BFM-decomp/docs/decision-log.md

# Decision & Pivot Log — the "why" behind BFM's strategic turns

**Purpose.** An append-only record of the project's STRATEGIC pivots, dead-ends, and reversals — the
*judgment* behind major direction changes, captured WHILE FRESH (R31). This is the perishable layer the
other records don't hold:

| Record | Holds |
|---|---|
| `phase-ends/PhaseEnd_*.md` | per-phase build history + terse **Deviations** tables (WHAT changed, briefly) |
| `docs/matching-cookbook.md` | TECHNICAL idioms (asm↔C compiler quirks, tool recipes) |
| **this file** | STRATEGIC why: what we believed → what we tried that failed → why we turned → the hindsight "better path" |

**Why it exists.** It is the substrate for two future deliverables (see the Phase-25 discussion, 2026-07-08):
1. the project **retrospective** — "with hindsight, the best way to have done this"; and
2. the public **"how to AI-decomp a brand-new project"** wiki at the public flip.

The quantitative curve (fleet % over time) is safe in git + the PhaseEnds forever; the *reasoning* —
"we tried X, it was a dead end because Y, so we turned to Z" — is what evaporates between sessions. So it
is logged here as it happens.

**Discipline (R31 — confirmed by Drew 2026-07-08, Phase 25).** **Forward-only.** Do NOT backfill historical pivots from
compressed summaries — that reproduces exactly the low-quality reconstruction R30 exists to prevent (the
PhaseEnds/cookbook already hold what was captured fresh at the time). Log each NEW strategic pivot during
the session that produced it. Route TECHNICAL idioms to the cookbook; this file is for direction/judgment.

**Entry format:**
```
## [date] · Phase N — <pivot title>
- Context / belief: what we were doing and assumed.
- Dead-end (if any): what we tried that didn't work + the byte/measurement evidence.
- Pivot: what we changed to.
- Why: the grounded reason (bytes, measurement, constraint).
- Hindsight / for the wiki: the transferable lesson — what a from-scratch project should do.
```

---

## 2026-07-08 · Phase 25 — a fresh session nearly closed an OPEN phase (handoff-misread)
- **Context / belief:** resuming Phase 25, the `CURRENT_PHASE.md` handoff header read *"finish T7 → Close;
  do NOT start T4 yet."* I concluded the remaining tasks (T4/T5/T6 + the exemplar-cracking that is the
  phase's actual goal) were being **deferred to Phase 26**, and drove toward writing the PhaseEnd.
- **Dead-end:** I reframed the plan's core work as a "Phase-26 backlog" and presented a milestone-close for
  confirmation.
- **Pivot:** Drew corrected — T7 had been *pulled ahead* opportunistically (only its free/cheap MECHANICAL
  wins); *"not yet"* meant **finish those first**, not **defer to next phase**. The 127-draftable-family
  curriculum (T4→T5→T6→T7-cracking, the 6.7 MB byte-weight prize) remains this phase's substance. No PhaseEnd.
- **Why:** I anchored on the literal *"→ Close"* in a handoff header — a phrase written for one moment
  ("don't start T4 in THIS session") that I read as a permanent scope decision. The approved **plan of
  record** (`plan-mode-…-galaxy.md`) plainly lists T4–T7 as this-phase work with an *open-ended* milestone;
  reconciling the handoff against it first would have prevented the error.
- **Hindsight / for the wiki:** **the #1 failure mode of a long multi-session AI project is a fresh session
  misreading a compressed handoff.** What saved it: (a) an approved plan-of-record that outlives any single
  handoff, and (b) a human who caught it in one message. What to bake in: handoff headers must distinguish
  *"not this session"* from *"not this phase"* in unambiguous words; and **a resuming session must reconcile
  the handoff against the plan-of-record before ever concluding a phase is done** — never close on a handoff
  phrase alone. (This is why the plan-of-record and the two-gate model earn their keep.)

## 2026-07-08 · Phase 25 — the "mechanical" family sweep has a hard ceiling at TU-local type collisions
- **Context / belief (T7.2):** the h_norm structural-family remap is *mechanical* — crack one exemplar,
  lift the local types it references into the shared header, and every sibling overlay compiles + banks.
- **Dead-end:** lifting ALL overlay-split-file types into the fleet-wide `engine_types.h` broke the build —
  `typedef Buf` has a DIFFERENT layout in `_a.c` vs `_after.c` (they never clashed as separate `.o` TUs),
  and `_a.c` locally redefines PsyQ SDK names (`MATRIX`/`VECTOR`) that a fleet-wide lift would shadow.
- **Pivot:** lift only the collision-free subset (base + `_after` minus `Buf`); defer the conflicting
  families to per-type reconciliation (Phase 26). Banked 1,729 of the ~3,857 targeted — the rest is genuine
  work, not mechanical.
- **Why:** the overlay split files are independent translation units with independent local type
  namespaces; "share everything" is unsound across that boundary. (Technical detail → cookbook §40a.)
- **Hindsight / for the wiki:** **know the mechanical ceiling before launching a mass pass.** A cheap
  detect-collisions-first probe would have scoped the safe subset up front instead of discovering it via a
  failed build. General lesson: an automated bulk transform needs an explicit *soundness boundary*, and the
  byte-gate (not optimism) is what stops a partial success from masquerading as a full one.

## 2026-07-08 · Phase 25 — the local-7B tier is capacity-bound and off the endgame critical path (T4)
- **Context / belief:** the fine-tuned local drafter (`bfm-match-7b-v3`) was a core cheap tier; retraining **v4**
  on the much larger post-giant-campaign corpus (2,891→3,574 pairs, +994 medium + 597 large functions v3 never
  saw) should extend its band upward and make it a stronger drafter for the T5 wave.
- **Dead-end:** v4 **did not beat v3** — it was marginally WORSE. Gate-true head-to-head on identical held-out
  functions: easy 6-14 ins both 5/5; **medium 18-40 ins** v3's near-misses closer (one at `near-1`, permuter fuel)
  with 1 compile-fail vs v4's 4 — v3 closer on 9/12; **hard 45-85 ins** both 0/10. Crucially v4 scored 0/5 even on
  the 76-83 ins functions it TRAINED on (verified ~1.4-1.7k tok, well inside maxlen 2048 → NOT truncation → genuine
  capacity). (Note: a real corpus-prep flaw exists — functions >85 ins WERE truncated at maxlen 2048 → training on
  cut-off completions, likely the source of v4's slight medium regression — but it doesn't touch the decisive band.)
- **Pivot:** discard v4, **keep v3 (the frozen ceiling)**, and stop investing in the local-7B tier. Not retired
  (still a $0 mop-up for the ≤~15-ins setter/leaf tail), just no longer load-bearing and no more retrains.
- **Why:** the byte-gate A/B settled it directly (G3/P9). "**Corpus quality > size**" landed empirically: v2→v3
  gained from *better* data (the extern-capture fix); v3→v4 was just *more/harder* data and it didn't lift a
  capacity ceiling. Byte-matching's hard part is compiler-codegen REASONING (scales UP with model size), not
  language breadth (which a smaller model could shed) — so neither "more data" nor "a smaller RE-specialist" is the
  lever; the reasoning has to come from a large pretrained base or a frontier model, and the RE-smartness that IS
  small+deterministic already exists as **m2c** (rules, not weights).
- **Hindsight / for the wiki:** **the endgame engine is `frontier-crack → deterministic-propagate → byte-gate`,
  with the permuter softening near-misses — the local small model is a convenience on the small tail, not a
  load-bearing part.** For a *matching* decomp you already own the ground-truth compiler + a perfect verifier, so
  the ML task is candidate-PROPOSAL + search (proposal quality scales with reasoning/size; the check is free). A
  bespoke small "RE model" founders on data scarcity (the asm↔C-under-a-specific-compiler corpus only exists, tiny,
  in decomp git histories). The honest tiering: **m2c** for structure, a **frontier reasoner** for the byte-exact
  precision on the hard/byte-weighty band, the **permuter** for regalloc/schedule search, a **frozen small LoRA**
  only for the cheap ≤15-ins tail. Don't spend GPU-hours chasing band-extension on a 7B; rent a bigger GPU or use
  the frontier tier when the hard band is the target.

## 2026-07-09 · Phase 25 — the GIANT def-side wall is mechanically crackable → build the lever, don't just measure (T5b batch-2)
- **Context / belief:** T5b batch-2 (the 29 giants) was scoped as a pure *measure* wave — draft, `match_one`, map the
  frontier; the plan filed the def-side loose-typing wall as a T7 (post-curriculum) problem, expecting ~0 giant banks.
  Belief going in: giants would mostly near-miss and feed Fable5/permuter; any that isolation-MATCHed would bank via
  the gate's existing `sig_unify`/`cast` transforms.
- **What was tried that failed:** all **16** R14-verified isolation-MATCH giants banked **0/16** through `gate_stage`
  AND through raw `harvest_verify`. Root cause (dug out by placing one and reading the cc1 error, not trusting the
  gate's summary): `conflicting types for func_X` — the drafters wrote Ghidra-*typed* sigs (`void f(u32*, s16*)`) that
  clash with the TU's **canonical** sig, which lives *inside a `DEFINE_func_*` macro* in `engine_core.h` where
  `sig_unify` (a file-scope-extern rewriter) can't see it. First reconcile attempt ALSO failed twice: an `s32/s32-args`
  form conflicted with the engine_core.h `void/void*` canonical; then intermediate cast-locals (`u32 *a0 = (u32*)arg0`)
  *compiled* but produced the WRONG bytes (`70ff4748`) — a fresh pseudo shifted regalloc.
- **Pivot:** stop treating the def-side wall as a future-T7 abstraction and **build the lever now**
  (`tools/canon_sig_reconcile.py`): strip ambient-dup typedefs/externs, rewrite the def to the engine_core.h canonical,
  and **cast each changed param AT ITS USES, never via a local**. That banked `func_8013B274` byte-identical, then
  **5/16** giants total; 3 swept ×134. Batch-2 turned from "measure + backlog" into "prove + partly-automate the
  phase's #1 lever," pulling a chunk of T7 forward on real data.
- **Why (byte/measurement-grounded):** the cast-local vs at-use difference is a *measured* byte fact (`70ff4748` wrong
  vs `d19c9580` right), not a style call — an intermediate local is a new pseudo gcc-2.7.2 may color differently; an
  at-use cast is free. The whole-binary byte-gate stayed the sole arbiter throughout (G3/P9): every one of the 5 banks
  is byte-identical, and R22 clean-fleet is the backstop — it caught my *own* buggy R22 harness (an unexpanded
  `$(OVERLAY_BINARIES)` that only extracted 2 of 136 binaries) before any false "136/136" could be reported.
- **Hindsight / for the wiki:** when a whole *class* of candidates fails the gate identically, **read the raw compiler
  error on ONE placed candidate before concluding "hard / defer"** — the summary ("0 banked") hid a *mechanical*
  declaration conflict behind what looked like an intractable matching wall. The giant tier was never a matching
  problem; it was a **plumbing** problem (the body was already right in isolation). General lesson for
  frontier-crack→propagate→byte-gate: distinguish *codegen* residuals (permuter/Fable5) from *TU-integration* residuals
  (a deterministic reconcile) early — they look identical at the gate ("0 banked") but have completely different levers,
  and the integration ones are cheap ×134 wins hiding as "hard giants." Residual: non-identical ambient types and
  macro-local data symbols are the genuinely-hard remainder (real `reconcile_decls`/rename work), and a reconciled body
  doesn't `family_sweep` cleanly to siblings (per-sibling re-reconcile needed) — both logged for T7.

## 2026-07-09 · Phase 25 — T6: the def-side wall was ~71% tool-shaped; the endgame's frontier shrank from 95 to 33 (Fable5 curriculum session)

- **Context + belief going in (the T5 handoff):** the 95 still-stub family exemplars decomposed as
  "~19 clean-canonical mechanical via canon_sig_reconcile; ~32 harder implicit-int fallback; 31 genuine
  near-misses; 11 hard walls (non-identical types / macro-local data); the 3 `_o0` giants need a special
  -O0 reconcile; the frontier 'match' statuses carry un-verified agent claims (spot-check 3/5); reconciled
  bodies don't sweep (func_8016DC20 = 133 sibling failures)." T6's job was to author the crack curriculum
  from that map, with the expectation that most of the 95 needed per-fn Fable5/permuter work.
- **What was tried that failed / surprised:** (1) R14 re-verification of ALL 95 (not 5) flipped the
  caution — 62/95 genuinely isolation-MATCH; the spot-check pessimism didn't generalize. (2) Probing all
  62 through the REAL TU (splice + full pipeline + masked in-TU byte-compare) showed v1
  canon_sig_reconcile itself was the wall for most: six probe iterations (v1→v3.1) flipped 10→20→37→44
  fns to BANKABLE as five tool defects fell (scalar-typedef strip; preprocessed-TU canonical; block-scope-
  move-not-strip; collision RENAME; decl-line cast protection). (3) Two of my own v2/v2.1 policies
  (ambient-type rewrite of data externs; canonical-text re-emission + blanket casts) BYTE-DRIFTED or
  parse-broke dozens — caught only because every iteration re-probed all 62 (the probe program, not the
  first diagnosis, was the method). (4) A byte-perfect gate run reported MISMATCH because I hand-typed
  `--good-sha` from memory — the check file is the only source of a hash.
- **The pivot:** T7 is now mostly DETERMINISTIC EXECUTION, not discovery: tiers M1 (44 reconcile-banks,
  4,254 ins, 13 giants) → M2 (sweep ×134 via remap + PER-SIBLING re-reconcile — proven 6/6, dissolving
  §41's "sweep fragility") → M3 (6 fns: no-proto rewrite of engine_core.h macro externs — arity conflicts
  with a visible typed prototype are the one thing no draft transform can fix) → M4 (8 fns: §33 TU
  stale-decl retypes) ≈ **+2.2% fleet for ~0 agent tokens**; the true Fable5/permuter frontier is 33 fns
  (31 verified nears + 4 perturbs, minus overlaps) + 2 trivial drafts. The "-O0-specific reconcile" need
  was refuted outright. Deliverables: `docs/phase25-t6-curriculum.md`, `.run/t6_worklist.json`,
  cookbook §41a, canon_sig_reconcile v3.1.
- **Hindsight / better path:** a "wall" verdict is only as good as the TOOL REVISION it was measured
  against — re-derive wall taxonomies after every tool change (three iterations moved 34 fns from
  "wall" to "mechanical"). And validate the gate-proxy against the real gate before trusting either
  direction (the probe caught real walls the drafts hid; the gate caught my SHA typo). The general
  law: when a residual class's members share an error SHAPE (`conflicting types`, `undeclared`,
  parse-at-decl), suspect the pipeline before the compiler.

