BFM-decomp/phase-ends/CURRENT_PHASE.md

CURRENT PHASE — Phase 30: The Recovery & Concentration Campaign (overlays to their ceiling)

Opened: 2026-07-30 · Effort: Max (Fable 5; saved defaults) · Generation: Gen2 (22nd phase of the arc) · Baseline: PhaseEnd_Phase29 (v1.28.0) · Roadmap: docs/roadmap-to-100.md v2 §3 P30 Plan approved (P3 gate 1): 2026-07-30, in-session (Drew's explicit waiver of the fresh-session step at 28% context). Full plan mirror: ~/.claude/plans/continue-dazzling-newt.md (out-of-repo; this file is the in-repo authority + crash-recovery log).

Fleet at open: 87.5% instr-weighted / 78.0% distinct-code / 92.00% fn-count · 140/140 byte-identical · 0 NON_MATCHING · dedup 1886/0 · tools-health RC=0.

Drew's Phase-Start decisions (2026-07-30):

• In-session start (fresh-session step waived at 28% context).
• The Ghidra-C prefetch batch pulled forward from P31 → T0.5 (tail + main's 1,034 stubs; headless, background).
• T2's primary path = the two-file atomic o0b substitution driver (the log's costed no-splat route); Arm-A splat research is the fallback only.
• ROI floor for T3: two consecutive sessions each < +0.3pp instr with all lanes exercised (Drew-adjustable).

⚠️ PRECONDITION (R6): Drew makes the milestone-close commit + push (PhaseEnd_Phase29.md + phase-ends/logs/Phase29.md, both in the tree) before P30's first commit lands. P30 work proceeds; the first P30 commit waits on it.

---

The numbers this phase must pin FIRST (T0 — do not consume until re-derived, R35/R14)

• family_hseq 29,961 vs progress.py 28,296 remaining instances — an unexplained R32 gap.
• The -O0 population: "10 families / ~1,287 raw distinct" (checkpoints) vs "123 families / 121,264 ins / 3,100 distinct" (the log's R14 correction) vs "129 distinct on the table" (the costed o0b route) — three numbers, one T0 job.
• The zero-crack family list (map said 61 on 07-29) and the concentration table (top-20/top-100 shares).

Standing invariants (every task)

• Whole-binary byte-gate is the sole arbiter (G3/P9); match_one/closeness/masked = candidates (§52b).
• R22 clean-fleet per banked batch: make clean && make extract-all && make check-all → 140/140.
• make tools-health green + fail-closed before matching (corpus · cdecl · audit-binaries(R36) · report/lint/dedup).
• Blast-radius typing on every integration write (T0 draft-only / T1 binary-local / T2 fleet-shared ⇒ R22 mandatory).
• Probe before costing (R37-candidate discipline): probe one member before pricing a job; derive attribution from corpus.stubs before/after; diff the artifact to prove an edit ran (§120).
• Flywheel: idioms → cookbook + tooling in-session (R16/R30); pivots → decision-log (R31).
• Effort/model transitions prompted, never assumed (R26/R27) — STOP + WAIT for the toggle.
• One commit per task/sub-repair after this file's update; Drew pushes (R6/R20). R23 no-op on db.*.gbf churn.

---

Task checklist (current-task pointer = ▶)

• T0 — Frontier regen + instrument repair [Max] ✅ (all six letters; see per-task log; frontier report = docs/frontier-p30.md, report point #1): (a) gate_stage.py snapshot/restore (the ladder destroys good drafts) · (b) rtu_match.py surface the real cc1 error in the verdict · (c) reconcile the family_hseq↔progress instance gap (R32) · (d) purge since-banked rows from the backlog ledger + still-a-stub filter from corpus.stubs (R33) · (e) refresh .run/autopsy/residuals.jsonl + fix the 21-file absolute-include defect · (f) regenerate the frontier (report / family_hseq / worklist / fuel manifest) + pin the three contested populations → the one-page frontier report (report point #1). Verify: tools-health green; digests committed; velocity derives from digest git history (R33).
• T0.5 — Fleet Ghidra-C prefetch batch [background] ✅ (SESSION-27: 124/124 programs, 7,716 files, 335 min unattended; 95.2% of crack-needing families seeded, main 2,001/2,002, resident 14/14. Was left unticked through S27 — corrected S28.) — ImportOverlay.java (~130 missing programs) + DecompileFunctions.java over every remaining distinct stub (overlay tail + main's 1,034; skip LINKED) → .run/ghidra_c/. R23 lock discipline (stop the hook-launched MCP server first or route through it). Verify via build_fuel_manifest.py counters + 5 spot-reads.
• T1 — Integration-recovery sweep ✅ DELIVERED as T1a (report point #2), then CLOSED both ways. 108 fresh autopsy-MATCH strandeds / 50 binaries through the tiered driver → +18 banked (derived: fn-count +18 / distinct +12; the commit-message sum of 19 was a gate-report double-list — the metric wins), R22 140/140. The measured verdict: the S16 "39%" prior did NOT generalize (it was measured on FRESH same-wave drafts); on this stored-backlog population — the class decision-log A10 had already byte-proved non-bankable by plain re-gate (0/958) — the driver's reconcile/de-macroize lifted 18 of ~111 ≈ 16%, a real but bounded delta. The ~90 that did not bank stay match_one-MATCH but are integration-DECAYED: they route to T3's REDRAFT lanes, not another recovery pass. No Ultracode was needed (the lever was deterministic).
• T2 — The -O0 cluster [Max, deep] ✅ ROUTE PROVEN, TOOLED, AND SWEPT FLEET-WIDE (S28). Report point #3 = this entry. DEVIATION from the phase-start plan, byte-justified — BOTH named routes were refuted: the PRIMARY "two-file atomic o0b substitution driver" mis-places by construction unless the target abuts the o0b object's END (the ×1 probe), and the FALLBACK premise was also wrong — the Arm-A splat %lo +0x20 does not bite (a pure -O2 sub-split of a jr object is byte-neutral, probe 1). The real constraint is §126: an address range is not an optimization region — it contains already-matched -O2 bodies, so the bound is (range MINUS matched) and K islands ⇒ K+1 regions (probes 3–4). Shipped: tools/o0_subsplit.py · Makefile -O0 glob _o0b→_o0? · cookbook §126/§126a. Swept: 135/135 overlays, 0 refusals; 140 new -O0 region files; o0_sources() 137→277; 0 invisible to the oracle; 2,200 open stubs now in a genuinely -O0 TU. 6 fns banked (ov_SC03_014/015). R22 140/140; tools-health OK. Population correction (R14): my "275 stubs / 18 overlays" was an 8× under-count from a bare except swallowing R32 during a rebuild; truth 2,184 / 138, which vindicates the T0(f) "2,192" pin I had wrongly called stale (§126a). Handoff → T3: drafting the 2,200 is crack-wave work; seeds cached; match_one --o0.
• T3 — The standing crack-wave loop ✅ RAN across S26–S29, all four lanes exercised, with a measured per-lane verdict (the ROI floor is a LANE property, not a phase property): · Lane B (top-mass fresh families) PAYS and is not exhausted — S27 waves 1–4 (8/14 · 15/19 · 48/48 · 69/70) + 7 confirmed behemoths at 700–970 ins; S28 the jtbl OVER-SPAN fix + the 710-ins func_80191C50; S29 the jr-pair = frontier families #3 and #4 by templatable weight, banked ×138 each for ~0 agent tokens after the instrument repair. · Lane C (tail mass, ×2-reach cached band) is AT the floor — S28 measured 1.33M tokens → 12 banks → +0.00pp headline. That, not the phase, is what the ROI floor names. · Lane A (zero-crack propagation) + Lane D (PINS/W4/permuter) exercised; 29 zero-crack families remain as propagation-only fuel behind matched exemplars. (Original text:) — lanes interleaved, propagate behind every crack same-session: A zero-crack (~61, propagation-only) · B top-mass fresh families (no size cap; jr via jtbl_family_bank; gate_stage call-site-casts, never bulk header edits) · C tail mass on prefetched seeds (local v3 ≤15 → GLM/Haiku ≤~50 → cheap-Opus mid) · D PINS (19) + W4 bounded diagnoses + permuter backlog via the T4-fixed grinder. Per-session checkpoint + velocity + 3 metrics. ROI floor: 2 consecutive sessions < +0.3pp instr each with all lanes exercised.
• T4 — Carried-tool resolution [xHigh] ✅ (S28, commit:1279). The 2 "Phase-22 grinder bugs" were struck as STALE in S27 (already fixed). This session: --fix-def-sig posture audited clean (defaults False, one consumer, no caller passes it) — and the audit found the live hazard the earlier pass missed: docs/decision-log.md still recommended making it default-on, which T84/§119 byte-refuted; struck in place with a superseding note (R31). Grinder ILS warm-restart WIRED — permuter_ils.py had sat unused beside grinder.py since Phase 24, so every grind was a COLD search; now --cycles (default 4) warm-restarts from the best byte-waypoint, --cycles 1 = old behaviour. Justified by measurement before building (the lane looked dead: 0 banks since Phase 21) — 665 open near-misses in the permuter-tractable band, 157 at close 1–4, incl. func_8016BA68 close=1 reach=134. ⚠️ Yield UNPROVEN on this backlog — wiring only; the ILS evidence is one Phase-24 fn.
• T6 — The measured-frontier continuation [approved by Drew 2026-08-01, in-phase re-plan (P5d)] — P30 stays OPEN. Order derived from the S29 frontier regen, ranked by the family map's TEMPLATABLE weight (see the pricing finding below), NOT by worklist.md: S1 zero-crack propagation sweep — 29 families / 67,470 ins, ~0 agent tokens (head 0x8014032C 183×137 = 25,071 jr · 0x8017BEBC 12,376 · ~8 ×4-straggler families) · S2 the LAST two reach-138 fresh cracks — func_80176734 (371×138 = 51,198, no jr, cached) + func_8016EC0C (88×138 = 12,144) · S3 the close=0 stored drafts as a diagnostic pass (§132 ladder per draft), not a blind re-sweep (T1a's ladder measured 16%; S29 resolved 2 of 2 "gate refuses" to tooling) · S4 PINS bounded wave (14 fns / 44,279 ins; the pin-crash wall dissolved in P27) · S5 the ×10-133 mid-multiplicity families (142,527 ins). Deliberately excluded: the 2 GIANT walls (close=110/91 — P32's track), the ×2-9 mass (339,742 ins, worst multiplier), the singleton/non-substantial residue (635,805 ins, ×1). ⚠ THE PRICING FINDING (R14/R35): worklist.md ranks by h_exact reach, so a per-location PURE family is priced ×1 — it under-prices the frontier head by up to 138×. Byte-proof: S29's pair was priced 272 and 198 ins and delivered 37,536 + 27,324. func_80176734 sits at rank ~50 in worklist.md and is the single largest item on the board. Rank by .run/family_hseq.json templatable weight for family work; the worklist is correct only for genuinely h_exact-reach functions. ⚠ THE STRUCTURAL SIGNAL: after S2 the ×138 era ENDS — those are the last two crackable fleet-wide families; everything after is ≤133 members and mostly ≤9. That rise in cost-per-point, not a session count, is P30's honest ROI floor and the trigger for T5.
• T5 — Phase close [Max] — burn-down from digest history; P7 milestone walk; Roadmap delta (including the v2 gap this analysis exposed: bucket W3's remainder has NO owner phase after P30 — P31 is T/M/R, P32 is walls/behemoths); gate 2; PhaseEnd_Phase30.

Milestone: overlays at their measured ceiling — ≥95% instr fleet, or every remaining overlay stub on a named wall/behemoth/queue ledger — 140/140 byte-identical throughout.

Blockers

• (none) — the precondition (Drew's milestone-close commit) landed at commit:1213.

---

⏳ SESSION-38 IN FLIGHT (2026-08-04) — WAVE 6 IS RUNNING, DO NOT make clean

A wave is IN FLIGHT. Workflow wf_4eabc086-c4c (.run/w6.js, 16 targets / 50,596 open templatable ins — 3× wave 5). Agents write .run/w6/<ov>/func_*.c ONLY; they read src/, asm/ and run match_one. The tree lock is FREE and must stay that way until the wave lands — a make clean/extract under a running wave is the S-era 63/140 incident. Check /workflows before assuming idle. Effort: ultracode (Drew enabled, regular high).

If this session dies mid-wave: the drafts on disk are the asset. .run/ is gitignored, so git add -f .run/w6/ them before anything else (the wave-5 precedent), then resume by GATING — never by re-running the wave (resumeFromRunId is same-session-only).

Resume/land sequence (unchanged from the ▶ block below): tools/treelock.sh g .venv/bin/python .run/s6f_gate.py '.run/w6/*/func_*.c' → capture any failure (.run/s36_capture.py <ov>:<fn>) → reconcile per the §138 direction rule → make sig-overlays + tools/family_hseq.py + family_sweep --hseq --band all --only <banked addrs> → R22 → commit → append a row to docs/wave-metrics.md.

New this session (both reusable, both committed with the wave):

• .run/w6_pool.py — the pool derivation, finally a SCRIPT instead of inline python. Ranks by open templatable weight (corpus.stubs over the member list, §138 rule 4), carries the walls/ledgered-residual exclusion set, drops families whose address was attempted-and-still-open in any prior wave (83 addrs), and enforces one target per (overlay, TU) so a wave cannot manufacture its own §138-rule-2 conflict. Pool: 2,928 fresh families / 476,611 open templ ins.
• Two derive-don't-assert fixes to the manifest itself (R37): model is routed by measured band (Sonnet is measured 81–100% over 125–793 ins; func_8017C974 at 947 is outside it → Opus directly, rather than paying a sonnet chain plus an escalation chain on the wave's biggest target), and seed is now derived from .run/ghidra_c/ existing on disk — it was being carried as an arbitrary i % 2 alternation, i.e. the prompt was telling ~half the agents a cached Ghidra seed existed without checking. 9 of 16 actually have one.

---

🛑 SESSION-33..37 CHECKPOINT (2026-08-04, wave 5 BANKED) — FRESH SESSION SAFE HERE

NOTHING IS RUNNING. Tree lock FREE. Tree CLEAN but for the R23 db.*.gbf churn — never stage. Effort: ultracode. R22 clean-fleet run TWENTY times this session, 140/140 every time. Last CODE commit commit:1393 (wave-5 bank); this checkpoint commits on top of it. Drew's standing decision: NO phase close — keep grinding. VERIFY THIS BLOCK IS CURRENT BEFORE TRUSTING IT (it goes stale, and stale is worse than absent): git log --oneline -3 should show this checkpoint at or near HEAD, and grep -E 'FLEET (fn-count|instr|distinct)' docs/progress.fleet.md must agree with the FLEET line below. If they disagree, the digest wins — re-derive before scoping anything (R35). Nothing is owed. The 16 wave-5 drafts flagged in the previous checkpoint are BANKED.

FLEET — R22 140 passed / 0 failed of 140

96.28% fn-count · 94.1% instr-weighted · 88.7% distinct-code (77,765 uniq) · dedup 1910/0 · C1 241216/241216 · 0 NON_MATCHING (G4). Session opened 96.01 / 93.6 / 88.0 ⇒ +0.27 / +0.5 / +0.7pp. Phase opened 92.00 / 87.5 / 78.0 ⇒ +4.28pp fn, +6.6pp instr, +10.7pp distinct.

WHAT THIS SESSION DID

lane	result
SC07 EXTEND	0/36 → 31/36
PROPAGATE head	0 → 18,545/18,545 ins (5/5 classes)
wave 1 (17 tgt)	15 banked + 65 members
wave 2 (13 tgt)	10 banked + 18 members
wave 3 (13 tgt)	13/13 + 21 members
wave 4 (14 tgt)	14/14 + 26 members
wave 5 (16 tgt)	16/16, ZERO reconcile + 26 members
reconcile lane	21/22 lifetime

📊 NEW: docs/wave-metrics.md — the wave performance log (Drew asked for this)

Wave-by-wave targets / tokens / wall-clock / parallelism / bank rate, with the derivation commands so future rows are COMPUTED not transcribed (R33). Append a row after every wave. Four findings, each with its caveat recorded rather than as a bare number:

1. The prompt is the lever, and the agents write it — 76→77→100→100→100% with models and gate held constant. The jump was STEP 0 (magic-literal grep of src/, ahead of engine_core.h), which came from a wave-2 agent's index_gap report. Caveat: waves 3–5 targets also trended easier; the mechanism is the durable claim, not the exact %.
2. pipeline() beats batched parallel() — 136 min/14 tgt → 82 min/16 tgt, 2.5×→3.8×. The two-batch design was a hard barrier (37–50 min dead gaps); the harness already caps at 16. Copy .run/s37w.js's execution block for every future wave. Floor: the slowest agent is still ~50 min of real match_one iteration ⇒ the lever is target SELECTION, not concurrency.
3. Economics — ~170–300k tokens/banked head, but the unit of value is head + propagated members, and sweep yield is bimodal not average (21/21 vs 18/165) because it is a property of the FAMILY.
4. A perfect gate means the prompt rules landed — waves 1–4 each needed 1–2 post-gate reconciles; wave 5 needed none. The reconcile lane is the fallback, not the plan.

