Measured the h_exact free pool from the bytes rather than trusting the frontier report's
numbers (R14 — its whale claim was 3/4 wrong: it said the whale was open in all four SC07
overlays; three were already banked and I closed the fourth earlier this session).
MEASURED: 215 open function-instances / 8,763 instructions are byte-identical (h_exact,
including reloc payloads) to an already-matched function. ONE class is 86% of that pool:
func_801758FC — 55 ins, same address in all 138 overlays, matched in ov_SC01_000 only,
OPEN in the other 137 => 7,535 instructions.
h_exact means identical INCLUDING jal/lui/%lo reloc immediates, so the matched body compiles
byte-identically at every member with NO remap (dedup_extend's correctness argument, §14).
dedup_propagate --addr authored it once as DEFINE_func_801758FC() in engine_core.h and
instantiated it at all 137 open sites in address order.
[ OK ] 138 overlays byte-identical after propagation; 1 new group in config/dedup.us.yaml
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet instr 12411467 -> 12419002 = +7,535 EXACTLY; fn-count +137; instr-weighted crosses to
94.5%. distinct-code unchanged BY DESIGN -- the class was already matched in ov_SC01_000, so
the 137 add fleet instructions but no new DISTINCT function. audit-digest OK. 0 NON_MATCHING.
Note this function had been sitting in the stored-draft backlog for ov_SC06_030 and
ov_SC07_010 and re-gated "no" earlier tonight -- because gating a DRAFT is the wrong move for
an h_exact class. The right move is propagating the already-MATCHED body. Same function, two
routes, and only one of them is free.
Remaining free pool after this: 78 instances / 1,228 ins across 32 classes.
The 39 draft-exemplar families all have their exemplar still OPEN in ov_SC01_077 -- a
draft-exemplar family cannot propagate until its head banks. Re-gated the newest stored
ov077 draft per head, in 4 small batches (§61: a wide harvest broke a TU in S38).
Set aside the top 4 heads (65% of the pool's weight, all known-hard): func_801412A8 +
func_80178004 ARE S6's two giant walls (198x138 + 165x138 = 50,094 ins riding on 2 cracks),
func_8017C974 is today's byte-proven close=47, func_8017C294 its 246-ins neighbour.
batch 0 1/9 batch 1 3/9 batch 2 4/9 batch 3 0/8 = 8/35 (23%)
BANKED: func_8017EC7C func_8018281C func_801820DC func_80182988 func_80183BAC
func_80183AF0 func_80183CF4 func_80182E7C
(+474 ins x1 now; ~1,441 ins of templatable weight behind them once their families propagate.)
CALIBRATION REFINEMENT (docs/calibration.md, S39): this population re-gates at 23%, vs 8%
for the general stored pool and 4/6 for fresh post-repair drafts. Three different populations,
three different rates -- which is exactly why the rule is "re-gate what a repair plausibly
touched", not "re-gate the ledger". ov_SC01_077 is the split-heaviest overlay, so the S38
alias-deletion repair plausibly touched all of these.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12410275 -> 12410749 instr (+474), distinct +474 / +8 uniq, fn-count +8. audit-digest OK.
0 NON_MATCHING (G4).
Tested whether decision-log A10 ("stored drafts re-gate at 0/958", measured in T1) survives
S38's tool repairs. Three populations, plain re-gate, no draft edits:
fresh wave-6 drafts (diagnosed "blocked on a class") 4/6
stored pool, unbiased sample (every 96th of 1,155) 1/12 <- hit was in a REVERTED overlay
the two REVERTED overlays, targeted 3/17
A10 BROADLY STANDS. ~8% on the general stored pool is not a harvest, and a 1,155-wide sweep
(= 1,155 whole-binary builds) is not justified by it. Do NOT generalise the fresh-draft rate
(4/6) onto the stored pool -- different populations. The honest rule is narrower and cheaper:
after a tool repair, re-gate the drafts THAT DEFECT plausibly touched, targeted by its
blast radius -- not the whole ledger. (R35 applied to the backlog, not just to metrics.)
BANKED (+146 ins): ov_SC06_030 func_80161208 + func_80162CCC; ov_SC07_010 func_801506A4 +
func_8016F0AC. R22 clean-fleet 140 passed, 0 failed of 140 -- which also proves byte-neutral a
fleet-shared engine_core.h edit the bank required (extern s32 func_801506A4(s32,s32) -> the
no-prototype form), reaching all 138 overlays (T2 blast radius).
Fleet 12410129 -> 12410275 instr; distinct +95 / +1 uniq; fn-count +4. audit-digest OK.
Also documents the LEDGER MECHANICS in calibration.md (Drew asked): .run/backlog.jsonl is
append-only and nothing is deleted on bank -- open-ness is DERIVED from corpus.stubs at every
read (load_best drops now-banked rows per-binary, P9) and `make report` runs `backlog.py prune`.
Membership is therefore self-maintaining and currently clean: 863 rows, 0 already-banked, 14
duplicate-addr (was 6,867 rows / 98% banked before Phase-29 compaction). What pruning does NOT
re-validate is the VERDICT on surviving rows -- closeness + residual class are as old as the
tooling that wrote them (Phase 28 found a corrupt one: func_80178004 close=0 -> 91). That is
the staleness that matters, and it is exactly what this probe measured.
Propagation behind the crack banked this session. jtbl_family_bank.py over the 4 open
h_seq siblings of func_801878E8 (513 ins each):
ov_SC03_001 BANKED ov_SC03_124 BANKED
ov_SC04_019 BANKED ov_SC05_017 BANKED
ROUTE NOTE (§53, worth keeping): family_sweep --hseq REFUSED this family by design --
has_mid_jr => it needs the jtbl carve, not the remap sweep, and the interlock says plainly
that "a 0% from this path would be a TOOL artifact, not a wall". Taking the refusal at face
value and using the named tool banked 4/4 first try. This is the same lesson as the rest of
the session from the other side: the instrument told the truth about its own limits.
jtbl_family_bank also enforces a CLEAN tree (it reverts from HEAD per sibling, so an
uncommitted prior bank would be destroyed) -- which is why the ×1 banks committed first (H4).
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140
(each sibling carves its own jtbl => config changed => fleet blast radius). Fleet
instr-weighted 12408077 -> 12410129 = +2,052, exactly 4 x 513; distinct +1,539 / +3 unique
fns (the 4th sibling shares an h_exact class already matched); fn-count +4. audit-digest OK.
0 NON_MATCHING (G4).
Session running total: +4,727 instructions (whale 770 + 4 drafts 1,905 + family 2,052),
12405402 -> 12410129, every step R22 clean-fleet 140/140.
The 6 still-open wave-6 drafts were triaged against S38's own diagnosis table; 4 banked,
R22 clean-fleet 140/140.
func_801919A0 ov_SC06_032 710 ins (was: undefined ref func_8018B878 -- "alias class")
func_80189030 ov_SC03_001 557 ins (was: undefined ref func_80186F88 -- "alias class")
func_801878E8 ov_SC04_018 513 ins (was: undefined ref func_801848DC -- "alias class")
func_8018A564 ov_SC02_027 125 ins (was: CC1-FAIL Error 33)
THE FINDING: all four banked with NO change to the drafts. S38 recorded them blocked on a
class that needed cracking ("cracking this one class frees 6 drafts at once"); they had
ALREADY been freed by S38's own tool repairs -- the jr_isolate_all/overlay_src_split
alias-DEFINITION-deletion blindness and harvest_verify._reload_corpus. The drafts were
correct all along; the instruments were failing them. That is the FIFTH recorded "wall"
this phase to resolve to our own tooling.
=> RE-GATE STORED DRAFTS AFTER ANY TOOL REPAIR before treating a stored verdict as a
fact about the code. A verdict is only as current as the instrument that produced it
(R35 applied to the backlog, not just to metrics).
Each bank also performed a jtbl carve, so config/ changed => fleet blast radius => full R22
(clean + extract-all + check-all) = 140 passed, 0 failed of 140.
Metrics move exactly as the model predicts: instr 12406172 -> 12408077 = +1,905, the exact
sum of the four (710+557+513+125); distinct +1,905 / +4 unique fns; fn-count +4.
audit-digest OK. 0 NON_MATCHING (G4).