## 2026-07-10 · Phase 25 — T7 executed the curriculum; the "mechanical tier" was ⅓ probe-over-counted (Opus-Max)

- **Context + belief going in:** the T6 Fable5 curriculum projected 58 fns / ~2.5 MB as MECHANICAL (M1 44
  reconcile-banks, M3 6 no-proto, M4 8 reconcile_decls-retypes), sized from an in-TU OBJECT probe (compile the
  fn in the real TU, mask jal/%hi/%lo, byte-compare). Drew: "run T7."
- **What was tried / surprised:** executing against the WHOLE-BINARY gate (`harvest_verify`) revealed the object
  probe systematically OVER-counts, because it is blind to three things it cannot see: (1) **rodata** — 4 "M1"
  jump-table fns have byte-perfect `.text` but a switch table in rodata diverges (this also REFUTED the T6 "Q3
  -O0 REFUTED" claim — the `_o0` giants are jump-table fns); (2) **link** — 3 "M1" fns are the only asm referencer
  of a scratch data symbol, so C-ifying them drops splat's auto-symbol → `ld undefined`; (3) **in-TU codegen
  perturbation** — all 8 "M4" fns are byte-correct in ISOLATION but drift 8–69 in the real TU (scheduling order,
  `volatile`-loss), and `reconcile_decls` banks 0/8 (4 have no data-decl conflict at all). Plus 4 M3 residue
  (arity/loose-typing). Net: 19 of the 58 "mechanical" were per-fn F-band work.
- **The pivot:** bank the TRULY-mechanical core and re-tier the rest honestly. M1 37 + M3-clean 2 = 39 exemplars,
  reconciled (`canon_sig_reconcile` v3.2 — a type-name-uniquify fix unblocked the struct-collision giants) and
  swept ×134 via `family_sweep --reconcile` (the Q5 per-sibling re-reconcile law: plain remap banks 0, per-sibling
  reconcile banks 94%). **~4,694 fleet fns, fleet 72.29 → 73.66% (+1.37%), R22 136/136, ~0 agent tokens.** The 19
  over-counts → F-band (permuter/§31/Fable5) or specialist workflows (jump-table-in-rodata, manual undefined-syms).
- **Hindsight / better path:** an in-TU OBJECT probe is a necessary FILTER but must never SIZE a "mechanical"
  tier — gate a full sample on the whole binary first, and budget ~⅓ of any object-probe "drift/fail" bucket as
  genuine per-fn work. The mechanical ×134 sweep is the real economic engine (4,655 members from 39 cracks); the
  curriculum's value was concentrating the crack effort onto the 39 exemplars, not the tier-size projection.
  Reinforces R14 (verify vs bytes) at the tier-classification level and cookbook §41b/§41b-addendum.

## 2026-07-10c — Propagation-recovery (task b) is NOT a simple --edit-remap gap; it's a family_remap LIMITATION on lever-heavy F-band cracks

**Context+belief:** After waves 3/4 banked exemplars but dropped ~1,200 siblings, I projected task (b) as a cheap
`family_sweep --edit-remap` enhancement (carry the exemplar's //@EDIT file-scope edits per sibling) — a ~0-token bulk win.

**What failed / the diagnosis:** Probing the 7 dropper families showed the drops are HETEROGENEOUS and dominated by
**`family_remap` symbol-pairing FAILURE ("133 remap-fail")**, NOT the decl/byte layer an --edit-remap would fix.
`family_remap` was built for MECHANICAL (reloc-only, T3) families; the F-band cracks carry heavy levers (register
`__asm__` pins, density dead-reads, phantom-frame `frame_pad`, pointer casts, //@EDIT flips), whose disassembly
doesn't positionally pair to the sibling image cleanly → remap aborts. func_8014FE60 (engine_core void→s32) also
remap-fails even after the correct global flip. The families that DID sweep (func_80166690/8017B238/80131B14/
8016CF04/8014FBC0) have remap-clean bodies; the droppers don't.

**The pivot:** recovering the ~1,200 dropped siblings needs EITHER (a) a `family_remap` upgrade that pairs
lever-heavy bodies (structure-aware, not positional-reloc-only), OR (b) a per-sibling re-crack (m2c+reconcile+rtu_match
fan-out ×133/family — expensive). Both are a focused follow-up, not a quick enhancement. Backlogged.

**Hindsight better-path:** the ×134 economics assumed family_remap propagates any crack; it only propagates
remap-clean ones. Future crack-then-sweep waves should CHECK remap-ability of the exemplar body BEFORE counting the
×134 (a cheap `family_remap --dry` per exemplar), so the frontier map's leverage estimate reflects propagate-able
families, not all same-address families.

## 2026-07-10c (CORRECTION, same day) — the "family_remap limitation" was a MISDIAGNOSIS; the real bug was canon_sig_reconcile's def-finder (R14)

**Correcting the entry above.** I concluded the propagation drops were a `family_remap` limitation because `family_sweep`
reported "133 remap-fail". **That label was misleading.** Running `family_remap` directly on all 7 droppers SUCCEEDED
(it paired 4–6 symbols each). The None that `reconcile_remap` returns — which `family_sweep` counts as "remap-fail" —
actually came from **`canon_sig_reconcile.reconcile` raising "no definition of func_X found in draft"**: its def-finder
regex required a leading `\n` (`r'\n(<type> fn(...)){'`), but stripping the `//@EDIT` lines left the fn definition on
LINE 1 of the raw draft (no leading newline) → not found. The swept-clean families happened to have a leading `// @class`
comment, so their def had a `\n` before it.

**FIX (1 char, low-risk):** def-finder regex `\n` → `(?:^|\n)` (also match a def at the draft start; only ADDS matches).
**Result:** func_8014FE60 fully recovered — 133/133 siblings banked (fix + its engine_core.h void→s32 global flip).

**Residual (the GENUINE --edit-remap work):** 4 families (func_8016DF5C/80136334/8013D9B0/80156044) now RECONCILE but
BYTE-DRIFT per sibling — their crack levers (s32↔void return flip, array-decay pointer `//@EDIT`, no-proto `//@EDIT`)
aren't carried/re-derived per sibling. Recovery = carry the exemplar's `//@EDIT` per sibling (symbol-remapped) + apply
the return-type flip to the shared engine_core.h decl once. Still a focused follow-up, but SMALL and well-understood
now — NOT a family_remap rewrite.

**LESSON (R14):** a tool's failure LABEL can misattribute the failing STAGE. "remap-fail" was actually a
reconcile-def-finder throw. Trace the real exception (`reconcile_remap` swallows it) before concluding a limitation.

## 2026-07-11 · Phase 25 — task B: `--edit-remap` BUILT, but 4/6 byte-drift families are cc1-crash-walled (~266, not ~800)

**Context + belief (from the 2026-07-10c handoff):** the 6 byte-drift `//@EDIT` families were framed as "SMALL and
well-understood — recover ~800 fns by carrying the exemplar's `//@EDIT` per sibling + a once-global engine_core.h
flip." Drew locked B first on that basis (scoped, mechanical). I built `family_sweep --edit-remap MANIFEST` to do
exactly that.

**What the byte-gate revealed (probe-before-invest, R14):** the families are NOT one bucket. Only the **2 array-decay
pointer-flip** families (`extern s32 D_x[];`→`extern s16 *D_x;`) recover — `func_80136824` + `func_80136334` banked
**266/266 siblings byte-identical (0 failed), full ×134**. The other **4 are register-pin-heavy** (`func_80133AB0`'s
exotic `register int zr __asm__("$0")`; `func_8016DF5C`/`func_8013D9B0`'s GTE 20-pin bodies; `func_80156044`'s
inline-asm trampoline) and **cc1-2.7.2 SIGABRTs (`make` Error 134) compiling the SIBLING TU** — the identical body
compiles fine in ov077. Universal (func_80133AB0 crashed 3/3 siblings tested). The hand pins are ov077-TU-context-
specific: cc1's fixed-table 1996 register allocator aborts on the pin pattern in a different overlay's surrounding
function set. func_80156044's engine_core.h `int`→`void` flip IS byte-neutral (verified) — the wall is its body, not
the edit.

**The pivot:** ship the 2 tractable families (266 ×134, R22 136/136 green, fleet 74.40→74.48%), backlog the 4 crashers
as exemplar-only (×1) / per-sibling permuter-Fable5 fuel, and move to A (the 7 giants — all remap-clean 133/133, ~938
fns high-byte-weight, the real ROI). The `--edit-remap` tool is reusable for future array-decay-class cracks; its yield
must be sized by that subset, never by "family has an //@EDIT."

**Better path (hindsight):** the "~800" estimate counted `sibs × families` without asking "does the CRACK compile in a
sibling TU?". A hand crack that banked in ov077 by exotic register pins does not generalize — a 30-second single-sibling
`make build` probe per family would have sized B honestly up front. **LESSON (R14):** an exemplar match proves the crack
in ITS TU only; the ×134 claim needs a sibling-TU compile probe, because pins are TU-context-specific and cc1 *crashes*
(not just drifts) on the ones that don't transfer. Corollary: rtu_match/match_one are blind here — their neutralized/
isolation compiles crash too (harness artifact); only the real `make build` is the arbiter.

## 2026-07-11 · Phase 26 — the "reach-1 tail" is largely a reloc-tracker blind spot, not unique code (Task 1)

**Context + belief (from the Phase-25 close):** the h_seq reframe had already shown the "36k unique tail" collapses
~90% into per-location families. The open question entering Phase 26 was HOW the families differ — the megaplan framed
immediate-substitution as the central new problem (families "differ in immediates, so are NOT free dedup").

**What the design pressure-test found (byte-verified before any scaling — R14):** the dominant difference is NOT
immediates — it is a **tracker blind spot**. `norm_stream`/`reloc_targets` dropped the lui-hi on every R-type write,
but gcc-2.7.2's indexed-global idiom `lui;addu $idx;lw %lo($at)` preserves it. So `D[i]`-indexing functions were
*mis-normalized per overlay* → they inflated the "h_norm reach-1 tail," and `family_remap` silently dropped their
indexed `D_` symbols → those families couldn't bank even though they are pure per-location templates. On the
substantial tail the classification is **PURE-same-addr 62 fams / 1.55M ins · PURE-cross-addr 103 / 0.10M · genuine
IMM only 8 / 0.10M** — i.e. ~95% of the byte-weight is reloc-only, fixable by a ≤15-LOC tracker change, and the
immediate engine shrinks to an escalation tier for ~8 families. A second latent bug surfaced alongside: `remap`'s
sequential substitution corrupts chained/permuted maps (harmless on h_norm, breaks the imm engine).

**The pivot:** front-load the tracker fix (Task 1) as the load-bearing change, demote the immediate engine to a
diff-driven 3-tier escalation (Task 3), and add a **free validation corpus** — 63 families / 0.31M ins already have a
MATCHED exemplar and only failed earlier sweeps from this bug → they bank with zero cracking the moment the fix lands
(Task 5 V2), simultaneously measuring the real template success rate before any Fable5 spend.

**Better path (hindsight):** the tracker's own design note already said "conservative: can miss a match, never forge
one" — but a *missed* reloc in a REMAP tool isn't harmless the way a missed h_norm match is; it silently produces a
wrong-but-compiling sibling body that only the byte-gate catches. When a normalization/remap tool is REUSED for code
generation (not just clustering), its conservative-miss becomes a correctness bug. **LESSON (R14):** before treating a
"unique/unmatchable" population as intrinsic, re-run the *grouping and the remap* under a corrected fingerprint —
here the "reach-1 tail" and the "unremappable family" were the SAME artifact of one dropped register-tracking case.

## 2026-07-11 · Phase 26 — Task 5 GO/NO-GO: the h_seq engine is 100% correct on clean families; the substantial matched band is type-dominated (a reconcile follow-on, not a machinery gap)

**Context + belief:** Task 5 was the validate-before-scaling gate — run the whole-binary byte-gate on the
matched-exemplar families and MEASURE the real template success rate before any Fable5 spend. Expectation:
a high mechanical bank rate on the "free-win" tracker-miss corpus.

**What the byte-gate revealed (R14):** the substantial matched-exemplar band banked **532/1507 non-pinned
members (~35%)** — but that aggregate is misleading, and the per-family breakdown is the real signal:
- **3 families banked 100% ×133** = 399 byte-perfect members — exactly the 3 tracker-miss PURE families
  (`0x8015d5e8`/`0x8015f118`/`0x801407f4`) the design pass named. The addu-hi tracker fix + the extern-carry
  fix (below) make these template flawlessly. **The machinery is 100% correct on clean families.**
- **9 of 13 clean families ZERO-banked** — all reference overlay-local custom struct types (`Work8016`,
  `Work8017`, `SV4_8017B368`, `Prim`, `E4`) → the templated body is `undeclared-type` in the sibling TU =
  the §41 def-side / type-lift RECONCILE class, which the existing `family_sweep --reconcile` /
  `build_engine_types` path already handles. NOT a machinery failure — a known follow-on (Task 8).
- 16 of 29 substantial matched families were pinned-exemplar (×1-only hard-reg cracks) → skipped by the
  new static pin guard → Task-7 pin-free re-crack.

**The load-bearing tooling fix (Task-5's "measure then fix", R16):** the first gate run banked only ~34%
because `extract_unit` grabs only *immediately-preceding* externs — a per-location body that indexes a
global (`(*D_x[i])()`) references symbols declared once at file scope elsewhere in the exemplar TU; templated
into a sibling TU that never declared them, they are `undeclared` at the gate. Added `remap_hseq.gather_externs`
(carry the file-scope externs for every body-referenced symbol, remapped). `func_8015F118` went from gate-fail
to BYTE-IDENTICAL; the 3 clean families then banked 133/133 each.