📌 THE DISTILLED RULES (cookbook §138 + index; memory wave-prompt-seed-step0-and-gaps)

1. Bucket a gate refusal by the (macro-shape, TU-shape) PAIR, not the SYMBOL.
2. Reconcile direction depends on WHERE the TU's decl is — ABOVE the splice ⇒ DELETE your duplicate (§100); BELOW ⇒ KEEP a decl in the TU's EXACT shape and cast at the use (§17a-1 D2). A wave can manufacture this for itself: two targets in one TU means the first to bank puts its type tags in the second's way.
3. STEP 0: grep -rn "<MAGIC>" src/ with a distinctive literal from the .s, ahead of engine_core.h — the only step that reaches a banked twin in ANOTHER overlay's TU.
4. Never rank off the family map's exemplar field — derive open sites from corpus.stubs (measured 16,696 ins → 41,023 on the same map).
5. Three carry variants hide in one "CARRY-FIXABLE" bucket — multi-line comment (fix the tool) · draft-local struct Tag (use the shared type) · file-scope static inline helper (hand-author + EXCLUDE the source overlay).

🔬 SETTLED — do not re-litigate

• func_801758FC / func_80132018 / func_8018A808's families are NOT templatable. Probed (.run/s6_diag.py, 2 builds): remapped member is BUILD OK + byte diff ⇒ genuine per-member codegen. The h_seq refusal ceiling, confirmed.
• _alias_decl_for was ONE function, not a class (91 alias decls fleet-wide, 90 already matched).
• "normalized distance 0" is NOT an h_exact guarantee — relocs are masked.

▶ RESUME HERE — wave 6, nothing blocked

1. Regenerate the map, then derive the pool the corpus.stubs way (rule 4). Pool at the last regen: 852 families / 225,217 ins. Exclude the 3 walls (0x801412a8, 0x80178004, whale 0x80144b9c) and the ledgered residuals (0x8017c294 close=12, 0x8017f7b4, 0x801898e4, 0x80186e24 close=187, 0x801758fc, 0x80132018, 0x8018a808).
2. Build the script from .run/s37w.js (pipeline block + the accumulated extra prose), 16 targets. Launch → gate (.run/s6f_gate.py '.run/<w>/*/func_*.c') → capture any failure (.run/s36_capture.py <ov>:<fn>) → reconcile per rule 2 → make sig-overlays + family_hseq.py + family_sweep --hseq --band all --only <addrs> → R22 → commit → append a row to docs/wave-metrics.md.
3. EXTEND's last 5: func_80144B9C ×4 needs the §38 -O0 shared-header route (and dedup_extend should refuse-and-name that class per R32); func_80149954 ×1 sits behind func_80147364's u16 params.

🧰 MY PROCESS ERRORS — one mechanism: the signal sampled was not the thing measured

1. nohup CMD & inside a backgrounded call ⇒ the harness signalled the WRAPPER; the fleet check stood at 63/140 and I nearly read it as a pass.
2. pgrep -x make is right for ONE make, WRONG for a campaign of sequential makes; pgrep -f SELF-MATCHES so that waiter never exits. Use the campaign's real argv or treelock.sh --status.
3. A corpus.stubs probe mid-rebuild returned garbage; R32's assertion refused to answer.
4. I predicted a fix without reading the macro in front of me (relaxed 42 (void) decls; that macro declares (u8*) — banked 0/1 in all 134) → the PAIR rule.
5. I violated §136a in my own capture tool — a narrow keyword filter reported "NO COMPILE ERROR" on a build failing with redefinition of struct PW8017C290. No bad bytes from any of them — the byte-gate and R22 caught everything.

---

🛑 (superseded) SESSION-33..37 CHECKPOINT (2026-08-04) — pre-wave-5-bank

NOTHING IS RUNNING. Tree lock FREE. Tree CLEAN but for the R23 db.*.gbf churn — never stage. Effort: ultracode. R22 clean-fleet run NINETEEN times this session, 140/140 every time. Drew paused here to restart Windows. Drew's standing decision: NO phase close — keep grinding.

⚠️ THE ONE THING THAT IS OWED: 16 UNGATED WAVE-5 DRAFTS

.run/s37/*/func_*.c — 16 drafts, all claiming MATCH, NONE gated, NONE banked. They are force-added to git (.run/ is otherwise gitignored) because they cost ~2.7M agent tokens and the Workflow resumeFromRunId cache is SAME-SESSION-ONLY, so it does not survive the restart. Resume by gating them — do not re-run the wave:

tools/treelock.sh g .venv/bin/python .run/s6f_gate.py '.run/s37/*/func_*.c'

then: capture any failure (.venv/bin/python .run/s36_capture.py <ov>:<fn> — it ranks HARD errors above warnings) → reconcile per the §138 direction rule → make sig-overlays + tools/family_hseq.py + family_sweep --hseq --band all --only <banked addrs> → R22 → commit. Manifest: .run/s37_wave.json (16 targets / 16,884 templatable ins). Script: .run/s37w.js.

FLEET (at the last commit, before the ungated drafts)

96.27% fn-count · 94.1% instr-weighted · 88.6% distinct-code (77,723 uniq) · dedup 1910/0 · C1 241216/241216 · 0 NON_MATCHING (G4). HEAD commit:1391 + this checkpoint. Session opened 96.01 / 93.6 / 88.0. Phase opened 92.00 / 87.5 / 78.0 ⇒ +4.27 / +6.6 / +10.6pp.

WHAT THIS SESSION DID

lane	result
SC07 EXTEND	0/36 → 31/36
PROPAGATE head	0 → 18,545/18,545 ins (5/5 classes)
wave 1 (17 tgt)	13 heads + 65 members + 2 reconciled
wave 2 (13 tgt)	9 heads + 1 reconciled + 18 members
wave 3 (13 tgt)	13/13 + 21 members
wave 4 (14 tgt)	14/14 + 26 members
wave 5 (16 tgt)	16/16 claimed — UNGATED, see above
reconcile lane	21/22 lifetime

🔑 THE RESULT WORTH KEEPING: the wave prompt is the lever, and agents write it

Bank rate, same models and same gate — the prompt was the only variable: 76% → 77% → 100% → 100% (wave 5 pending its gate). The jump came from STEP 0: grep -rn "<MAGIC>" src/ with a distinctive literal from the target .s, placed AHEAD of engine_core.h in §136c's search order. §136c's first two steps are same-TU/shared-header scoped and structurally CANNOT reach a banked twin in another overlay's TU — where the big template classes live. That step came from a wave-2 agent's index_gap report. By wave 4 most agents cited it by name and reported index_gap: none. → memory wave-prompt-seed-step0-and-gaps; cookbook §138.

⚡ THE pipeline() FIX — measured, not assumed (Drew asked why waves were slow)

The two-batch parallel() design was a hard barrier: batch 2 could not start until batch 1's slowest agent finished.

	wave 4 parallel()	wave 5 pipeline()
targets	14	16
wall-clock	136 min	82 min
parallelism	2.5×	3.8×
median agent	14 min	9 min
40% faster on 14% more targets. .run/s37w.js is the pipeline version — copy its execution
block for every future wave. (Slowest single agent is still ~50 min; that is real match_one
iteration, 200-300 turns, and is the irreducible floor now.)

📌 THE DISTILLED RULES (cookbook §138 + index; all measured this session)

1. Bucket a gate refusal by the (macro-shape, TU-shape) PAIR, not the SYMBOL — the same symbol conflicts in BOTH directions across the fleet.
2. Reconcile direction depends on WHERE the TU's decl is. ABOVE the splice ⇒ DELETE your duplicate (§100); BELOW ⇒ KEEP a decl in the TU's EXACT shape and cast at the use (§17a-1 D2). Picking wrong CREATES the next error. A wave can manufacture this for itself: two targets in one TU means the first to bank puts its type tags in the second's way (wave 4's only failure).
3. STEP 0 magic-literal grep (above).
4. Never rank off the family map's exemplar field — it is IN-FAMILY and can name an ALREADY-BANKED instance. Derive open sites from corpus.stubs: 16,696 ins → 41,023 on the same map.
5. Three carry variants hide in one "CARRY-FIXABLE" bucket — multi-line comment (fix the tool) · draft-local struct Tag (use the shared type) · file-scope static inline helper (hand-author + EXCLUDE the source overlay; gcc-2.7.2 accepts implicit decls so it passes compiles_standalone and only fails 137 gates later).

🔬 SETTLED — do not re-litigate

• func_801758FC / func_80132018 are NOT templatable. Probed with .run/s6_diag.py (2 builds): the remapped member is BUILD OK + byte diff ⇒ genuine per-member codegen. The h_seq refusal ceiling, confirmed. Same for func_8018A808's family (0/14).
• _alias_decl_for was ONE function, not a class (91 alias decls fleet-wide, 90 already matched).
• "normalized distance 0" is NOT an h_exact guarantee — relocs are masked.

▶ RESUME ORDER

1. Gate .run/s37/*/func_*.c (above) → reconcile → propagate → R22 → commit.
2. Wave 6: derive the corpus.stubs way (rule 4) from a REGENERATED map. Pool at pause: 852 families / 225,217 ins. Exclude the 3 walls (0x801412a8, 0x80178004, whale 0x80144b9c) and the ledgered residuals (0x8017c294 close=12, 0x8017f7b4, 0x801898e4, 0x80186e24 close=187, 0x801758fc, 0x80132018, 0x8018a808). Use .run/s37w.js's pipeline block.
3. EXTEND's last 5: func_80144B9C ×4 needs the §38 -O0 shared-header route (and dedup_extend should refuse-and-name that class per R32); func_80149954 ×1 sits behind func_80147364's u16 params.

🧰 MY PROCESS ERRORS — one mechanism: the signal sampled was not the thing measured

1. nohup CMD & inside a backgrounded call ⇒ the harness signalled the WRAPPER; the fleet check stood at 63/140 and I nearly read it as a pass.
2. pgrep -x make is right for ONE make, WRONG for a campaign of sequential makes (fires in a gap); pgrep -f <pat> SELF-MATCHES so that waiter never exits. Use the campaign's real argv or treelock.sh --status.
3. A corpus.stubs probe mid-rebuild returned garbage; R32's assertion refused to answer.
4. I predicted a fix without reading the macro in front of me — relaxed 42 (void) decls and asserted it unblocked the lane; it banked 0/1 in all 134 because that macro declares (u8*). → the PAIR rule.
5. I violated §136a in my own capture tool — a narrow keyword filter reported "NO COMPILE ERROR" on a build failing with redefinition of struct PW8017C290. No bad bytes from any of them — the byte-gate and R22 caught everything.

---

🛑 (superseded) SESSION-33/34/35 CHECKPOINT (2026-08-04)

Tree CLEAN but for the R23 db.*.gbf churn — never stage. Effort: ultracode. R22 clean-fleet run FIFTEEN times, 140/140 every time. HEAD commit:1388. Drew's standing decision: NO phase close — keep grinding, run waves all night. ⚠️ A wave (.run/s35w.js, 13 targets / 17,644 ins) may still be IN FLIGHT — check /workflows and .run/s35/*/*.c before assuming the tree is idle. Agents write .run/ ONLY.

FLEET — R22 140 passed / 0 failed of 140

96.24% fn-count · 94.0% instr-weighted · 88.4% distinct-code (77,655 uniq) · dedup 1910/0 · C1 241216/241216 · 0 NON_MATCHING (G4). Session opened 96.01 / 93.6 / 88.0 ⇒ +0.23pp fn, +0.4pp instr, +0.4pp distinct. Phase opened 92.00 / 87.5 / 78.0 ⇒ +4.24pp fn, +6.5pp instr, +10.4pp distinct.

WHAT THIS SESSION DID — the deterministic lane first, then two agent waves

lane	result
SC07 EXTEND	0/36 → 31/36
PROPAGATE head	0 → 18,545/18,545 ins (5/5 classes)
wave 1 (17 targets, 26,227 ins)	13 heads + 65 members + 2 reconciled = 80 instances
wave 2 (13 targets, 37,943 ins)	9 heads + 1 reconciled + 18 members
reconcile lane	19/20 lifetime

🔑 THE THROUGH-LINE: every blocker was TOOLING or a DECLARATION — not one was gcc

Five tool defects, four of them the SAME §134 multi-line-blindness class in THREE tools: dedup_propagate.find_site (was discarding a 3,288-ins fn) · overlay_src_split._split_macro_body (38 live instances in engine_core.h) · family_remap._alias_decl_for (the wrapped alias) · dedup_propagate blind to the asm-label alias form entirely. All now decide on cdecl._mask or reuse _alias_decl_for — one oracle (R33), never a second hand-rolled line test.

📌 THE DISTILLED RULES (all in cookbook §138 + index)

1. Bucket a gate refusal by the (macro-shape, TU-shape) PAIR, not by the SYMBOL. The same symbol conflicts in BOTH directions across the fleet; func_80146C3C cost this twice in one session.
2. Reconcile direction depends on WHERE the TU's decl is. ABOVE the splice ⇒ DELETE your duplicate (§100); BELOW ⇒ KEEP a decl in the TU's EXACT shape and cast at the use (§17a-1 D2). Picking wrong CREATES the next error. Grep the TU and compare line numbers with the stub FIRST.
3. STEP 0 of sibling-first: grep -rn "<MAGIC>" src/ with a distinctive literal from the .s. §136c's first two steps are same-TU/shared-header scoped and CANNOT reach a banked twin in another overlay's TU — where the big template classes live. Found func_80188C04 (328 ins) byte-identical to a banked twin in a different overlay, reused verbatim.
4. Never rank off the family map's exemplar field. It is an IN-FAMILY pointer and can name an ALREADY-BANKED instance, hiding the family. Derive open sites from corpus.stubs over the member list: measured 16,696 ins → 41,023 on the same map.
5. Three carry variants hide in one "CARRY-FIXABLE" bucket — multi-line comment (fix the tool) · draft-local struct Tag (switch to the shared type) · file-scope static inline helper (hand-author + EXCLUDE the source overlay). gcc-2.7.2 accepts implicit decls, so the third PASSES compiles_standalone and only surfaces as a byte DIFF 137 gates later.

🔬 MEASURED THIS SESSION, so nobody re-litigates it

• func_801758FC / func_80132018 are NOT templatable. They banked as heads but the sweep managed 18/165 members. Probed with .run/s6_diag.py (2 builds): 0x80132018's remapped member BUILD OK (byte diff) — it compiles clean and the bytes still differ ⇒ genuine per-member codegen, not plumbing. That is the known h_seq refusal ceiling, confirmed, not a new wall. (0x801758fc reported "(no family)" — the diag keys on exemplar.addr; find its real map key.)
• _alias_decl_for is ONE function, not a class: 91 distinct alias decls fleet-wide, the old per-line matcher resolved 90. The S6b note implied a class; it is not one.
• "normalized distance 0" is NOT an h_exact guarantee — relocs are masked. An agent's distance-0 claim is why dedup_propagate correctly answered reach<2.

▶ RESUME HERE

1. Gate .run/s35/*/*.c when the wave lands: tools/treelock.sh g .venv/bin/python .run/s6f_gate.py '.run/s35/*/*.c' → capture blockers for any claimed-MATCH failure (.run/s34_capture.py <ov>:<fn>, which now ranks HARD errors above warnings) → reconcile per rule 2 → make sig-overlays + family_hseq.py + family_sweep --hseq --band all --only <addrs> → R22 → commit.
2. Derive wave 4 the corpus.stubs way (rule 4). Pool at this checkpoint: 864 families / 250,799 ins. Exclude the 3 walls (0x801412a8, 0x80178004, the whale 0x80144b9c) and the ledgered residuals (0x8017c294 close=12, 0x8017f7b4, 0x801898e4, 0x80186e24 close=187).
3. EXTEND's last 5 (unchanged): func_80144B9C ×4 needs the §38 -O0 shared-header route, and dedup_extend should refuse-and-name that class per R32; func_80149954 ×1 sits behind func_80147364's u16 params.

🧰 MY PROCESS ERRORS — all ONE mechanism: the signal sampled was not the thing measured

1. nohup CMD & inside a backgrounded call ⇒ the harness signalled the WRAPPER; the fleet check stood at 63/140 and I nearly read it as a pass.
2. CORRECTION to the S10 rule: pgrep -x make is right for ONE make, wrong for a campaign of sequential makes (it fires in a gap). And pgrep -f <pat> SELF-MATCHES, so that waiter can never exit. Wait on the campaign's real argv or treelock.sh --status.
3. A corpus.stubs probe mid-rebuild returned garbage; R32's assertion refused rather than answer.
4. I predicted a fix without reading the macro in front of me — relaxed 42 (void) decls and asserted it unblocked the lane; it banked 0/1 in all 134 because that macro declares (u8*). The PAIR rule (#1 above) is the distillation.
5. I violated §136a in my own capture tool — filtered the build log for error|conflicting|undefined reference, so it reported "NO COMPILE ERROR" on a build failing with redefinition of struct PW8017C290. A narrow keyword filter is how a real error goes unseen. No bad bytes from any of them — the byte-gate and R22 caught everything.

---

🛑 (superseded) SESSION-33 CHECKPOINT (2026-08-04)

Nothing is running. Tree lock FREE. Tree CLEAN but for the R23 db.*.gbf churn — never stage. Effort: ultracode. R22 clean-fleet run FOUR times this session, 140/140 every time. HEAD commit:1380. Drew's standing decision: NO phase close — keep grinding.

FLEET — R22 140 passed / 0 failed of 140

96.10% fn-count · 93.7% instr-weighted · 88.0% distinct-code · dedup 1908/0 · C1 240807/240807 · 0 NON_MATCHING (G4). Phase opened 92.00 / 87.5 / 78.0 ⇒ +4.10pp fn, +6.2pp instr, +10.0pp distinct.