LEFT ON THE BACKLOG as genuine codegen residuals, not forced (P9):
func_8017C974 (ov_SC01_077, 947 ins, close=47, REGALLOC-PERM, 12 permuter variants inert)
func_80188C68 (ov_SC03_124, 551 ins, close=370, the only target with no twin anywhere)
NEXT: the func_801878E8 family (4 open siblings x 513 ~= +2,052). family_sweep --hseq
correctly REFUSED it via the §53 interlock (has_mid_jr => jtbl carve route; "a 0% from this
path would be a TOOL artifact, not a wall"), and jtbl_family_bank.py requires a clean tree
because it reverts from HEAD per sibling -- which is why this commit lands first.
Closes the first of S38's two reverted R22 failures. ov_SC07_010 was the lone overlay
still shipping func_80144B9C (770 ins) as INCLUDE_ASM while the other 137 banked it.
Its _jr_80140608 object ran 0x184b0..0x2c2f4 straight through the whale; the sibling
ov_SC07_006 carves the same span into _o0d (0x1ca44) + _jr_801457A4 (0x1d64c). Note
0x1ca44 + 0x80128158 = 0x80144B9C exactly.
tools/o0_subsplit.py ov_SC07_010 --lo 0x80144B9C --hi 0x801457A4
-> 1 unmatched stub, 0 ALREADY-MATCHED in range (so no §126 island; K=0 => 3 regions)
-> split BYTE-NEUTRAL first (d7b5875d), then banked via ../shared/func_80144B9C.h
The S38 cause ("its -O0 split reused an EXISTING _o0c instead of a fresh _o0d") did NOT
recur: o0_subsplit.free_letters derives the unused suffix (_o0c is free in THIS overlay).
Two decl conflicts on the way, enumerated with `cdecl` in ONE pass (R33) rather than one
build at a time — of the whale header's 94 symbols the §8b carried layer re-declares 3,
and 2 conflict: D_801274D0 (layer `s32 (*)(s32)`) and D_801274CC (layer `void *`) vs the
header's canonical `s32`. Dropped both: nothing in the region uses them, they are carried
from an earlier region of the old object, and 0 of the 137 other whale-including files
carry either. Decls emit no code => byte-neutral (§8c), and byte-gated.
ov_SC06_030/func_8017E120 needed NO work — it is already banked (defined at
ov_SC06_030_jr_8017C8D0.c:3491). S38 reverted the surrounding batch, not that function.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140
(config changed => fleet blast radius, R22 mandatory). Fleet instr-weighted
12405402 -> 12406172 = +770, exactly the whale's size. distinct-code unchanged by design:
that h_exact class was already matched via the other 137, so the 138th adds fleet
instructions but no new DISTINCT function. audit-digest OK (the new S1e gate, on its first
real use). 0 NON_MATCHING (G4).
Metric note (R30, same class as S1e): a body banked by #include-ing a shared header is
invisible to fn-count's NUMERATOR (the definition is not in the .c) while its stub leaves
the denominator -- 341186/353718 -> 341186/353717. The weighted metrics counted it
correctly because they derive from corpus.stubs, not re-parsed C. Trust the weighted pair.
Still open from S3: the 61 SC07 -O0 members (untouched).
The S38 checkpoint gated the phase's best lever ("do NOT scale the alias lever") on
distinct-code falling 89.3 -> 89.2. It never fell.
PROOF (each commit's metric recomputed from its OWN committed tree, 0 unresolved):
commit:1426 TRUE : instr 12394533 distinct 5022306 (77895 uniq)
commit:1426 COMMITTED: instr 12402412 distinct 5029324 (78025 uniq) <- stale
HEAD TRUE == COMMITTED: instr 12405402 distinct 5025082 (77952 uniq)
=> true delta 843->HEAD: instr +10869, distinct +2776 ins / +57 uniq. ALL ROSE.
The 843 digest was generated from a working tree still holding work REVERTED before the
commit landed (+7,879 ins / +130 uniq overstated) and never regenerated, so the next
HONEST digest read as a fall. => THE ALIAS LEVER IS UNGATED (scale it, §61 small batches).
Both recorded leads were wrong (R14): progress.py:423's SIG regex feeds fn-count ONLY
(neither weighted metric sees a C identifier — both derive matched = sig - corpus.stubs),
and "the harvest reverted functions to INCLUDE_ASM" died on one grep (483 removed, 0 added).
The 3-grep proof: identical sigs + unchanged tools/ + zero +INCLUDE_ASM => HEAD's stub set
is a strict subset => both numerators are FORBIDDEN to fall.
THREE INSTRUMENT DEFECTS, all one class (a bare except around a fail-CLOSED oracle):
- progress.py stub_addrs wrapped corpus.stubs in `except Exception: return set()`. An empty
stub set means "could not answer", not "no stubs", so matched = sig - stubs credited EVERY
function. Byte-witnessed: instr 100.00% / distinct 100.00% in a tree with no asm/. Now
propagates.
- cast_call_sites.tu_for + reconcile_tu.tu_for had the identical swallow, falling back to the
default <ov>.c instead of the jr/-O0 split TU — silently reinstating the exact bug
cast_call_sites' own docstring says it exists to fix. A wrong-TU reconcile fails the gate,
and this phase's base rate is ~24k PLUMBING vs 4,917 DIFF, so it presents as a codegen wall.
Now propagate CorpusError; ValueError fallback for curated names preserved; derived-TU path
re-verified (a _jr_ split stub resolves correctly, both tools agree).
NEW GATE (R34 — the byte-gate is a null oracle for DOCUMENTS; check-all stays 140/140 over a
stale digest forever): tools/audit_digest.py + `make audit-digest`, wired into tools-health
after report. Recomputes the three headline metrics from the current tree and fails if the
committed digest disagrees. Compares INTEGERS, not percentages — the +7,879-instruction
staleness printed as "94.4%" on both sides. Negative-control-proven against the stale 843
digest (fails, exit 1) and green on HEAD.
Verified: make report exit 0 (dedup-check 1910 validated / 0 failed, C1 coverage
241216/241216); audit-digest OK; cookbook-index OK (398 sections); metrics unchanged by the
fix (94.40% / 89.18%). No src/ or config/ edits — no bytes touched, nothing banked.
cookbook §140 · decision-log 2026-08-04 · SETUP.md inventory (R21) · R14/R32/R34/R35.
Refreshes a checkpoint that had gone stale (v3 predated S1d) — stale is worse than absent.
FLEET 96.46 / 94.4 / 89.2, +37,166 instructions this session, ~0 agent tokens after the opening
wave. R22 run thirteen times: 140/140 on eleven, TWO REAL FAILURES (ov_SC07_010, ov_SC06_030), both
caught by the clean-tree rebuild after passing their per-binary gate, both reverted and recorded.
Records the session's biggest find: the §37/§124 DEFINITION-SIDE ASM-LABEL ALIAS is a CLASS lever,
not a one-off. It cracked the 208-conflict narrow-parameter class 138/138 after cast_call_sites,
--normalize-self-decls and --fix-def-sig were each eliminated BY MEASUREMENT. S33 proved it once and
it was never generalised.
Carries the unresolved accounting anomaly prominently (new task #11 / S1e): distinct-code FELL
89.3 -> 89.2 across the alias harvest while fn-count ROSE, which no pure naming artifact explains.
The bytes are proven; the yield number is not. Next session starts there, before scaling the lever.
Also records eleven tool defects fixed (nine of ten "walls" were our own instruments, two of them
mine), that §134 has now appeared in SIX tools and wants cdecl._mask rather than a seventh patch,
and seven process errors of my own including piping away a gate summary I then could not report.
docs/progress.fleet.md is the authoritative metric source the checkpoint's staleness self-check
compares against — committing it keeps that check meaningful for the next session.
Families that returned 0/N before the 895-type lift now bank: 97 member-matches across 137 families
(1,138 failed; skipped 186 STRUCT-class by design, 106 unresolved-immediate, 6 not-stub).
R22 clean-fleet 140/140. Fleet 94.2 -> 94.3% instr / 88.9 -> 89.1% distinct / 96.29 -> 96.32% fn.
RESIDUE PRICED FROM THE SWEEP'S OWN .classified.txt PAYLOADS, not inferred: in the 400 most recent
failure records, 133 are genuine DIFF and the clear majority are `conflicting types for <sym>` —
the §103/§20 extern-conflict class, across ~12 overlays. That confirms S1b (wire reconcile_tu /
cast_call_sites into the --hseq path) is the correct next lever, and it is now justified by
measurement rather than by the plan's projection.
Method note worth keeping: those per-member diagnoses have been written on every sweep run for a
month and were never read — including by me, until after I had spent two probes and a manual
--stage-only round rediscovering one of them. Every remaining task now starts by reading the payload.