**The pivot / verdict: GO.** The zero-crack h_seq machinery (tracker + imm + cross-address + extern-carry) is
byte-proven correct. Scale it (Task 8: the mid/tiny bands + the reconcile pass for the type families). The
Fable5 window (Task 6/7) goes to the big PURE cores that have NO matched exemplar (890/562/536… — they need a
crack before they can template).

**Better path (hindsight):** the raw aggregate rate (35%) nearly read as "the engine is weak"; the per-FAMILY
breakdown showed it is "the engine is perfect on the families it targets; the rest are a different, already-
solved problem." **LESSON:** when measuring a mechanical harvest, stratify by family/class before judging the
rate — an aggregate mixes 100%-clean, 0%-type-blocked, and skipped-pinned populations that demand different
follow-ons.

## 2026-07-11 · Phase 26 — Task-8 pipeline-validation slice (pre-Fable5-window de-risk): reconcile→bank works; templating reconcile-class cracks needs per-sibling re-reconcile in --hseq

**Context + decision (Drew):** before spending the closing Fable5 window (Task 7), validate that the
reconcile→gate→template pipeline actually banks an isolation-crack end-to-end — else the window's output
(same isolation-crack format) could pile up un-bankable. Ran optimal-order step 1 only ("bank the wins,
pause before Task 8").

**What the byte-gate revealed (R14), in two halves:**
1. **Reconcile→bank WORKS.** The 23 triage closeness-0 cracks gate **0/23 raw** (they carry standalone
   `struct Obj`/scalar typedefs + Ghidra sigs → §41 def-side wall). Run through `canon_sig_reconcile` v3.2
   (strip ambient dups, canonicalize the sig, cast callees at use) they bank **4/15** into ov077
   (`func_801506A4`/`func_8016A73C`/`func_80167540`/`func_80155800`, byte-identical). The 11 residual fails
   are a data-extern-typing gap (e.g. `conflicting types for D_801891B8`, a fn-ptr array the seed types
   differently than the TU) the reconcile's pt-9 data-extern handling doesn't fully cover for these seeds.
2. **Templating a RECONCILED body ×133 FAILS (0/4).** The reconciled ov077 body is TU-SPECIFIC — its
   canonical-sig casts + `Name_<addr>` collision-renames fit ov077, not the sibling TUs (each has its own
   ambient types/sigs). Plain `remap_hseq` copies the ov077-reconciled body → re-hits the def-side wall in
   every sibling. This is the decision-log 2026-07-11 lesson again: an exemplar match proves the crack in
   ITS TU only; ×134 needs per-sibling work.

**The implication (the point of validating first):** the PURE tracker-miss families template cleanly via
plain `--hseq` (Task 5: 399 banked). But the **type-using families — the triage cracks AND the 61 Fable5
cores — are reconcile-class**: their cracks bank as ov077 exemplars but need **per-sibling re-reconcile**
to template ×134. That machinery EXISTS for h_norm (`family_sweep --reconcile` / `reconcile_remap`, the
Phase-25 M2 4,389-bank path) — it just needs porting into the `--hseq` path (over `remap_hseq`, i.e. with
cross-address + imm). **So the Task-8 prerequisite before the Fable5 window is productive: wire per-sibling
reconcile into `hseq_sweep`.** Otherwise Fable5 output stalls at ×1 (ov077-only).

**Outcome:** kept the 4 real ov077 exemplar banks (byte-verified). Paused before building the per-sibling
reconcile wiring (that IS Task 8, per Drew). **LESSON:** the validation slice paid for itself — it converted
"the pipeline works, go spend the window" into "reconcile→bank works, but templating reconcile-class needs
one more wiring step first," a decision that would have been very expensive to learn after the window closed.

## 2026-07-12 · Phase 26 — the crack-harvest has TWO tooling gaps + the rtu_match-vs-whole-binary lesson (Fable5 batch-1 processing)

**Context:** processing the Fable5 batch-1 cracks + the 23 triage isolation-cracks through the whole-binary
gate revealed the "closeness-0 / rtu_match-MATCH" counts were optimistic. Whole-binary reality (G3/P9):

**1. The rtu_match blind spot on jump-table functions.** Both Fable5 cracks (`func_80159C84`, `func_8015444C`)
rtu_match-MATCH but FAIL the whole-binary gate. rtu_match neutralizes `INCLUDE_ASM` (excluding the §8
jump-table rodata `.s`) and compares only the masked INSTRUCTION stream — it never verifies the jtbl rodata
data bytes. The code is right; the jtbl rodata isn't confirmed. **rtu_match is NOT a sufficient sole arbiter
for jr-functions** — the whole-binary gate is (as always, G3). Pattern is clean: all 6 whole-binary-banked
cracks are no-jtbl; every jtbl crack (2 Fable5 + the 2 jtbl triage cracks) fails.

**2. Two distinct harvest gaps, both fixable Task-8 tooling:**
   - **§8 jtbl-rodata gap:** replacing an `INCLUDE_ASM` jr-function with C needs the compiler-generated jtbl to
     byte-match + land in the right rodata slot (the §8 dotted-`.rodata`-subseg + ld_interleave). The overlay
     splits don't have this per-cracked-jr-function setup → every jtbl crack fails the binary. **This blocks the
     jtbl-heavy Fable5 window** (most top cores are jr giants).
   - **reconcile data-extern gap:** ~15 of the 21 no-jtbl triage cracks fail canon_sig_reconcile on a
     `conflicting types for D_x` (fn-ptr-array / typed-global the seed declares differently than the TU) — pt-9
     data-extern handling is incomplete for these. Only 6 no-jtbl reconcile-clean cracks bank (729 members:
     463 committed + 266).

**The implication for the Fable5 window:** cracking a jtbl giant with an rtu_match-only agent produces an
UNVERIFIED result — the §8 rodata must be handled + the whole-binary gate must be the arbiter. So the window
is only productive on jtbl cores AFTER the §8-overlay-jtbl tooling exists (or with serial whole-binary
verification). **Better path (hindsight):** the Fable5 crack prompt should have required the whole-binary gate
(or an rtu_match variant that includes the jtbl rodata) as the bar, not plain rtu_match — for jr-functions the
two diverge. LESSON: an indicator that MASKS a byte-region (rtu_match masks relocs + excludes neutralized
INCLUDE_ASM rodata) cannot arbitrate a match whose difference lives in that region.

## 2026-07-12 · Phase 26 — §8 unblocked the HEAVIEST byte-weight lever (switch functions), reframing the endgame priority
- **Context / belief:** built the §8 overlay jtbl-rodata tooling + the ×134 automation (jr-functions can
  now bank as C, proven func_8012ACE0 ×133), then recommended continuing with "the 45 small jr families"
  as the next mechanical lever.
- **Dead-end (the off-plan recommendation, Drew caught it):** that rec optimized for *mechanically easy*
  (small families template cleanly) instead of the endgame's actual objective — **heaviest byte-weight
  first**. Measured: the 45 small jr families = **129,028 templatable ins** (trivial).
- **Pivot:** re-target the heavy tier. The frontier byte-weight (the instr-weighted metric lever):
  **jr (switch) substantial families = 191 fams / 5,534,884 ins**; non-jr substantial = 1,168 / 7,328,348;
  all substantial = 12.86M. And decisively — **9 of the 10 heaviest unmatched family cores are switch
  (jr) functions** (func_80178D40 890×134 = 477K ins alone; func_8017BEBC 952×113; func_8015AE2C 562×134…).
- **Why:** switch functions were UNBANKABLE before §8 (the jtbl-rodata duplicate). §8 didn't just enable a
  small mop-up — it **unlocked the single heaviest chunk of the remaining game** (the byte-weight is
  dominated by big switch cores). The endgame plan (heaviest-byte-weight-first via crack-core → template
  ×134) is intact; §8 was its key enabler, and the correct follow-through is the HEAVY jr cores, not the
  light tail. This needs Task 7 (Fable5) un-paused: Fable5 cracks the giant switch core → §8 + the ×134
  automation bank it fleet-wide → the now-10×-faster R22 verifies.
- **Hindsight / for the wiki:** when a build-mechanism unblocks a whole *class*, re-rank the endgame by
  the class's byte-weight, not by which member is easiest to bank next. "Easy and on-metric" ≠ "easy";
  the plan's objective (heaviest byte-weight) must gate the next-target choice, or you grind the light tail
  of a newly-opened heavy vein. (Drew's steer: "the endgame plan is 1st smartest play to unlock the
  heaviest byte-weighted remainder — does your recommendation follow it?" — it did not; this corrects it.)
- **Sequencing refinement (Drew, same day):** do the 45 SMALL jr families first as a **de-risk
  preamble** (NOT for byte-weight — ~+1% instr), THEN the heavy 191. The decisive reason isn't
  size: `jtbl_carve` only built the single-jtbl carve, but func_8012ACE0 is now matched in all 133
  siblings, so family #2 forces the **multi-jtbl address-ordered carve** — build & prove THAT on
  cheap 30-ins targets before a Fable5-cracked 890-ins core depends on it. Also needs no Fable5
  (progress without burning its limits). **Guardrail:** the small tier is a MEANS (harden the
  pipeline), not the objective — pivot to the heavy 191 once multi-jtbl is proven; don't grind the
  light tail because it "feels productive." Wiki lesson: when a newly-built mechanism has an
  un-built sub-case that the expensive targets will hit, force that sub-case out on the cheap
  targets first — de-risking and building-the-missing-piece are the same move.

### 2026-07-13 — the jr-core ISOLATION wall: mechanical TU-splitting breaks gcc-2.7.2's lenient scoping

- **Context + belief:** Stage 2 of the multi-jtbl campaign (heavy jr cores → template ×134) needs each
  matched jr-function ALONE in its own code subseg so its jtbl carves without a same-subseg collision.
  Drew's steer: build the **scalable "isolate-ALL-jr-per-sibling" upfront resegment** (one-shot multi-cut
  per overlay) so every Stage-2 core bank is a trivial fill during the closing Fable5 window. Belief going
  in: this is mechanical source-splitting — partition the overlay `.c` at jr boundaries, repoint config +
  carves, rebuild byte-identical.
- **What was built + PROVEN:** `tools/overlay_src_split.py` — an overlay-`.c`-aware partition (header =
  includes + Phase-17 canonical-sig layer; each addressed item = its preamble + body; robust
  definition/declaration/K&R/`DEFINE_func`/`SETTER`/`RETCONST` classification). **Fleet-validated 404/404
  overlay `.c`, 341,902 items, round-trip exact / 0 unresolved / 0 non-monotonic.** `tools/jr_isolate_all.py`
  — multi-cut resegment (config split at jr boundaries, source repartition + INCLUDE_ASM path repoint,
  banked-jr carve repoint, -O0-object skip). **SINGLE-cut isolation byte-identical** (isolate func_8013FFD8
  in the simple `main` object → clean `make build` = `d19c9580`, R22).
- **What FAILED (byte-verified):** the FULL 54-jr isolation on ov_SC01_077 hits a **long tail of C-scoping
  edge cases**, culminating in the decisive one: **`func_801734BC` uses `D_80126B3E` with no local decl;
  `D_80126B3E` is declared `extern s16` ONLY inside `DEFINE_func` macros in `engine_core.h`.** The original
  `_after.c` compiles because **gcc-2.7.2 lets a block-scope `extern` (from an earlier `DEFINE_func` macro
  expansion) persist to file scope for the rest of the TU** — splitting `_after` separates the core from the
  earlier macro that declares the symbol → `undeclared`. Earlier tail members (all fixed incrementally, in
  order): block-scope externs must not be hoisted (per-fn type shadows — `D_80115118` is `unsigned short`
  in most funcs but the struct `S115118` in one); file-scope decl ORDERING across a cut (`D_80115110` used
  above its in-region decl); **ambient decl context** (a region needs the file-scope decls that lived in
  earlier regions of the object — solved: prepend, original order, shadow-safe because a file-scope-declared
  symbol can't carry a *different*-typed block shadow or the original wouldn't compile); file-local-typed
  externs (`extern Vec8 D_…;`) can't hoist above their typedef.
- **The why (root):** these overlay TUs are hand-matched against a compiler that treats a block-scope
  `extern` as declaring the symbol for the WHOLE TU. Mechanical splitting into per-jr TUs breaks that
  invisible cross-function dependency, and the dependency is carried through **`DEFINE_func`/`SETTER` macro
  expansions in `engine_core.h`**, not just visible col-0 decls — so no amount of *col-0* ambient-carry
  fixes it.
- **The candidate fix (not yet built):** **declaration-completion** — build a global symbol→type map from
  `engine_core.h`'s macro `extern`s + all overlay col-0 decls, and for each region emit a file-scope
  `extern <type> <sym>;` for every `D_`/`func_` symbol the region USES, EXCLUDING type-inconsistent symbols
  (the `D_80115118` shadow set, kept block-scope in bodies). This makes every region self-contained
  regardless of where the original declared the symbol. Est. ~40–60 LOC on top of the proven parser; the
  whole-binary byte-gate arbitrates. **Owner decision pending (Drew): invest in declaration-completion vs
  a different Stage-2 approach** — surfaced this session before sinking more time (P5a: repeated failures,
  distinct root cause each).
- **Hindsight / for the wiki:** "mechanical source split" of matching-decomp overlay code is NOT mechanical
  — the C is written against a specific compiler's lenient scoping (block-scope-extern TU persistence,
  macro-injected decls, per-function type shadows). Splitting a TU means REBUILDING each fragment's full
  declaration environment from a global symbol map, not relocating text. The parser (structure) was the
  easy 20%; the declaration environment (semantics) is the 80%. Prove the mechanism on the SIMPLE object
  first (it passed) but budget for the dense object's scoping tail before committing to upfront-×134.

