- match_one --emit-streams (additive; stdout-identity NC'd): word streams for the
len lane
- family_align.addr_true_rel: reloc-vs-constant range discriminator — FULL
conservative set kept for pair semantics (NC-1 157/157 regression), address-
true subset for indel eligibility only (a constant li-cluster must not read as
reloc-in-indel); synthetic probes green both directions
- tools/len_tells.py: aligned classification + §172b tell tagging (EXTPAIR/
SELECT/NOP) on target-side indels; detectors imported from atlas_features
(R33); cookbook text embedded in cards
- tools/lenmiss_route.py: pool-parallel (A8) — 587 audit LEN rows re-verified
live + routed in 24s: redraft 345 (frac>0.35, APPEND-ONLY backlog
reclassification — near-miss metrics stop lying) / permuter-length 49 (grinder
fuel) / cards 192 incl 14 tell-tagged (the audit's own detectors had emitted
ZERO) / mechanical 0 — an HONEST NULL: stored drafts rarely get constants
wrong; LEN drift is shape, family_align's value here is classifier/detector
- R32 accounting 587/587
- tools/family_align.py (NEW module — classify_member's return contract untouched,
the remap_hseq silent-pass trap avoided by design): SequenceMatcher alignment
over FC.tok streams; li-cluster reconstructor (lui/lui+addiu/lui+ori/li-from-$0
chains, split-cluster absorb for the rs-changed addiu partner); verdicts
LEN-LI/LEN-NOP/LEN-JTBL/LEN-STRUCT/STRUCT-ALIGNED/PURE/IMM; aligned imm engine
mirroring imm_map_tier1 (ordinal deliberately out in v1)
- NC-1 verdict-equivalence 157/157 banked pairs — the NC caught two real gaps:
R-type non-shift sa diffs are STRUCT; registers tested BEFORE the reloc skip
(a reloc-slot word with a different register is STRUCT). NC-2 parity 21/21
- R37 PROBE REFUTED the planned mechanical driver before it was built: 0/26
LI-ONLY cards classify mechanically (regfields x19) — cousins are 0.85-similar
DIFFERENT functions; §168 law 1 re-derived by measurement; no driver written
- family_align re-scoped: its consumer is T8's LEN+N near-miss pile (draft vs
its OWN target = same function); reloc-vs-constant range discriminator parked
for T8. decision-log entry (R31)
- tools/plumbing_groups.py: derives the honest still-open pool from the classified
ledgers (R38) — '1,217 PLUMBING' collapsed to 237 (SELF 109 / CALLEE 48 / OTHER
48 / DATA 32)
- recover_integration: PER-GROUP ISOLATION (git-checkout binary TUs between groups
— one TU-stage edit was poisoning every other group's whole-binary gate with a
phantom shared error; per-group banked_from_source capture) + new stages
'macro-externs' (§121 draft-tier, via family_sweep.macro_def_sig_map, R33) and
'tu-scope' (§103 STU binary-tier, the sweep-only lever)
- the probe (ov_SC03_107): raw 0/14 -> root-caused (poisoning + stale seed
symbols; rtu_match MATCHes them — blind to reloc names, R34) -> symfix-first
-> 9/14 BANKED (64%)
- sweep finding (Law 3): the no-draft majority (ov_SC02_037 44/44, most of
ov_MAIN_012) had verdicts from transient sweep remaps never persisted — family-
lane fuel, not recovery fuel; the stored-draft class is consumed
- cookbook §173 (symfix-first / per-group isolation / verdicts-without-drafts);
index 518 green; R22 clean fleet 213/213; phase total 17 banked @ 0 agent tokens
- decision-log: the P31 re-charter entry (organize-before-grind; R37/R38/R39
ratified at gate-1) per R31
- harvest_verify.py: import guard — a bare import now RAISES loud instead of
running a full gate (CLI unchanged, verified both directions)
- sig-resident: bootstrap boundary artifacts fixed (fused +0 data word with
func_800CEDFC; func_800D33E0 dropped past a glued tail) -> ELF-seeded per the
S45 pattern, exactly 145 fns; true denominator confirmed 145 (progress was
right); audit-corpus 0 PHANTOM + 0 TRUNCATED; all three oracles agree
- family maps regenerated at HEAD commit:2161: 11,025 open non-main members
reconciles EXACTLY with 12,059 - main 1,034 (102 stale phantoms cleared);
adapt cards 704, aprop cards 204 (full emission)
- main fuel-gap finding: 2,001/2,002 main stubs already have cached Ghidra-C
(only func_80049600 missing) — the roadmap '0/2,096' note was stale
- tools-health OK (dedup 2,063/0; C1 254,521/254,521; audit-digest green)
- tools/cc1_dumps.sh: run the pinned cc1 with -dr/-ds/-dj/-dc/-dl/-dg and count standalone
(use (reg)) insns in the .combine dump — each is one 8-byte never-referenced reload slot.
The NEAR-2 residual = 12 such orphans vs the target's 16, now ENUMERATED pseudo-by-pseudo
instead of inferred from ablations.
- The orphan mechanism, exact (combine.c:10835 + mips.md force_not_mem): every short-mem read
is a movhi+shift-pair triple; the fold orphans its ashift temp iff the HI reg carries an
extra HImode use and the death-note walk hits a label/jump. Single-use loads never orphan;
the head can never orphan (walk reaches insn 0).
- The zero-code +4 factory is byte-refuted: 18 new probe families x 3 placements (s16
respellings, cast truncations, <<16>>16 factories, placement sweeps, loop t/n, dossier-file
re-sweep) all land at vars=224-same-bytes or drift. p_optr reproduces the target's exact
16-orphan frame at +7 insns — opacity that defeats cse equally blinds num_sign_bit_copies;
fold-const closes the tree level. Three canonicalizers, one wall.
- Dossier NOTES.md updated with the S50 ledger; backlog klass -> WALL(P32) for both names.
- NOT banked: no 0-closeness draft exists. The dial draft (NEAR 2) remains the floor.
Three carries a mechanical seed-body draft needs beyond the symbol rebase (cookbook §171b),
each found from one compiler verdict:
- DATA DEFINED INSIDE THE MEMBER'S OWN .s must be DEFINED, not externed — it vanishes with
the stub it lives in. Re-initialised with THIS member's bytes (the 0x801F1CD8 family carries
4 distinct 8-byte patterns across 42 members). Flat-byte-list initializers only; refuse the
rest rather than mis-initialise something the gate would reject unexplainably.
- SHARED TYPES the destination cannot see: MATRIX/SVECTOR live in engine_types.h, which md_*
TUs never include. `parse error before 'm1'` was the only thing between 4 of 9 members and a
bank. Carried brace-matched, vetoed by the destination.
- A POSITIONAL LITERAL MAP where imm_map_tier1 gives up: it refuses a value that also appears
at a non-differing slot (0x10 collides with the struct offsets), but the 8 differing slots
map 1:1 onto the C's call sites in order. Asserted ([C literals] == [seed slot values])
before substituting — that assert is the whole safety argument. 10/10 refused -> 9/9 banked.
- ROOT-CAUSE FIX: body_text matched `extern void func_X(...);` at column 0 and returned the
NEXT function's body — silently shipping wrong seed bodies, visible only as "no definition
after rename" skips. A definition is now confirmed by a `{` with no `;` before it.
R22 clean rebuild: check-all 213 passed, 0 failed of 213. Stubs 12,161 -> 12,111.
Fleet 95.3% instr / 90.0% distinct / 96.66% fn-count.
Final S50 state: 307 instances banked, stubs 12,468 -> 12,161, fleet 95.3% instr / 90.0%
distinct / 96.65% fn-count. R22 clean rebuild 4x, check-all 213/213 every time.