WHAT S33 DID — lane 2 of the S10 checkpoint (the propagation lag), taken FIRST on the

standing "probe the cheap lever before funding the expensive one" doctrine. ~0 agent tokens.

• SC07 EXTEND: 0/36 → 31/36. Commits commit:1377 · commit:1378.
• PROPAGATE head: 7,398 of 18,545 ins banked — func_80147364 ×137 (4,110) + func_8012A598 ×138 (3,288). Commits commit:1379 · commit:1380.
• Distilled to cookbook §138 (+ index regenerated) — the full triage is there, not here.

🔑 THE FINDING: a gate refusal in these lanes is a DECLARATION, not codegen

Not one of the 36 EXTEND blockers was compiler codegen. Nine failures per binary reduced to 4 distinct symbols repeated across all four. Two masqueraded as walls:

1. The 4 "undiagnosed DIFF"s in dedup_extend's own header are a volatile. Its correctness argument says an h_exact match makes a DIFF impossible — and the contract HELD (original bytes sha1-identical across the failing and working overlay, af1aceb2…, checked in one command with no build). The host TU declared extern volatile s32 D_80127090/94/98 at FILE scope, which no working overlay's copy does ⇒ the macro's three stores became a scheduling barrier ⇒ addu could not sink into the jal delay slot ⇒ built body emitted it early plus a nop, one instruction longer. Tell: a positional shift with a nop at a delay slot is an ORDERING constraint, not a wrong body. Fixed by the §37/§124 DATA asm-label alias. 16/16 on retry.
2. func_80147364: definition (u16,u16) vs 4,046 fleet decls (u16,s32). u16 promotes, so the () escape is ILLEGAL. The DEFINITION-side asm-label alias is the only zero-radius escape (1,725 in-tree precedents). Banked ×137 first try. Before relaxing a decl to (), grep the fleet's decl shapes for that symbol — () is illegal only against a default-promotion param. 4,020 decls of func_80146C3C were all (void)/()/(u8*) ⇒ safe ⇒ 8 of 36 for ONE token.

🧰 TOOL FIX — §134 multi-line blindness, now in a SECOND tool

dedup_propagate.find_site's preamble backscan had the SESSION-18 fix for blank////single-line comments and still halted on a multi-line block comment. Same class S6b fixed 3× in family_remap. Now decides on cdecl._mask (one oracle, R33) with an R32 length-preservation assertion. Monotone + byte-gate-fed ⇒ can fail to bank, never falsely bank. Paired with switching the func_8012A598 exemplar off its draft-local struct BigCopy164 to the shared struct BigCopy (engine_types.h L312, already used identically at engine_core.h:16158). Either fix alone leaves the function written off.

🔬 S33b — the PROPAGATE remainder DIAGNOSED: it is func_80146C3C again

Ran the one probe the checkpoint asked for, in the order §138 prescribes:

1. Originals byte-identical across ov_SC07_006 / ov_SC06_025 / ov_SC01_000 / ov_SC01_077 / ov_SC03_001 for BOTH 0x8016BA68 and 0x8012F274 (one command, no build) ⇒ the registry is sound, the cause is TU context.
2. One build in an excluded overlay named it: conflicting types for func_80146C3C — the SAME symbol as the EXTEND lane, the same (void)-vs-(u8*) shape, the same one-token lever. ⇒ Relaxed all 42 remaining extern void func_80146C3C(void); in engine_core.h to (). Measured safe before editing: every fleet decl of that symbol is (void)/()/(u8*)/(u8 *a0) (no default-promotion param anywhere, so gcc-2.7.2's () rule cannot bite), and every use in the header is either a no-arg call or already cast ⇒ codegen-neutral. dedup_propagate cannot finish this one — the 4 SC07 members are now macro sites, so find_site never returns a def and the auto-source scan errors. dedup_extend is the tool for an already-macro-ized group. Probe: exactly 1 extendable group per ordinary overlay (so the fleet has no hidden wiring backlog beyond this) ⇒ run it across the 134 non-SC07 overlays.

✅ S33b RESULT — the PROPAGATE head is 15,257 of 18,545 ins banked (82%)

class	ins	outcome
func_80147364	4,110	×137 — definition-side asm-label alias
func_8016BA68	3,886	×134 via dedup_extend (not dedup_propagate — see the tool boundary)
func_8012F274	3,973	×136 — hand-authored macro, source overlay excluded
func_8012A598	3,288	×138 — the cdecl._mask backscan fix + the shared-type switch
func_801466F0	3,288	OPEN — the wrapped-alias regex, measured as exactly ONE function

THREE DISTINCT CARRY VARIANTS, each needing a different response (they were all filed as one "CARRY-FIXABLE" bucket before):

variant	what extract_unit cannot carry	response
a MULTI-LINE comment halts the backscan	the externs above it	fix the tool (cdecl._mask)
a draft-local struct Tag {…} in the preamble	the type	switch the exemplar to the shared type
a file-scope static inline helper	the helper	hand-author the macro and 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. And instantiating that macro in the SOURCE overlay is a
duplicate definition (its file-scope helper is still there) — `--source-overlay X --binaries
is the shape;--binaries` alone removes the source from the scan pool and errors.

_alias_decl_for measured, NOT assumed (R37): 91 asm-label alias decls exist fleet-wide, the current single-line regex matches 90, and the one miss is func_801466F0. The S6b note implied a class; it is one function worth 3,288 ins. Fix the regex (join-and-map, the §134 shape) or hand-author.

⚠️ MY WRONG PREDICTION, CORRECTED BY THE BYTES (R14/R37) — and it refines §138

Commit commit:1382 relaxed 42 extern void func_80146C3C(void); decls and its message implies that unblocked the PROPAGATE remainder. It did not. The re-run banked 0/1 in every one of the 134 overlays, with the same error — because DEFINE_func_8016BA68 declares that symbol (u8*), not (void), and the TU declares it (void). That is the MIRROR of the EXTEND-lane pair, and my relax only touched one direction.

Root cause (the R37 shape, again): I bucketed by SYMBOL and stopped there. §138 says "bucket by which symbol is named" — that is not sufficient. The lever is determined by the (macro-shape, TU-shape) PAIR, and the same symbol conflicts in BOTH directions across the fleet: · EXTEND lane: macro (void) vs TU (u8*) → relax the MACRO decl. · PROPAGATE lane: macro (u8*) vs TU (void) → relax the MACRO decl of the other form. One awk over the specific macro's own body — which I ran for the EXTEND macros and NOT for this one — would have shown it before the 134-build run. Read the decl of the macro you are actually fixing, not of its sibling.

The 42-decl relax is still byte-neutral and still useful (it removes a real conflict class and R22 proved it 140/140) — it simply did not do the thing I predicted. Relaxing the remaining 2 (u8*) decls is the actual fix for this lane. §138 to be amended with the PAIR rule.

🧹 THE §134 SCANNER SWEEP — sized, one clear target

Hand-rolled comment tests still outside the cdecl._mask oracle: · tools/overlay_src_split.py:345 (_split_macro_body) — the IDENTICAL single-line-only form I just fixed in dedup_propagate (not s or s.startswith("//") or (s.startswith("/*") and s.endswith("*/"))); it decides where a macro body's file-scope externs END, so a multi-line comment there silently truncates the extern set. Same defect, same consequence. Highest value. · tools/overlay_src_split.py 239/242/520/523 — same file, other scanners (these DO track a block-comment state; lower risk). · tools/split_src_region.py:71-73 and family_remap.py:829 already handle the multi-line form. Already on the oracle: conform_decls · normalize_self_decls · reconcile_tu · scope_tu_externs · dedup_propagate · family_remap.

▶ RESUME HERE — three named items, none diagnosed against a build yet

1. The PROPAGATE head remainder — 11,147 ins. func_8012f274 (3,973, dropped), func_8016ba68 (3,886, 4 of 138), func_801466f0 (3,288, no source found — the S6b D4 gap: its def at ov_SC01_077_after.c:495 carries a wrapped __asm__ alias decl that _alias_decl_for's single-line rx.match cannot see, with a file-scope typedef struct Rec801466F0 behind it, §100). --recover is NOT a retry — it banked 4 of 138 here, and the reconcile lane that is 16/16 lifetime on drafts does not transfer to propagation. Probe ONE excluded overlay's build output (§136a) before re-running any lever.
2. The 41-class PROPAGATE tail — 2,316 ins total. Low value; batch it or ledger it.
3. EXTEND's last 5: func_80144B9C ×4 — the whale's registry func field is a bare name, not a DEFINE_ macro, so write_drafts emits a CALL (hence "undefined reference" in ov_SC07_010); it needs the §38 -O0 shared-header route, and dedup_extend should refuse-and-name the class per R32 rather than feed the gate a draft that cannot pass. Plus func_80149954 ×1, blocked behind func_80147364's u16 params. Then: lane 1 (S10's 17 unbanked targets, 26,227 templ ins, Sonnet at ~12-16 concurrency — the manifest .run/s10.json / script .run/s10.js are still valid) and lane 3 (the ×2-9 grind).

🧰 MY PROCESS ERRORS THIS SESSION — all ONE mechanism, all now in §138

The signal I sampled was not the thing I was waiting for, three times:

1. nohup CMD & inside a backgrounded call ⇒ the harness signalled the wrapper's exit; the fleet check stood at 63/140 and I nearly read it as a pass.
2. CORRECTION to the S10 checkpoint's own rule. It says use pgrep -x make. Right for ONE make, wrong for a campaign of sequential makes — it fired in a gap and called a live campaign done. And pgrep -f <pattern> SELF-MATCHES, so that waiter can never exit (two spinning shells, killed). Wait on the campaign's real argv or on treelock.sh --status.
3. A corpus.stubs probe run mid-rebuild returned garbage — R32's coverage assertion refused to answer rather than hand me a wrong stub set. A measurement taken during a rebuild is not a measurement (S27's law, re-earned). No bad bytes from any of them — the byte-gate and R22 caught everything.

---

🛑 (superseded) SESSION-31/32 CHECKPOINT (2026-08-03)

Nothing is running. Tree lock FREE. Tree CLEAN but for the R23 db.*.gbf churn — never stage. Effort: ultracode. R22 clean-fleet run TWELVE times, 140/140 every time. HEAD commit:1375. Drew's standing decision: NO phase close — keep grinding (task #15).

FLEET — R22 140 passed / 0 failed of 140

96.01% fn-count · 93.6% instr-weighted · 88.0% distinct-code (77,550 / 87,459 uniq) · dedup 1905/0 · 0 NON_MATCHING (G4). Phase opened 92.00 / 87.5 / 78.0 ⇒ +4.01pp fn, +6.1pp instr, +10.0pp distinct.

SESSION TOTAL — ~2,320 function-instances; 167/167 drafted targets banked earlier (100%)

Campaigns: wave 4a (33) · wave 4b b1/b2/b3 (111) · S8-3 x10-99 (23) · S6c jr (12 members) · x2-9 calibration (22 heads + 67 members) · S10 Sonnet wave (13 heads + 57 members). Lane records: reconcile 16/16 lifetime · redraft 15/15 · §136b closed 15/15. A first-pass gate failure is a ROUTING DECISION, not a wall (capture -> reconcile/redraft).

🔬 THE SONNET FINDING (Drew's call — override my routing; he was right)

	Opus (S8-3, 121-328 ins)	Sonnet (S10, 125-793 ins)
first-pass bank rate	10/13 = 77%	13/16 = 81%
30-wide concurrency	fine	SERVER-throttled, 14 of 30 never ran
The §136i ~120-ins Opus threshold is TOO LOW — it was my extrapolation, never measured. Sonnet
is at least as capable to ~800 ins. The limiter is CAPACITY, not capability: run Sonnet waves at
~12-16 concurrency, not 30. Not naming a new cliff number — 16 samples under a throttle confound
can't. Task #12 (controlled A/B) is how that gets fixed properly.

▶ RESUME HERE — three lanes, all measured, none blocked

1. S10's 17 unbanked targets = 26,227 templ ins (14 server-throttled + 3 gate-failed). Re-run in TWO batches of ~9 at ~12 concurrency. Manifest .run/s10.json, script .run/s10.js (already all sonnet), banked list .run/s10_banked.txt. Heads: func_8017D174(793x5) func_80186E24(611x5) func_8017E2EC(296x7) func_8017C8FC(210x7) func_8017F2E8(207x7) func_80181EE0(198x7) func_8017CDF0(693x2) func_8017E9F4(221x6) func_8017D318(184x7) func_801829BC(150x8) func_8018F060(397x3) func_8017F624(170x7) func_80190144(128x9) func_801813D0(163x7) func_80187DD0(174x6) func_8018457C(160x6) func_8018A390(136x7).
2. The propagation lag — 26,006 ins, ~0 agent tokens, PARTLY BLOCKED (task #14):
   • EXTEND shape (14 classes / 48 stubs / 5,169 ins): dedup_extend --binaries ov_SC07_006,ov_SC07_007,ov_SC07_010,ov_SC07_011 planned 36 groups and banked 0/36 — 6 PLUMBING / 4 DIFF, and the PLUMBING conflicts repeat the SAME ~5 symbols across all four binaries: func_80146C3C (the same one that blocked func_8017C120), func_80144B9C, func_8014F4C0 (both self-axis => §37 asm-label alias), func_8012E5CC, D_80126CC4. ~5 declaration fixes unlock most of the 36. Reconcile-lane shaped (16/16 lifetime). The 4 DIFFs are ODD — dedup_extend's contract says h_exact identity guarantees byte-identity including relocs, so a DIFF should be impossible. Probe one (overlays-share-VAs trap?).
   • PROPAGATE shape (45 classes / 780 stubs / 20,837 ins): NEVER PROBED. Needs dedup_propagate --addr from the single matched source. Probe ONE before costing (§136h).
3. The x2-9 grind (task #15): 411 of 1,872 families cover the 212,594-ins gap to 95% instr. Measured ~22,937 templ ins/wave at the head; mean ins/family decays 1,844(top-25) -> 525(top-400) -> 193(band-wide). Est 19-30 waves. Work highest-templatable-weight-first.

⚠️ P32's BEHEMOTH LEDGER IS STALE — REBUILD BEFORE PLANNING IT

roadmap-v2 claims 0x80183814 = 5,122 ins in ov_SC07_006, "largest fn in the game". Measured: it is ~108 ins in ov_SC02_000/003. The largest unmatched function fleet-wide is 969 ins (0x8017ef68, ov_SC06_000); NOTHING >=1,000 ins remains unmatched. Of 45 distinct live stubs

=400 ins: 21 are already cracked elsewhere (propagation lag, 24,354 instruction-instances) and 24 are genuinely never-cracked (22,065). ~Half of "P32 wall tonnage" is not wall at all.

✅ MAIN IS NOT FREE DEDUP — checked two ways, both negative

main's 1,525 stubs vs 39,917 banked overlay classes: 0 h_exact hits AND 0 h_norm hits (h_exact alone would be meaningless — different vram base, so relocs differ by construction). Resident 14 stubs: also 0. Consistent with Phase-11's EXE-vs-resident finding, now extended to overlays. Main's stubs are all genuinely fresh work — P31 scope, not a dedup opportunity. (Caveat: main's sig is Ghidra-derived from June 14 — valid since the binary is unchanged, but the relocation-masked xsig tool is the instrument built for cross-base comparison if you want proof.)

🧰 MY PROCESS ERRORS TODAY — the corrected patterns (all ONE root cause)

Trusting a proxy instead of the thing itself, three times:

1. §136h — priced a 141-family sweep off byte_weight_templatable; it banked 1 of 1,781. That column counts what COULD template if bodies reproduced — the very question. Probe one member per family. (Also: I pointed family_sweep --hseq — which REMAPS — at an h_exact- identical population, where the correct tool is dedup_extend/dedup_propagate and no remap is needed at all. Wrong tool, not a wall.)
2. §136a — classified blocker captures on EXIT STATUS, then on error PROSE. make build runs check, so a clean-compiling byte-DIFF also exits non-zero; and too many arguments matched no regex. Derive from the invariant: did the compile produce an object?
3. The waiter — until [ -s <output> ] fires at the FIRST LINE, not completion. It fired mid-propagation and I ran make clean on a live family_sweep. Use pgrep -x make (exact process name; pgrep -f <pattern> also SELF-MATCHES and leaked 4 spinning shells earlier). No bad bytes from any of them — the byte-gate and R22 caught everything.

🔑 THE FOUR FINDINGS THIS SESSION PRODUCED (cookbook §136 · §136a · §136b · §136c · §136d)

1. §136 — the LOCAL-VARIABLE lever. In the 60–120-ins band most "regalloc residuals" are decided by how many C locals you declare and at what scope, not by register pins (local-alloc.c:472 promotes any pseudo with REG_N_DEATHS > 1 to a global allocno, which loses the low register). One case explicitly REFUTES the pin as the lever for a redundant copy.
2. §136b — a "genuine byte-DIFF" verdict is a fact about ONE DRAFT, closed at 8/8. Every DIFF-ledgered function banked on redraft: wave 3's four, the three I classified myself, and one from batch 1. Never retire a target on a DIFF verdict; route it to REDRAFT. And re-GATING an unchanged draft is not a retry (that is why wave 4a's 3 DIFFs survived one gate).
3. §136a CORRECTED, by an agent, 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. 2 of 3 second-round PLUMBING drafts had a real codegen residual behind the conflict. Also: classify on the build's OUTPUT, never its exit status (make build runs check, so a clean-compiling byte-DIFF also exits non-zero).
4. §136c — SIBLING-FIRST is a DERIVATION shortcut, not just a conflict fix. Search order: engine_core.h DEFINE_* near-twin → same-TU banked sibling → the .s → the Ghidra seed LAST (byte-proven an entirely different body twice this session). Produced first-draft matches. §136d adds four new levers, each with its refuted axis recorded.