---

## 2026-07-13 (session 6) — the §8b scoping wall RESOLVED: rebuild the decl environment, don't map symbols

**Context + belief going in.** Session 5 hit a wall isolating jr cores: the full 54-jr split of `ov_SC01_077`
failed with `D_80126B3E undeclared`, and I logged the cause as **"gcc-2.7.2 block-scope-extern TU-persistence"**
— i.e. a non-conformant compiler quirk where an `extern` inside one function body leaks to file scope for the
rest of the TU. The proposed fix (Drew-approved) was **declaration-completion**: build a global symbol→type map
and emit a file-scope `extern` for every symbol a region *uses*, minus a heuristic "type-shadowed set".

**What was actually wrong (R14 — the hypothesis was incorrect).** There is no gcc quirk. `DEFINE_func_80173460()`
expands **at file scope** to `extern void func_801734BC(...); extern struct S80126B38 D_80126B38; extern s16
D_80126B3E; void func_80173460(...) { … }`. Those externs are *genuinely file-scope* — they are merely
**textually invisible in the `.c`**, because they live in `engine_core.h`. Any col-0 scan of the source can
never see them. The wall was a blind spot in our own tooling, not a compiler eccentricity.

**The pivot — and why the approved design was the wrong one.** Chasing "declare every used symbol from a global
type map" would have been actively harmful. The engine is loosely typed: `func_80173544` is *defined* at file
scope as `s32 f(void *)` while `func_801734BC`'s body declares `extern void f(void);` — contradictory, and legal
only because the block-scope decl never meets the definition. Hoisting "every used symbol" lifts that shadow to
file scope, **creating** a conflict that then needs the heuristic shadow-set to dodge. Instead I **reconstructed
the original TU's file-scope declaration environment and carried it strictly forward**. That is conflict-free
*by construction*: every carried decl already coexisted with every definition in the one original TU, and decl
compatibility is order-symmetric. Shadows stay inside bodies and travel with them. No heuristic, no shadow set.

**What the bytes taught (found by gating, not by reasoning).** Three decl sources were lost, not one — and I only
found #2 and #3 because the byte-gate kept failing with a *new* error class each time:
1. `DEFINE_func_*` macro leading externs (3,929 lines / 1,462 symbols) → `D_80126B3E undeclared`.
2. **A definition is itself a declaration** for everything below it in its TU → `func_8012B2CC undeclared`.
3. File-local typedefs used by a carried prototype → `parse error before '*'` (`Vec3s`).

**Result.** Full 54-jr isolate-all on `ov_SC01_077` → `d19c9580` byte-identical, **R22 clean-fleet 136/136**.
Two latent bugs fell out and were fixed: `func_subseg` derived the owning subseg from the *asm tree*, which
`make extract` never prunes — so after an isolation it returned the STALE owner and silently re-created the
collision the isolation had just removed (now derived from the config); and the sweep's revert **deleted** the
shared `overlays.mk` carve var unconditionally, which would have destroyed a *committed* carve (all 134 overlays
have one) on any failed sibling (now restored to its committed value).

**Upfront vs lazy (new information for the owner).** Drew chose lazy isolation when isolate-all was *failing*,
to avoid ~7,200 region files. Isolate-all is now byte-proven at 136/136, so upfront is available — but lazy is
strictly cheaper (pay only for cores we bank) and is what shipped: `jtbl_family_bank` catches `jtbl_carve`'s
`NON-CONTIGUOUS` fail-loud → isolate that one core → re-carve. Proven on `func_80178D40` (890×134, the heaviest
core): blocked → isolated (byte-neutral) → carve lands in its own subseg. **The heavy-jr harvest is unblocked.**

**Hindsight / for the wiki.** Two lessons. (1) *A wall's stated root cause is a hypothesis until the bytes
confirm it* — I recorded a compiler quirk that did not exist, and the "fix" it implied would have introduced
real conflicts. Re-derive the mechanism before building on it. (2) *Splitting a translation unit is a semantic
operation, not a textual one.* The parser (structure) was the easy 20%; the declaration environment (semantics)
was the 80% — and the correct move is to **reproduce the environment the original had**, never to invent a new
one from a global map. Faithful-forward-carry needs no heuristics; "declare everything used" needs a growing
pile of them.

## 2026-07-13 (session 8) — the ×133 sweep blocker was OUR tool, not the compiler: the R17 triage rule, applied

**Context / prior belief.** Session 7 banked `func_8015AE2C` (562 ins, reach 134) ×1 but its ×133 sibling sweep
failed on `conflicting types for D_801812A4`, and the checkpoint diagnosed it as `reconcile_decls` resolving
against a *fleet-majority* canonical oracle instead of the type the TU can actually see. Drew had just asked the
routing question and we had committed the rule: **"wrong BYTES" → read the gcc source (R17); "won't COMPILE" →
read our Python.** This was the first real test of that rule, and it held — but the diagnosis underneath it was
only half right, and the half that was wrong is the interesting part.

**What the bytes taught.** Reproducing one sibling by hand (rather than trusting the handoff — R14) produced a
much sharper picture than the checkpoint's:

1. The **isolated region compiles and builds `[ OK ]` *without* the body.** So §8b isolation was never implicated.
   The conflict is introduced *entirely* by the templated body.
2. `D_801812A4` was the **only** hard error in the whole build. All 27 carried *function* externs were fine raw —
   `cast_call_sites` was not needed at all. (The checkpoint's "cast_call_sites already fixes the function half"
   was true but irrelevant; it also implied ~4 data symbols needed reconciling. Eight were demoted; none needed
   a type reconcile.)
3. The real mechanism is an **ordering asymmetry**, both halves byte-proven:
   `BLOCK(int) → BLOCK(struct*) → FILE(void*)` builds; `FILE(void*) → BLOCK(int)` is a hard error.
   `family_remap.gather_externs` prepends carried decls at **file scope**. For a per-location symbol the sibling
   declares only at *block* scope inside its own later functions, that carried decl **establishes a global
   declaration the TU never had** — and every later block-scope `extern` of it must now agree. In loosely-typed
   engine code they never do. `D_801812A4` is one fn-ptr dispatch table declared **four incompatible ways** in a
   single region and the TU is perfectly happy — until we add a fifth decl *at the top*.
4. `reconcile_decls` was the wrong instrument **twice**: its oracle answers "what does the fleet call this
   symbol" when the question is "what can *this TU* see" — and its `DATA_DECL_LINE_RE` **cannot parse the
   fn-ptr-array form** `extern void (*D_x[])(void *);` at all, so it silently skipped precisely the symbols that
   were failing. (This is the same "reconcile fn-ptr-extern gap" logged on 2026-07-12; it had been filed as a
   *separate, smaller* lever and was in fact the blocker itself.)

**The pivot.** Don't teach `reconcile_decls` a TU-visible oracle (the checkpoint's plan, and a much bigger,
riskier change to a proven path). Instead **don't change the TU's decl environment in the first place**:
`tools/scope_data_externs.py` demotes a carried `D_` extern to **block scope inside the function body** whenever
the TU has no file-scope decl of it above the insertion point. It then declares no global, nothing below can
conflict, and the environment is preserved exactly. Byte-neutral (an `extern` emits no code; the declared type
and every access opcode are unchanged), and *strictly never worse than raw*, so it needs no type comparator, no
fn-ptr parser, and no oracle. It also **restores fidelity** — the original source declares these symbols at
block scope in exactly this way. Wired as the `scoped` stage (raw → scoped → recovered → reconciled).

**Result.** First sibling byte-identical on the first try; the 133-sibling sweep run to completion.

**Hindsight / for the wiki.** Three lessons. (1) **The R17 triage rule paid for itself immediately.** The
temptation with a `conflicting types` failure on a 1997 compiler is to assume the compiler is being exotic. It
was not — gcc was correctly rejecting plain C89, and every minute spent in `cse.c` would have been wasted. *Ask
which half of the compiler is complaining: the front end (our C is invalid → our bug) or the back end (our C is
valid but the bytes differ → read the source).* (2) **A tool that no-ops on the failing input looks exactly like
a tool that has nothing to fix.** `reconcile_decls` reported success while skipping the only symbol that
mattered, because its regex couldn't see fn-ptr arrays — a silent-skip class we have now been bitten by three
times (`find_site` braces, `overlay_files` splits, this). Prefer transforms that *fail loud on unparsed input*.
(3) **The cheapest fix was to do less, not more.** The instinct was to make our reconciler smarter (a TU-visible
oracle, a fn-ptr type comparator, a cast-at-use taxonomy). The correct move was to stop perturbing something we
had no business perturbing. When a transform breaks a TU, first ask what it is *changing* that it needn't.

## 2026-07-13 (session 8, Fable5 Max) — func_8017BEBC closed: the allocno-tie class is a DIAL, not a wall

**Context / prior belief.** The 952-ins jr core (reach ×113, the largest unmatched function in the game) sat at
close=2 — two transposed preheader `addiu`s. The session-7 Fable5 agent had localized it to `global.c`'s
allocno-priority tie and prescribed a §45-B gdb-on-cc1 read of `allocno_live_length`; the permuter had run 25
minutes without closing it. The residual class: allocation order and emission order are COUPLED (both follow
creation order), but the target needs them to DIFFER — the shipped draft could have either correct, never both.

**What the bytes taught.** The dumps alone settled it — gdb was never needed. `.lreg` gave the two pseudos'
ground truth: refs 13/13, live lengths 783/782 → `pri = int(390000/L)` = 498/498, an EXACT int-truncation tie
(the agent's remembered "270000/L" had the wrong refs count — reading beats recalling, R14). The quantization
boundary sat one insn away: +1 on both lengths → 497 vs 498. And the split direction is FORCED: the later-created
pseudo always has the shorter live range, so a split always hands it the earlier allocation — precisely the
"allocation ≠ creation" the target requires.

**The pivot.** Rather than hunting an L-shifter that survives cse (the agent's proposed hunt), the map's own
zero-byte-asm toolkit already contained the dial: `__asm__ volatile ("")` placed BETWEEN two existing GTE
volatile asms adds no new cse/sched barrier (one is already there) — it is purely +1 static insn at
global-alloc time, zero bytes emitted. Natural operand order restored (emission correct), one slider inserted →
MATCH 952/952 first try. Whole-binary gate BYTE-IDENTICAL (jr function — the §8a trap respected); one TU-visible
decl reconcile en route (`D_800B9A02`, §8d sub-class b). Banked ×1; the ×113 sweep is IMM-class Task-8 work.

**Hindsight / for the wiki.** (1) *An "irreducible" tie is often a measurable quantization accident* — the
formula is public, the dumps print its inputs, and the fix is one insn of live-range arithmetic. Before
declaring a register-order residual intrinsic, READ THE PRIORITY NUMBERS. (2) *The dumps-first discipline
scales:* .lreg/.greg gave everything gdb would have, at a fraction of the setup. gdb remains the tool for
DYNAMIC questions (which reg find_reg actually grants when hand-modeling stalls), not for static quantities the
dumps already print. (3) *The zero-byte toolkit compounds:* the slider now joins the density dial and the
lifetime-extender as the third allocation dial that emits nothing — and the "adjacent to an existing volatile
asm" placement rule makes it safe in GTE-heavy renderers, which is exactly where the remaining jr cores live.

## 2026-07-14 (session 8) — the silent-skip class: promote the lesson from a rule to a MECHANISM (Drew approved)

**Context / prior belief.** Six silent-skip bugs surfaced in one session (`scope_data_externs`' file-scope
placement; `extract_unit` decl-vs-def; `_body_open_brace`'s own-line brace; `SIG_IN_BODY_RE`'s 10% oracle hole;
`revert()`'s config residue; `jr_isolate_all`'s empty region 0) — and THREE were the same brace-placement class,
the same class as the Phase-15 `find_site` bug and the Phase-24 `overlay_files` bug. Each was written off at the
time as a one-off parser slip.

**What the bytes taught.** They are not one-offs; they are a *structural* blind spot in how this project is
built. Every instance has the identical shape: **a scanner extracts N items from a corpus, the true count is
M > N, and nobody ever compared N to M.** The whole-binary byte-gate (G3/P9) is a perfect guard on
*correctness* — it never once accepted a wrong match — but it is **blind by construction to work that was never
attempted**. A tool that silently no-ops on input it cannot parse is indistinguishable from a tool that had
nothing to do. That is why these survived 26 phases: nothing in the system was looking.

The cost is not hypothetical. `SIG_IN_BODY_RE` hid **186 of 1801 (10%)** of the shared-callee signatures, which
is why nine byte-exact cores from the crack wave would not bank — the draft kept its guessed signature, hit
`conflicting types`, and the recovery pass truthfully reported nothing to fix. It read exactly like an intrinsic
wall. One character class turned it into a zero-hand-edit bank.

**The pivot (Drew, "agreed").** Promote the lesson from a *rule* to a *mechanism*, and do NOT audit by reading
regexes — that is precisely the failure mode that wrote them. Instead **measure coverage**: for each scanner,
build a deliberately OVER-APPROXIMATING candidate detector, run both over the corpus, and report
found-vs-candidates; every gap must be classified as a real skip or a justified exclusion. Going forward, a new
text scanner ships with a coverage assertion or it does not ship. (Rule candidate for PhaseEnd, P10.)

**The bigger prize (the uncomfortable part).** Several verdicts we have treated as settled physics were reached
*on top of* the broken oracle: the **def-side loose-typing wall** (§20/§41, "triple-confirmed" in Phase 23), the
**159 arity/narrow-param conflicts** (Phase-15 "documented dead-end"), the **3,098 type-heavy tail**, the 9
zero-bank type-using families. Each was diagnosed as "no C declaration exists satisfying both the definition and
the call site" — but the tool computing the call site's canonical signature was blind to 10% of them. Phase 16
byte-proved that genuinely contradictory typings DO exist, so the wall is real in part; but **"some of it was our
tooling" is now the prior, not the long shot.** Re-test the cheap ones against the repaired tools.

**Hindsight / for the wiki.** The deepest lesson of the phase, and it generalizes far past decomp: *an
incorruptible correctness gate creates a false sense of completeness.* It tells you everything you banked is
right. It tells you nothing about what you never tried. Pair every correctness oracle with a **coverage** oracle,
or you will spend phases mistaking your own parser's blind spots for properties of the problem.