- tools/aprop_autodraft.py + tools/draft_prechecks.py: seed body + symbol_map + a MINIMAL
synthesized preamble. The seed's decl layer never travels — that layer is family_sweep's
dominant failure (331 of 458 S49 verdicts). 256 banked at zero agent tokens, against the
~20M the same work would have cost as a wave.
- Macro seeds (567 of 1196 members, all 3737 de-macroize) take the DEFINITION only; the block
stays the decl source. Pasting it whole measured 28% vs inline's 68% — func_8016AB6C's macro
is 1,891 lines of which 108 are the function.
- IMM is a second engine, not a wall: T2a's imm_map_tier1 resolves a per-location LITERAL like
symbol_map resolves a per-location SYMBOL. 131 of 275 IMM members resolve.
- draft_prechecks negative-controlled against ALL 205 banked drafts: zero false positives,
catches 39 of 67 known failures. That control found two bugs in the checks themselves —
C89 `f()` declares UNSPECIFIED parameters (not zero), and a member's own definition read as
a call to itself. Conservative by design: a pre-check that discards good drafts is worse
than one that lets a few builds fail.
- The A-prop pool is now priced exactly: PURE 437/37,376 ins, IMM 275/8,849, STRUCT 238/4,259.
- Cookbook §171a; SETUP rows; CURRENT_PHASE S50 FINAL checkpoint.
Measured on the 489-draft run: inline seeds banked 145/213 (68%), macro seeds 77/276 (28%),
and the macro failures were decl collisions to a fault — 'parse error before *', 'too few
arguments to function'. Cause: a de-macroized DEFINE_ block is decls + definition, and
pasting it wholesale reintroduces the exact decl-agreement failure the minimal-preamble
design exists to avoid. func_8016AB6C's block is 1,891 lines of which 108 are the function.
- seed_body() now returns the DEFINITION as the body and keeps the macro block as the DECL
SOURCE, so only the externs the definition actually references are emitted.
- aprop_symfix.body_in_text(): text-based sibling of body_text (R33 — one implementation).
Indented-definition fallback for macro blocks, confirmed by a '{' with no ';' before it,
so an indented CALL can never be mistaken for a definition.
A lane-A member's body IS its matched sibling's body with the per-location symbols rebased.
Two mechanisms had each left that population on the table: family_sweep remaps but CARRIES
the seed's decl layer (decl-agreement = 331 of 458 S49 failures), and the agent wave
re-derives the same body at ~80k tok/fn. This does neither — seed body + symbol_map
(positional reloc zip) + a minimal synthesized preamble.
- Byte-proven before the tool existed: a hand-written draft of exactly this shape banked
func_801E2858@md_SC03_132, in a family the mechanical sweep had refused. Then 9/9 banked
on the first generated batch.
- THE SELECTOR, measured: only PURE members (every diff site a RELOC) are drafted. An IMM
(per-location literal) or STRUCT (register/opcode drift) site cannot be reached by a
rename — the first 5 gated 0/5, all IMM or STRUCT. classify_member now decides that
BEFORE a build is spent, not after. --allow-impure to override.
- Macro seeds supported: DEFINE_<fn>() de-macroizes back to a plain self-contained block
(3737/3737 extract). They were the largest skip bucket — 567 of 1196 members.
- 592 drafts emitted (304 macro / 288 inline) from 1196 members; every skip is named (R32).
- aprop_symfix: asm_syms_ordered() so an n:m card hands over the target's reference ORDER.
- Two defects caught by probes before they scaled: a typedef regex that stopped at the `;`
INSIDE the struct braces, and a continuation walk that started on the empty pre-newline
element and returned an empty body.
- The blocker was carried as "one missing file-scope extern gates 83 PURE members". Both
halves were wrong (R14): corpus.stubs says 4 open members, and D_801ED98C is a DEFINED
const Blk8 whose rodata lives inside the member's own nonmatchings .s — replacing the stub
deletes the data with it. gather_externs can carry an extern DECL, never a DEFINITION,
which is why it reported "no file-scope decl" for a symbol md_SC05_023 defines on line 114.
- Fix: paste typedef + const definition + body per sibling (data bytes verified identical
across md_SC05_024/025/028/029). 4/4 banked.
- aprop_symfix: new `local-only` class — draft-DEFINED identifiers that merely carry a
vram-looking suffix (Blk8_…, S8_…, L_call_…) are not stale symbols. Measured: that is every
non-clean case in the whole wave-7a/7b stored-draft residue, which holds ZERO stale-symbol
recoveries (a clean negative result — the defect was A-prop-specific).
- cookbook index regenerated (tools-health fails closed on a stale index — it caught §171).
- R22 clean rebuild: check-all 213 passed, 0 failed of 213. Stubs 12,445 -> 12,441.
- REFUTES §170's open hypothesis (batched cards concentrate members into one TU ⇒ §169
collision): 5-draft groups banked 5/5; 11 of 35 unbanked drafts were already one-per-TU;
and the two "concentrated" groups banked 12/12 and 10/10 once the real defect was fixed.
- The cause: a per-location data symbol carried out of the seed body unrebased. match_one
compares instruction ENCODINGS and is blind to a relocation's target NAME, so it scores
MATCH standalone and dies at link in the host TU. 24 of 24 concentrated failures, all 1:1
rewritable at one constant vram delta (0x4128).
- tools/aprop_symfix.py: audit + --fix, emits a gate_lane-shaped slate; deterministic and
build-free, so it runs BEFORE the gate. The R34 second oracle for the class match_one
cannot see.
- family_cousins.py --aprop-cards: members now carry sym_map, the explicit {seed -> member}
renames, read from the seed's C BODY (a matched seed has no .s of its own) vs the member's
.s. Two case-mismatch defects fixed while wiring it (sig lowercase vs splat uppercase).
- 23/24 banked. Stubs 12,468 -> 12,445. Fleet 95.2% instr / 89.9% distinct / 96.57% fn.
R22 clean rebuild: check-all 213 passed, 0 failed of 213. dedup 2,043/0.
- A-prop's true conversion is 87% (79/91); the 320 batched members are unblocked.
- Cookbook §171 + §170 struck in place; SETUP row; decision-log (R31).
- checkpoint block refreshed for a fresh session (R30/checkpoint-before-pause): the TU-spread
test is the named FIRST action, the regen chain and gate contract are spelled out, and the
>=16 head's three open items are listed with their evidence.
- promote gate_lane.py into tools/ (it was scratchpad-only): accepts slate OR confirmed shapes,
explicit draft paths, R32 coverage assertion (refuses to report 0 as a result), dirty-tree
abort, no outer timeout, per-function propagation after.
- session: 179 commits, 177 instances banked, 5 R22 clean-fleet gates all 213/213.
- NEW family_cousins.py --aprop-cards + tools/wave/aprop_wave.js: lane A (1,700 open fns /
76,419 ins) had NO card type — cousin diffs are empty for h_seq-identical members, so the card
is a positional WORD diff vs the matched sibling, grouped BY FAMILY (one agent, N drafts).
Head cards: 13 families / 433 members, median TWO differing words each.
- calibration 9 batches / 108 members: 98 agent-MATCH (91%, best of any wave) -> 56 BANKED (57%),
~80k tok/banked fn vs 157k (cousin card) vs 400k+ (crack wave). R22 213/213 BYTE-IDENTICAL.
- HONEST GAP (R14): 91% agent -> 57% gate is the worst conversion measured; 14 groups banked 0.
Hypothesis TESTABLE not proven — family batching concentrates members per destination TU, the
§169 collision. Re-gate unbanked ONE PER TU before scaling the remaining 320.