▶ RESUME HERE

1. Gate batch 3 if it was in flight (see the header) → propagate → R22 → commit.
2. The 4 remaining stubs of 107: func_8017C120(SC02_026 — classified UNKNOWN, neither a compiler error nor a SHA mismatch; owes a direct look at its build output, not a guess) · func_8017CD9C(SC04_003) · func_801848DC + func_80185904(SC04_018). Capture then route — and per §136b none of these is a wall on one attempt.
3. The wave-4b remainder: batch 3 was the last of the 111-family volume lane. After it, the B-shape lanes are spent — re-derive the queue from a freshly regenerated family_hseq before scoping anything new (the ×138 era already ended; the next tiers are ×10-99 and the zero-crack non-jr residue at 120 families / 65,946 ins).
4. T5 close — only after re-deriving; the checklist's ROI-floor trigger is REFUTED and must not be used to close the phase.

---

🛑 (superseded) SESSION-30 CHECKPOINT (2026-08-01, S6 complete through wave 3)

Nothing is running. Tree lock FREE. Tree CLEAN (0 non-ghidra entries; the db.*.gbf churn is R23 restart-noise — never stage). HEAD commit:1352. Effort: xHigh → ultracode (Drew enabled mid-session for the waves). make tools-health green at session open; frontier REGENERATED at HEAD after the last bank, so every number below is current (not carried forward).

FLEET — R22 clean-fleet 140 passed / 0 failed (run 9× this session, 140/140 every time)

95.42% fn-count · 92.4% instr-weighted · 85.5% distinct-code (75,926 unique fns) · dedup 1905/0 · 0 NON_MATCHING. Phase opened 92.00 / 87.5 / 78.0 ⇒ +3.42pp fn, +4.9pp instr, +7.5pp distinct.

WHAT S6 DID — ~3,500 member-matches in 11 commits

commit:1343 S6a 842 · commit:1344 S6b +740 · commit:1345 S6e fix · commit:1346 B +268 · commit:1347 S6f calib +65 · commit:1348 S6f wave-1 +544 · commit:1349 S6g wave-2 +342 · commit:1351 S6h wave-3 +673 · commit:1352 cookbook §135 (+ 2 checkpoint/doc commits).

Two distinct engines ran, and BOTH paid:

1. Tooling repair (deterministic, ~0 agent tokens). The pre-fix tools scored the zero-crack population at 842 members; with three scanner fixes the SAME population scored 1,582. Six defects found, five fixed — all one root cause: a multi-line C construct read one line at a time, or a guard firing on a condition it never verified. → cookbook §134.
2. Agent waves (two-lane: cheap drafts + Opus reconciles). Bank rate 83% → 93% → 89%; the reconcile lane is 12/12. → cookbook §135 (6 codegen idioms + 4 integration rules).

⚠️ TWO THINGS A FRESH SESSION MUST NOT INHERIT UNCHECKED

1. The T6 ROI-floor trigger is REFUTED. Do not close P30 on it. The checklist says T5 fires when "the ×138 era ENDS after S2". That was wrong twice today: two ≥126-member families were open (33 ins of C → 266 members), and a 36-family / 28,829-ins pool of the same shape sat unexamined behind them. Three §133-class misses now share one mechanism — a conclusion drawn from a view that silently excluded the thing being looked for. Re-derive from a freshly regenerated family_hseq before any close.
2. Token economics, measured today (re-check before scaling wave 4): deterministic sweeps ≈ 0 tokens for +0.5pp instr; wave-1 ≈ 2.0M for +0.1pp; wave-3 ≈ 4.1M for +0.3pp. The agent lane is still paying, but the tooling lane paid better — the same finding Phase 15 recorded. Probe the cheap lever before funding the expensive one.

📊 THE LIVE QUEUE — re-derived at HEAD after the final bank

lever	families	templ ins	note
B-shape ≥20 memb, ≤60 ins	0	0	EXHAUSTED (was 36 / 28,829)
B-shape ≥10 memb, 61-120 ins	33	34,895	best per-target value; Opus ≥90 ins
B-shape 10-19 memb, ≤60 ins	111	33,554	volume lane, Haiku
zero-crack non-jr residue	116	61,230	S6-diagnosed; bimodal residue
zero-crack has_mid_jr (S6c, untouched)	9	9,364	needs jtbl_family_bank, 26 slots, ~0 tokens
fresh ×2-9	1,877	363,873	worst multiplier — deprioritised
fresh ×1 singletons	3,642	207,945	×1 — deprioritised
Unmatched fleet-wide: 15,161 instances / 935,927 ins.

🔁 THE WAVE RECIPE (cookbook §135 has the full version)

Manifest derived from corpus.stubs + family_hseq, paths verified on disk, pasted from the file — never typed (a hand-transcribed list cost wave 1 three agents' time). Draft lane: Haiku ≤89 ins (≡ Opus at ≤~50, ~4.8× cheaper), Opus direct ≥90, Opus escalation between. Then: gate whole-binary via .run/s6f_gate.py '<glob>' (derives every TU/split; agents' match_one MATCH is NOT a bank) → capture each failure's compiler error → Opus reconcile lane → gate → family_sweep --hseq --band all --only <addrs> → R22 → commit. If N reconcile targets share ONE TU, forbid agent builds.

▶ RESUME HERE

1. Wave 4 over the two B-shape slices (144 families / 68,449 ins) — recipe above, §135 in prompt.
2. S6c — the 9 jr zero-crack families (9,364 ins, ~0 agent tokens) via jtbl_family_bank, probing ONE sibling first (the S29 discipline: 1 build, not 137).
3. The 4 carried failures: func_80184A94 (reconciled to match_one MATCH, still gate-refused)
   • 3 genuine byte-DIFFs func_801845B0, func_8017BEBC@ov_SC02_026, func_8018480C.
4. T5 close — only after re-deriving; see the refuted trigger above.

---

🛑 (superseded) SESSION-29 CHECKPOINT (2026-08-01, refreshed ~11:2x)

Supersedes the earlier S29 block below (which said "AT GATE 2"; Drew then directed P30 stays OPEN and approved the T6 continuation). Nothing is running. Tree lock FREE. Tree clean but for the R23 db.*.gbf churn (never stage). HEAD commit:1335. Effort: Max → xHigh (Drew).

FLEET — R22 clean-fleet 140 passed / 0 failed (run 7× this session, 140/140 every time)

94.43% fn-count · 91.4% instr-weighted · 84.0% distinct-code (73,787 unique fns) · dedup 1905/0 · 0 NON_MATCHING. Phase opened 92.00 / 87.5 / 78.0 ⇒ +2.43pp fn, +3.9pp instr, +6.0pp distinct. Session start (S28 close) was 93.25 / 89.2 / 80.5 ⇒ +1.18 / +2.2 / +3.5pp today.

THE UC WAVES (Drew opted in mid-session) — 33 heads banked, ~3,560 members propagated

• Wave 1 (14 agents, 2 lanes): 14/14 match_one MATCH → 8 heads banked by three mechanisms (plain gate · a NEW targeted decl reconcile that edits ONLY the line the compiler names · the gate_stage ladder) → 1,370 members propagated, 0 failed.
• Wave 2 + 3 (19 + 11 agents): 19/19 heads banked — the best conversion of the phase → 2,192 members / 411 stage-but-DIFF residue.
• THE RECONCILE LANE IS THE RESULT WORTH KEEPING. 17 of 19 wave-2 heads needed a declaration fix and not one needed a header edit. Two escapes did all of it: the §37/§124 asm-label alias (<ret> aF<ADDR>(...) __asm__("func_<ADDR>")) when the conflict is on the function ITSELF (return-type or arity), and conform-the-decl + cast-at-use when it is on a DIFFERENT symbol the draft declares (func_80142A80, RotTransSV, func_80146C3C, D_8011D030, D_800AF634). This is the direct answer to the morning's failure: the documented fix for that class was a fleet-wide decl widen, which touched 2,046 files and still did not build. The alias sidesteps the collision at ZERO blast radius.
• Agents cannot run the gate, so a blocker-capture pass (splice→build→revert) had to precede each reconcile fan-out and embed the exact compiler error per agent. Without it they are blind on the one axis that matters. (R14 on my own capture: the note: lines I passed through were pre-existing SHB macro-redefinition NOISE — four agents independently called it and worked the real error.)

WHAT S29 DID

1. JR-PAIR-IN-ONE-O0-OBJECT RETIRED — two instrument defects (§132), not a compiler wall: jtbl_carve blind to a merged-double span + no .DELETE_ON_ERROR (a truncated .o outlived its own loud compile error and the NEXT build linked it). Pair banked ×138 = 276 instances.
2. S1 zero-crack tier — 186 members (head func_8014032C 137/137 = 25,071 ins via the new --span-rel, §132b; func_8017BEBC 13/13; + jr tier). My routing error recorded: 13 of 28 families are not jr and carve-failed by construction (§123's own law).
3. S2 — both remaining reach-138 heads cracked. func_8016EC0C (void-return delay-slot idiom) ×138; func_80176734 (371 ins — the largest single item in the frontier) banked first try and propagated. TWO agents reached byte-identical solutions independently via different lever sets.
4. S3 measured, not scaled — 77% of stored drafts have DECAYED (7/30 still verify); the survivors are def-side-walled. A fleet sed over decls touched 2,046 files and failed ⇒ reverted (§29 law).
5. UC wave 1 (14 agents): 14/14 match_one MATCH → 8 banked heads by three different mechanisms (plain gate · a NEW targeted decl reconcile that edits only the line the compiler names · the gate_stage ladder). 1,370 member instances propagated, 0 failed.
6. Cookbook §132 · §132a · §133 (the DEFAULT-FILTER class — three times in one session a tool silently answered a narrower question than the one asked).

📊 THE POST-WAVE FRONTIER (re-derived at HEAD after the 2,192-member propagation — R35)

Overlays: 94.7% fn / 91.8% instr / 85.0% distinct; unmatched 18,635 instances / 1,069,466 ins / 13,658 distinct classes → 2,378 families + 3,767 singletons.

lever	scale	cost	verdict
Zero-crack propagation (source-agnostic)	190 families / 129,997 ins	~0 agent tokens	DO FIRST (S6)
×10-99 fresh-crack band	243 families / 239,058 ins	agent waves	next (S5)
STRUCT class (ov077 surface)	29 fns / 39,892 ins	agent waves	after S5
GIANT	4 fns / 100,188 ins	2 are the known walls (close=110/91)	P32's track
×2-9 fresh-crack	1,941 families / 410,671 ins	biggest mass, worst multiplier	deprioritised
×1 singletons	3,767 / 230,367 ins	×1	deprioritised
S4 PINS	101 ins (not 44,279)	—	RETIRED, see below

TWO PRICING CORRECTIONS (both §133, in OPPOSITE directions — this is why the table is re-derived):

1. S4 PINS is retired on measurement. The plan priced it at 44,279 ins from worklist.md, which prices by h_exact reach (×138/fn). Re-priced against the family map after the waves, the PINS functions' families are almost all already matched/propagated: 24 fns / 101 templatable ins. Running the planned wave would have bought ~0.001pp. The same mis-pricing that once under-valued the frontier head by 138× over-valued this by ~440×.
2. 106 of the 190 zero-crack families have exemplars OUTSIDE ov_SC01_077 (ov_SC03_014 17, ov_SC06_018 10, ov_SC01_084 9, ov_SC02_027 7, ov_SC02_011 7, …). Every sweep this project has run passed --source ov_SC01_077, so those families were structurally unreachable — not hard, not walled, just never addressed. Probe-confirmed: func_8012B77C is DEFINED in ov_SC07_006 and still an INCLUDE_ASM stub in ov_SC01_000.

▶ RESUME HERE (T6: S1–S3 done, both UC waves done; nothing blocked, nothing running)

• The ×138 redraft band is now CONSUMED. All 33 fresh-crack ≥100-member families are banked except the 2 GIANT walls (func_801412A8 close=110, func_80178004 close=91 — P32's track).
• S6 (NEW, highest ROI): source-agnostic zero-crack propagation — 190 families / 129,997 ins, ~0 agent tokens. Drive each family from ITS OWN exemplar overlay, not always ov_SC01_077; route by tier (§123) and pass --band all (the S1 trap). Probe one sibling per family first.
• S5: the ×10-99 band — 243 families / 239,058 ins, the proven 3-lane workflow. Expect the ROI floor here: the multiplier is 10-99, not 138.
• S4 PINS RETIRED on measurement (101 ins, not 44,279 — see the pricing corrections above).
• The 411 stage-but-DIFF propagation residue — per-overlay variants whose remapped bodies do not reproduce; individually gate-rejected, ledger material not a lever.
• Reusable machinery from today (all committed): the 3-lane workflow scripts under .claude/…/workflows/scripts/p30-uc-wave{1,2,3}-*.js · .run/uc_capture.py (blocker capture, the precondition for any reconcile fan-out) · .run/uc_reconcile.py (compiler-named decl edits only) · tools/jtbl_family_bank.py --span-rel (§132b) and its like_arg guard (§132a).
• T5 close fires on the measured ROI floor. The ×138 lever is spent; S4/S5 are the remaining in-phase levers, and their multiplier is materially lower — that is the honest floor signal.

---

🛑 (superseded) SESSION-29 CHECKPOINT (2026-08-01 ~00:40) — AT GATE 2

Supersedes SESSION-28 below. Nothing is running. Tree lock FREE. Tree clean but for the R23 db.*.gbf churn (never stage). HEAD commit:1303. Effort: Max all session. make tools-health RC=0 at close; R22 clean-fleet run TWICE this session, 140/140 both times.

FLEET — R22 clean-fleet 140 passed / 0 failed

93.33% fn-count · 89.6% instr-weighted · 81.6% distinct-code (72,416 unique fns) · dedup 1905/0 · C1 240496/240496 · 0 NON_MATCHING. Overlays-only: 93.6% fn / 90.1% instr / 82.5% distinct. Phase opened 92.00 / 87.5 / 78.0 ⇒ +1.33pp fn-count, +2.1pp instr, +3.6pp distinct this phase.

WHAT S29 DID (all committed, all R22-verified)

1. JR-PAIR-IN-ONE-O0-OBJECT RETIRED — the S28 class and its §81-isolate escape are BOTH refuted. Two instrument defects: jtbl_carve's merged-double blindness + no .DELETE_ON_ERROR (a truncated .o outliving its own loud compile error). Both fixed; cookbook §132.
2. The jr pair banked ×138 each — 276 function-instances. Frontier families #3 and #4 by templatable weight (37,536 + 27,324 ins). The -O0 cluster is now COMPLETE fleet-wide.
3. §132a — the --like role-name over-transfer (ov_SC07_010), guard shipped.
4. §100 type-locality — a file-scope typedef in a draft is an exemplar-only bank: it passes the gate and silently cannot travel. Caught by probing ONE sibling before 137 builds (R37).
5. T1/T3 closed with measured per-lane verdicts (see the checklist), tools-health + fresh frontier.

▶ RESUME HERE — P30 STAYS OPEN; T6 approved (Drew, 2026-08-01)

Gate 2 was reached and deliberately not taken: the ≥95% branch is out of reach (89.6%) and the ledger branch IS met, but closing now would strand roadmap-v2 bucket W3 (the overlay family mass) with no owner phase — P31 is scope-complete/main/resident, P32 is walls/behemoths. So P30 continues on the T6 order in the checklist above (S1 → S2 → S3 → S4 → S5), and T5 fires on the measured ROI floor (the end of the ×138 era after S2), not on a session count.

Start here: S1's head — 0x8014032C (183 ins ×137, jr, per-location, exemplar matched in ov_SC01_077) via jtbl_family_bank, probing ONE sibling first (the S29 discipline: 1 build instead of 137). Then 0x8017BEBC, then the ×4-straggler families. Frontier data: .run/family_hseq.json (regenerated at HEAD commit:1303, same-tree as the digests).

---

🛑 (superseded) SESSION-28 CHECKPOINT (2026-07-31 08:1x)

Supersedes SESSION-27 below. Nothing is running. Tree lock FREE. Tree clean but for the R23 db.*.gbf churn (never stage). HEAD commit:1293 (+ this doc commit). Effort: opened xHigh → Max for the jr re-measurement and the T2 probe ladder → back to xHigh for the driver + sweep. make tools-health RC=0 at session open. R22 clean-fleet run THIRTEEN times this session, 140/140 every time (one 139/140 and one 137-file false-pass, both MINE, both caught and reverted — see the honesty ledger).

FLEET — R22 clean-fleet 140 passed / 0 failed (verified THIRTEEN times this session)

93.25% fn-count · 89.2% instr-weighted · 80.5% distinct-code (72,154 / 87,459 unique fns) · dedup 1905/0 · C1 240496/240496 · 0 NON_MATCHING. Phase opened 92.00 / 87.5 / 78.0 ⇒ +1.25pp fn-count, +1.7pp instr, +2.5pp distinct this phase.