## 2026-07-14 (session 8) — the coverage audit's biggest finding was REAL, and my reading of it was WRONG. The correction is the lesson.

**Context / prior belief.** Six silent-skip bugs in one session led to the coverage-oracle rule (Drew: "agreed").
The audit's headline came back alarming: *progress.py under-counts by ~243k instructions because classify()
reads a K&R definition as a forward declaration.* I verified the MECHANISM against the bytes (it is real:
`s32 f(arg0)` / `s32 arg0;` / `{` — the `;` precedes the brace, so the scan calls it a declaration and drops the
function into NO bucket), measured 400 banked instances / ~190k instructions in that shape, and told Drew our
headline numbers had been under-reporting our own progress.

**What the bytes taught — I was wrong, and the null result caught me.** After fixing it, old-vs-new on the same
tree moved the headline numbers by **+376 instructions**, not +190,000. A null result where a large effect was
predicted is not noise; it is a refutation. Reading the code: **`weighted_metrics()` never calls `classify()`.**
It determines "matched" as `func not in src_stubs(binary)` — and because the fleet is 136/136 byte-identical,
anything NOT wrapped in `INCLUDE_ASM` is necessarily compiled C emitting the exact original bytes. It never
parses a definition, so it is **structurally immune** to the bug. The published **instr-weighted (65.6%) and
distinct-code (44.9%) were CORRECT ALL ALONG**; only the secondary REAL count and fn-count % were wrong.

I had done the R14 thing (verify the mechanism against the bytes) and still got the conclusion wrong, because I
verified the DEFECT and not its BLAST RADIUS. The auditor conflated "classify() is blind" with "the metrics are
wrong", and I propagated it — to the owner, as fact, in the same breath as lecturing about unverified oracles.

**The pivot.** Both bugs are still worth fixing (they corrupt the REAL/fn-count report, and the phantom-dedup
over-count double-counts 532 stubs) and a coverage assertion now guards classify(). But the strategic conclusion
inverts:

> **A metric DERIVED FROM A PROVEN INVARIANT beats a metric that RE-PARSES THE WORLD.**
> `weighted_metrics()` leans on the byte-gate — "not a stub ⇒ byte-exact, because the build is byte-identical" —
> and *inherits its correctness for free*. `classify()` re-derives the same fact by parsing C, and inherited a
> bug instead. Two tools, one question, and the one that refused to re-derive is the one that was right.

**Hindsight / for the wiki.** Three lessons, and the third is the real one.
(1) *Verify the blast radius, not just the defect.* "This tool is broken" and "this number is wrong" are
different claims needing different evidence. A confirmed mechanism proves nothing about consequence.
(2) *A null result where you predicted a large effect is a refutation — chase it.* The +376 delta was the whole
story, and it would have been trivially easy to wave off as noise or as "the fix worked, the numbers moved".
(3) **The coverage-oracle rule is right but incomplete.** Auditing parsers is treating the symptom. The cure is
to STOP PARSING where an invariant already answers the question. Our byte-gate proves a strong property
(byte-identical build); every fact derivable from it should be *derived*, not re-computed by regex. Before adding
a coverage assertion to a scanner, ask the better question first: **why is this scanner re-deriving something the
build already guarantees?**

## 2026-07-14 (session 8 close) — Drew: the TOOLING-INTEGRITY AUDIT gates further matching work, and gets its own phase

**Context / prior belief.** Session 8 was the most productive of the project: 13 cores cracked (incl. the four
heaviest functions in the game), a 12-agent wave at 11/12 first-pass MATCH, fleet 63.0→65.6% instr-weighted /
39.1→44.9% distinct-code, 136/136 byte-identical throughout. The natural next move was obvious: bank the six
blocked cores (~1.2 MB, all plumbing), then run the next wave.

**Drew's call:** *"I feel like we should do T14 now, before the rest of the work. but not in this phase."*
The tooling-integrity audit **gates** the remaining matching work, and it is substantial enough to deserve its own
phase rather than being squeezed into Phase 26.

**Why this is right (and why I would not have prioritised it as hard).** The session found **seven silent-skip
tool bugs**, and the instinct is to treat them as a tax — annoying, fixable, keep moving. That instinct is wrong,
for a reason that only became clear at the end:

> The byte-gate is a perfect CORRECTNESS oracle and a **null COVERAGE oracle**. It never once accepted a wrong
> match — and it is blind *by construction* to work never attempted. It has been green since Phase 5, when 0% was
> decompiled, because `INCLUDE_ASM` pastes the ORIGINAL assembly: **a green byte-gate is compatible with any
> decomp percentage.** So every silent skip is invisible to the one instrument we trust absolutely.

The cost is not wrong answers. It is **invisible work, and walls that aren't there.** A single 10% hole in the
callee-signature oracle (`SIG_IN_BODY_RE`, a `\s` that could not match a line-continuation backslash) made **nine
byte-exact functions look like an intrinsic compiler wall** — and we would have written them up as such. How many
of the walls we have already "byte-proven" across 26 phases were lookup misses wearing a wall's clothes? The
def-side loose-typing wall, the 159 arity conflicts, the type-heavy tail — all were diagnosed on top of that hole.
**That is the question the audit answers, and it is worth more than the next 1.2 MB.**

Auditing after more matching would compound the problem: every wave run on broken selection tooling produces more
"walls" we would then have to re-litigate.

**Scope discipline (do NOT audit all 82 tools).** 19 were audited (23%), chosen by risk. The filter for the rest
is: **does it PARSE something, and does it GATE or SELECT work?** (~15 tools.) Priority order:
`dedup_integrate.py` (a fail-closed validator that can print a FALSE GREEN — *"1813 validated, 0 failed"*) →
`jtbl_family_bank.py` (3 bugs found by hand this session, never audited) → the SELECTION tools (`family_hseq`,
`wave_targets`, `exemplar_miner` — a hole here makes work invisible to *planning*, the worst kind) →
`masked_diff`/`match_one` (the closeness oracle every agent trusts).

**And apply R33 to each, first:** *why is this tool re-deriving something the build already guarantees?*
`harvest_verify` is the model — it derives from `make build` + SHA1, so a parse hole makes it **conservative, not
wrong**. Tools that lean on the invariant inherit its correctness for free; tools that re-parse inherit bugs.
The best audit outcome is not a fixed regex — it is a **deleted scanner**.

**Hindsight / for the wiki.** The owner saw this faster than I did. I had just spent the session proving that
every wall was our own tooling, had written the coverage-oracle rule, had *corrected myself* about a metric — and
my instinct was still "bank the 1.2 MB first". The lesson: **when your measurement layer is suspect, more
measurements are not progress.** Fix the instrument before taking more readings.

---

## 2026-07-14 (session 9) — The audit runs as an INSERTED HALF-PHASE, not as Phase 27

**Context.** Session 8 closed by gating the tooling-integrity audit ahead of all further matching work (entry
above) and left the phase-boundary shape as an explicit Tier-1 question for the owner: **(a)** close Phase 26
early with a PhaseEnd and open the audit as Phase 27, or **(b)** run it as an inserted half-phase (the Phase-3.5
precedent) and return to Phase 26 afterwards.

**Drew's call:** *"audit as an inserted half-phase in the current phase and then resume phase 26."* (Effort: Max.)

**Why (b) is right.** Option (a) reads as the tidier choice — a clean PhaseEnd, a fresh phase, a fresh context
window. But it would have closed Phase 26 on an **unmet milestone**. Phase 26's milestone is *structural
completion*, and the audit is not a **successor** to that goal — it is a **prerequisite** to reaching it: the
family engine's own numbers are what the audit found broken (93 of 218 "matched" exemplars are phantom; 1,834
clean member templates never attempted; 407 of 811 overlay files invisible to propagation). Closing the phase
would have forced a PhaseEnd that reported the milestone as abandoned, when in fact the tooling that *measures*
the milestone was the thing at fault. The half-phase keeps the goal live and fixes the instrument under it.

The Phase-3.5 precedent is exact: a spike inserted mid-arc, on the owner's directive, to answer a question that
gates the work either side of it. It closed with a go/no-go, not a PhaseEnd, and the roadmap resumed.

**The structural insight this rests on (worth repeating, because it is the whole reason the audit exists).**
A scanner extracts N items from a corpus; the true count is M > N; **nobody ever compared N to M.** The seven
silent-skip bugs were not typos — they are that one blind spot, seven times. And the byte-gate cannot see it:
it is a perfect correctness oracle and a **null coverage oracle** (green since Phase 5 at 0% decompiled, because
`INCLUDE_ASM` pastes the original asm). So the audit's ordering rule is **R33 before R32** — before adding a
coverage assertion to a scanner, ask whether the scanner should exist at all. **The best outcome is a deleted
scanner, not a fixed regex.**

**First finding, immediately (A1).** `dedup_integrate.py` — the fail-closed byte-honesty validator, and the
audit's #1 priority precisely because a silent skip there prints *a false green from a gate* — has **three**
false-green paths, all confirmed within minutes of opening it: the 7 stale registry groups name a `DEFINE_func_*`
macro with **zero hits in `src/`** and still print `[ OK ]`; an **absent** `.run/sig.*.jsonl` yields
*"0 validated, 0 failed"* and **exit 0** (on a fresh clone the gate validates nothing and passes); and it never
checks that a member is **actually banked** rather than still `INCLUDE_ASM` — which is exactly the invariant the
build already proves. The tool that guards byte-honesty was the one least able to prove its own.

**Hindsight / for the wiki.** The owner's framing — *fix it inside the phase, don't ceremonially close the phase
around it* — avoided a subtle honesty trap. Writing a PhaseEnd that says "milestone: not met, closing anyway"
when the real story is "our measuring tape was short" would have been technically true and substantively
misleading. **Phase boundaries should follow the work, not the paperwork.**

---

## 2026-07-14 (session 9, A2) — The audit found the endgame plan was majority-fiction

**What we ran.** 6 auditor agents over the 18 unaudited PARSE+GATE/SELECT tools, each finding handed to an
independent skeptic told to REFUTE it. 38 agents, 2.24M tokens. 32 findings raised → **28 survived**, 4 refuted,
16 downgraded, and **40 scanners measured clean**. The skeptic pass earned its keep: it killed four claims and
corrected magnitudes in both directions.

**The root cause is singular, and it is not a regex.** Almost every finding is the same defect:

> a hand-maintained model of the corpus layout — a file allowlist, a single-`.c` assumption, a `func_`-only
> symbol regex, a `REGION_SUB` dict — sitting on top of a **filesystem that already answers the question**.

An overlay's source is spread over up to 14 `.c` files (`<ov>.c`, `_a`, `_o0`, `_o0b`, `_after`, and the Phase-26
`_jr_<ADDR>` carves). Tools written when there was one file still believe there is one file. **The decay is
measurable:** `.run/fuel_manifest.json` from Jul 8 recorded 130 live stubs; the same tool run today returns **30**.
The Phase-26 splits moved ~100 stubs out from under a dict literal last edited in Phase 22. **Nobody noticed,
because a target that is never nominated produces silence, not an error.**

**Why this is worse than a wrong answer.** 91.6% of all remaining project gain is invisible to the target-selection
layer (994,633 instructions of real work; the manifest sees 83,305). 117 of the 127 reach-134 functions — the
entire high-ROI band — are never nominated by anything. We were about to run Task 7's crack waves against that.

**Three results overturn things we had written down as settled:**

1. **"The permuter's fuel is exhausted" (Phase 22) is unsafe.** The grinder banks through `harvest_verify`, which
   can only see one translation unit — and **1,290 of the grinder's own 1,298 queued functions live in a different
   one**. 99% of its queue could never have banked, however good the permuter's output was. "7 all-time banks, 0
   since Phase 21" is *equally consistent* with "the tool could not bank" as with "there was nothing to bank."
   We concluded the latter and moved on. **Re-test before repeating it.**

2. **The Phase-25/26 endgame plan is majority-fiction.** `docs/family-manifest.md` — the document the whole
   structural-family endgame is planned from — advertises 2,758 multi-member families holding 11.0 MB of hidden
   leverage. **1,071 of them (6.80 MB, 62% of the advertised byte-weight) are already fully matched.** The
   matched-set oracle scans a single overlay. So the byte-weight *ranking*, which is the entire purpose of the
   file, is sorted mostly on dead work, and the real targets are buried under phantoms.

3. **A corpus defect the byte-gate cannot see, and never could.** `config/symbols.us.txt:981` declares
   `listCdBuffer = 0x80180000` — a main-EXE **data** symbol — and every overlay's splat config loads that file. In
   overlay space, 0x80180000 is **code**. splat therefore cuts 97 real functions in half and invents 96 phantom
   ones: **193 slices that cannot be matched by anyone**, across 97 of 134 overlays. You cannot write C for a
   function that ends on a `lui` with no return, nor for one that begins by reading the assembler temp `$at`. They
   sit in the harvest queue as ordinary work items, so agents burn on them indefinitely and the failure reads as an
   intrinsic compiler wall. **And the full-binary byte-gate stays green the entire time**, because the `.s` halves
   are pasted back verbatim in original order.

   This is the purest instance of the thesis that motivated the audit: *the byte-gate is a perfect correctness
   oracle and a null coverage oracle.* And note precisely what rescued us — **`sig_image` was right.** Its
   independently-computed function boundaries agree with spimdisasm on 58,524 of 58,621 functions, and on all 97
   disagreements sig_image is demonstrably correct. **A second, independent oracle is the only reason the defect
   was visible at all.** That is a design lesson worth more than the fix: when one oracle is structurally blind to
   a class of error, the answer is not a better assertion inside it — it is a second oracle that can disagree with it.

**The fix follows the root cause: ONE derived corpus oracle, and ~10 deleted scanners.** Not ten fixed regexes.
`tools/corpus.py` answers — from the filesystem and the proven invariant, with coverage assertions baked in —
*which files make up a binary*, *which stubs are live*, *which functions are matched* (sig − stubs, derived, never
re-parsed), and *where a function's asm lives* (globbed, because splat already wrote the truth). Then the allowlists,
the `REGION_SUB` dict, the single-TU regexes, and `census_conflict_callees` in its entirety all get deleted. This is
the "best outcome is a deleted scanner" rule (R33) applied at scale.