FLEET 96.29% fn / 94.2% instr / 88.9% distinct, R22 140/140 twice. Session banked 7 wave-6 heads
+ 25 propagated siblings = +10,616 instructions (12,368,236 -> 12,378,852; the digest delta matches
the hand-derivation exactly).
wave-metrics.md gains Findings 5 and 6:
- RANK WAVES BY INSTRUCTIONS, NOT HEADS. Wave 6 banked ~45% more instructions than wave 5 while
banking less than half as many heads, because a bigger head carries more instructions AND its
family propagates at the same cost per sibling (func_8017FEE0: ONE 299-ins head -> 4,485 ins
across 15 siblings, ~0 agent tokens). The metric to beat is POOL REALISATION (21%), not bank
rate — and wave 6's bank rate is NOT comparable to waves 3-5 because the difficulty knob moved
deliberately (median target 438 ins vs 143, mostly has_mid_jr).
- A WAVE TALLY IS A COVERAGE CLAIM and needs its own assertion.
The 9 unbanked drafts are diagnosed and preserved, and 6 of them are ONE class worth cracking
first: `undefined reference` to a sibling that IS defined in the overlay but only via a
definition-side __asm__ alias (§37/§124) — the carve repartitions the object and separates the call
site from the alias definition. One fix frees 6 already-paid-for drafts and will recur in every
carve-heavy wave. The other two are genuine near-misses (close=47 REGALLOC-PERM, close=370).
Tooling committed for reuse: w6_pool.py (the pool derivation, now a script), w6_diag.py (runs the
REAL gate path — s36_capture.py splices without the carve and is wrong for jr targets),
w6_jtbl_prop.py (carve-path propagation with the per-family commit jtbl_family_bank requires).
The wave-6 gate printed BANKED 5 / FAILED 1 over SIXTEEN drafts and nothing said so. Captured with
the three corollaries that outlive the bug (R30 — written in the session that produced them):
(a) the byte-gate is a perfect CORRECTNESS oracle and a NULL COVERAGE oracle (R34 again), so every
wave-level tally is a coverage claim and needs its own assertion;
(b) a pipeline's exit status is the LAST command's — "family_sweep ... | tail" hid a non-zero exit
in this same session; use pipefail/PIPESTATUS or the tail IS the error handler;
(c) a reverted CONFIG needs a re-extract before the next measurement means anything (the Phase-20
R22 corollary; skipping it made three genuinely-banked functions read as failures).
Plus the inverse-lookup trap: a just-banked head LEAVES its family's members list and becomes
exemplar.kind='matched', so a member-list lookup reports "no family" for exactly the heads you just
banked. Distinct from §138 rule 4, which governs target SELECTION.
Index regenerated: 393 sections, 3 new symptom keys.
load() globs .run/sig.ov_*.jsonl, so main and the resident are absent from every denominator —
the numbers run ~0.3-1.0pp above the authoritative `make report` fleet (measured today: 96.6/94.6/
89.7 vs the digest's 96.28/94.1/88.7). docs/family-hseq.md has always carried the "(overlays)"
qualifier; the stdout print did not.
That asymmetry matters because stdout is the channel a session actually reads and transcribes into
a checkpoint — a right number under a wrong label is how a wrong number propagates (R35: the
measurement was fine, the instrument's LABEL was the defect; R14: a checkpoint that disagrees with
the digest must lose, and it can only lose if the disagreement is visible).
Print-only; the metric itself is unchanged and correct for its scope.
Fleet 96.27 -> 96.28% fn-count / 94.1% instr / 88.6 -> 88.7% distinct (77,765
uniq). R22 clean-fleet: 140 passed, 0 failed of 140. dedup 1910/0.
16 targets / 16,884 templatable ins. 19 agents, 2.71M tokens, 82 min wall.
The gate banked 16/16 — the FIRST perfect gate of the session, and the first
needing NO reconcile at all. Sweep: +26 members / 3 failed across 17 overlays.
NEW: docs/wave-metrics.md — the wave-by-wave performance log, with the derivation
commands so future rows are COMPUTED, not hand-transcribed (R33). Four findings,
each recorded with its caveat rather than as a bare number:
1. THE PROMPT IS THE LEVER, AND THE AGENTS WRITE IT. Bank rate 76 -> 77 -> 100
-> 100 -> 100% with models and gate held constant. The jump was STEP 0 (a
magic-literal grep of src/, ahead of engine_core.h) — which came from a
wave-2 agent's index_gap report. Caveat recorded: waves 3-5 targets also
trended easier, so the mechanism is the durable claim, not the exact %.
2. pipeline() vs batched parallel(): 136 min/14 targets -> 82 min/16 targets,
parallelism 2.5x -> 3.8x. The two-batch design was a hard barrier with 37-50
min dead gaps; the harness already caps at 16 so it bought nothing. Floor
recorded honestly: the slowest agent is still ~50 min of real match_one
iteration, so the lever there is target SELECTION, not concurrency.
3. ECONOMICS: ~170-300k tokens per banked head in the stable regime — but a head
is not the unit of value. Head + propagated members is, and sweep yield is
BIMODAL not average (21/21 vs 18/165), because it is a property of the FAMILY.
Averaging those two predicts nothing.
4. A perfect gate is a signal the prompt rules landed. Waves 1-4 each needed 1-2
post-gate reconciles; wave 5 needed zero. The reconcile lane is the fallback,
not the plan. Lifetime 21/22.
Fleet: R22 clean-fleet 140 passed, 0 failed of 140. dedup 1910/0.
14 targets / 16,844 templatable ins. 17 agents, 3.0M tokens. Claimed 14/14; the
whole-binary gate banked 13, the 14th on reconcile. Sweep: +26 members / 0 failed
across 18 overlays. Reconcile lane 21/22 lifetime.
STEP 0 HAS BECOME THE AGENTS' DEFAULT MOVE. Nearly every wave-4 verdict cites the
cross-overlay magic-literal grep BY NAME, several reporting `index_gap: none`
because it resolved the target in one pass with no cookbook derivation needed:
- func_8018BED0: grep 0xE100000A -> func_80188C04 (ov_SC03_089), verbatim, MATCH first try
- func_8018BAB4: grep D_800A6610/D_800B9A02 -> func_801887E8, verbatim + callee swap
- func_8017ED54: grep named all 5 family members -> reused func_8017D9F0's body, 14 data remaps
- func_8017C290: grep found byte-identical twins ALREADY banked in two other overlays
Bank rate by wave, same models + same gate, prompt the only variable:
76% -> 77% -> 100% -> 100%.
THE ONE FAILURE IS THE §138 RECONCILE-DIRECTION RULE, in its purest form:
`redefinition of struct B16_8018A758` — the agent copied its sibling's struct tag
verbatim, and that sibling had banked into the SAME TU earlier in THIS wave. Decl
ABOVE the splice => DELETE the duplicate (do not rename it). Worth noting the
mechanism: a wave can create its own reconcile work when two targets share a TU.
func_8018A808's own family swept 0/14 — its members are the per-location kind
that do not template (the settled h_seq ceiling), not a plumbing failure.
Fleet 96.24 -> 96.25% fn-count / 94.0% instr / 88.4 -> 88.5% distinct.
R22 clean-fleet: 140 passed, 0 failed of 140. dedup 1910/0.
13 targets / 17,644 templatable ins. 14 agents, 2.7M tokens. Claimed 13/13;
the whole-binary gate banked 12, the 13th on reconcile (another §37/§124
SELF-axis alias — the TU declares `(void)`, the def takes an s32). Reconcile
lane 20/21 lifetime. Sweep: +21 members / 0 failed across 16 overlays.
THE FLYWHEEL, MEASURED ACROSS THREE WAVES (same models, same gate):
wave 1 baseline prompt 15/17 claimed -> 13 banked (76%)
wave 2 + the S33 rules 11/13 -> 10 (77%)
wave 3 + S34 magic-grep as STEP 0 13/13 -> 13 (100%)
Multiple wave-3 agents report the cross-overlay magic-literal grep landing the
answer on the FIRST search. One found a banked twin whose own header comment
already documented it as byte-identical to the new target, so the body
transferred verbatim with only file-local type suffixes renamed. That is the
wave-2 discovery paying off one wave later (R16).
Sweep quality also differed for a reason worth keeping: 21/21 here vs 18/165 in
wave 2. Wave 2's two big families are the per-location kind I then probed and
ruled out (BUILD OK + byte diff = genuine per-member codegen, not plumbing);
wave 3's are genuinely templatable. The sweep rate is a property of the FAMILY,
not of the wave.