- >=16 head diagnosed: 3 of 4 blockers are plumbing — the --band substantial default hid 5 of 13
families from every prior sweep; one missing file-scope extern (D_801ED98C) gates 56 PURE
members; dedup_extend is macro-only. Only func_8017C294 is a genuine crack.
- fleet 96.56% fn / 95.2% instr / 89.9% distinct; stubs 12,535 -> 12,468; dedup 2,043/0.
- cookbook §170.
- thresholds relaxed to <=6 blocks/<=16 tokens UNION edit-fraction <=0.20: cards 518 -> 721,
MIXED 310 -> 50 skeletons; the 753-ins func_8017BEBC (0.987 sim) became reachable.
- 59 cards -> 48 agent-MATCH (81%) -> 44 BANKED (92% MATCH->bank, 75% end-to-end), 6.9M tok.
- FINDING (the actionable one): 7b's bank rate crushed 7a's because it SPREAD 48 drafts over 35
destination TUs; 7a's failures were per-TU declaration collisions between sibling drafts.
Cookbook §169 updated with the spread law.
- R22 213/213 BYTE-IDENTICAL from clean; fleet 96.55% fn / 95.2% instr / 89.9% distinct;
stubs 12,584 -> 12,535; dedup 2,035/0.
- incidents 3 & 4 recorded: an agent wrote a TRACKED header (guard caught it, prose is not
enforcement); my own gate_lane filtered on the wrong key and printed 'gating 0 drafts' as a
result (R32 silent skip) — fixed with a coverage assertion that refuses to report 0.
sig_image emits lowercase func_<hex> names; splat writes the .s with UPPERCASE hex, so
corpus.asm_path returned None for every address containing a hex letter — 60% of the cards
would have shipped agents a 'null/func_x.s' path. resolve_asm() tries the sig spelling
(authoritative for CURATED names), then both hex cases; a card with no resolvable .s is
now DROPPED, not emitted (R32). 518 cards, 0 bad paths.
- family_cousins.py --adapt-cards: per seeded-unit member, drift classified vs the seed
(LI-ONLY 27 / SMALL-EDIT 491 / MIXED 310 excluded); cards carry the seed C location + the
aligned diff blocks with the member's raw words + disasm (the new constant is readable in
the card). 518 cards / 1,101 instances / 23,820 ins; 514 haiku-band.
- tools/wave/adapt_wave.js: the EDIT-contract wave (crack_wave contracts preserved: per-agent
dirs, sha1-last, UNVERIFIED != refuted); symbol surface from the TARGET .s; haiku<=60/sonnet.
- regen chain absorbed the 48 lane-A banks (A-prop open ins -6,475 == the report's instr
delta exactly — two independent derivations agree); pilot slate .run/wave7a_pilot.json (30).
- R37: pilot before scaling to the 518-card pool.
- FINDING (Drew's smell, byte-verified): the '4,513 unique singletons' picture is substantially
an h_seq exact-hash artifact — 86/120 near-pairs in the 0.85-0.99 band differ by PURE
insertion/deletion (li-expansion tell in 25). Specimen: ov_SC06_010:0x8017bebc (753 ins,
'singleton') is 0.987-similar to a MATCHED fn in the same binary.
- NEW tools/family_cousins.py: distinct open skeletons -> shingle index -> >=0.85 union-find ->
matched-seed attachment -> .run/family_cousins.json + docs/family-cousins.md. R32 BOTH ways
(independent stub recount fails loud on a stale map — negative-control-proven; partition
assert). Reproduced the probe within +-1%; totals EXACT (11,627 inst / 584,448 ins).
- Unit table: A-prop 197u/68,729ins · seeded 418u/50,422 · cousin-multi 1,552u/249,799 ·
cold 3,240u/215,498 — the genuinely-unique tail is 37% of the remainder, not 90%.
Main's 'structurally barren' HOLDS at the similarity tier (94% mass <0.70).
- --targets wave slate: .run/wave7_targets.json = 40 targets / 33,304 unit ins (+33% vs
family-ranked), 9 resolved seed C paths, size-routed 2 haiku/20 sonnet/18 opus.
- LAWS (§168): a cousin is a SEEDED CRACK never a remap; rank waves by UNIT weight; discount
short-fn similarity. Byte-gate stays the sole arbiter (G3/P9).
- docs/family-hseq.md: this session's frontier regen (post-S48 propagations) rides along.
- cookbook §168 + SETUP inventory row (R16/R21/R30); CURRENT_PHASE S49 entry.
gate_stage labelled every "standalone MATCH / whole-binary DIFF" with
"(declaration/TU plumbing)". The tool never checked for a declaration conflict —
that was a GUESS printed as a diagnosis, and func_8017F2D4 carried it through
SEVEN attempts across five waves while every agent hunted codegen. The body was
byte-correct from the first attempt; the notes had simply named the wrong
destination TU (a file holding only a caller + prototype), and splicing there is
a no-op that leaves the INCLUDE_ASM bytes in place.
- gate_stage now says only what is true (the two oracles disagree) and hands
over the check that resolves it, instead of naming a cause it did not measure.
- §166a banks the oracle: asm/<ov>/nonmatchings/<TU_stem>/<fn>.s => the
INCLUDE_ASM is in src/<ov>/<TU_stem>.c. The third path component IS the TU
stem, derived from the split config, and it beats any prose citation — a grep
for the function name also hits callers and prototypes in OTHER TUs and reads
exactly like a destination hit.
- Plus the two probe gotchas that cost wave-5/6 agents real time: the wrong
--aspsx-version fakes ~32 ori-vs-addiu mismatches, and a collateral-drift
check must filter to sized symbols (nm -S) or the zero-size .NON_MATCHING
aliases all report false drift.
cookbook_index.py: 506 sections.
A markdown backtick inside the prompt's template literal terminates the string
and the workflow dies at parse time. Cheap (0 agents, 0 tokens) but it has now
cost two launches, so the file carries a warning at the top: use double quotes
for inline code in prompt prose.
Wave 5 wrote its drafts into .run/wave4/ because the script was built by
sed-ing the wave-3 copy. Harmless (per-function dirs) but wrong. Pass
{wave:'waveN'} on any target; defaults to '.run/wave/' so a missing field can
never silently reuse a prior wave's directory. Meta name/description are now
generic too — this is THE crack-wave script, not wave 4's copy.
A prior attempt that was GATE-REFUSED (standalone MATCH, whole-binary DIFF) has
an in-TU residual, so the standalone gate cannot see it. A wave-5 agent closed
exactly this by splicing into the real destination TU, running the pinned
triple end-to-end, and masked-diffing the function out of the WHOLE-TU object
(0 mismatched, 279/279). That is now instruction, not luck.
Carries both gotchas it paid for: the wrong --aspsx-version produces ~32
spurious mismatches all of the ori-vs-addiu li-form shape (the fingerprint of a
version mismatch, not a codegen residual), and engine_core.h resolves relative
to the including file so the spliced TU needs a directory + shared symlink.
Also warns that two wave-5 agents found their predecessor's file/line citations
wrong while its idioms were right — re-derive the TU, don't trust quoted lines.
crack_wave.js classified anything without verdict_check.confirmed as 'refuted'.
When the usage limit killed 22 verifiers mid-wave, the result read
'refuted: 22' — 22 good drafts reported as rejected, with evidence 'verifier
died' as the only tell. Acting on that would have discarded the wave.
Same disease as the no-diagnostic classifier (S47) and this session's
poisoned-tree 0/17: a tool stating a conclusion it never reached. UNVERIFIED is
now its own outcome, returned with the draft path and sha1 and the instruction
'VERIFIER NEVER RAN — re-verify, do not discard'.
A workflow script is plain JS; the cookbook citation I added used markdown
backticks inside a template string, which terminated it. Caught at launch
(parse error, 0 agents, 0 tokens) rather than mid-wave.