🎯 THE THREE REACH-138 TARGETS + THE SWEEP RESIDUE — 270 members banked, 1 wall

✅ func_8013C08C 0/137 → 137/137 (commit:1286) — §94 type-carry; E_13C08C was a MULTI-LINE typedef at FILE scope and extract_unit's backscan cannot carry one (it presents its CLOSING line first). Fix = §100 draft-local. Byte-neutral. ✅ 0x8013BC7C 133/133 (commit:1291) — the sweep residue. My hypothesis (split tracks the -O0 file) was REFUTED (SC07_006/007/011 were sub-split today and banked anyway), so I staged one member and READ the error: (1) a typedef textually identical to engine_types.h's, which C89 still rejects — already handled by harvest_verify's cdecl.strip_provided_typedefs; (2) §57 self-decl divergence ⇒ --normalize-self-decls. One flag, 133/133. ⛔ func_8013B83C + func_8013BD74 — NOT BANKED, class JR-PAIR-IN-ONE-O0-OBJECT. Both bodies byte-correct (--o0 MATCH, 272/198 ins). Two jr fns sharing one -O0 object: carving both gives a valid span (pads [0,4,4]) and an INCREMENTAL build says BYTE-IDENTICAL, but the CLEAN build cannot link — undefined reference to $L105 (local label from the C-emitted jtbl) + undefined reference to func_8013C938 (a previously-MATCHED fn). Untested escape: §81 step 1 — isolate one into its own code subseg so each object owns exactly one table. Drafts preserved at .run/s28w/.

▶ RESUME HERE — FRESH SESSION SAFE. Nothing running, tree clean, lock FREE.

Drew's directive at the S28 close: items 1–3 done; item 6 wants Max (prompted, not started); items 4–5 are for a FRESH SESSION (this checkpoint is that handoff).

1. [T1, Ultracode] Integration-recovery sweep — the ~90 stored drafts that are match_one-MATCH but integration-decayed, + the 10 named SESSION-16 blockers. Prior measured recovery 39%. Report point #2. Prompt for /effort ultracode and WAIT (R26/R27).
2. [T3, Ultracode] More crack waves — 3,873 fresh cached cores across 119 binaries (.run/p30w4_pool.json). ⚠️ S28 ROI evidence: a 15-target wave cost 1.33M tokens for 12 banks and +0.00pp headline. Only worth resuming against HIGH-REACH targets; ×2-reach drafting is not where the leverage is. Deal ACROSS binaries so the gate fans out.
3. [Max] #9 — jtbl_carve diverges on ov_SC06_018 ✅ SOLVED + BANKED (S28, commit:1293). Root cause: a jtbl OVER-SPAN. sltiu 0xC names 12 entries, the object emits 12 words, the carve reserved 13 — the 13th being ordinary NON-ZERO data spimdisasm ran into the dlabel, which the zero-word trim cannot see. The piece under-filled ⇒ every later symbol slid −4 ⇒ 812 %lo immediates changed. Fixed with an sltiu-authorized clamp (refuses loudly if a surplus word looks like a real entry). func_80191C50 (710 ins) BANKED, R22 140/140. Cookbook §131. ⇒ the JTBL-CARVE-BREAKS-BYTES class is RETIRED — it was a real bug, now fixed.
4. [Max, xHigh-able] JR-PAIR-IN-ONE-O0-OBJECT — try §81 step 1 ✅ SOLVED + BANKED (S29). The class is RETIRED and the §81-isolate escape was never needed: it was TWO instrument defects (a merged-double span jtbl_carve could not see + no .DELETE_ON_ERROR, so a truncated .o survived its own loud compile error and the NEXT build linked it). Both fns banked, R22 140/140. Cookbook §132. Sibling sweep (2 × 137) unblocked — see the S29 log entry.
5. [T5, Max] Phase close. ⚠️ Milestone reality: 89.2% instr vs a ≥95% bar. The remaining overlay volume is small; the bulk is main's 1,034 stubs + the 39 type-1 modules, both P31 scope. P30 realistically closes on the milestone's ledger branch ("every remaining overlay stub on a named wall/behemoth/queue ledger"), which is an explicit either/or in the approved milestone — Drew's call, to be made deliberately at T5, not discovered.

⚠️ S28 HONESTY LEDGER (read before trusting any S28 conclusion)

Five wrong calls, every one caught by an oracle, none committed:

1. Two of three jr "wall" verdicts RETRACTED — a grep-of-the-build-log gate in a driver with no revert-on-abort contaminated its own measurements (§125).
2. An 8× population under-count (275 vs 2,184) — a bare except: continue swallowed R32's coverage assertion during a rebuild; it also made me wrongly call a good T0 pin "stale" (§126a).
3. "15 contiguous -O0 fns" — an under-count from scanning asm/**/*.s, which cannot see matched functions (§126).
4. Broke the fleet gate to 139/140 — a negative control that corrupted the REAL tracked file and restored it with tangled escaping (§128a). A negative control must mutate a scratch copy.
5. Reported "BOTH BANKED — BYTE-IDENTICAL" from an INCREMENTAL build that the clean build cannot even LINK (§130). The sharpest lesson of the session: an incremental byte-gate result can be actively FALSE, in the most convincing direction. Standing consequence: only make clean && make extract-all && make check-all counts, and a tool's exit status is never the oracle — read the output.

📌 THE 4th -O0 REGION — verified, sized, and correctly BLOCKED (do not re-litigate)

My own byte scan (not the agent's claim — R14/R37) over all 156 fns of ov_SC03_014_jr_8017EB7C: the -O0 frame-pointer prologue covers exactly the contiguous run 0x80183CF0..0x80184868 (15 fns), with -O2 neighbours 0x80183A54 below and 0x80184920 above — a clean cut. Prize: 30 fn-instances / 1,504 ins — the region is h_exact-identical in ov_SC03_014 + ov_SC03_015 ONLY; the 3 other overlays holding a func_801846E4 (SC03_090, SC04_005, SC05_018) are DIFFERENT code at the same VA (overlays share VAs — the §-index warning, live). Blocked structurally: all 15 live in an -O2 split (corpus.o0_sources confirms), and the append-to-_o0b route fails BY CONSTRUCTION (the existing _o0b covers the whale at file 0x1CA44..0x1D64C; this region is far above it, so appended bytes land at the wrong VA — the same reasoning that killed the T2 ×1 append probe). ⇒ needs the re-carve; it IS T2.

⚠️ TWO CORRECTIONS TO THE SESSION-27 CHECKPOINT (mine, R14)

1. The 137 were SKIPS, not failures. S27 folded them into the 401 gate-failures; 137+22+10 = 169 ≠ 401. They are a separate skipped{} bucket.
2. The cause was NOT "the exemplar banked in a different binary." It is banked in ov_SC01_077 under the §37/§73 asm-label alias (int aF8016191C(...) __asm__("func_8016191C")), and extract_unit only matched a head literally named func_<ADDR>. One exemplar × 137 same-address members — a tool lookup miss, the 4th consecutive "structural" residual to resolve to our own tooling (R35). Fixed T0-only; the alternative (widen engine_core.h, drop the alias) was REJECTED — it fixes only §73's RETURN axis while the decl/body also disagree on PARAMS, and it is a T2 fleet-shared edit. §124.

🧪 A METHOD NOTE WORTH KEEPING (cost me one probe this session)

I ran a corpus.stubs probe while make extract-all was mid-flight and got 4 binaries' worth of garbage. corpus.py's R32 coverage assertion caught it and refused to answer instead of returning a wrong stub set — the assertion earning its keep. A measurement taken during a rebuild is not a measurement (SESSION-27's law, re-earned). Re-derive after the campaign exits.

---

🛑 (superseded) SESSION-27 CHECKPOINT (2026-07-31 01:25)

Supersedes SESSION-26 below. Nothing is running. Tree lock FREE. Tree clean but for the R23 db.*.gbf churn (never stage). HEAD commit:1258.

FLEET — R22 clean-fleet 140 passed / 0 failed (verified at HEAD)

92.67% fn-count · 88.3% instr-weighted · 78.7% distinct-code (70,590 / 87,459 unique fns). Phase opened at 87.5 / 78.0 / 92.00 → +0.8pp instr, +0.7pp distinct, +0.67pp fn-count so far.

WHAT LANDED THIS SESSION (all committed, all R22-verified)

• T0 instrument repair (a–f), 6 commits — incl. gate_stage stage-0 (§122) and the fix_arity_callers per-edit journal undo.
• T0.5 fleet Ghidra-C prefetch COMPLETE — 124/124 programs, 7,716 files, 335 min unattended. 95.2% of the 6,070 crack-needing families are seeded; main 2,001/2,002; resident 14/14.
• T3 waves 1–4 + behemoths: wave1 8/14 · wave2 15/19 · wave3 48/48 · wave4 69/70 banked (70/70 drafted) · 7 behemoths CONFIRMED at 700–970 ins (4 banked; 3 jr deferred).
• Propagation: 548 + 911 members via family_sweep --hseq; h_exact leg 14 propagated / 42 benign skips / 0 failures.
• New instruments: tools/treelock.sh · tools/prefetch_fleet.py · tools/cookbook_index.py
  • docs/cookbook-index.md (324 sections, symptom-keyed, in tools-health).
• Cookbook §122, §123.

▶ RESUME HERE (nothing blocked, nothing in flight)

1. func_801846E4 (ov_SC03_014) — the 1 of 10 resumed drafts that did NOT bank (agent reported MATCH; gate refused). Draft at .run/drafts-p30w4/ov_SC03_014/func_801846E4.c. Its agent flagged it as the -O0 class (SIZE-MISMATCH/short + frame-pointer prologue) — likely needs --o0.
2. The 137 "no matched unit for func" sweep-routing gap — family_sweep looks for the exemplar's banked unit in the wrong binary. Cheap, worth ~137 members.
3. 3 jr behemoths (func_80191C50, func_8018057C, func_80181CDC; staged in .run/beh-gate/<binary>/) + group B's carve (13-member IMM/jr @952, 103 already matched → jtbl_family_bank.py). Both run make extract — run ALONE under tools/treelock.sh.
4. T2 -O0 carve-within-a-carve (13 families / 1,746 members). Newly cheaper: wave-4 agents documented the match_one --o0 flag AND found a 4th -O0 region beyond the 3 known ones.
5. Next wave: 3,873 fresh cached cores across 119 binaries remain (.run/p30w4_pool.json); deal them ACROSS binaries so the gate fans out.

🔁 THE LOOP THAT WORKS (proven 4 waves running)

draft (agents, .run/ only) → --no-propagate gate, per-binary IN PARALLEL → sig+map regen → propagate per-function, routed by tier (§123) → one R22 → commit. Wrap every tree-writing campaign in tools/treelock.sh.

⚠️ THE THREE LAWS THIS SESSION COST US (all mine, all now mechanised)

1. Guard the CAMPAIGN, not the process — a pgrep poll cannot see a campaign made of sequential processes; mine started in a gap → 63/140 broke. → treelock.sh.
2. A killed process performs no undo — a fixed 3600s propagation timeout killed the driver mid-fleet-write. → timeout scales with bank count AND is caught.
3. A measurement taken during corruption is not a measurement — I reported a "per-binary bank-rate cliff" and theorised a cause; re-gated clean it was 6/6 everywhere. Both incidents cost build cycles and ZERO work, because agents only ever write .run/.

📊 THE MEASURED FINDING WORTH KEEPING

Bank rate by wave: 57% (no index) → 79% → 100% → 99% (index-first). Same models, same gate. The bottleneck was discoverability of our own knowledge, not the compiler. Agents also self-report index_hit/index_gap, and their gap reports are now curated symptom lines — the flywheel is closing on itself.

---

🛑 (superseded) SESSION-26 CHECKPOINT (2026-07-30 23:05)

Supersedes the in-flight block below. Paused for the 5h usage limit; a one-shot cron is set for 00:47 (2026-07-31) to resume — but cron is SESSION-ONLY and dies with this session, so if the session ended, just follow "RESUME HERE" manually.

HEAD commit:1254. Fleet at last verified point: 92.59% fn-count · 88.2% instr · 78.7% distinct · R22 140/140 (commit commit:1252).

⏳ RUNNING AT PAUSE (check these FIRST)

• w4-gate under tools/treelock.sh (pid at pause 1313761): gating wave-4's 68 drafts across 14 binaries in parallel, then sig+map regen, then R22. Log: .run/w4gate_run.log + .run/w4gate/<binary>.log. If it did not finish: tools/treelock.sh w4-gate bash .run/p30w4_gate.sh. Nothing from it is committed yet — commit only after R22 shows 140/140.
• Tree carries Lane A's 15 banks (sweep finished: 15 banked / 712 failed / 910 already-not-stub — the population was largely consumed by the earlier sweep, so Lane A is now nearly spent).

▶ RESUME HERE (in order)

1. Finish/verify the w4-gate → R22 140/140 → commit.
2. RESUME THE 12 UNFINISHED AGENTS (.run/p30w4_remaining.json) — Drew's explicit instruction, cron 799f00d8 set for 00:49 on 2026-07-31. Wave 4 launched 70, 58 returned (ALL 58 match_one MATCH), 12 were killed mid-run: func_8017EA1C(SC02_041) · func_8018A150(SC02_027) · func_80185680(SC02_017) · func_8017D900,func_801846E4(SC03_014) · func_80187940,func_80185B5C,func_801885C0(SC04_011) · func_801832D4,func_80183CBC,func_80183258,func_8017D7A4(SC02_035). Workflow scriptPath /tmp/claude-1000/-home-musashi-bfm-decomp/f1a9b030-5924-4b17-b102-3de2ca684a6f/scratchpad/p30w4.js, args = the 12 targets as an ARRAY (a JSON string breaks pipeline()). ⚠️ 10 of the 12 have a PARTIAL draft on disk from the killed run — a starting point, NOT a verified result; re-verify with match_one before any MATCH claim. (Counting note, R14: my first count said "2 remaining" because I measured draft-FILE existence. An agent writes its draft early and then iterates, so a file proves nothing about completion — the journal's result records are the truth. Drew's "12" was right and my measure was wrong.)
3. The h_exact propagation leg that halted at 0x801466F0 (rc=1, stopped by design).
4. The 137 "no matched unit for func" sweep-routing gap — the exemplar banked in a DIFFERENT binary than family_sweep looks in. Likely cheap and worth ~137 members.
5. 3 jr behemoths (func_80191C50, func_8018057C, func_80181CDC, staged in .run/beh-gate/<binary>/) + group B's carve (13-member IMM/jr family @952, 103 already matched → jtbl_family_bank). Both run make extract — run them ALONE under the lock.
6. T2's -O0 carve-within-a-carve (13 families / 1,746 members walled) — the deep arc.

⚠️ THE TWO LAWS THIS SESSION COST US (both self-inflicted, both now mechanised)

• Guard the CAMPAIGN, not the process (tools/treelock.sh). A pgrep poll cannot see a campaign made of sequential processes; mine started in a gap and 63/140 binaries broke.
• Gate with --no-propagate, then propagate PER FUNCTION, routed by tier (§123). A fixed 3600s propagation timeout killed the driver mid-fleet-write; a killed process performs no undo.
• Corollary proven twice: agents must only ever write .run/ — that is why both incidents cost build cycles and zero work.

📓 NEW THIS SESSION

Cookbook §122 (gate raw before transforming; undo belongs to the writer as a per-edit journal), §123 (propagate a family with the tool its TIER needs + routing table), docs/cookbook-index.md (226→324 sections, symptom-keyed, in tools-health), tools/treelock.sh, tools/prefetch_fleet.py, tools/cookbook_index.py.

---

🛑 (superseded) SESSION CHECKPOINT (2026-07-30, in-flight)

Written while two background campaigns run; REFRESH before any pause (the stale-checkpoint trap).

Running in background: (1) T0.5 full prefetch batch (prefetch_fleet.py, 126 programs — imports ~3 min each + define+decompile; cache was 3,680 at launch; task baw61zyz6; resumable — re-run the same command if interrupted). (2) T1a recovery sweep (.run/recover/t1a_runner.py → .run/recover/t1a.log; 108 autopsy-MATCH fns / 50 binaries; driver commits banks per binary via gate_stage's scoped add; OWED AFTER IT LANDS: a config sweep-up commit — jr-bank preps edit config/overlays.mk + splat yamls which gate_stage --commit does NOT stage — then R22 clean-fleet

• report point #2).

HEAD at checkpoint: commit:1223. Fleet: 87.5 / 78.0 / 92.00 (pre-T1a-banks). Tree: T1a's in-flight src/config edits (theirs); .run/ artifacts.

**T0 ✅ (6 commits) · T0.5 ▶ (batch running; main+3 overlays done: +2,241 cached) · T1a ▶ · T2 ▶ (two probes done, arc re-scoped — see log; NEXT session opener: sub-split carve + shift-diff

• symbol-pin probe on ov_SC07_007) · T3/T4/T5 pending.**

If resuming fresh: read docs/frontier-p30.md (report point #1, with the T2 R14 correction) + the T2 log entries; check both background tasks' outcomes first (git log for t6-recover commits; ls .run/ghidra_c | wc -l); the Ultracode prompt for T3's first wave is still OWED to Drew (R26/R27) — T1b agent-residue wave folds into it.

Per-task log

▶ S11 — the propagation lag: EXTEND 0/36 -> 31/36, and every blocker was a DECLARATION (2026-08-03/04)

Lane 2 of the S10 checkpoint ("26,006 ins, ~0 agent tokens, PARTLY BLOCKED"), taken first on the standing doctrine that the cheap deterministic lever is probed before the expensive agent one.