TOOLING: an agent left 8 scratch files (test_licm*.c) in the drafts dir and the
gate driver died on `int('full', 16)`, taking the whole gate with it. Hardened to
treat a non-conforming filename as a NAMED, COUNTED skip rather than a crash
(R32) — a drafts dir is agent-writable by design, so it must not be trusted to
contain only deliverables.
Fleet 96.24% fn-count / 93.9 -> 94.0% instr / 88.4% distinct. R22 clean-fleet:
140 passed, 0 failed of 140. dedup 1910/0.
WAVE 2: 13 targets / 37,943 templatable ins. 19 agents, 4.5M tokens. Claimed 11
MATCH; the whole-binary gate banked 9, +1 on reconcile (func_8017E5D0 via the
§37/§124 DEFINITION-side alias — the TU declares it `(void)`, the byte-true def
takes a pointer). Reconcile lane now 19/20 lifetime. 18 members swept.
THE FINDING (an agent caught a hole in our own procedure). §136c's search order
— engine_core.h near-twin -> same-TU banked sibling -> the .s — is entirely
SAME-TU or SHARED-HEADER scoped, so no step can reach a banked twin in a
DIFFERENT overlay's TU. But the large template classes live cross-overlay by
construction. func_80188C04 (328 ins) turned out byte-identical to an
already-banked func_801833F0 in ov_SC02_028, and ONE command found it:
`grep -rn "E100000A" src/` — a magic word lifted from the target .s. The body was
then reused verbatim, only file-local suffixes renamed. Promoted to STEP 0 of
§136c, ahead of engine_core.h.
That compounds with the manifest finding this session: the family map's
`exemplar` is an IN-FAMILY pointer, so a family whose twin is banked elsewhere
looks un-cracked — and the pointer can itself name an ALREADY-BANKED instance,
hiding the family from any ranking built on it. Derive open sites from
corpus.stubs over the member list instead. Measured on this wave: ranking off the
map's exemplar gave 16,696 templatable ins; deriving from corpus.stubs gave
41,023, including a 55-ins family open in 138 overlays and a 46-ins one in 133.
HONEST ON THE SWEEP: those two big families templated 18/165. That is the known
h_seq refusal ceiling, not a new wall. One agent reported "all 10 members
distance 0" — that is NORMALIZED distance, not h_exact, which is why
dedup_propagate correctly answered reach<2. Do not read a normalized-distance
claim as an h_exact guarantee.
LEDGERED (real residual, not paperwork): func_8017F7B4 — needed its sibling's
type names AND a data asm-label alias for a u8-shaped symbol, and still refuses.
Plus func_8017C294 (DIFF close=12: 4 register/schedule permutations + a frame
where I can get the 0x138 size OR pEnd's slot at 0x108, not both) and
func_801898E4.
Fleet 96.23 -> 96.24% fn-count / 93.9% instr / 88.3 -> 88.4% distinct.
R22 clean-fleet: 140 passed, 0 failed of 140. dedup 1910/0.
func_8017D318 (184 ins) + func_80181EE0 (198 ins) both banked, + 6 members swept
(6 per-overlay variants failed — ledger material, not a lever).
THE RECONCILE DIRECTION DEPENDS ON WHERE THE TU'S DECL IS, and picking wrong
CREATES the next error (-> cookbook §138):
- decl ABOVE the splice point -> DELETE the draft's duplicate (§100).
func_8017D318: the TU defines MATRIX_/SVECTOR_8017C290, D_801EA8C0 AND a
`struct PW8017C290` tag above it; I missed the tag on the first pass, so it
took two rounds.
- decl BELOW the splice point -> KEEP a decl in the TU's EXACT shape and cast
at the use (§17a-1 D2). func_80181EE0: I removed its decl assuming the TU
provided one; the TU's `extern int func_80143C74(short *, int);` is at L5082,
~180 lines BELOW the splice at 4901, so the identifier went undeclared.
Grep the TU for the symbol and compare line numbers with the stub line first.
MY OWN §136a VIOLATION, recorded: the blocker-capture filtered the build log for
`error|conflicting|undefined reference` and reported "NO COMPILE ERROR" on a
build that was failing with `redefinition of struct PW8017C290` and
`'func_80143C74' undeclared` — neither phrase matched. A narrow keyword filter is
exactly how a real error goes unseen, which is the thing §136a exists to say.
Widened to keep any line naming a source position.
Fleet 96.21 -> 96.23% fn-count / 93.8 -> 93.9% instr / 88.0 -> 88.3% distinct
(+73 unique fns). R22 clean-fleet: 140 passed, 0 failed of 140. dedup 1910/0.
THE WAVE. Re-ran S10's 17 unbanked targets (26,227 templatable ins) at LOW
concurrency in two batches of ~9 — S10's finding was that 14 of 30 agents were
SERVER-throttled, i.e. the limiter is capacity, not capability. Targets
re-derived against corpus.stubs first (R35): all 17 still live, paths verified.
25 agents, 6.87M tokens, ~2.8h. Claimed 15 MATCH; the whole-binary gate — the
sole arbiter (G3/P9) — banked 13, then family_sweep propagated 65 members across
38 overlays.
HEADLINE: func_8017D174 (793 ins) — the largest single crack of this phase. Its
agent closed two compiler-internal residuals jointly: a §137 allocno-priority tie
between &g.sz0/&g.sz1 (R=7, L=607 vs 606 -> 230/231) that spilled the wrong one
and cost a load-delay nop in BOTH switch arms, and a sched2 rotation in the
outer-loop head block that survived 470+ statement orderings. Fix was four
zero-byte asms: two `"=r"/"0"` re-ties splitting wz's live range, plus two
volatile sliders placed in a DIFFERENT basic block so they lift the live-length
count without perturbing the head schedule.
THE 4 NON-BANKS SPLIT CLEANLY (§136b — none is a wall on one attempt):
- func_8017E2EC (close=20) and func_80186E24 (close=187): honest DIFF verdicts,
real codegen residuals, ledger material.
- func_8017D318 and func_80181EE0: claimed MATCH, gate refused -> the known
match_one->gate gap, which is DECLARATION plumbing (agents cannot run the
gate, so a TU-level conflict is invisible to them). Routed to the reconcile
lane, not retired.
AGENT-REPORTED INDEX GAPS worth acting on (the flywheel closing on itself):
- no symptom key for "schedule rotation at a loop-head block that NO statement
permutation reaches" — the index's nearest line points at §76 regalloc, and
the decisive doc was gcc-2.7.2-map/sched.md, which no scheduling symptom
cross-references.
- §137 is written as a two-compile arithmetic on ONE contender pair; the real
fix here was an N-zero-byte-insn budget that ties only for N in {1,3,4} and
splits the WRONG way for N=2, so a naive "add one slider, add another" walk
silently regresses.
- no key for "gcc hoists a loop-invariant SYMBOL_REF base out of a loop the
target keeps in the `sym(reg)` macro form" (~105 of func_80186E24's 187).
Fleet 96.17 -> 96.21% fn-count / 93.8% instr / 88.0% distinct; dedup 1909 -> 1910
groups, 0 failed, C1 241216/241216. R22 clean-fleet: 140 passed, 0 failed of 140.
The head is now 5/5 classes, 18,545 templatable ins, all banked this session from
a standing start of 0.
func_801466F0 had sat since S6b behind THREE separate blockers, each of which
looked sufficient on its own to explain the failure:
1. Its definition is under a §37/§73 ASM-LABEL ALIAS (`aF801466F0` in C, bound to
the real symbol by `__asm__`), and dedup_propagate.find_site anchored its head
regex on the literal `func_<ADDR>` — structurally blind to the form, returning
None, which every caller reads as "not matched". Now reuses
family_remap._alias_decl_for rather than growing a second matcher (R33).
2. That matcher was itself blind to the WRAPPED (multi-line) declaration — the
§134 shape, third tool. Fixed by matching over the joined text and mapping the
offset back to the decl's FIRST line (extract_unit carries from there).
Regression control: the single-line form still resolves. Fleet census after:
2,768 of 2,768 alias sites resolve, 0 missed.
3. Its record type was a draft-local typedef, so the body failed
compiles_standalone. Lifted Rec801466F0 to src/shared/engine_types.h INSIDE
the include guard (the SESSION-19 double-include note) and switched both the
macro and the exemplar to it — byte-neutral, gate-proven.
Probed on ONE member before the fleet run: byte-identical 9052dc0e first try.