**Hindsight / for the wiki.** The strategic error was not writing any one of these tools badly. It was **letting the
corpus layout become a fact that lived in ten places**. Each split was a correct, well-gated change to the *build*;
none of them updated the ten private models of the tree, and nothing existed to notice. **A derived fact has no
maintenance cost and cannot rot; a hand-maintained copy of it is a liability that grows with every structural
change.** And the reason it stayed invisible for four phases is the deepest lesson of the audit: *we had no
instrument that could report absence.* Every gate we owned answered "is this right?" — none answered "is this all?"

---

## 2026-07-14 (session 9, A4) — A corpus defect the byte-gate could never have caught

**The defect.** `config/symbols.us.txt:981` declared `listCdBuffer = 0x80180000`. That is a correct,
Phase-3-derived name for a main-EXE RAM buffer (the LIST.CD cache). But 0x80180000 lies **outside main's
image** (0x80010000–0x80074800) and **inside the overlay slot** (0x80128158–~0x801DAB30) — and every
overlay's splat config stacks `symbols.us.txt`. High RAM is *reused*: an address that is a buffer to
main is live **code** to an overlay.

So splat saw a symbol boundary in the middle of overlay code and, across 97 of the 134 overlays:

* **cut 97 real functions in half** — leaving a head that ends on a `lui` with no return, and
* **invented 96 phantom functions** — a tail that begins by reading the assembler temp `$at`.

**193 slices that nobody can ever match.** Not "hard". Not "a compiler wall". *Unmatchable by
construction* — there is no C you can write for either half. And they sat in the harvest queue as
ordinary work items, so agents would burn on them indefinitely and the failures would be filed as
intrinsic compiler residuals.

**Why no gate caught it, and why that is the important part.** `INCLUDE_ASM` pastes the two `.s` halves
back **verbatim, in original order**, so the image is byte-identical either way. The full-binary
byte-gate — the instrument this project trusts absolutely, and rightly, because it has never once
accepted a wrong match — **was green the entire time and always would have been.** It is a perfect
*correctness* oracle and a **null coverage oracle**. No assertion added *inside* it could ever have
found this.

What found it was a **second, independent oracle**: `tools/sig_image.py` derives function boundaries
from the ORIGINAL bytes without splat, and it *disagreed with the corpus*. It agrees with spimdisasm on
58,524 of 58,621 functions and is demonstrably correct on all 97 disagreements. That is the whole
lesson, and it generalises well past this bug:

> **When one oracle is structurally blind to a class of error, the answer is not a better assertion
> inside it. It is a second oracle that can disagree with it.**

`make audit-corpus` now *is* that second oracle, standing.

**The evidence that makes it concrete.** The phantom `listCdBuffer.s` in ov_SC01_005 begins:
`lw $ra, 0x10($sp)` / `addiu $sp, $sp, 0x18` / `jr $ra`. splat cut a function immediately before its
**epilogue** and called the epilogue a function. You cannot write C for a routine that restores a return
address it never saved.

**And it had already contaminated real work.** In `ov_SC03_031` the cut happened to land where the
epilogue was exactly `jr $ra; nop`, so the Phase-26 ×134 sweep innocently "matched" it as
`void listCdBuffer(void) {}` — byte-correct, gate-green, and completely fictitious — while leaving
`func_8017FFC4` permanently unmatchable. A phantom got *banked*.

**The rule, which nobody had written down.** R13/R15 say overlay-derived symbols are overlay-region only
and must never be merged into `symbols.us.txt`. The mirror is equally true and was never stated:

> **A symbol whose address falls inside ANOTHER binary's vram window must never enter that binary's
> symbol stack.**

Fix: `config/symbols.us.ram.txt` — main-scoped symbols that live outside main's image — stacked **only**
by `config/splat.us.exe.yaml`. Main keeps the name (its asm carries 10 `%hi` / 11 `%lo` references and
rebuilds `143dbb89` byte-identical); the overlays never see it. Exactly one symbol was in scope
fleet-wide, and the resident window was clean.

**Hindsight / for the wiki.** We had *two* oracles all along and never made them argue. The byte-gate and
`sig_image` were both trusted, both correct, and silently disagreeing about the shape of 193 functions
for four phases. The cheapest possible check — *do our two independent views of "where does this function
start and end" agree?* — was never run, because each oracle was individually green and nobody thought to
ask them the same question. **Redundancy is only worth what you spend comparing it.**

---

## 2026-07-14 — `cdecl`: parse the grammar, do not enumerate the shapes

**Context & belief.** The audit's own prescription for the fifteen broken declaration scanners was a
*shape-aware alternation* per tool: add an `(fn-ptr|sized-array|scalar)` branch to `DATA_DECL_RE`, mirror
it in `DATA_DECL_LINE_RE`, add a `fnptr` kind to `parse_data_decl`, add a fn-ptr arm to
`_uniquify_draft_types`, and so on — roughly fifteen coordinated regex edits, each with its own
suggested coverage assertion.

**Why I did not do that.** The audit had *already proved* that fifteen independent hand-maintained models
diverge: two tools in ONE pipeline disagree today about whether `extern s32 D_a, D_b;` is a declaration.
Patching fifteen regexes is fifteen fresh chances to diverge again, and an alternation only ever covers
the shapes somebody remembered — it is the same hand-maintained model, one shape wider. The real problem
was never the character class. It was that **the thing being scanned has a grammar, and nobody was
parsing it.**

C's declarator grammar is small, closed, and **total**. It describes fn-ptr arrays, sized and 2-D arrays,
multi-declarators, fn-ptr parameters, and K&R identifier-lists *without being told they exist*. A
250-line recursive-descent parser is **less** code than the fifteen regexes it deletes, and it is
exhaustive by construction rather than by anyone's memory. That is R33 in its strongest form: the best
outcome is not a fixed regex — it is a deleted model.

**The measurement (not a belief).** Three oracles, whole corpus: coverage (**2,952,246 depth-0 statements
→ 2,731,521 declarators, 0 parser defects**), the real cross-gcc (**50,405 distinct declarations
round-tripped, 0 rejected**), and a differential against the incumbents (0 symbols they see at file scope
that `cdecl` misses; 26 in `engine_core.h` they cannot see; 6 they wrongly promote from *block* scope).

**Two design decisions worth keeping.**
1. **The candidate set is derived, not hand-written.** At file scope C admits nothing but declarations, so
   the over-approximating detector R32 demands is *every depth-0 statement* — supplied by the grammar,
   with no second model to rot. (LAW 4.)
2. **gcc adjudicates my own coverage gap.** When 40 statements would not parse, deciding for myself which
   "don't count" is grading my own homework — the precise habit that wrote the fifteen bugs. gcc decides
   instead: a statement it *also* rejects is not C (my rejection is correct, the input is corrupt); one it
   *accepts* and I do not is my defect. All 33 residual came back NOT-C, all in dead scratch. (LAW 5.)

**Hindsight / for the wiki.** The near-miss is the lesson. Those 33 corrupt drafts were written by a
*recovery tool* that prepended `extern` to an `if` statement, and I was one step from reporting a live
tool bug. Checking the blast radius instead (R14) showed the source defect was fixed back in Phase 19 —
today's oracle emits **0 garbage over 300 signatures**. *Mechanism confirmed, consequence nil.* But note
what it cost while it was live: a draft that cannot compile fails the byte-gate, and the failure reads
downstream as **an intrinsic compiler wall.** That is the audit's whole thesis in one artifact — and the
new parser is what finally makes the guarding assertion expressible: *every canonical signature the
callee oracle emits must PARSE as a C declaration.* Before `cdecl`, nothing in the repo could tell a
signature from garbage.

**Scope discipline (deliberate).** This commit lands the parser and its proof and changes **no consumer**
— so it cannot move a byte, and `check-all` is 136/136 by construction. That is not timidity: the audit
explicitly warns that *making the parser see more ARMS dormant downstream transforms* — the moment
`reconcile_decls` can parse a fn-ptr decl, its `data_access_subs` would happily mangle `D_1[i]()` into
`((u8 *)D_1)[i]()`. Consumer migration is therefore one tool at a time, each byte-gated.

---

## 2026-07-14 — Probe the compiler; and the adjudicator must BE the compiler

**Context.** Building `cdecl.compatible()` — *"will cc1 accept these two declarations of one name?"*, the
question every recovery pass in this repo actually asks and four of them half-implement. I wrote the rules
from the C standard, then validated them against a compiler.

**What happened.** The compiler contradicted me — and then the *right* compiler contradicted the first one.
Validating against modern `mipsel-linux-gnu-gcc` and against the real gcc-2.7.2 `cc1` gives **three
different answers** (with the standard as a third): typedef redefinition is an error in C89, accepted by
C11 gcc, and rejected by cc1; a qualifier mismatch is an error to modern gcc and **accepted** by cc1; the
no-prototype/narrow-param rule is an error to both — and **accepted by cc1 in one direction.**

**The decision.** `--compat` adjudicates with `tools/bin/gcc-2.7.2-psx/cc1`, the front end that actually
arbitrates the build. Now 1,485/1,485 live corpus pairs agree. **Validating a compiler rule against a
compiler that is not the one compiling your code is not a shortcut — it is the same class of error as the
five phases we spent reading `gcc-papermario` believing it was 2.7.2. It was 2.8.1.**

**The prize (→ A10).** Phase 15 closed the "159 arity/narrow-param conflicts" as *"no clean deterministic
fix — it is simply C's default-promotion rule."* **cc1 disagrees.** The rule is order-dependent:
`void X(s16); void X();` compiles; only `void X(); void X(s16);` fails. The wall's stated cause does not
hold. Four three-line probes, 90 seconds, zero tokens.

**Hindsight / for the wiki.** *Probe the compiler for FACTS; read its source only for LEVERS; byte-validate
both.* Reading source is inference and can be wrong (it was, for five phases). Probing is ground truth,
because it IS the compiler — and it is orders of magnitude cheaper. We have the exact binary sitting in the
tree and spent 26 phases reasoning about it instead of asking it.

**And the discipline that saved this from being an over-claim.** Fixing the wrong-TU bug (95.1% of drafts
canonicalized against a TU that would never compile them) took the callee-conflict repair from 8 to 58 of
196 drafts — 7× reach — and banked **exactly zero** functions, because the historical tail fails on codegen,
not plumbing. The real gain is narrower and still worth having: **52 drafts moved from "won't compile" to
"compiles, N instructions off"** — from an invisible failure that reads as a compiler wall into a scored
near-miss the permuter can act on. Three times in one session a confirmed mechanism produced a null
consequence. *"This tool is broken" and "this number will move" are different claims, needing different
evidence.*

## 2026-07-15 (session 13, A10) — the wall re-test verdict: the broken tools WERE the walls, and the payoff was banked by the FIXES

**Context + belief.** The audit set out to answer one question (its own thesis): *how many of the walls we
byte-proved across 26 phases were lookup misses wearing a wall's clothes?* Going in, the honest prior — set by
the `SIG_IN_BODY_RE` finding, where one 10% oracle hole made nine byte-exact functions look like an intrinsic
compiler wall — was *"some of it was our tooling."* A10 was gated ahead of all matching to test that at scale.

**What we found.** The payoff did not come from A10's own re-gating — it came from the **fixes**, and it was
already banked by the time A10 ran: retiring the fleet-majority oracle (A3d → `reconcile_tu`), the all-TU gate
(A3e), the h_seq callee oracle (`SIG_IN_BODY_RE`), the fn-ptr blindness (A9a), and the `build_engine_types`
blocker (A7) collectively moved the fleet **66.5 → 68.6% instr**, with the flagship proof being A9b —
`func_8017A4AC` (536 ins × 134), a "blocked on plumbing" wall since session 8, banking ×134 the moment the
oracle it tripped was fixed. **The walls named in this audit's thesis were, in the parts that moved, our tooling.**

**The pivot inside A10 (measure, then adapt — R14).** The obvious A10 move — brute re-gate the 958
`closeness==0` backlog drafts through the fixed gate — was tested first on one overlay (0/14 bank) and then
settled at fleet scale: **0 of 958 bank across 135 binaries.** `match_one closeness==0` (isolated,
reloc-masked) systematically overstates whole-binary bankability; the fixed gate recovers **none** of them.
So rather than a Workflow fan-out that would have burned agents confirming a null, A10 ran it as a
deterministic parallel job and reported the number. **The closeness-0 residual is genuine codegen — a
re-confirmed wall (P9), which is as valuable as a dissolved one: it tells the endgame where NOT to look.**

**The better path, in hindsight.** The single most load-bearing lesson is upstream of any specific fix:
**a green byte-gate is compatible with any decomp %, so it can never tell you what you failed to attempt.**
Every wall this audit dissolved was invisible to the one instrument we trusted absolutely — not because the
gate was wrong, but because it is a correctness oracle with a null coverage dimension (R34). Had a coverage
oracle (R32) and a *second, disagreeing* oracle (R34) existed from Phase 6, most of these walls would never
have been written down as walls. The audit's real deliverable is not the ~15 fixes — it is the three rules
(R32/R33/R34) and the derived-oracle pattern (`corpus.py`/`cdecl.py`) that make the *next* 26 phases unable
to manufacture a wall out of a lookup miss.

**Handed forward:** #4 the type-heavy harvest (~1,200 members; `build_engine_types` unblocked but not yet
wired into the family path — Phase-26 Task-8 integration, not a re-test). Substrate for the retrospective + the
public "how to AI-decomp" wiki (R31): *audit your instruments before you trust their silence.*

### 2026-07-15 — Phase 26 Task 7 resume: re-baseline corrections + the flagship regalloc wall (single Fable5, likely intrinsic)

**Re-baseline before cracking (R14).** Task 7 resumed at xHigh after the audit. Regenerating the target
frontier from the FIXED tools corrected the handoff twice, both byte-grounded: (1) the "~1,200-member
type-heavy harvest" from the audit ledger was **already banked** by A3h's post-fix h_seq re-run (+2,675) — the
current STRUCT-excluded tail is ~6 substantial members, not 1,200. (2) The real remaining deterministic lever is
bigger and different: **128 matched-sibling families ≈ 2.64M templatable ins**, and **register pins are NOT a
banking blocker** (`func_8017A4AC` banked ×134 with 4 pins/sibling) — so most of that is *un-run ×N sweeps*, not
a wall. Also: all 97 reach-134 tractable cores are already walled, and re-gating the close 1-5 seeds through the
fixed pipeline banked **0/13** — Task 7 is a near-miss CLOSING campaign, not fresh cracking.