EXTEND shape — 31 of 36 banked (was 0). Every failure was captured from the compiler's own output (§136a) rather than guessed, and the 9-per-binary failure set reduced to 4 distinct symbols repeated across all four SC07 binaries plus the 4 "DIFF"s:

• func_80146C3C (8 of 36) — one token. Two macros declared it (void); the SC07 TUs declare (u8 *a0). Both macros already CAST at the call site, so the prototype is codegen-irrelevant -> relaxed to (). Measured BEFORE editing: all 4,020 fleet decls are (void)/()/(u8*)/(u8 *a0) — no default-promotion param anywhere, so gcc-2.7.2's () dead-end (which needs a promoting param) cannot bite. T2 but 2 lines.
• func_8014F4C0 (self-axis, 4), D_80126CC4 (3), func_8012E5CC (1) — T1, binary-local. Conformed the SC07 decls to the fleet-canonical form. Each has zero uncast uses in its TU (verified per file), so they were splat boilerplate with no codegen effect.
• One conflict hid another: fixing func_8012E5CC revealed func_80147364 behind it — the S29 law that a declaration conflict ABORTS the compile, so one error says nothing about the next.

The 4 "DIFF"s are a volatile declaration, not a codegen wall. dedup_extend's own header records them as UNDIAGNOSED while its correctness argument says an h_exact match makes a DIFF impossible. Both halves resolved against the bytes:

1. The contract HOLDS — func_80162FF4's original bytes are sha1-identical in ov_SC07_006 and ov_SC01_000 (af1aceb2…).
2. The cause is TU CONTEXT — the SC07 host TU (_jr_8015C32C.c:1177) declares extern volatile s32 D_80127090/94/98 at FILE scope; none of the 134 working overlays' copy of that TU does. Volatile makes the macro's three stores a scheduling barrier, so addu $a0,$s2,$zero cannot sink into the jal func_80146D30 delay slot — the built body emits it early plus a nop, one instruction longer. Measured word-for-word: built +0x090 addu / lui,sw x3 / jal / NOP vs ref +0x090 lui,sw x3 / jal / addu-in-delay-slot. All 4 macros touch exactly those 3 symbols, which is why all 4 fail in all 4 binaries and nowhere else. Fix: the §37/§124 DATA asm-label alias inside the 4 macros — a distinct C identifier is immune to any TU's declaration and byte-neutral in the other 134. 16/16 banked on the retry.

Still open in this lane (both NAMED, neither a wall): func_80144B9C x4 — the whale's registry func field is a bare name, not a DEFINE_ macro, so write_drafts emits a CALL (hence "undefined reference" in ov_SC07_010); it needs the §38 -O0 shared-header route, and dedup_extend should refuse-and-name the class per R32. And func_80149954 x1, blocked behind func_80147364's u16 params.

func_8012A598 banked x138 (3,288 ins) — and it took TWO fixes, either alone insufficient. The tool skipped it as missing file-scope extern (CARRY-FIXABLE): D_801151D4, D_80126DB8_a, D_80127504.

1. TOOL (R33): find_site's preamble backscan had the SESSION-18 fix for blank / // / SINGLE-LINE /* … */ lines, but a MULTI-LINE block comment still halted it — its middle lines start with * and its last line ends */ without starting /*. The §134 multi-line-blindness class, which S6b already fixed three times in family_remap. Replaced the line-syntax test with a decision on cdecl._mask (the project's ONE masking oracle) — it subsumes every comment form and cannot be fooled by a /* inside a string — plus 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 (a bug here can fail to bank, 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 redefine it in a 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. Probed on ONE member first: byte-identical 9052dc0e…; then 138 overlays byte-identical.

PROPAGATE shape — the head is 5 classes, not 45. .run/s8_lag.json re-split: EXTEND 13 classes / 45 stubs / 5,145 ins; PROPAGATE 46 / 783 / 20,861 — but 5 classes carry 18,545 ins (89%) and the other 41 carry 2,316. Ranked: func_80147364 30x137=4,110 · func_8012f274 29x137=3,973 · func_8016ba68 29x134=3,886 · func_8012a598 24x137=3,288 · func_801466f0 24x137=3,288.

func_80147364 banked x137 (4,110 ins) via the DEFINITION-SIDE asm-label alias. Its byte-true definition is (u16, u16) while the fleet declares it (u16, s32) in 4,046 places — u16 is a default-promotion type, so the () escape is ILLEGAL and a decl conform would change caller codegen. The alias gives the definition 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 the full run: 137 overlays byte-identical. In-tree precedent for the form: 1,725 files already use it.

Named + sized, not yet banked: 0x8012a598 skipped missing file-scope extern (CARRY-FIXABLE): D_801151D4, D_80126DB8_a, D_80127504. 0x801466F0 reports "no source overlay has it matched" — the S6b D4 case, still unfixed: its def in ov_SC01_077_after.c:495 carries a wrapped __asm__ alias decl and _alias_decl_for is a single-line rx.match (the §134 multi-line-blindness class), plus a file-scope typedef struct Rec801466F0 that a macro body cannot travel with (§100).

My process errors this block, recorded: (1) I wrapped the first R22 in nohup … & inside a backgrounded call, so the harness signalled completion of the WRAPPER while the fleet check stood at 63/140 — I nearly read that as a pass. Wait on pgrep -x make, never on a wrapper's exit. (2) I ran a corpus.stubs probe while make extract-all was mid-flight and got garbage; R32's coverage assertion refused to answer instead of returning a wrong stub set. A measurement taken during a rebuild is not a measurement — S27's law, re-earned. (3) CORRECTION to the S10 checkpoint's own waiter rule. It says "use pgrep -x make (exact process name)". That is right for a single make, and wrong for a CAMPAIGN: dedup_propagate runs a SEQUENCE of make build BINARY=<ov> calls, so between any two there is a window with no make process — my waiter fired in one and reported "done" on a live campaign whose log had one line. This is the identical structure tools/treelock.sh's own docstring warns about ("guard the CAMPAIGN, not the process"; a pgrep poll is a sampling test on a gappy signal). Wait on the campaign itself — pgrep -f dedup_propagate / family_sweep — or on tools/treelock.sh --status, which is a statement of intent that spans the gaps. Three waiter failures now share one mechanism: the signal being sampled is not the thing being waited for.

▶ S6a — the source-agnostic zero-crack sweep: 842 banked, and the residue is OUR TOOLING again (2026-08-01)

Setup, measured before running anything (R37 — probe before costing). The 190 zero-crack families (129,997 templatable ins) were decomposed with zero builds by remapping each family's exemplar onto one member and type-checking the resulting draft statically: · 117 families / 92,368 ins / 2,735 member-slots — draft self-contained ⇒ sweepable now. · 17 / 24,332 ins — the §94/§100 class: the body uses a file-scope MULTI-LINE typedef, which extract_unit cannot carry (it carries single-line ones only). Head prizes 0x8012b77c (ov_SC07_006, 139 members, Ang2_8012B77C) and 0x80128c98 (ov_SC01_077_a, 275 members, CdFileLoc) = 14,937 ins. · 9 / 9,364 ins — has_mid_jr ⇒ the §53 carve path, not this sweep. · 47 / 3,933 ins — remap_hseq REFUSES (mostly unresolved immediates). Genuine, small.

⚠ PREMISE CORRECTION (R14) — the checkpoint's stated mechanism was wrong. The post-wave frontier says these families were unreachable because "every sweep this project has run passed --source ov_SC01_077". It is not the mechanism: --source ov_SC01_077 is the default, and hseq_sweep's override block only fires when source != EX_OV, so the manifest's own non-ov077 exemplar was already being used. The real gate is --band substantial (the default): of the 181 non-jr zero-crack families only 13 are substantial — 114 mid, 54 tiny. --band all is the unlock, exactly as the S1 trap note warned. (The 106 non-ov077 exemplars are real and interesting; they were not what blocked the sweep.)

Result: family_sweep --hseq --band all -j12 --only <117 addrs> staged 2,735 drafts / 1,239 groups, 0 skips → BANKED 842 / 1,893 failed. R22 clean-fleet make clean + extract-all + check-all → 140 passed, 0 failed of 140. Fleet 94.43→94.67% fn-count · 91.4→91.6% instr · 84.0→84.5% distinct.

The residue is bimodal, and that is the finding: 57 families banked ALL members, 52 banked ZERO, 8 partial. A per-member compiler wall does not produce that shape; one per-family blocker does. Eight zero-banked families were probed with a new generic blocker-capture (.run/s6_diag.py — the any-overlay sibling of S29's ov077-only .run/uc_capture.py: remap → splice → build THAT overlay → read the compiler's own error → revert; one build per family, not 137). Seven of eight are declaration/carry plumbing; one is a genuine byte DIFF. Two are proven defects in tools/family_remap.py, both located at source:

• D1 — the preamble backscan's {-guard does not exempt comment lines. The guard is a real T65 fix (extern void f(int); int g(){…} on one line), but it fires on any documentation line that merely mentions a brace, so the carry stops mid-comment and the sibling receives a dangling * … fragment → parse error before 'the' (0x80176144, 0x80146afc).
• D2 — _def_head_at returns True on "param list continues past this line ⇒ ANSI definition". False for a wrapped multi-line DECLARATION (extern void func_801466F0(s32 a0, …,\n s32 sp8);), so extract_unit accepts line 457 of ov_SC01_077_after.c as a definition head and returns a 16-line fragment that contains no function body at all (DEFINE_func_801466B4() + #include "common.h" + a comment), terminated by a brace pair inside that comment. Verified directly: _def_head_at('extern void func_801466F0(s32 a0, s32 a1, s32 a2, s32 a3,', idx) -> True. This is the §110/R35 class again — a silent 0/137 wearing a compiler wall's clothes.

S6b — the fix belongs in the TOOL, not in 17 exemplars (R33). The plan was to hand-move each MISSING-TYPE exemplar's typedef into its function body (§100, byte-neutral, 17 edits + 17 verifications). Probing the class first showed it is the SAME root cause as D1/D2 — a multi-line construct read one line at a time — so it was fixed once, in family_remap, for every future family instead:

• D5 — _typedef_block_start(): on a } T; closing line, walk up to the matching typedef (both the typedef struct { and brace-on-its-own-line forms) and carry the WHOLE block. Only blocks that literally begin with typedef are carried (a struct VARIABLE closes identically but would be a duplicate global in the sibling); safe because harvest_verify already strips a typedef the sibling TU provides. The old code comment claimed these functions "route through the engine_types.h lift" — measured, they routed nowhere: 17 families / 24,332 ins were simply dropped.
• D4 — the wrapped asm-label alias (_alias_decl_for is a single-line rx.match) was MEASURED (1 exemplar, 0x801466f0, 137 members / 3,288 ins, with a second blocker behind it) and deliberately left unfixed. It now returns None, so the sweep reports skipped {'no matched unit for func'} instead of failing 137 times. A known gap that announces itself is not the same defect as a silent one (R32).
• Also hardened: the forward body scan now counts braces in cdecl._masked text (R33 — one masking oracle), so a brace inside a comment or string can never close a function body.

Blast radius verified, not assumed (R14): extract_unit output diffed against the pre-S6 tool over all 181 zero-crack exemplars — 157 byte-IDENTICAL, 24 changed, every change in the intended direction (more preamble carried; one → None, the D4 case that previously produced garbage).

Measured payoff of the repairs, one sweep each: · D1+D2 → +323 members recovered from families that had banked zero. · D5 → +417 members, including func_8012B77C 139/139 (8,062 ins — the head prize) and func_80128C98 137/275; func_80146AFC 135/137, func_80176144 99/137. · S6 total: 1,582 member-matches. The pre-fix tool scored this same population at 842. R22 clean-fleet make clean + extract-all + check-all → 140 passed, 0 failed of 140. Fleet 94.43→94.88% fn-count · 91.4→91.9% instr · 84.0→84.6% distinct; phase arc from open (92.00 / 87.5 / 78.0) = +2.88pp fn · +4.4pp instr · +6.6pp distinct.

Family state now: 70 ALL-banked · 17 partial · 94 zero-banked; 64,050 templatable ins still open. Distilled to cookbook §134 (the MULTI-LINE BLINDNESS class + the bimodality tell: 57/52/8 is a tooling signature, not a codegen one — read ONE compiler error before believing the compiler).

✅ S6h — wave 3 (38 targets, size-routed) + an 8/8 reconcile lane (2026-08-01)

38 targets / 44,297 templatable ins — the 11 high-multiplier families at 61-120 ins plus the top 26 of the 10-19-member band, model-routed (Haiku ≤89 with Opus escalation, Opus direct ≥90). 52 agents, ~4.1M tokens, 30 min → gate banked 27/38 (71%). All 11 failures were captured and classified: 8 declaration/link plumbing · 3 genuine byte-DIFF. An 8-agent Opus reconcile wave then fixed 8/8 (7 banked; 1 re-failed) → wave-3 total 34/38 = 89%, propagation +639 members / 1 failed / 83 overlays. R22 clean-fleet 140/140. Fleet 95.23 → 95.42% fn · 92.1 → 92.4% instr · 85.0 → 85.5% distinct.

The reconcile lane is now 12/12 across two waves and is the most reliable stage in the pipeline. Design note (S27 law, applied BEFORE it bit): six of the eight reconcile targets share ONE TU, so this wave forbade agents any build — six concurrent splice-builds would have clobbered a tracked file. Wave 2's reconciles were allowed one build each because they were spread across TUs.

The agents out-diagnosed the blockers I handed them — three findings worth keeping: · func_801848B0: an agent REJECTED MY PREMISE. I told it the draft was byte-correct and only declaration-blocked; it ran match_one first, found a real 1-instruction DIFF, and fixed both. R14 coming back at me from a subagent, correctly. · func_8017C5F0: the "invented symbol" was an INTERIOR ADDRESS. D_801DA0F0 has no symbol because it is offset 0x6C into D_801DA084 (which spans 0x801DA084..0x801DA103, asm/ov_SC02_027/data/tail18.data.s). The lui/addiu pair builds an interior pointer — a class the drafters will hit again. · func_8018A860: the TU declares memcpy THREE times with incompatible signatures (L81, L1518 builtin-compatible, L1524), and a latent byte bug sat behind it. A TU that declares one symbol three ways is a defect waiting for the next draft that touches it.

Carried (4): func_80184A94 (reconciled to match_one MATCH, still gate-refused) + the 3 genuine byte-DIFFs func_801845B0, func_8017BEBC(ov_SC02_026), func_8018480C — real codegen residuals, ledger material.

✅ S6g — wave 2: 83% → 93% bank rate from ONE batch of learning (2026-08-01)

15 targets (11 fresh Haiku + 4 gate-failed reconciles on Opus), 15 agents, ~0.74M tokens. Gate banked 14/15 (93%) — up from wave 1's 20/24 (83%) — and all 4 reconciles banked. Propagation +328 members / 0 failed / 76 overlays. R22 clean-fleet 140/140. Fleet 95.13 → 95.23% fn · 92.0 → 92.1% instr · 84.9 → 85.0% distinct.

Three changes between the waves, each fixing a wave-1 failure — this is the S27 finding reproducing (57→79→100% there, 83→93% here):

1. Args pasted from the DERIVED manifest, not typed. All 30 paths verified present on disk before launch (asm 15/15, tu 15/15).
2. Blocker-capture before the reconcile fan-out (the S29 law: agents cannot run the gate, so a match_one-MATCH draft that dies on conflicting types for D_800A5E88 looks like a codegen problem to them). Each reconcile agent got the exact symbol + line, plus the two byte-neutral levers (conform-and-cast-at-use; the §37/§124 asm-label alias).
3. Wave-1's Opus DISCOVERIES became wave-2's Haiku INSTRUCTIONS (the ori-vs-addiu unsigned destination rule; the store-sinking scheduler rule). The flywheel closing on itself.

The reconciles produced better root causes than my captured blocker did — worth keeping: · func_80189B78: my capture named ONE conflicting symbol; the agent found the draft had invented prototypes for six the TU already declares, two of them AFTER the splice point where cc1 had not yet reached. Fixed by copying all six from the TU verbatim and pushing every type disagreement to a cast at the call site — zero bytes changed. · func_8018584C: lever (A) was blocked in both directions — the draft cannot use the TU's tag (it must also compile standalone for match_one) and cannot redefine it. Closed with the data form of the asm-label alias: extern struct B16_8018584C aD800A5E88 __asm__("D_800A5E88"); (in-tree precedent at ov_SC06_008_jr_80135D20.c:1434). · func_80180A4C: the draft declared extern s32 D_801A9DA0[], the TU says extern u8 D_801A9DA0[] eleven lines after the splice point — conform to u8, the use site already casts. One character class, byte-neutral.

✅ S6f — the B-shaped wave: Haiku drafts, Opus closes, the gate arbitrates (2026-08-01)

The pool. Derived from the regenerated map: 36 families / 28,829 templatable ins that are kind=modal (no member matched ANYWHERE, so no sweep could reach them), ≥20 members, ≤60 ins, non-jr, and not one exemplar in ov_SC01_077. Calibrated by hand first (Phase-15/18 discipline): 3/3 one-shot — func_8017CDD8 (17 ins ×142), func_8017CE7C (16 ins ×126), func_8017E934 (29 ins ×65) = 330 members from 62 instructions of C, ~0 agent tokens.

The wave (Drew enabled Ultracode; Haiku drafters + Opus escalation, 24 targets). Haiku was the right tier by the cheap-tier-ab-validated measurement (≡ Opus at ≤~50 ins, ~4.8× cheaper); this pool is 16-53 ins. 31 agents, 0 errors, ~2.0M tokens, 12.5 min. · agents claimed 24/24 MATCH · whole-binary gate banked 20/24 (83%) · propagation +524 members / 0 failed / 91 overlays. 17 of the 20 banks were Haiku, 3 Opus. R22 clean-fleet 140/140. Fleet 94.97 → 95.13% fn · 91.9 → 92.0% instr · 84.7 → 84.9% distinct.