MEASURED, NOT INHERITED (R37): the S6b note frames the alias-regex gap as a CLASS
of missed work. It is ONE function — 91 distinct alias decls fleet-wide, the
per-line matcher resolved 90. Recording it so a future session does not scope a
phase against a class that does not exist.
cookbook §138 extended with the alias-form tool boundary and the three-blocker
story; index regenerated.
Fleet 96.10 -> 96.17% fn-count / 93.7 -> 93.8% instr / 88.0% distinct.
dedup 1908 -> 1909 groups, 0 failed, C1 241078/241078.
R22 clean-fleet: 140 passed, 0 failed of 140.
func_80147364 4,110 x137 definition-side asm-label alias
func_8016BA68 3,886 x134 dedup_extend + the MIRROR decl relax
func_8012F274 3,973 x136 hand-authored macro, source overlay excluded
func_8012A598 3,288 x138 cdecl._mask backscan fix + shared-type switch
func_801466F0 3,288 OPEN the wrapped-alias regex — measured as ONE function
THREE DISTINCT CARRY VARIANTS were hiding in one "CARRY-FIXABLE" bucket, and
only one is a tool bug (-> cookbook §138):
- a MULTI-LINE comment halts the preamble backscan -> fix the tool (cdecl._mask)
- a draft-local `struct Tag {…}` -> switch the exemplar to the SHARED type
- a file-scope `static inline` helper -> hand-author, EXCLUDE the source overlay
The third is the sneakiest: gcc-2.7.2 accepts implicit function declarations, so
the extracted body PASSED compiles_standalone with the helper undeclared and the
miss surfaced only as a whole-binary byte DIFF 137 gates later. Instantiating
that macro in the SOURCE overlay is a duplicate definition (its file-scope helper
is still there), so the shape is `--source-overlay X --binaries <all-but-X>`;
`--binaries` alone removes the source from the scan pool and errors.
TOOL BOUNDARY: once a group's members are DEFINE_func_*() sites, dedup_propagate
cannot extend it (find_site never returns a `def`). dedup_extend is the tool for
an already-macro-ized group — and `dedup_extend --check-only` across ordinary
overlays is a cheap fleet-wide wiring census (measured: exactly 1 group per
overlay, so no hidden backlog).
MEASURED, NOT INHERITED (R37): the S6b note frames _alias_decl_for's single-line
regex as a CLASS of missed work. It is not — 91 asm-label alias decls exist
fleet-wide, the regex matches 90, and the single miss is func_801466F0. Worth
3,288 ins, but a one-function fix. Correcting the expectation so a future session
does not scope against it.
HONEST CORRECTION to commit:1382. That commit's message implies the 42 `(void)`
relaxes unblocked the PROPAGATE remainder. They did NOT: the re-run banked 0/1
in all 134 overlays with the same error, because DEFINE_func_8016BA68 declares
func_80146C3C `(u8*)` — the MIRROR of the EXTEND-lane pair — and my relax only
touched the `(void)` direction.
Root cause is the R37 shape a third time: I bucketed by SYMBOL and stopped. The
lever is set by the (macro-shape, TU-shape) PAIR, and the same symbol conflicts
in BOTH directions across this fleet. One awk over the macro I was ACTUALLY
fixing — which I ran for the EXTEND macros and not for this one — shows the pair
before a 134-build run. §138 amended with the PAIR rule; correction logged in
CURRENT_PHASE.md rather than rewritten out of history.
The 42-decl relax still stands: byte-neutral, R22 140/140, removes a real
conflict class. It just did not do what I predicted.
THIS commit relaxes the 2 remaining `(u8*)` decls (uses are cast; `()` is
compatible with the (void)/()/(u8*) forms the fleet carries and no decl of this
symbol has a default-promotion param). R22 clean-fleet: 140 passed, 0 failed.
ALSO: tools/overlay_src_split.py `_split_macro_body` — the §134 sweep's one real
target, fixed. It carried the identical single-line-only comment test, and it
decides where a macro body's file-scope externs END, so a multi-line comment
truncated the extern set. SIZED FIRST: 38 live lines in engine_core.h macro
bodies hit it today. Now decides on cdecl._mask (one oracle, R33) with the
length-preservation invariant asserted (R32). Proven both directions by a
control: pre-fix it stopped at `/* multi` carrying 1 of 2 externs and treated the
comment as the definition head; post-fix both externs carry and the def head is
correct. Not in the gate path (only o0_subsplit + jr_isolate_all import it).
§138 — "the propagation lanes: a gate refusal is a DECLARATION, and which lever
you owe depends on blast radius". The durable content of this session:
- the 4-lever triage table ranked by blast radius, and the rule to grep the
fleet's decl shapes BEFORE relaxing to `()` (illegal only against a
default-promotion param — 4,020 decls measured, 8 of 36 banked for one token)
- `volatile` in the host TU is a SCHEDULING BARRIER that masquerades as a codegen
wall; the tell is a positional shift with a `nop` at a delay slot, and the
h_exact contract is confirmed/refuted in ONE command with no build
- the DEFINITION-side asm-label alias as the only zero-radius escape when the
fleet canon disagrees on a promoting param (1,725 in-tree precedents)
- rank a lane by measured concentration, not class count (5 of 45 classes carried
89%), and `--recover` is not a retry (16/16 on drafts, 4/138 on a propagation)
- §134 multi-line blindness recurring in a second tool; decide on `cdecl._mask`
- the waiter rule CORRECTED: `pgrep -x make` is wrong for a campaign of
sequential makes, `pgrep -f` self-matches, `nohup … &` signals the wrapper
Checkpoint refreshed and placed below the task checklist (fresh-session safe):
fleet 96.10 / 93.7 / 88.0, dedup 1908/0, R22 140/140 four times this session,
nothing running, lock free. Three named next items, none yet diagnosed against a
build — the PROPAGATE head remainder (11,147 ins), the 41-class tail (2,316),
and EXTEND's last 5 (the whale needs the §38 -O0 route; dedup_extend should
refuse-and-name that class per R32).
Fleet 96.06 -> 96.10% fn-count / 93.7% instr / 88.0% distinct; dedup 1907 -> 1908
groups, 0 failed, C1 240807/240807. R22 clean-fleet: 140 passed, 0 failed of 140.
func_8012A598 (3,288 templatable ins) was being written off as CARRY-FIXABLE.
It took TWO fixes; either alone leaves it skipped.
1. TOOL (R33) — find_site's preamble backscan. The SESSION-18 fix handled blank,
`//`, and SINGLE-LINE `/* … */` lines, but a MULTI-LINE block comment still
halted the walk: its middle lines start with `*` and its last line ends `*/`
without starting `/*`. So the three externs above the body were dropped and
the body then failed compiles_standalone on now-undeclared data. This is the
§134 multi-line-blindness class — S6b fixed the identical shape three times in
family_remap (D1/D2/D5) and this copy was never reached.
Fixed by deciding skippability on `cdecl._mask` — the project's ONE masking
oracle — instead of on line syntax: it subsumes every comment form at once and
cannot be fooled by a `/*` inside a string, with an R32 assertion on the
length-preservation invariant it rests on. Strictly monotone (it can only
carry MORE preamble), and dedup_propagate is a byte-gate feeder, so a bug here
can fail to bank but never falsely bank.
2. EXEMPLAR — the body also declared a draft-local `struct BigCopy164` tag, which
the tool refuses by design (two macros defining one tag would redefine it in a
single TU). The shared `struct BigCopy` (engine_types.h L312) is the identical
layout and is ALREADY used this exact way at engine_core.h:16158, so switching
the exemplar to it is byte-neutral and drops the alias too.
Probed on ONE member before scaling (R37/S29): byte-identical 9052dc0e first try;
then 138 overlays byte-identical.
PROPAGATE head accounting after this: 7,398 of 18,545 ins banked (func_80147364
4,110 + func_8012A598 3,288). Still open, each with a NAMED cause and none yet
diagnosed against a build: func_8012f274 (3,973, dropped), func_8016ba68 (3,886,
4/138), func_801466f0 (3,288, the S6b D4 wrapped-alias gap).
Continues the S11 lane. Fleet 96.01 -> 96.06% fn-count / 93.6 -> 93.7% instr /
88.0% distinct; dedup 1905 -> 1907 groups, 0 failed, C1 240669/240669.
R22 clean-fleet: 140 passed, 0 failed of 140. 0 NON_MATCHING (G4).
EXTEND (SC07): the 16 volatile-blocked DIFF slots banked on retry after the
data asm-label alias -> lane total 31/36.