**The single-Fable5 test ("can we crack not on Max?" — Drew).** One Fable5 (xHigh orchestration, 477k tokens,
93 min) on `func_80178004` (165 ins ×134, the regalloc-order class exemplar shared by 12 siblings). Result: an
**honest wall** (P9). A pin-free draft driven to structure-exact (163/165); residual = pure register identity,
reduced by a gdb-on-cc1 oracle to **three compiler-internal integers**, each ruled unreachable under every legal
C construct with file:line evidence. Byte-verified: `match_one` 126/165, the 126 dominated by one `$s0`↔`$s2`
swap. **R14 ledger correction:** the historic "pinned MATCH / close=0" was a myth — the seed was never a match
(best historic permuter score 5, pinned). Likely intrinsic to gcc-2.7.2; ONE untested lever remains
(`qty_n_refs` tie-shape, local-alloc.c:1869). The parallel permuter could not crack `func_801325B8` either (best 25).

**Payoff despite the wall (R16 flywheel).** The pass produced 6 byte-proven "walker-family" levers + the
skeleton idiom (cookbook §52) that transfer to the 11 regalloc-order siblings: **Fable5 DISCOVERS the skeleton,
cheap-Opus APPLIES it.** So a walled exemplar still fed the flywheel.

**Pivot (pending Drew's Max call, R27).** Per Drew's instruction the failed single-Fable5 → prompt for Max.
Recommendation to be logged on decision: Max-grinding `func_80178004` itself is low-EV (Fable5 already went to
depth; one untested lever); the higher-ROI use of the perishable window is the §52 sibling-idiom wave
(cheap-Opus) + the deterministic matched-sib harvest (2.6M-ins ceiling, pins OK) — both higher-certainty than
grinding an intrinsic wall.

### 2026-07-15 — Phase 26 close: the mechanical-harvest thesis is byte-proven exhausted → close + Phase-27 fresh scans

**The §52 flywheel worked — and then the byte-gate closed the phase.** The single-Fable5 failure on `func_80178004`
still distilled the walker-family idiom (§52), and two cheap-Opus waves applied it to bank **5 pin-free regalloc
cores ×134 = 670 instances** (68.6→68.9% instr), confirming *Fable5 DISCOVERS, cheap-Opus APPLIES* and the
crack-then-template loop for CLEAN families.

**The correction (P9/R14 — walking back my own prior entry).** The entry above recommended "the deterministic
matched-sib harvest (2.6M-ins ceiling, pins OK)" as *higher-certainty*. **That was wrong.** Three byte-gate
probes returned 0% — tiny-IMM 0/241, PURE reach-134 0/134, and pinned-PURE-templated-WITH-pins 0/133 (so the
`func_8017A4AC` pinned-×134 precedent does NOT generalize). The manifest's ~13,075 "templatable" member-slots
are an h_seq *prediction* the whole-binary gate refuses (collision / register-drift / pin-crash). A3h + the wave
propagations already banked everything cleanly templatable. The mechanical/templating thesis is spent at
68.9% instr / 49.2% distinct.

**The pivot (Drew, 2026-07-15).** Close Phase 26; open Phase 27 with a byte-gate-honest re-scan. Why it matters:
the 26-A audit fixed the TOOLS, but the megaplan's frontier was scoped by the PRE-audit (buggy) scans, and even
the post-audit manifests over-predict templatability (h_seq ≠ bankable). The remaining work is a different shape
— hard-wall cracks + hand-decomp of the genuine unique residue — and deserves a plan built on a gate-validated
frontier, not a manifest that over-promises phantom members.

**Hindsight better-path — the phase's most transferable lesson:** *a scan is a hypothesis; the byte-gate is the
truth. Validate a "templatable"/"matchable" count against the gate with one small probe BEFORE scoping a whole
phase around it.* The megaplan's "986 families / 2.6M ins" set an expectation the gate then had to walk back
twice this session; a probe up front would have sized the real yield. This is R14 applied at planning scale, and
it is exactly why Phase 27 opens with a gate-validated re-scan.

### 2026-07-15 — The Road-to-100 roadmap adopted (Phase 27+); the megaplan superseded

**Context.** Phase 26 closed on the honest pivot (mechanical harvest byte-proven exhausted; instruments made
trustworthy by the 26-A audit). Drew directed a plan-mode session (Fable5, Max): not a Phase-27 plan but a
**full roadmap from 68.9% instr / 49.2% distinct to game-code 100%**, grounded in the audit. Method: all 26
PhaseEnds + 3 scout distillations (audit / frontier / strategy) + a 16-defect red-team pass. Deliverable:
**`docs/roadmap-to-100.md`** (P27 farewell-sprint+honest-frontier → P28 engine → P29 families → P30 mass+main
→ P31 behemoths+walls → P32 verify+flip+Gen2-exit).

**Drew's four contract decisions (the values calls, 2026-07-15):** (1) **game-code TRUE 100%** — no completion
declaration while any game-code stub remains; walls re-attacked each phase boundary / model generation until
they fall (the §45/§52 model-relativity history). (2) **PsyQ LINKED = complete**; libs-from-source (sotn
precedent) recorded as a far-future side note only. (3) **Public flip AT 100%** — over the 2026-07-01 strategy
review's near-term recommendation ("community labor is how every peer crossed the hard middle"). The tension is
resolved by a **standing velocity checkpoint**: every phase close reports instr-%/session; if the trajectory
stretches beyond what solo+agents can credibly finish, the flip-timing question is re-surfaced with the numbers
— the decision stays falsifiable, not assumed. (4) **Fable5 window ~7/19**: the discovery sprint is P27's FIRST
task (recon-done seeds, 1–2 fresh top cores, the qty_n_refs lever, the pin-crash SIGABRT characterization).

**What the red-team caught (the why behind the roadmap's shape).** My own draft carried numbers past their
invalidation events — reconcile_tu "needs wiring" (already wired, PhaseEnd26), a "~300k free plumbing win"
(really ≈123k across 3 cores; func_8017A4AC already banked; func_8013F350 is a real class), "top-20 = 52%"
(34% post-audit), worklist "needs regen" (already regenerated) — exactly the R14-at-planning-scale failure the
Phase-26 close named. Fixes: every number in the roadmap traces to a committed post-audit artifact; §0 mandates
that every PhaseEnd re-baseline the roadmap via a standing **"Roadmap delta"** line; and P27 re-derives whatever
it consumes. The red-team also surfaced two real finds the plan now owns: **the `0x8017BEBC` family (952×~112,
~106k ins, exemplar MATCHED) was never covered by the exhaustion probes** — the IMM-scattered class gets a P27
gate-probe as possibly the largest cheap win left — and **the R34 second oracle covers only the overlays**, so
main/resident (exactly where the 100% flags plant) get the oracle extension before any 100% claim.

**Supersession.** `docs/family-endgame-megaplan.md` is superseded by `docs/roadmap-to-100.md` (banner added,
content preserved). Its h_seq reframe survives — it produced Phase 26 — but its scan-derived numbers and the
"crack ~986 exemplars → template ×120" thesis are byte-proven spent.

**Hindsight better-path.** A roadmap "grounded in the audit" nearly shipped with pre-audit numbers in it. The
transferable rule: when authoring any forward plan, red-team it against the committed artifacts *of the same
day*, and make the plan self-expiring (the Roadmap-delta line) rather than self-perpetuating.

## 2026-07-15 (Phase 27) — three strategic findings: the disc is bigger, a "wall" was our tool (again), and a "cheap win" is dead

**Context + belief.** Phase 27 opened to build the endgame on measured reality (the roadmap adopted the day before, red-teamed against same-day artifacts). Three beliefs going in, all now revised by the byte-gate:

**(1) The binary count is 136 — REVISED to 140 + a 39-module backlog.** The disc-completeness audit (T7) found 4 code-bearing SC07 overlays invisible for a month (code at PAC entry 1, not 0; `new_overlay.sh` hardcoded `0.4.dec`), onboarded byte-clean → 140. Then the type-sweep found **39 more un-onboarded type-1 code modules** (resident-class, in MAIN.CD), byte-confirmed code (jr $ra density ~3%, vs 0% for the data types that decode as valid-looking noise). They load at unknown addresses, so they are NOT mechanically onboardable — game-code TRUE 100% now spans 140 binaries PLUS ~39 modules pending runtime load-address RE. **Why it matters:** the completion contract's denominator was wrong, and the honest re-baselining LOWERED the headline (68.9→67.0% instr) because the SC07 overlays added mostly-unmatched code. The prior number was measured over an incomplete disc. Hindsight better-path: a disc-completeness sweep belongs at Gen2 *start*, not Phase 27 — the byte-gate is blind to un-onboarded code (R34), so "what did nobody onboard" must be asked explicitly and early.

**(2) The §42e "pin-crash wall" is intrinsic — REFUTED; it was `extract_unit` dropping macros.** For phases the project recorded that register-pin-heavy families "SIGABRT the sibling TU… ov077-TU-context-specific… fixed-size allocator tables… NOT ×134-recoverable." The wave-2 Fable5 characterization (`.run/giants/pin_crash_sigabrt.md`) located the abort exactly (`sched.c:2725 create_reg_dead_note`, a sched1 REG_DEAD-note conservation bug) and proved the trigger is `family_remap.extract_unit` dropping the body's file-scope `#define` macros (the T5 bug): the dropped gte_* macros became implicit-declaration CALLS, pushing caller-saved pins into the one fatal shape. Only 1 of 4 families genuinely crashed; the other 3 were exit-33 plumbing folded into one crash bucket by a shared gate-TU (R14, recursed). **Fixed (T5 `_carry_macros`), all four stage 133/133 clean.** Why it matters: **P31's pin-propagation route is OPEN** — a whole class of high-reach pinned cracks the roadmap wrote off as ×1 can now propagate ×133. This is the 26-A audit thesis a third time: the instrument, not the compiler, was the wall. Hindsight: every "intrinsic" verdict tied to a *tool's* behavior (a crash, a compile fail) deserves the R34 second look before it's recorded as a compiler limit.

**(3) The `0x8017BEBC` family is "possibly the largest cheap win left" (roadmap B2) — REFUTED, and the refutation is only trustworthy because the tool was fixed first.** Pre-T5, the probe would have read a fake 0% (112/112 CC1-FAIL on dropped macros) — a fourth phantom exhaustion proof. Post-fix: 106/112 stage, and the byte-gate says 0/8 (all genuine DIFF). The family is genuinely not byte-templatable; the h_seq structural match is necessary, not sufficient. **Why it matters:** it confirms Phase-26's mechanical-harvest-exhausted thesis extends to the families the roadmap hoped were cheap — and it is the cleanest demonstration of the phase's meta-lesson: **fix the measuring tool before you trust a measurement; a 0% from a broken tool and a 0% from a working one are the same number and opposite facts.**

**The through-line (the phase's transferable rule).** Every one of these was a case where an instrument — a scanner (`make report` swallowing gates, T2), a strip regex (six of them, T4), a boundary oracle blind to main/resident (T10), a staging step dropping macros (T5), an onboarding glob (T7) — silently mis-reported reality, and the fix changed the answer. The roadmap's own numbers were red-teamed; the *tools under them* were not, until this phase. R32/R33/R34 exist for exactly this, and Phase 27 is their first full application to the frontier the endgame plans against.

## 2026-07-15 (Phase 28 T0/T1) — the "families don't template" doctrine was a missing build step; B2 lives

**Context + belief.** P28 exists to measure the member-adapt close-rate — the roadmap's §6 "THE swing
number", with all P28/P29 yield projections deliberately withheld until it existed. Going in, the settled
belief (PhaseEnd_Phase26 + PhaseEnd_Phase27 + `calibration.md`'s "decisive P28/P29 input") was: **the
mechanical templating harvest is dead** — h_seq/h_norm structural families bank at **≈0%**, so P29's
arithmetic is "(cores cracked) × (reach)", not "(families) × 120", and B2 (`0x8017BEBC`, "possibly the
largest cheap win left") is byte-refuted.

**What failed.** All of it, and the failures compound:

1. **The ≈0% was measured with the wrong tool for the class.** `0x8017BEBC` is a **jr/switch** core. §47
   banked its exemplar as *"lazy isolation → carve (9-piece interleave) → splice → BYTE-IDENTICAL"* and
   called the fix *"×N template-safe."* The P27 T5 probe swept it with `family_sweep`, which stages C and
   gates and **has no carve step** — so gcc's jump table was never placed. The whole residual is **two
   words** (`classify_member` → **PURE, ndiff=2**, idx 343/345 = `lui/lw %hi/%lo(jtbl_801EC44C)`);
   `overlays.mk:112` carves the table for the exemplar, `:134` does not for the member. **Re-run through
   `jtbl_family_bank.py --raw` (which carves per sibling): 8 of 8 BANKED**, 4 same-address + 4
   cross-address, `make clean` + extract-all + `check-all` → **140/140**.
2. **n=1, on the least representative family in the population.** `has_mid_jr` is **3 of 163**
   matched-exemplar families (120 of 13,232 members). The rarest class was generalized to the whole frontier.
3. **Its corroborating evidence was pre-fix.** The three Phase-26 exhaustion probes (tiny-IMM 0/241,
   PURE 0/134, pinned 0/133) all predate `_carry_macros` (P27 T5, `ee4b3a02e`). P27's decision-log calls its
   own re-probe *"a **fourth** phantom exhaustion proof"* — it named the mechanism that would have faked the
   first three and never re-ran them. **The ≈0% doctrine now has no surviving post-fix evidence.**
4. **A second instrument was lying underneath.** `family_remap.img_path` hardcoded `0.4.dec`, so the 4 SC07
   overlays P27 onboarded returned `None` → `stream_words` → `None` → `classify_member` → `("LEN", [])` —
   **silently** classified "not templatable" AND poisoning their family's `diff_class` to MIXED. Fixed by
   deriving from `config/splat.<bin>.yaml`'s `target_path` (R33 — what the BUILD reads) + raising (R32).
   Negative control: `ov_SC07_006` **None → `1.4.dec`**; all **233** shared substantial fns between
   ov_SC07_006 and ov_SC01_001 classify **PURE**, every one of which the old tool called LEN.