The wave script encodes contracts paid for in real failures (per-agent output
dirs after wave 2 lost 21 verified drafts; sha1-last verification; prior-notes
seeding at 7/9 then 10/12; size routing). It was living only in the workflow
scratch dir, so each wave rebuilt it from memory — which is how the cookbook
citation list went stale: it still named only §162 while §163 and §164 (87 more
entries) had been banked in between.
Updated the citation list to §160-§164 and pointed agents at §164z, the 28
REFUTED claims, so they do not re-derive a disproved mechanism.
Wave 4 (wf_05895a19-121, 75 agents, 7.9M tok): 40 targets -> 35 agent-MATCH,
0 refuted, 5 NEAR, 0 FAIL, 0 drafts lost. 18 banked so far on the whole-binary
gate across 9 binaries; the other 17 are re-gating on a clean tree (see below).
PRIOR-NOTES SEEDING HELD AT SCALE: 10 of 12 seeded targets confirmed (wave 3
was 7 of 9). func_8017C294 — the x16 family, the largest single item on the
board — is now NEAR at **2 ins** (18 -> 11 -> 2 across three seeded attempts).
THE 4th COMMENT-BLINDNESS DEFECT OF THE SESSION, and the first with blast
radius. A crack agent annotated a decl in its draft:
extern void func_801842DC(s32 a0); /* TU:4023 INCLUDE_ASM (no decl) */
`corpus._INCLUDE_ASM_CAND` only skips lines that BEGIN with a comment marker,
so it read `INCLUDE_ASM (` out of the trailing PROSE, found no quoted path, and
refused the whole binary's stub oracle — correctly, by its own R32 contract.
That then failed gate_stage for every LATER binary in the run, because they all
walk the corpus: 12 binaries banked, then 5 were blocked by one comment in a
13th. Fixed the same way as the other three today: decide candidacy on
cdecl._mask'ed text, PARSE FROM THE ORIGINAL (the mask blanks string content
and would erase the asm path). Verified on 4 binaries incl. main (2,002 stubs).
AND THE DAMAGE IT LEFT: gate_stage raised out of the CorpusError BEFORE its
revert, stranding failed drafts spliced in src/. The 17 solo re-gates that
followed all read 0/17 — they were building a POISONED TREE, not judging their
own drafts. Residue reverted here; the 17 re-gate clean next.
The pattern is now sharp enough to state: any scanner that greps C source for a
token must mask comments and strings FIRST — and agent-authored drafts make it
far likelier, because their prose mentions the exact tokens our tools hunt for.
Today a 28-agent wave lost 21 adversarially-verified drafts to a shared output
directory, and I wrote them off before Drew asked whether the workflow results
could just be analysed. They were all recoverable, for zero agent tokens.
Encodes the method that worked 21/21, including the two shortcuts that do NOT:
- taking each Write's content recovers only single-write drafts (8/21 — agents
refine);
- taking an Edit's new_string as a file yields a FRAGMENT, not a file.
So it replays the mutation history per (agent, file_path), snapshots after
every mutation, emits newest-first, and also scans Bash heredocs (the 21st
draft never used Write/Edit at all). --gate runs match_one newest-first and
keeps the first MATCH.
Self-test on wf_d804f25a-f6f: 3/3 including the heredoc case.
`jtbl_family_bank` fed every module jr member to `jtbl_carve`, which died with
`jtbl_… not found in the raw data asm`; `harvest_verify` turned that into
CARVE-REFUSED and never built. So the verdict named the TOOL, and 12 slots in
the wave-1 propagation read as a carve bug. Probing one member to the byte
level shows it is a LAYOUT the carve model does not cover:
A module binds `.rodata` at 0x0 to the SAME subseg as its code (§154-A), so
the object's rodata order IS the C file's include chain — INCLUDE_RODATA
pieces, then each INCLUDE_ASM'd function's MIGRATED table, in address order.
That reproduces the island exactly while the function is a stub. Matching it
PRUNES its .s, its table leaves the chain, and cc1 re-emits it at the END of
the object's .rodata: build 43,768 vs 43,760 bytes, first diff at 0x144
inside the island's own pointer table.
`JTBL_PADS` does not reach it either — `jtbl_rodata_pads` refuses the object
outright ("unexpected rodata content .include ... D_801EF468.s"): the carve
model covers jump tables, not an island of mixed included data.
- `migrated_tables()` detects the layout by EVIDENCE (table absent from the
data asm, present as a dlabel in the function's own .s), refuses loud with
the measurement and the design that would work (isolate the jr function into
its own subseg so its .rodata is a separate OBJECT, then ld_interleave — the
§8 machinery re-aimed at a LEADING island instead of a data tail), and
refuses a mixed carve set rather than half-carving (R32).
- Regression-checked both ways: overlay stubs classify [], modules classify
migrated.
SIZED (R37): 70 module binaries, 42 with this layout; 1,345 open module
member-slots in sibling families, of which only 44 are jr. The island work is
worth 44 slots — it is NOT the module lane's main gate.
R22 clean-fleet: 213 passed / 0 failed of 213.
`_body_open_brace` ran BOTH its scans on unmasked text. A crack agent's draft
opens with a header comment that names the function and quotes C at it:
/* func_801EE8E0 (ov_MAIN_012 / jr_801789AC) — 188 ins, byte-exact vs …
* 3) The `do { } while (0)` around the loop-1 call is a REGISTER-ALLOCATION
so `sig` matched the COMMENT's first line and `find('{')` found the comment's
`do {`. Every carried `extern` was spliced into the comment — silently
commented out — and the gate reported `'D_8011511A' undeclared`.
The sweep classified that CC1-FAIL, so it read as a property of the SIBLING
(all 4 members failed identically) when it was a property of the EXEMPLAR'S
PROSE. It had nothing to do with module binaries, which is where I had filed
it. Every richly-commented agent draft is a carrier; the trigger is any brace
inside the header comment — so this would have grown with the campaign.
- both scans now run on `cdecl._mask`ed text and index the original by the
masked offsets (§134 / R33: one masking oracle);
- refuse outright if the length invariant is broken, rather than mis-place a
declaration into live code (R32).
Measured: family func_8017CBC8 -> its 4 md_ siblings went 0/4 -> 4/4 banked.
The test was `^\s*typedef\b[^\n]*\bNAME\b` — the name must sit on the SAME LINE
as the keyword. True of `typedef unsigned char u8;`, never true of the
multi-line form the preamble backscan actually carries:
typedef struct Foo { … } Foo;
so every multi-line typedef already in the unit was carried a SECOND time and
the unit reached the gate with two definitions of one tag. canon_sig_reconcile
uniquifies draft tags, so the duplicate is exact: `redefinition of struct
Foo_8013C0F8`.
- Extract the block capture as `_typedef_blocks(lines)` and derive the
already-carried set from it (R33: one parser, two callers), so the
multi-line form is recognised exactly as the single-line form always was.
Measured on the 0b population: func_8013C0F8 (3 slots) carried Foo+Bar twice;
func_8013B83C dropped a redundant file-scope copy of a typedef its body
declares at block scope. 6 other families byte-identical output.
`_proto_from_lines` starts `_strip` with in_block=False, but item boundaries
are `;`-terminated — so a declaration whose TRAILING comment wraps hands the
comment's continuation to the NEXT item, and the implied prototype came out as
extern * a prototyped (s32) decl is `conflicting types` … */ void func_80151664(void);
It compiled only because the hoist emits the opening `/*` line immediately
above it, so the garbage lands back inside a comment — but `_file_scope_decls`
then meets a col-0 `extern …;` whose base type is `extern` and REFUSES (R32).
That is the isolate-fail class: 23 of the 111 open 0b member-slots.