What Opus actually bought (it earned its slot on exactly the cases the A/B predicted): · a sh of a constant whose stored-width top bit is set needs a u16 destination — through s16 gcc-2.7.2 folds it sign-extended and li emits addiu; through u16, force_fit_type keeps it positive and li emits ori (func_8018A0F4). · a schedule-reorder closed by STATEMENT ORDER, not the permuter: gcc's list scheduler preserves relative order of disambiguable stores, so two field-zero stores had to be written ABOVE a three-halfword copy to sink correctly — 9 mismatches → 0 (func_8017CAB4). · three separate loose-typing fn-ptr casts (((void(*)(s32))func_80178D18)(a0)) where the TU declares (void) but the asm passes $a0 — a cheap drafter had misread all three as delay-slot / permuter-class residuals (func_8017EAAC, func_8017CBD4, func_8017D5C4).

⚠ MY ERROR, recorded (R37/R14). I generated the wave manifest to .run/s6f_wave_targets.json and then hand-transcribed the args into the Workflow call instead of pasting the generated file, pattern-filling _jr_8017BEBC across overlays where no such split exists (ov_SC02_027, ov_SC04_018, ov_SC03_002 — corpus.stubs says _jr_8017AE2C). Three agents lost time rediscovering the real paths and reported SPEC ERRATA. Cost: tokens, not correctness. The gate driver written afterwards (.run/s6f_gate.py) derives every TU/split from corpus.stubs per draft and asserts nothing — which is what the manifest should have done. Assert nothing you can derive.

The 24→20 gap is the known match_one→gate gap, not new: match_one compiles standalone and cannot see a TU-level declaration conflict (Phase-19 measured 88-92% match_one → 60-71% banking). 4 carried: func_8018584C, func_80180A4C (ov_SC02_026), func_8017CC80 (ov_SC02_027), func_80189B78 (ov_SC03_014).

✅ A + B — the ×138 era was NOT over: two 16-ins cracks bought 266 members (2026-08-01)

A (frontier regen at HEAD, R35). Sigs + family_hseq regenerated after S6's banks — required before pricing anything, and doubly so because .run/hseq_verified.*.txt has accumulated 22,841 files across every sweep ever run, so any per-family analysis that globs them over-counts. The regenerated map derives state from sigs + corpus.stubs (R33), which is the authority. Overlay-only frontier at HEAD: 95.2% fn / 92.3% instr / 85.5% distinct; zero-crack residue 106 families / 58,657 ins (was 181 / 120,633 at S6 open).

B — and the finding that matters. The S29 checkpoint's structural signal said "after S2 the ×138 era ENDS — those are the last two crackable fleet-wide families", and made that the trigger for T5/phase close. It is wrong, and this is the THIRD §133-class miss in a row. Re-derived from the family map, two fresh-crack families with ≥126 members were open the whole time: · 0x8017cdd8 ov_SC02_039 — 17 ins × 142 members = 2,414 templatable ins, PURE · 0x8017ce7c ov_SC03_114 — 16 ins × 126 members = 2,016 templatable ins, IMM Both are kind=modal — no member matched anywhere, so no sweep could reach them, and neither exemplar is in ov_SC01_077. They are invisible to exactly the two habits this phase already corrected.

Both hand-drafted off the .s, match_one MATCH on the FIRST try, ~0 agent tokens. Then:

• First gate attempt FAILED, class PLUMBING (not DIFF) — my draft declared extern void func_8017CFCC(s32 a0); while the TU defines void func_8017CFCC(void). The target passes $a0 anyway (the caller's incoming argument simply still sits in the register — loose typing, §16), so the byte-true C calls it with no argument and declares it exactly as the TU defines it. Re-verified MATCH, then gated byte-identical in both overlays.
• Propagated: 266 member-matches / 0 failed across 118 overlays — from 33 instructions of C.
• R14 on the seed: the cached Ghidra-C for func_8017CE7C decompiled an entirely DIFFERENT body (three calls that do not exist in the asm). Reading the .s is what made it one-shot; a drafter trusting the seed would have burned the target and possibly ledgered it as hard.

✅ S6e — --normalize-self-decls measured to ZERO, but only after fixing the guard that made the measurement impossible (2026-08-01)

The S6a blocker probe found conflicting types for func_X on the SELF axis, which is exactly what family_sweep --normalize-self-decls targets (drop the sibling's divergent decl of the function being defined, cast its in-scope calls). It is OFF by default, so it was the obvious next zero-token lever.

Run 1: 0 banked / 1,622 failed — and 909 of 909 groups reported ⚠ self-decl edit NON-NEUTRAL. A 100% rate is a statement about the mechanism, not about 1,622 different functions (the §134 tell, one day old). The three earlier sweeps over the SAME population reported 0 NON-NEUTRAL.

The defect (D6, family_sweep.hseq_sweep): the TU snapshot was taken unconditionally, one line before the if nfix: that decides whether to edit at all — so a TU that NSD merely INSPECTED still landed in tu_snapshots, and the phase-2 MISMATCH backstop then attributed ANY group failure to a "self-decl edit" that was never made, reverting the TU and reporting 0/N banked. Measured directly: NSD fires on ~25% of members (3 of 12 probed), so ~75% of those reverts were attributing a failure to an edit that did not exist. The §103 tu-scope path immediately below has always snapshotted inside its own if _rep["moved"]:; NSD was the odd one out. Fixed to match.

Run 2 (fixed): NON-NEUTRAL 909 → 303 (~33%, consistent with the measured fire rate) — the fix is confirmed — and still 0 banked. The 606 groups that now took the NORMAL path banked nothing, so the verdict is real: this residue is not self-decl-conflict-bound. Lever measured, closed, zero.

One open item, deliberately not chased (it is no longer decision-relevant). I byte-measured a firing case myself rather than trusting the backstop (R14): func_80162CCC in ov_SC01_000 builds to 9052dc0e… with and without the NSD edit — i.e. byte-NEUTRAL, so the surviving 303 verdicts are wrong too. Most likely the sweep applies NSD to several members of the SAME TU and the backstop attributes an accumulated-state mismatch to "the edit". Left as a named open item because the lever yields 0 either way; it would matter only if a future population makes NSD worth re-testing.

✅ S3 — the close=0 stored-draft DIAGNOSTIC pass: measured, classified, and correctly NOT scaled (2026-08-01)

Chartered as "classify, and only build a fix if ≥3 share a named class". Ran exactly that; the answer is that no cheap shared class exists, so nothing was scaled. The population is bigger than the plan thought and worth less than it looks:

• 30 ×138 family heads carry a stored draft = 182,850 templatable ins (the plan said ~12 drafts).
• Only 7 of 30 still verify (match_one MATCH). 23 have DECAYED — mostly LENGTH-DRIFT, a few SIZE-MISMATCH/WIDTH/ADDRESSING. ⇒ a stored draft's recorded closeness is not a current fact: 77% of this backlog no longer matches. Re-verify before valuing a draft (extends A10/T1a).
• Of the 7 survivors: 1 banked clean (func_801754A8, 37 ins ×138 = 5,106). The other 6 failed conflicting types — and gate_stage's ladder recovered 0/6.
• Those 6 split into: 1 §30#2 return-widen (func_8014C4AC) and 5 PARAMETER-signature conflicts (the Phase-16 def-side loose-typing wall, unchanged).

The §30#2 attempt, and why it was reverted (R14 + the anti-bulk-header law). func_8014C4AC's draft documents its own fix (extern void→s32 in engine_core.h, measured byte-neutral there) — but the note's claim that no split .c carries its own decl is wrong: the gate named _jr_80154C24.c, then _jr_80182268.c. A sed over every matching decl touched 2,046 files and STILL did not go green. That is precisely the "bulk header edits BREAK builds" pattern (§29 doctrine: call-site casts are the integration spine). Reverted; exemplar re-verified d19c9580. A fleet-wide decl reconcile needs a gated TOOL, not a sed — logged, not attempted.

Ledger: 23 decayed drafts → redraft lanes (they are not "nearly done"); 5 param-conflict → the def-side wall; 1 widen → needs the gated decl-reconcile tool. S3 is CLOSED as a measurement.

✅ S29 — JR-PAIR-IN-ONE-O0-OBJECT RETIRED: two instrument defects, both fixed; the pair banked (2026-07-31)

The S28 ledger class and its recorded escape (§81 step 1, isolate one fn into its own code subseg) are both refuted. Neither function needed isolation; neither is on a compiler wall.

Root cause 1 — jtbl_carve could not see a 4th table. ov_SC01_077_o0's carve at 0xb01a4 predates the §8e tables= persistence and is a merged double (func_8013C0F8 $L75 + func_8013C414 $L105); the second owner is MATCHED, so make extract pruned the stub .s that named its table. The tool's "single-table predecessor" inference therefore derived 3 starts where the object emits 4 tables, wrote JTBL_PADS := 0,4,4, and jtbl_rodata_pads refused mid-stream — correctly, with the exact message. Root cause 2 — the refusal left a corpse. as reads a pipeline, so it had already written a TRUNCATED .o (12 of 16 T func_; undefined $L57/$L59/$L63/$L75/$L76). make reported Error 1 and left the object on disk, newer than its .c — no .DELETE_ON_ERROR. The next build linked the corpse: that IS S28's undefined reference to $L105 + func_8013C938, one build downstream of a loud, correct compile error.

Fixes (both byte-gated). (a) Makefile: .DELETE_ON_ERROR: — negative-control-proven (make: *** Deleting file …), scratch invocation, tracked files untouched. (b) tools/jtbl_carve.py spec_from_starts: a R32 coverage assertion + payload recovery at the single choke point — every zero word inside a span is an original .align 3 pad (the tool's own axiom), so the word after it starts a table; recovered starts are logged. No-op where the structure is already known (the 134 sibling _o0c spans DO carry tables=+0x0,+0x70), so committed-green spans are untouched. Honest limit: only pad-separated boundaries are recoverable; a tight boundary still fails LOUD (short spec → filter refusal), never silently.

Byte proof. 4 tables at 0x801D8254 / 0x801D828C / 0x801D82FC / 0x801D836C (13/27/27/27 entries, each preceded by a zero pad); span 0xb00fc..0xb0280 = 388 B = 52+4+108+4+108+4+108 — closes exactly; derived spec 0,4,4,4 tables=+0x0,+0x38,+0xa8,+0x118. func_8013B83C (272 ins) + func_8013BD74 (198 ins) banked, ov_SC01_077 d19c9580, R22 clean-fleet 140/140 (make clean + extract-all + check-all — the incremental result was NOT trusted, §130). Fleet 93.25% fn-count / 89.2% instr / 80.5% distinct; dedup 1905/0; 0 NON_MATCHING. Cookbook §132 (incl. the reusable ladder: compile the ONE TU standalone and let .section .rodata + the nearest preceding .ent name every table owner — it found the 4th owner and attributed $L105 in one command, before any build or carve).

Sweep prep — the §94/§100 type-carry class, caught by probing ONE sibling (R37). The first sibling gate-failed. The §132 ladder named it in one command: the sibling object emitted only func_8013C0F8 + func_8013C414 tables — BD74's body never compiled (E_13BD74' undeclared). Cause: BD74's draft declared E_13BD74/P_13BD74 at FILE scope, and extract_unit/remap_hseq carry only the BODY — the exact class that held func_8013C08C at 0/137 earlier this phase. Fix = §100 draft-local types (moved both typedefs inside the function body; byte-neutral, exemplar re-gated d19c9580). B83C's draft was already draft-local, hence unaffected. A file-scope type in a draft is an exemplar-only bank: it passes the gate and silently cannot travel.

THE SWEEP — 137/137 both, i.e. 2 × 138 = 276 function-instances banked. jtbl_family_bank, BD74 first then B83C (order is load-bearing: B83C's table sits BELOW BD74's, so carving it while BD74 is still raw leaves an unmatched table inside the span — NON-CONTIGUOUS, byte-observed on the exemplar). ~8.5 s/sibling, zero-token. Both sweeps failed on the SAME single overlay, ov_SC07_010 — which is a property of that sibling, not noise, so I probed it instead of ledgering it and it returned 2 banks plus a real tool defect (§132a): jtbl_carve --like matches donor→recipient by the subseg's ROLE NAME; ov_SC07_010's -O0 region is named _o0, the same role as the exemplar's and the ONLY other overlay so named (the other 136 are _o0c, whose role never matched — the only reason the sweep worked at all). The exemplar had just banked 2 more owners than the sibling has, so the transfer derived SIX starts for THREE emitted tables. Guard shipped (jtbl_family_bank.like_arg: a sibling's own committed tables= is authoritative; --like is for a span with NO record — inert for all 136 already-banked siblings, fixes exactly the broken one).

Fleet after the sweep (R22 clean-fleet 140/140, make clean + extract-all + check-all): 93.33% fn-count · 89.6% instr-weighted · 81.6% distinct-code (72,416 unique fns, +260); dedup 1905/0; 0 NON_MATCHING. Phase arc so far: 92.00→93.33 fn / 87.5→89.6 instr / 78.0→81.6 distinct. That -O0 cluster is now COMPLETE across the fleet — the two functions were the last open stubs in every overlay's _o0* region.

✅ T0.5 COMPLETE — the fleet Ghidra-C prefetch: 124/124 programs, 7,716 files, 335 min, unattended

Zero-token, headless, resumable; imported ~120 overlay programs on demand and ran the DefineFunctions completion pass per program (raw-blob auto-analysis only finds the reachable subset — the Phase-10 finding, now automated). Coverage measured, not assumed:

• 5,777 of 6,070 families needing a fresh crack (95.2%) have a cached seed
• main 2,001/2,002 (100%) · resident 14/14 (100%) · 1 residue (one main addr with no fn)

What this changes: the drafting surface is now fully fueled with NO MCP dependency, so waves can run unattended against any part of the fleet — the precondition Drew asked about. It also un-blocks P31's main-EXE campaign early (its 1,034 game-code stubs are all cached), though that stays P31 scope. tools/prefetch_fleet.py is committed and re-runnable; the Ghidra programs are script-reproducible, so only the tool is tracked, not the DB bulk.

✅ REDO UNDER LOCK — 52 cores + 911 members banked; R22 140/140; the "cliff" was an artifact

Re-ran the whole banking sequence serially under treelock.sh: parallel 8-binary gate → non-jr behemoth gate → sig+map regen (a bank invalidates the map) → tier-routed propagation (§123) → one R22 over everything.

• Wave 3: 48/48 banked through 8 PARALLEL per-binary gates (the throughput unlock, proven).
• Wave 2: 15/19. Behemoths: 4 non-jr confirmed (incl. func_8017E120 884 ins ×14).
• family_sweep --hseq: 911 members banked / 401 failed across 137 overlays, 49 families.
• Fleet: 92.59% fn-count · 88.2% instr · 78.7% distinct (70,506 unique fns) · R22 140/140.

R14 CORRECTION, on my own earlier claim. I reported a "per-binary bank-rate cliff" (SC03_014 1/6, SC04_018 1/6, SC06_018 2/6) and theorised a cause — canonical-decl maturity per binary. It was an artifact: those gates ran against a tree that propagation was concurrently rewriting. Re-gated clean, every binary banked 6/6. A measurement taken during corruption is not a measurement, and I theorised before re-running. Same failure shape as the phantom 91.4% in the same incident.

Carried, honestly: the h_seq sweep's 401 failures split as 137 "no matched unit for func" (the exemplar banked in a DIFFERENT binary than the sweep expects — a routing gap, likely cheap), 22 unresolved-immediates, 10 STRUCT (refused by design). One dedup_propagate run stopped the h_exact leg early (0x801466F0 rc=1) — the loop halted as designed rather than risking a partial write; the h_exact leg is unfinished and owed.

⚠️ INCIDENT 2 (mine) — a POLL is not a MUTEX: concurrent tree writers broke 63/140; reverted clean

What I did. Wave-3 gated 8 binaries in parallel (correct — the byte-gate IS per-binary) while wave-2's propagation loop was still running, then I ran make clean (deleting asm/) on top of both. check-all → 77/140, and the corpus denominator itself moved (353,720 → 353,691), so the apparent "91.4% instr" was an artifact of a half-written tree, not a gain.

Root cause — the guard was structurally unsound, not merely unlucky. My gate waited on while pgrep -f dedup_propagate; do sleep; done. But a CAMPAIGN is a loop of short-lived processes — 15 sequential dedup_propagate invocations — so between every pair there is a window with no matching process. The poll sampled one of those gaps, read "clear", and started. Presence- of-a-process is a sampling test on a gappy signal; it cannot express "a campaign owns the tree."

Recovery (clean, nothing lost that mattered). Killed the writers; git checkout -- src/ config/ back to commit:1245 (the last R22-verified 140/140 commit); all 58 drafts survive untouched in .run/ because agents never write the tree — the discipline that made this cheap. Re-verifying baseline, then re-gating and re-propagating serially under the lock.

Fix shipped: tools/treelock.sh — an flock(1) mutex held for the WHOLE campaign, released by the kernel on exit or kill, with --status. Both drivers now refuse to run unlocked. The general law (worth a rule at close): guard the CAMPAIGN, not the process. Corollary, the second time this session a killed writer hurt: a killed process performs no undo, so a fleet-tier write needs a lock ABOVE it, not cleanup inside it.

