vs-sw awareness, K&R promotion, independently
derives the cast idiom) but banks only 1/7 reconciliations; mechanical relaxation 0/7. The def-side
wall is INTRINSIC (narrow-param + byte-level addressing defeat reconciliation) — Fable5 §3c re-test
CONFIRMS the wall holds even vs a frontier reasoning model aimed directly at it. func_80175184 banked,
check-all 136/136
- 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)
- 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
LM Studio was ejected, so serving is ours now: tools/serve_local.py serves base+LoRA via
Unsloth (.venv-train cu128) as an OpenAI endpoint — no llama.cpp build (this CPU has no
AVX-512, which SIGILLs the prebuilt llama-cpp-python CUDA wheels). api_draft/lora_grind hit
it unchanged.
PROMPT FIX (api_draft.LEAN_SYS + format_finetune.SYS, kept in sync): 'translate EVERY
instruction, never an empty body' — a prompt test took the small-leaf band 0/3 -> 2/3 MATCH
(the v2 corpus overfit an empty void f(void){} leaf pattern). Validated end-to-end: a fresh
ov_SC01_001 batch banked 3 via the local server + better prompt.
CORPUS-V3 (export_pairs + format_finetune):
- export_pairs now ALSO mines the 1623 DEFINE_func macro bodies in engine_core.h (the shared
setters/return-const/dispatchers extract_defs never saw -> 96.6% of v2 was overlay-unique,
the root of the empty-leaf overfit). Corpus 1312 inline -> 2891 (1312 inline + 1579 macros).
- format_finetune inlines engine_types.h structs in the compile-filter so struct-using bodies
are KEPT not dropped: train 2534/2591 (97.8%) compile standalone (v2 was 1111 total).
The grinder/backlog pipeline was ov_SC01_077-hardcoded 5 layers deep (same class as the
T7 lora_grind bug). Fixed all so the permuter grinder can process a non-077 near-miss:
1. gate_stage.append_record stores the source "binary"
2. backlog.FIELDS keeps it (else append_record dropped it)
3. backlog.load_best/_open_stubs is fleet-aware: a fn matched in ov_SC01_077 but
propagation-stuck stays OPEN in its overlay, so it surfaces via that record instead
of being dropped as "matched" (the grinder must SEE it to grind it)
4. p16_permute.setup takes the target binary's asm-subdir (was hardcoded 077)
5. grinder resolves per-binary asm + gates grouped by binary + allows unknown nins
Backward-compatible: legacy records (no binary) default ov_SC01_077.
Validated end-to-end: the 3 fresh reach-134 close=1 ov_SC01_000 fns now surface, resolve
to ov_SC01_000's asm, and gate via ov_SC01_000.
TWO byte-evidenced findings (redirect the fuel strategy):
- the reach>=2 close=1 fuel is MODEL semantic-misses, not permuter fuel: func_8012E27C's
target is "return 1" but the 7B drafted an empty "void f(void){}" (corpus overfit
empty-leaf); func_8012BF4C/AD64 are trivial sw/sh setters drafted empty. A corrected
draft banks them (+3 byte-identical via the fixed gate, @commit:0326); the permuter cannot
add a missing return/store. Lever = corpus-v3 leaf variety, not the permuter.
- x reach is propagation-capped: the 3 are inline-matched in ov_SC01_077_a.c (the stuck-
local cap) -> dedup_propagate "nothing to propagate" -> banked x1. Lever = dedup-collapse.
check-all 136/136 throughout. docs/gen2-mips-matching-model.md + CURRENT_PHASE updated.
Add a sig-based reach oracle + --min-reach N to lora_grind so the mass-run can prefer
SHARED functions (one bank propagates x reach — the fleet-% multiplier). The oracle reads
the same .run/sig.ov_*.jsonl dedup_propagate uses (validated: 0 mismatch over 60 stubs +
the func_8017CE24=2 ground truth), so a reach>=N target is exactly one dedup_propagate
will stamp x reach after the bank.
Bounded reach>=2 mass-run (ov_SC01_000, 15 shared <=15-ins stubs): 0/15 banked, vs the
reach-1-heavy spot-run's 7/15. The model is weakest exactly on reach>=2 (shared) code:
(1) the corpus skipped the shared DEFINE_func macro bodies (export_pairs reads only src/
defs -> 96.6% of the corpus is overlay-unique), and (2) the shared engine fns are the
harder regalloc/schedule residuals. So reach>=2 model-only is NOT a fleet lever by itself.
BUT the reach>=2 drafts are high-value FUEL: 5/15 are close<=3 reach-134 near-misses
(3x close=1: func_8012E27C/BF4C/AD64) -> x134 each if closed. The real lever is
reach>=2 draft -> permuter-grinder close (x134), which needs the SAME per-binary fix T7
applied to lora_grind: grinder.py calls run_gate with no binary (-> 077) and the backlog
stores no binary field. That two-part fix is the next step. byte-neutral: check-all 136/136.
- docs/gen2-mips-matching-model.md: T9 RESULT
- phase-ends/CURRENT_PHASE.md: T9 done; next = grinder per-binary fix, then corpus-v3
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
Rotates over every binary with open small stubs, drafts with the fine-tuned model (api_draft LEAN, GPU
via LM Studio), banks via gate_stage, propagates fleet-wide periodically. Near-misses -> backlog ->
grinder.py permuter closes regalloc/schedule residuals (synergy). Writes a classified near-miss
histogram (the 'missing idioms' signal -> corpus-v3 priorities). STOP/heartbeat/stats like grinder.py.
Smoke OK; histogram shows struct(49)+schedule/regalloc(94) dominate the backlog.
Drew T