- overlay_src_split._proto_from_lines: apply family_remap's D1 backstop — a
`*/` with no `/*` before it means the chunk opened inside a comment; drop
that residue before parsing the header.
- Repaired the 16 already-emitted region files (the garbage line's live
payload was a redundant `extern void func_80151664(void);`).
Byte-gate after the repair: ov_MAIN_012 / ov_SC02_037 / ov_SC03_107 all
BYTE-IDENTICAL.
The 0b blocker was not "the pads line is left behind" alone — it fails two
different ways, and the second one is silent:
* bare isolate + `make build`: the stale line arms the pads filter on the
RESIDUAL object, which emits no jump table ->
`jtbl_rodata_pads: consumed 0 rodata .align(s) but 4 pad spec(s) given` (S47).
* isolate -> jtbl_carve (the jtbl_family_bank path): `set_pads_vars`
regenerates the block keyed by the CURRENT subseg names, finds no prior spec
under the new `_jr_<addr>` name, and DROPS the line. cc1's natural `.align 3`
then pads the span's non-8-aligned interior tables and the image shifts —
reported only as `built, bytes differ`.
- jr_isolate_all.repoint_overlays_mk: repoint the `build/src/<ov>/<sub>.o:
JTBL_PADS` target with the `--order` leaf whenever a carve moves; refuse
loud if the old object still hosts a .rodata piece (R32).
- jtbl_carve.set_pads_vars: second, disagreeing oracle (R34) — refuse when a
spec would vanish for a subseg no longer in the carve set (rename/merge
drift), instead of silently emitting a padless object.
R37 probe: func_801789AC -> ov_SC02_037 went `built, bytes differ` -> BANKED
on the whole-binary byte gate. ov_SC02_037's spec is 0,0,0,0 over tables
+0x0,+0x14,+0x34,+0x4c — load-bearing (span start is 4 mod 8).
Stage 0a's first defect, and the largest single zero-token bank of the session.
family_remap's kind test asked ONE question — is this address a function in the SIBLING'S OWN sig?
— and defaulted to `D_` on "no". But a body calls outside its image constantly: an overlay calls
resident helpers, an md_* module calls the overlay-range engine. Those addresses are absent from
the sibling's sig, so the test fell through and emitted a DATA NAME FOR A FUNCTION —
`D_800183E0`, `D_800D1EBC`, `D_80171A1C`. None exist anywhere in src/ or config/symbols.us.txt,
while `func_80171A1C` alone has 1,061 references. Measured: 611 member-rows across 45 symbols,
the largest named residue class. "Not in MY sig" means "not mine", not "is data".
Fix — three oracles, strongest first, never a blanket fallback:
1. the sibling's own sig (authoritative for its image; this is what preserves the Phase-29 T82
case where a slot is a function in the exemplar and DATA in the member — unioning every sig
would have re-broken the 251 members T82 fixed),
2. the always-linked images via extern_fn_addrs() — resident + main, 2,146 addresses whose ranges
cannot collide with an overlay's,
3. the exemplar reached it by `jal` — a call target is a function BY DEFINITION, which covers an
external address neither sig claims (0x80171A1C from an md_* module, 112 rows).
Only a non-call reloc no oracle claims still falls to `D_`.
Result: BANKED 205 member-matches, failures 670 -> 575, derived net = report = 205.
R22 clean-fleet 213 passed / 0 failed of 213.
Fleet 94.4% instr / 88.3% distinct / 96.27 -> 96.33% fn-count; stubs 13,563 -> 13,345.
THE RESIDUE HAS CROSSED OVER: DIFF is now the LARGEST class at 143 of 575 — real byte divergence
outranks plumbing for the first time this session (undefined-ref 611 -> ~8, PLUMBING-other
231 -> 81). The ~5:1 plumbing:DIFF ratio that justified "tooling beats volume" has inverted in this
queue, exactly as the frontier analysis predicted: the declaration-axis vein was one-time.
The new head class is `conflicting types for func_80175414` (27) — the same addresses this fix
started naming correctly, now surfacing the NEXT layer (the symbol resolves; its declared signature
disagrees). That is the conform axis, not the remap axis.
Note for anyone auditing this class: rtu_match MASKS HI16/LO16, so a wrong %hi/%lo symbol still
reports MATCH (the T82 comment records `MATCH (10 ins)` on a member the fleet gate refused). This
defect is invisible to the per-function tool by construction — only the whole-binary gate sees it.
The targeting oracle stamped its scope as "the N OVERLAYS only (no main, no resident)" while
load() has scanned the md_* modules and the resident since S44. Measured at this HEAD: 141 location
overlays + 70 md_* modules + the resident = 212 binaries. That is the §159 coverage law broken by
the file that documents coverage, on the repo's most load-bearing targeting instrument — and it is
how "main is structurally barren" survived two phases unexamined.
The COUNT beside it was already derived, with a comment saying "report the scope we ACTUALLY
scanned, never a hardcoded count". The PROSE describing what the count meant was hardcoded and
rotted. Both are derived now.
Caught while fixing it: my first cut read glob(".run/sig.main.jsonl") and stamped "main INCLUDED"
the moment that file existed — while load() still did not glob it. Same defect one layer down: a
stamp describing the filesystem instead of the run. Now derived from the loaded instances.
Also added a glossary line: "zero-crack" means n_matched == 0 (needs its FIRST crack) in this map,
and the OPPOSITE (a matched exemplar awaiting propagation) in roadmap §3 T3 — a ~30x mis-scope risk
for any session reading one against the other.
NOT DONE — main inclusion (0c) is still blocked on settling the attribution. Confirmed the
mechanism: main has 49 LINKED PsyQ subsegs, corpus.stubs('main') returns 2,002 INCLUDING them,
progress.py correctly excludes them and reports 1,034 game-code stubs. progress.linked_subsegs'
own docstring records this exact trap ("an importer then classifies ~1,300 already-byte-identical
LINKED library stubs as outstanding game-code work") — and my sig-main seeded from corpus.stubs,
so it inherited the LINKED rows, which is why G2's 207-family finding was inflated.
My partition probe is NOT trustworthy: 954 of 2,002 stubs returned no asm path from
corpus.asm_path, so 199 LINKED / 849 game / 954 unresolved does not reconcile with 1,034. Fix the
probe before trusting any main-scope number.
func_801EDC18 (md_SC05_023) is the largest multiplier remaining — 57 members. The wave agent
abandoned it at "closeness 6" with class SIZE-MISMATCH [redraft]. It was two lines from correct.
THE CODE (cookbook §160a): the target copies 8 bytes with lwl/lwr + swl/swr — gcc-2.7.2's
emit_block_move for a type with ALIGNMENT 1. The draft used a u32 copy (aligned lw/sw), which is
wrong by construction. `typedef struct { char c[8]; } Blk8; buffer = D_801ED98C;` reproduces it.
Six spellings were tried in parallel; two independent agents converged on the same one.
THE INSTRUMENT (§160b) — this is the part worth more than the function. The target .s bundles a
leading `.section .rodata` block (D_801ED98C as two .word) ahead of .text. Those lines carry the
same `/* off vaddr HEX */` shape as instructions, so masked_diff.insns_from_s counted them as TARGET
instructions, while insns_from_object (objdump -j .text) can never emit them. A byte-perfect draft
therefore read `mine=26, target=28, 26 mismatched` — every position shifted by a constant +2 — and
got classified as needing a redraft. 116 of 12,583 .s files in the corpus have this shape, one at
-29 instructions. Every one of them would report a false wall to any agent that tried it.
Fixed: insns_from_s tracks .section and counts only .text. Full-corpus control: 12,467 unchanged,
116 corrected, 0 regressions. Same artifact class as §129a (post-carve jtbl inflation).