PROPAGATE head, measured rather than projected. .run/s8_lag.json re-split: the
checkpoint's "45 classes / 20,837 ins" is really 5 classes carrying 18,545 ins
(89%) and 41 carrying 2,316. Per-class outcome:
func_80147364 30x137 = 4,110 BANKED x137 (definition-side asm-label alias)
func_8012f274 29x137 = 3,973 DROPPED — byte-diverges in ~130 overlays
func_8016ba68 29x134 = 3,886 4 of 138 banked; excluded from ~130
func_8012a598 24x137 = 3,288 SKIPPED, cause NAMED by the tool
func_801466f0 24x137 = 3,288 no source found — the S6b D4 gap, still open
func_80147364's byte-true definition is `(u16, u16)` while 4,046 fleet decls
say `(u16, s32)`. u16 is a default-promotion type, so the `()` no-prototype
escape is ILLEGAL (the documented gcc-2.7.2 dead-end) and conforming the decl
would change caller codegen. The DEFINITION-SIDE asm-label alias gives the def
a distinct C identifier while emitting the real symbol -- zero blast radius on
every caller. Probed on ONE member first (1 build, not 137 -- the S29
discipline): byte-identical 9052dc0e first try; then 137 overlays clean.
In-tree precedent for the form: 1,725 files.
MEASURED NEGATIVE, recorded not buried: `dedup_propagate --recover` banked only
4 of 138 on func_8016ba68 and dropped func_8012f274 entirely (137 [exclude]
lines). The caller-extern reconcile that is 16/16 lifetime ON DRAFTS does NOT
transfer to PROPAGATION of these two. Cause not yet diagnosed -- probe one
excluded overlay's build output before any further attempt (§136a), do not
re-run the lever hoping.
NAMED NEXT (cheapest first): func_8012a598 skips on `missing file-scope extern
(CARRY-FIXABLE): D_801151D4, D_80126DB8_a, D_80127504` -- the SESSION-18
preamble-backscan class. Its body is 2 statements and `struct BigCopy` is
ALREADY in the shared engine_types.h (L312) with the identical statement already
macro-ized at engine_core.h:16158, so a hand-authored macro (the func_80147364
path) should take it x137 for ~0 tokens.
Process errors recorded in CURRENT_PHASE.md, all three one mechanism -- the
signal sampled is not the thing waited for: (1) a `nohup CMD &` wrapper's exit
read as the fleet check finishing (it stood at 63/140); (2) a corpus.stubs probe
mid-rebuild, which R32's coverage assertion refused rather than answer wrongly;
(3) CORRECTION to the S10 checkpoint's own rule -- `pgrep -x make` is right for
one make and WRONG for a campaign of sequential makes (it fired in a gap and
reported a live campaign done), and `pgrep -f <pattern>` SELF-MATCHES so that
waiter can never exit. Wait on the campaign process or `treelock.sh --status`.
- DREW'S CALL (2026-08-03): route the 30-target x2-9 wave to SONNET instead of Opus. The §136i
>=120-ins Opus threshold was MY EXTRAPOLATION, never measured; this wave (125-793 ins) probes
exactly the region where there was no data.
- RESULT: 13 banked of 16 that ran = **81%**, vs Opus's 10/13 = 77% on the comparable S8-3 slice.
At least 8 banked SONNET-DIRECT (65 agents spawned: 60 sonnet, 5 opus escalations). Propagated
57 member-matches / 4 failed across 42 overlays. 70 instances. R22 clean-fleet 140/140.
FLEET 96.01% fn / 93.6% instr / 88.0% distinct (77,550 uniq).
=> **Sonnet is at least as capable as Opus on 125-793 ins. The ~120 boundary is too low.**
NOT rewriting it to a specific number yet: 16 samples under a throttle confound cannot name a
cliff. The controlled A/B (task #12) is how that number gets fixed properly.
- THE REAL LIMITER IS CAPACITY, NOT CAPABILITY: 14 of 30 agents were killed by SERVER-side
throttling ("Server is temporarily limiting requests (not your usage limit)") that 30 concurrent
Opus agents did not trigger. Practical rule: run Sonnet waves at ~12-16 concurrency, not 30.
The 14 unrun targets are listed in the checkpoint for a smaller-batch retry.
- Sonnet's work quality was not shallow — three examples: func_8018797C read local-alloc.c and
forced loads into an AGGREGATE to stop find_free_reg greedily taking 3 callee-saved regs;
func_8018D870 used §136c sibling-first for ~70% of the body then blocked a coalesce with a pin;
func_8017E3AC diagnosed an RC-3 callee-saved-order swap and noted the pin must be s32 or a stray
`andi 0xffff` appears.
- MY ERROR, RECORDED: `until [ -s <output> ]` fires at the FIRST LINE of output, not at completion.
It fired mid-propagation and I ran `make clean` on top of a live family_sweep, deleting asm/ and
aborting both the regen and the sweep (corpus's R32 assertion refused to answer rather than return
a wrong stub set — working as designed). No bad bytes: R22 verified 140/140 immediately after, and
the propagation simply re-ran clean. Correct waiter is `pgrep -x make` (exact process name), which
also cannot self-match the way `pgrep -f <pattern>` did when it leaked 4 waiter shells earlier.
Third instance today of ONE root cause: trusting a proxy instead of the thing itself (a weight
column vs a probe §136h; an exit status vs build output §136a; file-existence vs process exit).
- CALIBRATION (25 stratified targets: 12 head-by-weight + 13 sampled across the band, so the
measurement captures the DECAY, not just the head): 28 agents -> 22 claimed -> gate BANKED 22/25
(88%) -> propagated 67 member-matches / 13 failed across 38 overlays. 89 instances.
**22,937 templatable instructions banked in one wave.**
HEAD 9/12 -> 19,492 of 28,584 templ ins
BODY 13/13 -> 3,445 of 3,445 templ ins (the small ones are EASY; all 3 misses were 611-793 ins)
- FIRST SONNET DATA (§136i ladder's new middle rung): **Sonnet 6/6 · Haiku 6/6 · Opus 10/13.**
The two cheap tiers went 12/12 and Opus absorbed every hard failure — consistent with correct
size-routing rather than luck. Small n; the controlled A/B stays parked (task #12).
- THE PROJECTION for the >=95% instr bar (Drew's decision input): 411 of 1,872 families cover the
212,594-instruction gap = ~19 waves optimistic, 20-30 realistic. Mean templ ins/family decays
1,844 (top-25) -> 1,046 (top-100) -> 525 (top-400) -> 193 (band-wide), so early waves look like
this one and later ones bank MORE functions for FEWER instructions.
- DECISION (Drew): NO phase close — keep grinding. Campaign tracked as task #15.
- Carried failures -> next lanes: func_8017D174 (793 ins, closeness 5 after ~90 variants; diagnosed
a backward-scheduler priority race -> permuter, correctly NOT ledgered a wall), func_80186E24
(611 ins, 133 of 139 diffs pure register numbers -> a natural §137 test), func_8017E2EC.
- §136c sibling-first paid again: func_8017DF84 (766 ins) MATCHED because a banked byte-matched twin
existed in the same TU; its 697 index-diffs traced to ONE root cause (a bare 0xFFFFFF literal that
loop.c hoisted to the OUTER preheader, stealing $s3) — closed by binding it to a local declared as
the FIRST statement of the inner loop body. Verified via rtu_match (real-TU), not just match_one.
- S8-3 (23 fresh x10-99 families, 121-328 ins — the hardest band this session): draft 16/23 ->
capture (1 PLUMBING / 6 DIFF) -> reconcile 1/1 -> redraft 6/6 => **23/23 (100%)**.
Propagated 206 + 81 = 287 member-matches across 80+54 overlays. R22 clean-fleet 140/140.
FLEET 95.97% fn / 93.4% instr / 87.5% distinct (77,404 uniq); dedup 1905/0; 0 NON_MATCHING.
- §136b CLOSES AT 15/15 — no function ledgered "genuine byte-DIFF" survived a redraft, all session.
- §137 (NEW, the session's most reusable result): REGALLOC-PERM — a clean 2-register swap — is a
TWO-COMPILE ARITHMETIC PROBLEM. global.c:allocno_compare ranks by floor_log2(R)*R/L*1e4*size;
read R and L out of `cc1 -dl -dg` for BOTH contenders AND their ranked neighbours to get the
admissible priority WINDOW, then place a zero-byte `__asm__ __volatile__("" ::"r"(v))` so L lands
inside it. func_801833F0: contenders ONE unit apart (1297 vs 1296), window (1228,1296), five
placements probed, only L=219 -> pri 1232 worked. R and L are FORCED BY THE EMITTED CODE (L is
recomputed post-sched1), which is exactly why source-reordering is a dead end for this class.