T3 BEHEMOTH WAVE (10 agents, 3.23M tok) — 7/10 CONFIRMED byte-matches at 700–970 ins, 0 disputed

Drew asked for up to 10 agents on the behemoths; grouping analysis first showed the ≥700-ins band is structured, not 29 singletons: one 16-member family @947, one 14-member @884, a 5-member @793, a 3-member @710, a 2-member @728, 5 true singletons — plus two groups that are TOOL jobs, not agent jobs (the 13-member IMM/jr @952 with 103 already matched → jtbl_family_bank carve; two 134-matched families with 4 stragglers each → sweep). All 10 agents therefore went to genuinely uncracked families/singletons, and each MATCH claim was adversarially re-verified by a second agent.

CONFIRMED (7): func_8017E120 884×14 · func_80191C50 710×3 · func_8017FA5C 728×2 · func_8018057C 897 · func_80181CDC 769 · func_8017CAD4 755 · func_8017E35C 719. NEAR (3), all with byte-grounded residuals: func_8017EF68 969 ins at closeness 2 — two transposed instructions, sched2 rank_for_schedule LUID tie-break, everything else (frame 0x228, spill map, every register) byte-exact · func_8017D174 793 at 82 (allocno tie + pin damage, length EXACT) · func_8017C974 947×16 at 812 (a pure regalloc cascade from ONE extra reserved reload reg; 598/957 regions already exact, no shape errors) — the last is the biggest family left and now has a full structural map + the exact gcc mechanism written down.

MY ERROR (recorded): I hand-typed the asm paths into the workflow args instead of passing the ones I had already DERIVED from corpus.asm_path in the same session — so several task cards named a split that does not exist (_jr_8017AE2C where the truth was _jr_8017C8D0 / _jr_8017C294). The agents located the real paths themselves and reported the discrepancy. Deriving then discarding the derivation is worse than never deriving it: it looks authoritative. Pass the artifact, never a retyped copy of it.

Gating plan (the jr/non-jr split matters): 4 confirmed are non-jr → gate freely. 3 are jr → their jtbl carve runs make extract, which rewrites asm/ under the 48 running wave-3 agents, so they are DEFERRED until wave 3 lands. Staged per-binary in .run/beh-gate/<binary>/ for parallel gating (7 distinct binaries).

T4 (partial) ✅ — the two "Phase-22 grinder bugs" are STALE carry items; the real fixes were elsewhere

Verified against the code, not the list (R14 on our own defect ledger):

• "split-file-blind lookup" — ALREADY FIXED. asm_subdir_for globs nonmatchings/*/ which matches every split; verified on a _jr_ fn, an _o0 fn, and an absent fn. Struck.
• "churn-without-blacklist" — ALREADY FIXED. The blind tried.clear() was replaced by input-signature gating in the T5 targeting work (re-open only fns whose draft mtime/closeness changed), and the blacklist exists. Struck.
• Real fix taken instead (R33): asm_subdir_for was a PARALLEL implementation of corpus.asm_path that silently took g[0] on multiple matches — now derived from the oracle.
• --fix-def-sig posture: flag correctly defaults OFF, but its help still advertised "Byte-neutral; gate arbitrates" — the exact claim T84 refuted (it imposed a signedness-wrong header decl over a byte-correct draft, slti/sltiu, and held 137 members at 0 until DROPPED). Help now carries the §119 warning. The posture was right; the documentation was the defect. Lesson for the phase close: the carried-defect list had 2 of its entries already fixed. A defect ledger nobody re-verifies decays into busywork — verify before scheduling (R35's sequencing applied to the backlog of our own bugs).

T3 wave 1 (Ultracode, 14 agents, 1.20M tok) — 14/14 match_one MATCH → 8/14 BANKED; the gate's propagation TIMEOUT left the fleet half-written (caught, reverted, tool fixed)

The wave: 14 fresh reach-138 cores in ov_SC01_077 (117,162 gain-ins in play), one agent each, drafting against the §31 map + cookbook with match_one/rtu_match self-verification. All 14 returned closeness 0, most cross-verified in the real TU. Whole-binary gate (the arbiter, §52b): 8 banked — func_80175820 (×276!) · func_8012E014 · func_80133298 · func_8015FBE0 · func_8016E9EC · func_80151C54 · func_8012F49C · func_80151B98. Bank truth derived from source (INCLUDE_ASM presence), never the gate report (§55b trap 4). 57% conversion = the §52b law holding exactly; the 6 rejects are all declaration/integration work their agents had already named.

THE INCIDENT (recorded, not buried). gate_stage hardcoded timeout=3600 on dedup_propagate. Ample for one bank; with EIGHT (each rewriting ~138 overlays) it blew, and TimeoutExpired propagated out and killed the driver mid-write: 313 files modified (alphabetically ov_SC01_000..ov_SC03_013 — it died partway through the fleet), config/dedup.us.yaml never updated, engine_core.h half-edited. check-all 124/140. Caught by running the oracle before trusting the state; git checkout -- src/ reverted all of it (§61 — and note a KILLED process performs no undo at all, so the driver's own restore logic was never reached). Nothing was committed at any point; zero contamination. Fix shipped: timeout now SCALES with bank count (1800 + 1800×banks, capped 6h) AND is CAUGHT — on expiry the driver reports TREE DIRTY, REVERT REQUIRED and returns cleanly instead of dying with the fleet open. Recovery path (also the new standing practice for multi-bank waves): re-gate --no-propagate, then propagate PER FUNCTION via dedup_propagate --addr — bounded, resumable, verifiable between steps — rather than one monolithic fleet write.

Three byte-grounded idioms to distill (R16/R30, owed): (1) void→s32 return is NOT byte-neutral — an s32 return keeps $v0 live-out and stops dbr stealing an addiu into the loop-back delay slot; this CORRECTS cookbook §3a-1's neutrality claim (func_8016EC0C). (2) The asm-label alias extern s32 SYM_w __asm__("SYM") — reads a u8-declared shared global as a word WITHOUT *(s32*)&SYM, which force_regs the address and drags 3 extra la pseudos across calls; found independently by two agents; a general unblock for the u8-vs-s32 loose-typing collisions (func_8012E014, func_80133298). (3) Zero-byte __asm__("") as a delay-slot fence — reorg.c stop_search_p halts the eager filler on an asm insn (func_8015FBE0).

T1a ✅ — the deterministic recovery sweep: +18 banked; the stored-draft question CLOSED (report point #2)

Population: the 108 fresh autopsy-MATCH strandeds (+ the 012 abort re-run, 0/3) across 50 binaries, 47 min, 17 driver commits. Metric truth (R14, derived): fn-count +18 instances / distinct +12 unique fns (commit messages summed 19 — gate-report double-list; the metric wins). R22 clean-fleet 140/140 after the sweep + residue revert. Fleet: 87.5 / 78.0 / 92.00 (+384 w-ins). The honest read: the S16 "39%" prior did NOT generalize — it was measured on FRESH same-wave drafts; this stored-backlog population is the class decision-log A10 already byte-proved non-bankable by plain re-gate (0/958). The driver's reconcile/demacroize lifted 18 of ~111 (≈16%) over A10's 0% — a real but bounded delta. The stored-draft recovery question is now CLOSED both ways: cheap wins taken; the ~90 not-banked stay match_one-MATCH but their drafts are integration-decayed — they route to T3's REDRAFT lanes (A10's "fresh re-drafts only"), not another recovery pass. §61 residue class found + reverted (2 orphan carve .c + overlays.mk/yaml edits from FAILED jr attempts). Two live gaps for T3 pre-work: (1) gate_stage --commit add-scope cannot stage NEW carve files / overlays.mk / yaml — moot at 0 jr banks here, live the moment a T3 wave banks a jr fn through it; (2) my own checkpoint edit mid-sweep tripped the driver's blast-radius assert (012's abort) — tracked-file writes and a tree-writing campaign cannot share a repo; queue docs until the campaign exits.

T2 finding 2 — the Arm-A reproduction target has MOVED: the fleet is jr-carved now (2026-07-30)

rollout_o0_cluster.py ov_SC07_007 → SKIP ("cluster fns not in base .c"): since Arm A (07-16), the P29 jr campaign carved every overlay (007 now: 16 _jr_*.c splits) and the -O0 range (0x8013B568..) sits INSIDE <ov>_jr_80135D20.c (23 stubs there in 007). The Arm-A tool's single-file premise is stale fleet-wide — the rollout is now the predicted carve-within-a-carve: sub-split _jr_80135D20.c (pre / o0 / post) per overlay + the Makefile -O0 wildcard for the new class + the +0x20 shift-diff at the NEW seam. Next concrete steps (deep arc, next session's opener): (1) extend the carve tool to sub-split jr files; (2) probe on ov_SC07_007 (zero T1a collision; baseline %lo/%hi inventory saved: .run/t2_007_baseline_syms.txt, 951 operands); (3) on shift: diff the generated asm symbol inventory → pin the drifted D_ symbols at pre-carve addresses in the overlay symbol file → re-extract → per-binary sha (the config-only fix hypothesis); (4) the R31 whale-shape hypothesis (no INCLUDE_ASM in the -O0 split) as fallback. ov_SC07_010's carve is still in-tree + byte-neutral (the Arm-A keeper) — the 010-vs-007 asm diff is a free shift oracle.

T2 ▶ driver generalized; ×1 probe REFUTED the append route for the remaining cluster (2026-07-30)

tools/rollout_o0.py (generalizes T85's rollout_801457a4_o0.py): map-derived members (cross-address-safe), stub located in any non-_o0 split, jr+o0 composites routed out loudly, atomic two-file + per-binary SHA gate + restore-both. ×1 probe (0x8013b6a0@ov_SC01_000): gate-reject — and the diagnosis corrects the frontier report (R14 on my own doc): T85's family banked by APPEND because 0x801457A4 abuts the o0b object's END; the remaining 18 families (0x8013Bxxx–0x8013Cxxx, below the whale) mis-place on append by construction, and per-fn isolation IS the Arm-A re-carve. o0b-bearing ≠ o0b-adjacent — my "the wall never has to fall" inference was wrong; caught at 1 build instead of 2,131. T2's real substance = the Arm-A +0x20 defect. First probe: identify WHICH symbol shifts on a failing overlay's re-carve; hypothesis: splat re-GUESSES a data-symbol boundary the carve moves → pin it in config symbols and re-probe.

T0.5 ▶ launched + probed (2026-07-30)

tools/prefetch_fleet.py (committed commit:1220): 126 programs / 7,966 uncached representatives. Probe 1 (main): +1,525/2,002 cached in <1 min — the decompile leg is CHEAP; 477 main addrs have no defined function in the program (main's 06-14 auto-analysis gap) → follow-up: a DefineFunctions.java completion pass over main from splat entries (the Phase-10 mechanism), then re-run (resumable). Probe 2 (import leg): FAILED then FIXED — my missing-program matcher checked "not exist"/"NOT_FOUND"; Ghidra's actual phrase is "Requested project program file(s) not found" → the self-healing import never fired (+0/371, +0/315). One-line trigger broadening; re-probe running. T1a (deterministic recovery sweep, 108 fns / 50 binaries, driver-committed banks, no-propagate — reach≥2 winners ride T3 Lane A) launched concurrently; commit-scope interaction VERIFIED safe (driver delegates to gate_stage's scoped git add -u src/ — ghidra/ churn cannot leak into bank commits).

T0(e2)+(f) ✅ — autopsy refreshed; frontier regenerated + the three populations PINNED (2026-07-30)

• Autopsy (collect -j12 + report): 1,349 rows — 1,178 near / 63 nobuild / 108 match_one-MATCH (T1's stranded-draft fuel, 3× the S16 sample), 0 classifier errors, second-oracle agree. docs/autopsy.md regenerated.
• Worklist/manifest regen: 101 live stubs / 261,789 gain-ins (stale 160/583k gone); R14 checks inside the manifest: 28/28 giants verified reach-138.
• THE -O0 PIN (headline): 18 families / 2,192 open members — 2,131 o0b-route-eligible / 61 SC07-no-o0b; 15/18 exemplar-matched. The "WALLED ~1,287" framing dissolves: the bulk never needs the Arm-A splat fix — T2 = generalize the proven rollout_801457a4_o0.py (130/130). Composites flagged: 0x8013c414/0x8013c0f8 are jr+o0 (carve inside the o0b TU — probe ×1 first).
• Zero-crack roster: 114 families / 3,166 open members / ~204k weighted ins (28 substantial); its head IS the -O0 set (T2 and Lane A open together).
• Concentration: top-20 = 24.0% / top-100 = 44.8% of open family weight (head consumed vs S13's 53% — expected post-campaign flattening).
• Report point #1: docs/frontier-p30.md (all numbers same-tree at one HEAD, post-repair — R35).

T0(e) — part 1 ✅: the 21-file absolute-include defect fixed (2026-07-30)

All 21 carriers were jr-carve split files including engine_types.h by absolute path (/home/musashi/bfm-decomp/..., some twice); swapped to the house-relative #include "../shared/engine_types.h". grep -rln /home/musashi src/ include/ → 0. R22 clean-fleet from a genuinely clean tree: 140 passed, 0 failed of 140. Part 2 (autopsy refresh) next.

T0(d) ✅ — backlog ledger: already clean; the scanner deleted (R33) (2026-07-30)

Findings, byte-verified: (1) the still-a-stub filter already existed (load_best drop + prune, built 07-24) and the tracked jsonl was already compacted — the tree's pre-existing uncommitted .run/backlog.jsonl edit was late-S25's un-committed prune output; prune today: 1350 → 1350, dropped 0. (2) The stale rows I'd flagged (func_80151944 etc.) live in worklist.md (regen at T0(f)), not backlog. (3) _open_stubs (private regex) vs corpus.stubs (the oracle): 0 divergence across all 131 ledger binaries — after fixing MY probe, which compared names against int vram addrs (R35 pointed at my own instrument). Hardening: _open_stubs now derives from corpus.stubs (STUB_RE deleted — R33's "best outcome is a deleted scanner"); hex-case canonicalized on both sides of the membership test (a lower-hex record can't silently drop, R32); count parity proven post-change (load_best 1350 == 1350). §83's doctrinal caveat (backlog ≠ work queue; worklist is the queue) STANDS — that's classification quality, not staleness.

T0(c) ✅ — family_hseq↔progress gap: ZERO definitional gap; stamped against recurrence (2026-07-30)

Same-tree regeneration of BOTH digests: family_hseq (overlays) 27,248 instances == progress fleet stubs 28,296 − main 1,034 − resident 14 = 27,248 — EXACT. The carried "29,961 vs 28,296" was a cross-date, cross-scope misread: the 07-29 map snapshot was SESSION-25's open state, and 29,961 − 27,248 = 2,713 = exactly the session's banked total. The tools were never in disagreement — both already derive from corpus.stubs (the Phase 26-A conversion, family_hseq:42). Hardening: the family-hseq digest header now stamps scope ("138 OVERLAYS only") + generation HEAD + the shared-oracle note ("compare digests only at the same HEAD") — a stale or scope-mismatched comparison is now self-announcing. Roadmap v2 D-bucket corrected. Fresh frontier readings from the regen (full pinning still at T0(f)): 478 substantial families / 678,404 templatable ins · 28 zero-crack (was 61 on 07-29 — S25 consumed 33) · 3,626 PURE / 40 IMM / 6 STRUCT · overlays 27,248 instances / 17,629 distinct / 1,587,311 ins.

T0(b) ✅ — rtu_match surfaces the real cc1 error in the verdict (2026-07-30)

_diagnostics() filter (the SESSION-25 recipe, in-tool): drop warning:/In function/At top level/###/blank lines from the failing stage's stderr and print the FIRST 15 survivors — gcc-2.7.2 hard errors carry no error: prefix and the old [-2000:] tail was all warnings. Raw-tail fallback if the filter empties (ICE/signal formats) — never prints less than before. Applied to all four stages (CPP/CC1/MASPSX/AS). Verified on a real deliberate CC1 failure: the verdict now leads with t.c:3116: 'nonexistent_var' undeclared instead of a 2 KB warning tail. --stderr-out unchanged (full dump). Docstring updated.

T0(a) ✅ — gate_stage stage-0 + fix_arity_callers journal undo (2026-07-30)

The carried defect pair closed structurally, not by root-causing:

• gate_stage.py: stage 0 gates the RAW drafts before any transform (GATE_NO_STAGE0 to disable) — the destroyed-good-draft mode (SESSION-22 reproduction: _o0 pair + func_80138C60, ladder-FAILED / bare-VERIFIED) is now impossible by construction; the canon/cast/rc ladder + arity pre-pass run only on stage-0 failures. Root-cause hypothesis recorded in-code (transforms are batch---src-file TU-blind where the gate is per-draft TU-aware) — open, now harmless.
• fix_arity_callers.py: --journal (per-edit literal before/after) + --undo-journal [--keep <banked>] — the exact undo now lives in the WRITER, shared by ladder AND bare workflows (the 17-TU residue class). Replaces gate_stage's two-special-case file snapshot; undo moved to after stage 2 (closes the latent stage-2 parity gap: a stage-2 bank used to lose its arity edit before its own gate attempt). Stale-journal guard (_arity_rc is not None).
• Verification: negative control apply→undo → byte-identical tree; --keep retains exactly the kept edit; in-process flow test .run/t0a_flowtest/driver.py 7/7 PASS (stage-0-first, s1in = failures-only, accumulation, undo-guard, src/ untouched). Real-tree at-scale proof lands with T1's first sweep (R22-bracketed there). Cookbook §122.