THE OWNERSHIP LAW (§160c) — my own error, corrected by the gate. Four sites declare
`extern short D_801ED98C;` and nothing in src/ defines it, so I shipped an extern-only draft. The
gate refuted it: `undefined reference`. The .s block the draft REPLACED was the definition. The
variant emitting `const Blk8 D_801ED98C = {{...}}` banks clean. Never infer ownership from externs.
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
ALSO BANKED — the wave's idiom harvest, which had been sitting unwritten in workflow transcripts
(R16/R30 debt): §160d the ASYMMETRIC INDEX RELOAD (a just-stored narrow field read twice emits
reuse-then-reload; the C is deliberately asymmetric), §160e a stack-layout scheduling rule now
byte-proven on a SECOND independent function (promoting it from coincidence to rule), §160f the
address-only global store via array decl, §160g sibling-search keyed on the CALLEE SET as step 0 of
every wave prompt (one grep turned a 126-instruction crack into a copy-edit).
Cookbook index regenerated: 468 sections.
scope_data_fix detected conflicts by SCANNING TU TEXT, which cannot see a MACRO-INJECTED
declaration — and that is where these conflicts live: `extern Vec8 D_80114F24;` sits inside a
DEFINE_func_* macro body in engine_core.h while the overlay .c holds only `DEFINE_func_XXXX()`.
That declaration is a genuine file-scope decl of every TU invoking the macro, and it is what the
draft collides with.
Fix: family_sweep builds a per-TU map from cdecl.tu_scope (cpp-derived, cached — 467 data decls in
ov_MAIN_012 vs 0 findable by text scan) and passes it to scope_data_fix, which now takes an
optional tu_types. tu_scope is the repo's stated oracle for "what does this TU declare", and its
own docstring warns that the `above` form answers VISIBILITY, not CONFLICT — C requires
compatibility regardless of order. scope_data_externs was built on that wrong question.
Banked 5 -> 31; failures 730 -> 699. The whole data-symbol class is gone: D_80114F24 (12),
D_800AE620 (10), D_80126B58, D_80078EB4, D_800183E0 all cleared.
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
REGRESSION I CAUSED, AND FIXED: the reclassification showed 9 fresh `conflicting types for
aD800B9A02` — my own alias colliding. The group-level path suffixes the alias per function
precisely to prevent this; scope_data_externs did not, so two drafts aliasing one symbol declared
`aD800B9A02` twice with different types — re-creating the collision one level down. Both paths now
use _alias_name(sym, func). Verified: distinct names, each keeps its own type, both bind the real
symbol via the asm label.
family_sweep stages every member of an (overlay, split) group into ONE TU before gating, so two
templated bodies routinely carry different views of one address — D_80114F24 is `s32` in one body
and `Vec8` in another. scope_data_fix is handed one draft plus the pre-splice TU and structurally
cannot see the others, so that collision was invisible to it.
_alias_group_data_conflicts(): after staging, any data symbol a group's drafts declare with >=2
distinct types gets a PER-DRAFT §37 asm-label alias. The alias is suffixed with the function name
(aD80114F24_8017B880) — aliasing both drafts to a shared name would re-create the same collision at
one remove, which the unit test exists to catch. Each body keeps its own type: for data the declared
type drives the load (lh vs lhu), so canonicalising would silently change codegen for every other
view. Codegen unchanged — the asm label pins the emitted symbol.
83 conflicts aliased across 172 groups; banked 2 -> 5. R22 clean-fleet 213 passed / 0 failed of 213.
WHY THE HEADLINE SYMBOLS DID NOT MOVE — root cause now CONFIRMED, not inferred. D_80114F24 (12)
and D_800AE620 (10) are unchanged because the conflicting declaration is MACRO-INJECTED:
`extern Vec8 D_80114F24;` lives inside a DEFINE_ macro in engine_core.h (D_800AE620 has 6 such),
while the overlay .c holds only DEFINE_func_XXXX() invocations. Neither a scan of the staged drafts
nor a scan of the TU text can see it — that needs the preprocessed TU. The sweep already does
exactly this for CALLEES (cast_call_sites' canonical map is cpp-derived "so it sees macro-injected
declarations"); the data path never got it.
This collapses F2's remainder and F4 into one fix: memcpy's 26 failures are the same shape — task B
found nine `extern void *memcpy(...)` spellings inside those same DEFINE_ macros. A cpp-derived
declaration map feeds both, and the alias mechanism is already built and control-tested; only the
detection SOURCE is wrong. For memcpy the alias is the documented house solution, not a workaround
(engine_core.h:24480 hand-writes `extern void func_8005C324(...) __asm__("memcpy")`).
Four attempts on this class for 5 members: three mechanisms proposed before reading what the
compiler actually complained about. The mechanisms are correct; they targeted the wrong collision.
The 44 data-symbol conflicting-types failures are NOT the cdFileLocTable duplicate-typedef class.
They are genuine per-view type differences: D_80078EB4 is s16 at 2,409 sites and u16 at 1,341;
D_800AE620 is Blk20/s32/Mat32. For data the declared type drives the load (lh vs lhu), so
canonicalising would rewrite thousands of already-banked sites' codegen. The answer is one type
PER VIEW — the §37 asm-label alias the fleet already hand-writes for D_800AE620 (9 sites).
Shipped: scope_data_externs now auto-aliases a conflicting extern —
extern s16 aD80078EB4 __asm__("D_80078EB4");
keeping the draft's own type (byte-truth for that body) while the private C name makes collision
impossible and the asm label pins the emitted symbol, so codegen is unchanged. Fires only where
the TU declares that symbol with a DIFFERENT type text; same-type and already-aliased drafts are
untouched (4 controls, 2 of them negative). Applied on both scope paths — a staged draft's externs
arrive indented, so a demote-path-only fix reached 1 of 44.
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
WHY ONLY 2 BANKED — the collision is DRAFT-vs-DRAFT, not draft-vs-TU. The failing draft declares
`extern s32 D_80114F24;` and ov_MAIN_012.c declares that symbol nowhere; the error lands at the
splice point. family_sweep stages every member of an (overlay, split) group into one TU before
gating, so two templated bodies with different views of one symbol collide with each other.
scope_data_fix sees one draft plus the pre-splice TU and structurally cannot see the others.
The real fix belongs in the staging loop, which knows the whole group: alias any data symbol
declared with >=2 distinct types across the drafts staged together. Not attempted here.
Three wrong inferences on this one task before reading a failing draft: scoped as the typedef
class; aliased only the demote path against the file's own comment; targeted the wrong collision.
harvest_verify.classify_fail kept only stderr lines containing the word `error`. gcc-2.7.2 emits
no `error:` prefix on hard errors, so lines like
src/…/ov_SC02_037_jr_8013B83C.c:447: multiple storage classes in declaration of `tail_…'
src/…/ov_SC06_025_jr_8012ACE0.c:2217: `tbl_D_80187044' undeclared (first use this function)
never survived the filter, `errs` held nothing but make's `Error 33` wrapper, and every hard error
was labelled CC1-FAIL(no-diagnostic) — "the compiler failed and we cannot see why". Measured cost
this session: 132 siblings of func_80132018 classified that way by one missing declaration, which
reads as a codegen wall and gets a family deprioritised. rtu_match had the same blindness repaired
at T0(b); the fix was never propagated here.
Fix: a diagnostic is a POSITION, not a vocabulary — `<file>:<line>: <text>`, plus the assembler's
`{standard input}:<line>:` (_SRC_DIAG). Context lines carry no `:<line>:` and are skipped.
Five controls pass, including the two that guard against over-fixing: PLUMBING still wins on a
declaration conflict, and a warnings-only failure still returns no-diagnostic.