Converts a class the permuter banked 0 from all session into a deterministic calculation.
Companion: floor_log2 makes ref-count a STEP function (5/6/7 refs are worthless, you must reach 8)
— func_8017EFA8 closed 30 register-name mismatches by taking a pseudo 4 refs -> 8 with a dead read.
- §136j — the failure MIX FLIPS WITH SIZE: <=120 ins fails ~70% on declarations; 121-328 ins fails
86% on genuine codegen. Budget reconcile for the small band, redraft for the big one — and do NOT
read 70% on a big-function wave as a broken pipeline; that is the expected shape.
- §137a — a gate verdict has a TIMESTAMP. Two "DIFF" ledger entries were STALE (draft rewritten 28
min after the gate ran, never re-gated); both were already byte-perfect. Compare verdict time to
draft mtime before redrafting. Plus two offline oracles an agent built: a FULL RELOCATION RESOLVE
(catches wrong jal/%hi/%lo targets that match_one's mask hides) and a COLLATERAL CHECK (whole-TU
objdump with/without splice). Together they discriminate all three causes of "match_one says MATCH
but the overlay SHA differs" without running make.
- §136f addendum — the collider is often an already-banked SIBLING BELOW the splice; locate it by
arithmetic (draft grows the file N lines, so TU line L reports at L+N).
- cookbook-index 380 -> 382 sections.
- MEASURED BASIS (P30 S7, 144-target campaign): the two-tier rule from the 2026-06-29 A/B left the
~50-120-ins band unassigned, and every wave since defaulted it to Haiku-with-Opus-escalation.
Haiku-direct banked 3/8 on that band while Opus-escalation-after-a-Haiku-miss banked 10/11 — i.e.
Haiku was acting as EXPENSIVE TRIAGE (a wasted draft + a full Opus redraft), not a cheap drafter.
The original A/B only proved parity <=52 ins; everything above that was extrapolation.
- LADDER: haiku <=~50 ins · SONNET ~50-120 · opus >=~120 or escalation · fable5 for a genuinely NEW
wall class only. Never haiku->opus directly; never default a whole wave to opus because the band
"looks hard" (the same extrapolation in the other direction).
- WIRED, not just documented: s7_manifest.py routes by the new thresholds; s7_wave4b.js escalates
haiku->sonnet->opus instead of haiku->opus, and its meta/prose say so.
- Boundaries (~50/~120) are current best estimates — re-measure per-tier from the journal + the gate,
never from the workflow's by_tier (it counts claims, not banks — §136).
- Byte-gate remains the sole arbiter, so a weaker drafter is a throughput risk, never a correctness
risk (G3/P9). cookbook-index 378 -> 379.
- FINAL LANES: reconcile ×5 (5/5) + redraft ×1 (1/1) -> gate BANKED 6/6 -> propagated 50 members
across 28 overlays. **ALL 144 DRAFTED TARGETS BANKED (100%); zero stubs remain in the queue.**
R22 clean-fleet 140/140 (seventh time this session).
FLEET 95.88% fn / 92.9% instr / 86.5% distinct (77,106 unique fns); dedup 1905/0; 0 NON_MATCHING.
- LANE RECORDS: reconcile 15/15 lifetime · redraft 9/9 · §136b closes at 9 FOR 9 (every function
ever ledgered "genuine byte-DIFF" banked on redraft).
- THE CAPTURE CLASSIFIER, third and final defect (§136a): it decided PLUMBING by matching a regex
against cc1's PROSE, and cc1's vocabulary is open-ended — `too many arguments to function` matched
nothing, so a trivially reconcilable function sat UNKNOWN through two gate rounds. Now DERIVES the
class from the closed invariant (did the compile produce an object: `make ... Error N` +
`Deleting file`). Re-running it moved 5 PLUMBING / 1 UNKNOWN -> 5 PLUMBING / 1 DIFF, and BOTH
reclassified functions then banked. Three defects in one small tool in one session — an
unreachable exit-status branch, a missed phrasing, and the prose-matching design behind both —
each SILENTLY MIS-ROUTING REAL WORK. R33 in one line: if an invariant answers it, never re-parse.
- §136f — two declaration sub-cases: (1) a symbol you call may be DEFINED, not just declared, BELOW
your splice point (func_8017D540 is defined 275 lines below as int(int); the draft guessed
void(s32) from a bare jal); (2) an ARITY clash on the symbol you are DEFINING cannot be fixed by a
cast — use the §37/§124 asm-label alias (func_801848DC; in-TU precedent at :8872).
- §136g — TWO INDEX ROUTINGS BYTE-REFUTED (func_801863B4). The index sends BRANCH-POLARITY to §3-T4
(invert) and §34 (zero-byte fence); the agent tested BOTH at zero, read the gcc-2.7.2 source, and
found jump.c:1806 `if (foo) bar; else break` range-swap — which runs long BEFORE reorg, so a fence
CANNOT block it. Real lever: put a label between the if-join and the return label (wrap the loop
in the guard). Also: same-address lh+lhu is MIPS LOAD_EXTEND_OP==ZERO_EXTEND (mips.h:1163), and
combine collapses the pair unless the HImode pseudo has two reaching defs.
REFUTED ROUTINGS ARE RECORDED NEXT TO THE CORRECT ONE — otherwise the next agent re-runs them.
- cookbook-index 375 -> 377 sections (§136 .. §136g earned this session).
- BATCH 3 (37 targets, 41 agents, 2.75M tok -> 35 claimed): gate BANKED 35; family_sweep propagated
305 member-matches / 39 failed across 77 overlays (14 STRUCT skipped by design). 340 instances.
R22 clean-fleet 140/140 (sixth time this session).
FLEET 95.86% fn / 92.9% instr / 86.5% distinct (77,061 unique fns); dedup 1905/0; 0 NON_MATCHING.
- WAVE 4b COMPLETE: b1 32/37 + b2 35/37 + b3 35/37; with wave 4a (30/33) the whole 144-family
B-shape queue that opened this session is worked through — 138 of 144 drafts banked (96%).
- §136e — batch 3's two HONEST NEGATIVES, worth as much as the wins:
(1) §136c SIBLING-FIRST HAS A PRECONDITION. func_801899AC's family has all 13 members still
unmatched and no engine_core.h twin, so there IS no byte-verified sibling and the search is
pure cost. Check a banked sibling EXISTS before spending the greps.
(2) A loop increment in the loop-back DELAY SLOT + a compensating negative addiu is a SOURCE
SHAPE, not a reorg artefact — MIPS1 has no annulling, so reorg CANNOT invent the
compensation. Write `p += 2; if (t == cur) break; ... p -= 2;`. combine's reg_n_sets==1 guard
stops the addiu folding into the following lw. The index's delay-slot entries point at reorg,
which is a dead end for this class.
Plus a new §136-L1 application on the RETURN axis (an over-scoped temp became a global allocno and
swapped $v0/$v1 with the returned local, collapsing the target's `j` + `addu` return).
- COMPOSITION, demonstrated on func_8017D5F4 (46 ins): flat early-returns -> dead-local frame pad ->
s16 locals -> operand order -> 3 register pins -> 2 zero-byte re-ties -> permuter for the last 2.
THE PERMUTER IS THE LAST STEP ON AN ALREADY-PINNED BASE, not the first.
- cookbook-index 374 -> 375 sections. 6 stubs remain; per §136b none is a wall on one attempt.
- THREE LANES: wave 4b batch 2 (37 targets, 46 agents, 3.44M tok -> 35 claimed) + the reconcile lane
on 3 PLUMBING failures (3/3) + a REDRAFT lane on 4 DIFF-ledgered failures (4/4). Combined gate
BANKED 41; family_sweep propagated 431 member-matches / 29 failed across 79 overlays.
103 of 107 drafts banked (96%). R22 clean-fleet 140/140 (fifth time this session).
FLEET 95.77% fn / 92.9% instr / 86.4% distinct (76,824 unique fns); dedup 1905/0; 0 NON_MATCHING.
- §136b CLOSES AT 8/8: every function ledgered "genuine byte-DIFF" banked on redraft — wave 3's
four, the THREE I classified from wave 4a's capture, and one from batch 1. The classifier is
right about what it measures ("this draft compiles clean and differs in bytes"); reading that as
"this function resists matching" is the error. A DIFF verdict is a fact about ONE DRAFT.