Re-swept: 0 no-diagnostic remain. The 93 resolve to 35 redefinition-note, 11 D_801202A0
undeclared, 10 too-many-arguments, 4 too-few-arguments, 4 func_8001534C undeclared — every one a
cheap declaration/arity class, not a wall.
Residue now fully named (737): 207 undefined-reference across 42 symbols (a link-stage REMAP gap,
now the largest class), 138 DIFF (real divergence, 19% — the honest floor), 44 data-symbol
conflicting-types, 35 redefinition-note, 26 memcpy, 24 redeclared, 15 undeclared, 14 arity.
Found while scoping the jr carve for the 3 newly-onboarded binaries. My scoping said "one
unplaceable construct" — it was the first of four layers. Three are fixed here; the fourth is out
of this tool's scope and leaves the carve blocked.
1. asm_label_aliases: the scan could START inside a #define. `#define gte_SetRotMatrix(r0)
__asm__ volatile ("lw $12, 0( %0 );" ...)` is textually `ident(...) __asm__(...)`, and
cdecl._mask blanks string CONTENT — deleting the `;`s that would stop the greedy [^;{}]*.
The match ran 116 lines and swallowed the real `aF8012EFB8 ... __asm__("func_8012EFB8");`,
so the alias never entered the map and addr_of returned None. jr_isolate_all then refused to
carve (R32, correctly), which presented as 112 isolate-fails that looked like a per-binary wall.
Fixed: _mask_cpp_directives() — a preprocessor directive is the other place a match must not
start. Masking comments/strings fixed the comment case and left this one.
2. _split_macro_body returned a `static inline` internal HELPER as the macro's definition, so
_proto_from_lines hoisted `extern static inline void tail_8012F274(...);` into all 41 regions:
invalid C (multiple storage classes) AND the wrong function — the exported definition sits
below the helper and lost its implied declaration. Fixed: skip static definitions
brace-balanced on the masked body. A static helper needs no hoisted declaration at all.
3. A declaration that WRAPS across continuation lines was taken as one line, so half became a
`;`-less extern and the continuation was read as the definition header, producing
`extern __asm__(""); void aF801466F0(...);` in 22 regions. Fixed: accumulate until the
statement terminates, tested on the masked text. Same wrapped-declaration blindness
family_remap._alias_decl_for records fixing at S33 — never propagated here (§134/§139).
Not fixed, and why: jtbl_rodata_pads reports "consumed 0 rodata .align(s) but 4 pad spec(s) given
— table-count drift vs the carve". The carve moves jtbl-owning functions into _jr_ regions but
leaves the pad specs on the residual gap object. jr_isolate_all's docstring states this class is
NOT isolate-fixable; it needs JTBL_PADS repointing in overlays.mk. 122 jr member-slots stay blocked.
Regression-checked: 1,948 macros parse with 0 malformed externs; alias maps unchanged on three
already-carved overlays. No build impact (splitters run offline). Carve reverted, tree clean.
scope_data_externs §8d drops the draft's decl of any symbol the TU already declares at file scope.
It keys on the SYMBOL, but a §37 asm-label ALIAS binds a DIFFERENT C identifier to that symbol:
the TU declares `D_801851BC`, it does NOT declare `tbl_D_80187044`. Dropping the alias left the
body referencing an undeclared name, which cc1 reports with no `error:` prefix — so the sweep
classified all 132 siblings as CC1-FAIL(no-diagnostic), i.e. as a codegen wall.
The bitter part: the alias exists PRECISELY BECAUSE the TU declares that symbol with a conflicting
type (a `void (*[])(void)` dispatch table vs this function's 20-byte-stride view). The drop rule
fired on exactly the declarations written to survive it. Why 1 of 2 died was fully determined:
tbl_D_80187048's symbol is not in the TU, so it demoted normally.
Fix: is_asm_alias() — an alias is demoted into the body, never dropped (the identifiers differ, so
it cannot collide with the TU's decl). Control-tested 6 ways incl. self-labels and plain externs.
Measured: func_80132018 3/135 -> 135/135; full re-sweep +16 more. Total +148 members.
R22 clean-fleet 213 passed / 0 failed of 213. tools-health OK, dedup-check 1949/0.
Fleet 96.11 -> 96.15% fn-count, 87.8 -> 87.9% distinct; stubs 14,120 -> 13,972 = -148 (2nd oracle).
CORRECTION TO MY OWN CLAIM (R14): after the probe I said the 58% aggregate was concealing a broad
problem. The re-sweep refuted it — only 16 more banks fleet-wide. The alias class really was one
family; the first read ("outlier") was right and the correction was wrong.
875 sweep failures classified: 231 PLUMBING-other, 141 DIFF (real divergence, only 16%),
136 `conflicting types for cdFileLocTable` (ONE symbol — biggest single class left),
77 CC1-FAIL(no-diagnostic), 26 memcpy, 12 D_80114F24, 11 D_800AE620, 9 D_800183E0.
STILL UNFIXED, and the most dangerous instrument left: the sweep's failure classifier greps for
`error:`, which gcc-2.7.2 never emits on hard errors. Every hard error therefore reads
CC1-FAIL(no-diagnostic). That is how a missing declaration looked like a codegen wall across 132
functions. rtu_match was fixed for this at T0(b); this classifier was not.
C, unblocked by B's declaration conform. 23/23/16 banked across ov_SC03_107, ov_MAIN_012,
ov_SC02_037 — the first non-zero result on this population (S46 got 0/142, then 0/129).
R22 clean-fleet: check-all 213 passed / 0 failed of 213. tools-health OK.
dedup-check 1949 validated / 0 failed, C1 coverage 249295 (= 249233 + 62, independent
confirmation of the count). Fleet 93.9% instr / 87.2% distinct / 95.72% fn-count.
THE TOOL REPORTED "BANKED 0 / 129" AND WAS WRONG. gate_stage's ladder hands the same
--verified-out path to harvest_verify on every rung, and each rung opens it for write: stage 0
banked 23 and wrote them, then a later rung that banked nothing truncated the file to 1 byte.
The in-memory list uses += and stayed correct, which is why the JSON verdict listed all 23 names
while the file said nothing. dedup_extend read the file, printed BANKED 0, and took its
`if not banked:` branch — skipping add_members_surgical, so the registry was missing 62
memberships for functions already spliced in and byte-verified.
- Registry repaired by deriving the banked set from git diff (+DEFINE_func_*), not from the
broken file. Post-check: 0 missing.
- ensure_include_revert did NOT fire (added_include False, include already present) — the
P29-S19 defect that once stripped a load-bearing include from 135 binaries stayed closed.
- gate_stage now writes verified_out once at the end from the accumulated truth.
Caught only because bank truth is derived from source (§55b), never from the gate report.
Residue (67) is consistent with the symbols B deliberately left: memcpy 17, ApplyMatrixSV 12,
gte_SetRotMatrix 4, plus 21 CC1-FAIL and 3 DIFF. Not separated: how much of the 62 is B's
conform vs the ladder's own recovery rungs.
Task B, re-scoped from evidence. The 129 dedup_extend failures are 106 conflicting-types /
21 CC1-FAIL / 4 undefined-ref / 3 DIFF — real byte divergence is 2%, and memcpy is 17 of 106,
not the story. Direction reversed too: the byte-true DEF of func_80128ED8 is what the target
.c files already declare; engine_core.h's macro-local extern was the stub-era guess.
Conformed 8 axes to byte-truth (func_8012F14C 2843, func_8012E5CC 2052, func_8012F038 2214,
func_8014C568 1816, func_80128ED8 1524, func_8012C750 406, func_8012C0EC 50, func_80144A04 25).
R22 clean-fleet: check-all 213 passed / 0 failed of 213. Zero functions banked by design.
Tooling (R33/R35) — three guards that asserted completeness over a narrowed population:
- NEW tools/macro_draft.py: a deduped fn has no definition in any .c (body lives in a DEFINE_
macro), so conform_decls had been refusing the largest class it was built for.
- conform_decls skipped engine_core.h wholesale as "a defining TU": 10 stale externs survived
while 1,514 fleet sites moved, and it still printed "axis complete". Skip now scoped to the
defining macro's span.
- Return-axis compare was literal: typedef int/s32 and a missing `extern` faked a return change.
Now compares normalized types.
- §85 consumer scan under-reported (the dangerous direction): a cast between `=` and the call
hid `s0 = (s32 *)func_80144A04(...)`. Now classified by position, validated both ways.
Corrections to my own predictions (R14): the documented scalar-narrowing hazard was benign
across 2,052 sites; the breaks were arity (6 call sites, fixed with §17a-1 fn-ptr casts) and
the consumer-guard gap. A header-only first probe broke ov_SC01_000 — §85 is literal.
Not done, named: memcpy (builtin codegen), ApplyMatrixSV (no DEF), gte_SetRotMatrix (link bug),
func_80147364 (unparseable macro), D_800AE620/D_80126CC4 (data axis). Cookbook §159.
- BANKED: 11 functions at 400-952 ins from the cascade (func_8017D898 952, func_8017CE58 733,
func_801902EC 673, func_8018C2D8 673, func_8018A8D4, func_8017C6F4, func_800CBB38,
func_800CF3A4, +3). check-all 213/213 from a clean tree. 6 near = jr/switch (§53 separate
banking step), 1 failed. The cascade agents wrote 6 new cookbook sections incl. §158.
⚠️ tools-health UNVERIFIED at commit (stale cookbook index fixed, confirming re-run
interrupted) — run it first next session. check-all is the byte oracle and it is green.
- WASTE PREVENTION (Drew: "prevent this from ever happening again, however you need to"):
* tools/validate_targets.py (NEW) — names 5 defect classes (NO-ASM / MID-BODY /
OUT-OF-RANGE / ALREADY-DONE / NO-BOUNDARY), exits non-zero.
* WIRED INTO wave_snapshot so it fails closed — every wave passes through there for its .s
files, so no path from target list to spawned agents bypasses validation. Negative-control:
a 3-target bad list is refused with the exact mid-body offset (+72 bytes of 100).
* The cascade `done()` predicate now short-circuits on SKIPPED as well as MATCH. It tested
only MATCH, so a non-existent target fell Sonnet -> Opus -> Fable and three agents each
proved the same phantom absent: ~29 invalid targets x 3 tiers = 87 of 119 agents, ~9.7M
tokens. A tier that cannot act must END the pipeline, not escalate emptiness.
* docs/accelerators.md A9, including that wave_snapshot's own R32 assertion REFUSED that list
(24 of 57 found) and was routed around — the one instrument warning that was right and ignored.
- B RE-SCOPED (S46-10) and deliberately NOT done: the extend blocker is INTRA-HEADER, not
target-side. engine_core.h declares memcpy FOUR incompatible ways across its DEFINE_ macros;
two in one TU collide. NOT a safe cleanup — the in-tree note at ov_MAIN_012.c:14333 records
that `extern memcpy` disables gcc's builtin and turns an inlined block-move into a CALL, so the
declaration CHANGES CODEGEN. Probe one macro in one binary and byte-gate before any sweep.
- C (dedup_extend over the 129) stays blocked on B. Full context for both in the checkpoint.
- RECOVERY PASS A (wave residue): gate_stage over the 3 big-3 draft dirs recovered
6 sites the bare gate rejected — func_80168664 x3, func_80168F40 x2, func_8012B77C
x1. That is 6 of 19 PLUMBING = ~32%, matching the 16-39% range P29 measured. Batch 1
goes 27 -> 33 of 60. R22 213/213 + tools-health green.
- HONEST SIZING (correcting my own claim): ~32% is NOT "a one-time fix for a ~50% draft
loss". It moves the batch loss from 55% to 45%. Real and free; not transformative.
- WIRED (task 9): dedup_extend now gates through gate_stage (the ladder:
canon_resident_calls -> cast_call_sites -> sig_unify -> harvest_verify) instead of
harvest_verify verbatim. Its 142 candidates failed 0/142 with reasons 118 PLUMBING /
21 CC1-FAIL / 3 DIFF — ~1 in 50 a real byte divergence, the rest declaration conflicts
in the TARGET TU, which is exactly what the ladder reconciles.
GATE_NO_ARITY=1 is forced for the child: gate_stage's arity pre-pass writes the
fleet-shared engine_core.h BEFORE the gate and a failing draft can leave that edit
behind — the F1 defect that broke 141 of 213 binaries in S45. The ladder's other rungs
are draft-local. --ladder can be disabled to restore the old path.
- DOCTRINE (step 3): gate every wave with gate_stage, not bare harvest_verify. Batch 1
needed a second manual pass only because I used the bare gate first.
- LEVERAGE METRIC CORRECTED (R14/R35): the behemoth ranking must use LIVE reach
(unmatched sharers), not total sharers. func_80144B9C reads 770 ins x 141 = 108,570
by total, but 138 of those are already banked — its true weight is 770 x 3 = 2,310.
Same x134 over-count the cookbook records in §25; build_wave_args --rank live exists
precisely for this and I used the wrong ranking. Remaining >=400 ins: 57 distinct
functions / 77 live instances / 38,968 ins, reach ~1.35 => ~0.3% instr-weighted.
Drew: "make it more multi-threaded... I still see my cpu idle for far too long."
Measured, fixed, and regression-tested against the S46-3 bank as a KNOWN ANSWER.
- THE MEASUREMENT: 31s saturated (33 makes/48 cc1/load 27) then ~25s with ONE build alive
while 31 cores idled, repeating. Causes: ex.map starts in list order so the giants land
last, and apply/restore is single-threaded.
- gate_all -> gate_failures: return EVERY failure the sweep already computed (~138 rounds -> 1).
- Longest-first gate scheduling; results re-sorted into `changed` order so the verdict stays
bit-identical to the serial loop's.
- PER-OVERLAY INDEPENDENT SEARCH, IN PROCESSES. My first cut used threads and the box refuted
it: 0-4 builds alive at load 3, because the work is regex over 15k-line files and 138
"parallel" searches all queued on the GIL. Same logic in a ProcessPoolExecutor: 14-29 builds,
load 34.75, search phase ~100s. Safe because the shared header is written ONCE by the parent
and each overlay owns its own .c files + build/<bin>/. Seeded with one in-process search
first — a pool submitted at once gives every worker an empty suspect list and makes all 138
pay a full bisection. place_in_overlay extracted to module level so the worker and the
in-process apply cannot drift (R33); compiles_standalone's fixed t.c is per-call now.
- THE REGRESSION (the point, not the stopwatch): revert src/+config to pre-bank, re-run the
identical command -> 29 functions (same), 141 overlays byte-identical, 682s vs ~1440s, and
285 exclusions vs ~350 => +62 MORE member instances (249,161). The old prefix-based
necessity probe was OVER-EXCLUDING (charging 4 fns to 9 overlays that did not all need
them); the per-overlay shrink minimises per overlay. The faster path is also more correct —
a timing comparison would never have shown it. R22 213/213 + tools-health green.
- STILL SERIAL, now the actual wall-clock (neither is a build): ~3min setup before the first
gate (registered_addrs() yaml-parsing a 1949-group/249k-instance registry + 213 sig loads)
and ~2.5min of sequential reconcile_caller_extern after the search.
- Captured as defaults: docs/accelerators.md A8 + memory fleet-tool-parallelism-defaults.
cookbook index regenerated (my §155c append left it stale — the gate caught it, exit 1).
Drew T