- §136a CORRECTED (a reconcile agent refuted me against the bytes): I wrote "70% of gate refusals
are paperwork, not codegen". WRONG. A declaration conflict ABORTS THE COMPILE, so a PLUMBING
verdict says NOTHING about the body. Two of three second-round PLUMBING drafts had a real codegen
residual behind the conflict (func_80188694 DIFF/4 SCHEDULE-REORDER, closed with a §21 zero-byte
re-tie after six other variants failed; func_8018C638 DIFF/6 ADDRESSING/cse). Both agents ran
match_one on the untouched draft FIRST and rejected my premise — which is what §135 asks for.
- §136c SIBLING-FIRST IS A DERIVATION SHORTCUT, not just a conflict fix: grep engine_core.h's
DEFINE_func_* bodies for a byte-verified NEAR-TWIN before deriving from the .s. func_801859D8
found DEFINE_func_80185978 (identical offset chain, 3 differing constants), reused its expression
forms verbatim -> FIRST-DRAFT MATCH, and the twin generalizes to its whole 10-member family.
Search order: near-twin -> banked same-TU sibling -> the .s -> the Ghidra seed LAST (byte-proven
an entirely different body twice this session).
- §136d, four new gcc-2.7.2 levers from the redraft lane, each with its REFUTED axis recorded:
RC-12 $0-add opaque copy (cse.c canonical-copy promotion; do NOT pin the pair to real regs);
jump.c if-then-else -> conditional-overwrite collapse (defeat with TWO SEPARATE CALLS, not a
ternary); fix the STORE not the load for a load hoisted above a constant-address store (the
INDIRECT_REF reshape is the wrong half of the /s lattice, 2 -> 32 mismatched); a branchless flag
is -(a != b) & 0xFF, never a ternary.
- cookbook-index 372 -> 374 sections. Batch 3 staged with all of the above promoted into its prompt.
- WAVE 4b BATCH 1 (37 volume-lane targets, 10-19 members, <=60 ins; wave 4a's §136 idioms promoted
into the drafting prompt per the measured 83%->93% law): 50 agents / 4.35M tokens / 32 min ->
34 claimed MATCH -> gate BANKED 32 -> family_sweep propagated 365 member-matches / 4 failed
across 84 overlays (3 STRUCT skipped by design). 397 function-instances from 37 targets.
- §136b — THE FINDING THAT CHANGES THE BACKLOG: all FOUR functions wave 3 ledgered as "genuine
byte-DIFF" BANKED on redraft. The recorded causes were never codegen:
func_801845B0 a branch to the EPILOGUE misread as an inner early-exit -> the whole tail was
hoisted out of its enclosing if (control-flow misread)
func_80184A94 a declaration conflict on a symbol declared BELOW the splice point; closed by
copying an already-banked family sibling's decl forms verbatim (§71)
func_8017BEBC the cached Ghidra seed was an ENTIRELY DIFFERENT body and the prior draft
followed it; the .s was the only usable source
func_8018480C re-derived clean
=> a DIFF verdict describes THE DRAFT THAT WAS ATTEMPTED, never the function's matchability.
Never retire a target on one; route it to REDRAFT. And re-GATING an unchanged draft is not a
retry — which is exactly why wave 4a's 3 DIFFs stayed stubs through this gate (same bytes
resubmitted); they still owe an actual redraft and are now likely winnable.
Corollary: backlog entries carrying an old closeness/class are stale by construction (P29
measured 77% of stored drafts decayed) — re-verify before valuing one.
- The wave-4b prompt handed each retry its prior verdict EXPLICITLY LABELLED "a data point, not a
verdict — re-derive from the .s". Every retry agent did exactly that and refuted it.
- R22 clean-fleet 140/140 (fourth time this session). FLEET 95.63% fn / 92.8% instr / 86.3% distinct
(76,499 unique fns); dedup 1905/0; C1 240496/240496; 0 NON_MATCHING (G4).
- Orchestration: this batch's workflow script was GENERATED from the manifest files rather than
hand-pasted — transcription had already cost this session one dead launch (args-as-string) and
cost the prior session three agents' time (hand-typed _jr_* paths). Generate the artifact; do not
ask yourself to be careful. cookbook-index 371 -> 372 sections.
- RECONCILE LANE: all 7 PLUMBING failures FIXED and banked (329K tokens — ~13x cheaper than the
drafting wave's 4.44M). Propagated +76 member-matches / 0 failed across 51 overlays.
Wave 4a final: 30/33 heads (91%) + 327 members = 357 function-instances from 33 drafted targets.
- THE CAPTURE CLASSIFICATION WAS EXACTLY PREDICTIVE: all 7 PLUMBING banked, all 3 DIFF stayed stubs
(func_8017E978 / func_80184494 / func_80184960 -> redraft lane, their C is wrong). That is what
makes the ~10-build capture step worth running before any reconcile fan-out. The lane is now
19/19 across three waves.
- EVERY reconciled draft had a HIDDEN SECOND CONFLICT cc1 never reached (it reports only the first)
-> "grep the whole TU in one pass" must be in the RECONCILE prompt, not just the drafting prompt.
One agent additionally assembled the spliced TU and masked-compared its function IN TU CONTEXT
(67/67) — proving the casts byte-neutral in situ, not merely standalone.
- NEW HAZARD, agent-surfaced (§136a corollary): an agent chose a SHARED scratch path, a concurrent
agent overwrote it, and its verification silently compiled ANOTHER agent's TU and returned a
meaningless rc=0. It caught the swap only because the emitted .s lacked its own function. A shared
scratch path yields a CONFIDENT WRONG VERDICT, and no tool fix reaches it — the choice happens
inside the agent, so the PROMPT must mandate a process-unique path. This is the Phase-28
match_one fake-isolation defect recurring one level up.
- R22 clean-fleet 140/140 (third time this session). FLEET 95.52% fn / 92.7% instr / 86.1% distinct
(76,273 unique fns); dedup 1905/0; C1 240496/240496; 0 NON_MATCHING in any default build (G4).
- .run/s7_extra.txt: wave 4a's idioms compiled into the wave-4b drafting prompt (the promotion that
measured 83%->93% between waves 1 and 2).
- S6c (deterministic, ~0 agent tokens): 12 sibling banks across 3 of 9 jr zero-crack families
(func_80178D40 890ins 4/4, func_801734BC 4/4, func_8012ACE0 4/4). The other 6 are ledgered:
5 gate-fail (genuine byte DIFF) + 1 carve-fail (span table starts do not fit the span).
R22 clean-fleet 140/140 over the whole S6c series.
- FINDING: all 12 banks landed in ov_SC07_006/007/010/011 — the four overlays P27 discovered and
P28 made citizens (R36). P28 drained their h_exact backlog via dedup_extend; the jr/h_seq
propagation lane was still owed. R14 GUARD AGAINST OVER-READING IT: the quartet are the top four
overlays by remaining zero-crack residue (2,190-2,355 ins each vs 500-870 typical) but hold only
7% of the 2,114 remaining slots — a per-overlay priority signal, NOT a bulk lever.
- BLOCKER CAPTURE for the 10 wave-4a gate failures -> .run/s7_blockers.json: 7 PLUMBING (all
`conflicting types for func_X`) / 3 genuine byte-DIFF. 70% of "the gate refused" is declaration
paperwork. New tool .run/s7_capture.py (any overlay/any draft dir; reverts the TU in a finally:).
- MY DEFECT, FIXED AND DISTILLED (§136a): the capture tool first classified on the EXIT STATUS, so
its `rc == 0 => byte DIFF` branch was UNREACHABLE — `make build` runs `check`, so a draft that
compiles perfectly and merely differs in bytes also exits non-zero, and all 3 real DIFFs were
filed as "unknown". Now classifies on the OUTPUT ([FAIL]/got/want vs a non-warning error line);
the warning-exclusion matters because `conflicting types` also appears benignly for builtins.
- Also probe-discipline: my first S6c probe reported 1/9, which was 1 bank + 8 CORRECT REFUSALS —
jtbl_family_bank refuses on a dirty config/+src/ (its per-sibling revert restores from HEAD).
Driver now commits between families. A uniform failure across N functions is a statement about
the mechanism, not the functions (§134).
- cookbook-index 364 -> 371 sections, --check green. CURRENT_PHASE SESSION-31 checkpoint refreshed
with the queue re-derived at HEAD (the S30 ROI-floor trigger stays REFUTED — do not close on it).
Drew T