**The pivot.** The mechanical-templating thesis is **un-refuted, not vindicated** — and re-opened as the
phase's central question rather than its retired premise. P28 T3 now measures the rate over the population
that actually exists (from the fixed map): **1,418 matched-exemplar families / 21,889 unmatched members**,
PURE 17,024 (78%) · IMM 4,473 (20%) · **STRUCT 392 (1.8%)**. Note the roadmap sizes its swing number on
*register-drift* = **STRUCT = 1.8%** of the input; the mass is PURE+IMM.

**Why it matters (the number).** Regenerating the map exposed a **doubly-hidden** pool: **1,255 families /
6,268 members / 230,612 ins whose ONLY unmatched members are the 4 new SC07 overlays** (0 elsewhere), each
behind an already-matched, byte-proven exemplar — hidden once because P27 never regenerated the map after
onboarding them, and again because `img_path` would have called them all LEN. Total addressable behind a
matched exemplar: **937,248 ins = 21.7% of all remaining weight = 7.16pp of fleet instr**. `0x8017BEBC`
alone is 115 members × 952 ins ≈ **109,480 ins**. All of this is a **prediction** until T3's gate — h_seq
predicts, the whole-binary gate decides (G3/P9).

**The better path, in hindsight.** Three of the four failures above are one habit: **a probe inherits the
authority of the tool that ran it, and nobody re-runs a probe after fixing the tool under it.** P27 coined
R35 for exactly this and then, in the same phase, generalized a 0% from a carve-less sweeper on the rarest
family class in the population. The transferable rule is sharper than R35 as written:

> **Before a 0% retires a lever:** (a) did the probe run every build step the *exemplar's own bank*
> required? (b) is the probe family *representative* of the class being generalized to? (c) was the
> corroborating evidence taken through the same tool you just fixed? A negative result is a claim about a
> tool until each is answered.

Two further R14 corrections this session, both mine: the approved plan's own population figures (163
families / 13,232 members) came from the **stale** map; and "add `jtbl_` to `symbol_map`" was a **wrong fix
derived from a true diagnosis** — a compiler-generated switch table is never named in C, so there is no
token to substitute; the fix is placement, not substitution. Recorded in cookbook **§53**.

**Handed forward:** the remaining 107 members of `0x8017BEBC`; T3's stratified probe (SC07-only pool first —
its exemplars are already byte-proven, so a failure isolates the templating mechanism with no drafting
variable); and the roadmap's B1/B2/§2 numbers + the "(cores) × (reach)" arithmetic all need re-deriving at
the P28 close (Roadmap delta).

## 2026-07-16 (Phase 28 T3b) — the legacy h_seq swing number: ~3% as-tooled, but the failure mode is the tooling-vs-wall ambiguity that keeps resolving to tooling

**Context + belief.** T3-A measured the SC07 pool (h_exact + unwired, banked 95.6% via dedup_extend) but
that answered a *different* question than the roadmap's swing number, which is the LEGACY h_seq
templatability rate. Going in, the roadmap's belief (from Phase 26 + the pre-T1 calibration) was that
structural families bank at ≈0% — a belief this phase already refuted for B2 (jr+carve, 88.7%) by proving
the 0/8 was a missing build step.

**What T3b measured.** `family_sweep --hseq --chunk 1` over 6 legacy PURE non-jr families: **9 BANKED /
37 PLUMBING / 274 DIFF** = ~3% as-tooled. **Classified, unlike Phase 26's 0%.**

**The load-bearing nuance (R14/R35 on my own probe).** The 274 DIFF is NOT proof the families don't
template. The members are byte-level PURE (`classify_member` = reloc-only, 20/20 sampled), genuine h_seq
(all DIFF_BYTES vs the exemplar, so `family_sweep` is the correct tool, not `dedup_extend`), at the same
vram. A PURE family should reproduce once its relocs are remapped — so 274 non-reproducing members means
the **remapped exemplar body fails to recompile to the member's bytes**, which is one of: (a) an
incomplete `symbol_map` (the recurring jtbl/prefix bug — B2's 0/8 and T4's 12 DIFFs BOTH resolved to
tooling THIS phase), or (b) genuine TU-context regalloc divergence (a real wall).

**Why I did not resolve it.** Distinguishing (a) from (b) needs a region-by-region byte-diff of one PURE
DIFF member's staged output against the target — a Max-effort diagnostic, and I was at ~40% context after
a very long session. Rushing it is exactly how Phase 26 manufactured a wrong ≈0%. So the honest deliverable
is the CLASSIFIED measurement + the named next probe, not a verdict.

**The better path / handoff.** P29 opens by running the disambiguating probe BEFORE scaling its
"(cores)×(reach)" arithmetic on 3%: diff one PURE DIFF member's `family_sweep`-staged bytes vs the target
— mismatch AT a reloc position ⇒ (a) incomplete remap, fixable, and the legacy h_seq ceiling is far above
3%; mismatch in regalloc/schedule AWAY from relocs ⇒ (b) a TU-context wall and 3% is real. Given this
phase's scoreboard — every "structural wall" probed (B2, SC07, the pin-crash wall in P27) has resolved to
tooling — the prior should lean toward (a), but that is a prior, not a measurement. **The swing number is
~3% as-tooled, ceiling unknown; do not treat 3% as the ceiling until the probe runs.**

## 2026-07-16 (Phase 29 Task 1) — the swing number RESOLVED: the "~3%" was an -O0 compile-flag artifact, not a wall (the third structural wall to resolve to tooling)

**Context + belief.** P29 opened, as the Phase-28 handoff mandated, by running the disambiguating probe on
the legacy-PURE-non-jr "~3% as-tooled" swing number BEFORE scaling any "(cores)×(reach)" arithmetic on it.
The prior (from the phase scoreboard — B2, SC07, pin-crash all resolved to tooling) leaned (a) incomplete
remap; but that was a prior, not a measurement, and the honest state was "ceiling unknown."

**What the probe found (byte-proven).** Built `tools/diff_regions.py` (the deferred roadmap tool): it
remaps the exemplar EXACTLY as `family_sweep --hseq` stages it and compiles at the EXEMPLAR's real
optimization level. The two families supplying **~272 of the 274 DIFF** (`0x8013c964`, `0x8013c938`) are
**-O0 functions** (their exemplar is in `ov_SC01_077_o0.c`, the Phase-19 -O0 cluster), and `family_sweep
--hseq` stages the draft into the member's -O2 stub file — so it compiled an -O0 target at -O2, which can
NEVER match. Compiled at -O0 the remapped C masked-MATCHes (`func_8013C964`→MATCH(10),
`func_8013C938`→MATCH(11)). A 106-member sample across all size bands (nins 2..133): **O0-FLAG 45 ·
already-banked 29 · TEMPLATES 17 · type-lift-plumbing 15 · REGALLOC 0**. Zero codegen walls.

**Why the T3b measurement was neither wrong nor a lie — it was a measurement of a broken build step.** The
Phase-28 T3b probe honestly classified the 274 as "genuine gate-DIFF" and honestly refused to call 3% a
ceiling. What it could not see (at ~40% context, end of a long session) was that its own tool
(`family_sweep --hseq`) had no per-member opt-level awareness — the SAME shape as the §53 carve-law finding
(a family swept with the wrong build step reads exactly like an intrinsic wall). R35 again: a 0/near-0 from
a tool missing a build step, and a real wall, are the same number and opposite facts. The `has_mid_jr`
carve gap (§53) and this -O0 opt-level gap are two instances of one class: **`family_sweep` must reproduce
every build step the exemplar's own bank required — the carve AND the optimization level.**

**The pivot.** The swing is (a) tooling. P29's member track is NOT a low-ceiling per-member grind and does
NOT need `member_adapt.py` (the (b)-wall delta engine) for this pool. It is the mechanical **-O0 split
rollout** (the deferred "-O0 ×134", the whale `_o0b` precedent) + the type-lift sweep — Task 2a. The
~478k-ins legacy-PURE pool is back on the table.

**Honest caveat (not yet a bank).** The verdict is masked-MATCH — a candidate (§52b). The -O0-split
mechanism is independently byte-proven (the whale banks ×134; `ov_SC01_077_o0.c` banks byte-identical), so
confidence is high, but Task 2a whole-binary-gates it (and byte-gates EACH cluster member — `func_8013B7AC`
in this cluster was called "overlay-local" in Phase 20, so no blanket assumption).

**Better path, in hindsight.** The generalizable lever `family_sweep` is still missing: it should DERIVE
each member's required build steps (carve for `has_mid_jr`, -O0 for an -O0-cluster exemplar) from the
exemplar's own bank, and refuse to gate at the wrong build step rather than book a phantom DIFF. Task 2a
builds the -O0 arm of that; the carve arm (§53/`jtbl_family_bank`) already exists — they should converge
into one build-step-faithful sweep. That would have made both the Phase-26 "≈0%" and the Phase-28 "~3%"
impossible to manufacture.

### 2026-07-16 — P29 Task 2 Arm A: the swing verdict is now a BANKED FACT, but the fleet-scale -O0 carve hits a splat-integration wall (deferred, not a compiler wall)

**Context + belief.** Task 1 proved the swing pool is -O0-flag tooling (masked-MATCH at -O0), a *candidate*
(§52b). Arm A was to build the -O0-cluster split rollout (`tools/rollout_o0_cluster.py`, adapting the whale
`rollout_whale_o0.py`) and **whole-binary-gate one overlay to convert the candidate to a fact**, then roll
the ~478k-ins -O0 pool out fleet-wide.

**What was proven (byte-gated).** The tool carves the -O0 cluster (16 fns, vram 0x8013B568..0x8013C98C,
file 0x13410..0x14834) into a per-overlay `<ov>_o0.c` compiled -O0 (new Makefile `O0_CLUSTER_OBJS`
wildcard). On **`ov_SC07_010` the carve is byte-neutral and `family_sweep --hseq` banked 9/9 of the -O0
exemplar families' members whole-binary** (`func_8013B568/B7AC/B7F4/BC7C/BCDC/BD34/C360/C938/C964`), R22
clean-fleet 140/140. **So the swing verdict is confirmed as a FACT — these -O0 cluster members DO bank when
compiled at -O0, not just masked-MATCH.** (Phase 20's "`func_8013B7AC` is overlay-local" blanket claim is
also refuted at the member level — it banks in 010.)

**The wall (byte-proven, and it is TOOLING not the compiler).** Carving the SAME cluster in the other 3
sampled tail overlays (006/007/011) **byte-shifts the whole image** — a +0x20 data-symbol-address shift
(`lw v0,%lo(D_..3b6c)` → `..3b8c`), 34% of bytes differ — from a genuinely-clean `make clean &&
extract-all`. The boundary offsets are verified real fn-starts in every overlay's sig (identical 14-fn
layout), so this is **not** a wrong-boundary bug: it is a splat *re-disassembly* sensitivity — 3-way
splitting a code subseg makes splat resolve some `%lo` data references to a different auto-symbol. The
whale carve avoids it (single fn, a shared-header `_o0b` body, no INCLUDE_ASM in the split); the multi-stub
cluster carve triggers it on most overlays. This is the same "-O0 cluster split infra" that **Phase 20
built + reverted** — now characterized precisely (splat data-symbol resolution, not the compiler).

**The pivot (ROI-gated, honest).** The full -O0 fleet rollout (~1,233 members / ~0.6pp) is **deferred**: (1)
the splat-data-shift wall blocks 3/4 sampled overlays and debugging splat's `%lo` resolution is deep
splat-internals work; (2) the 134 whale-swept overlays have the cluster embedded INSIDE the
`jr_801380E0` carve (a carve-within-a-carve, even harder); (3) the remaining P29 levers are bigger and
cleaner — Task 6's tiny-IMM mega-pools (`0x80131eec` 2887×15 + `0x80130d0c` 2679×15 ≈ 5,566 members via
`imm_map`) and Task 3's core-cracks. The swing verdict's *strategic* claim (the ~478k-ins pool is real
matchable work, ceiling ≫ 3%) stands, banked-confirmed; only its *mechanical fleet harvest* is blocked on
the splat-carve integration, logged for a future session. The tool + the byte-neutral 010 carve are kept.

**Better path, in hindsight.** The whale's `_o0b` shape (a thin split whose body is a shared header
`#include`, no INCLUDE_ASM in the -O0 object) is splat-safe; the cluster rollout should likely mirror it —
route each overlay's -O0 members through a shared-per-member header rather than leaving INCLUDE_ASM stubs
in the split that splat re-disassembles. Testing that hypothesis is the cheap first move if/when the -O0
pool is revisited; it may dissolve the +0x20 shift the same way the whale never hit it.

### 2026-07-16 — P29 Task 6: the tiny-IMM mega-pools CRACKED (+4,801) — a def-signature conflict, after THREE byte-gate-corrected mis-diagnoses (R14/R35)

**Context.** The two tiny-IMM mega-pools (`0x80131eec` 2887 + `0x80130d0c` 2679 members, ~15-ins jump-table
dispatchers repeated per location) were the biggest unbanked pool (~5,116). `family_sweep --hseq` banked
**1/4966 (0.0%)** — a total block.

**Three wrong diagnoses, each refuted by the byte-gate/build (the R35 lesson, live, three times).** (1) I
first read `diff_regions`'s `O2:MATCH(0)` as "just a symbol-definition gap" and committed that finding
(`53ee97494`) — WRONG: masked_diff masks `%hi/%lo`, so a masked-MATCH cannot prove the reloc target resolves.
(2) I traced it to a "splat-local undefined symbol" — WRONG: the symbol (`D_801815EC`) is a defined `dlabel`
in the data tail. (3) The failure is a **compile** error, not a link/symbol issue: substituting one member
draft gave `conflicting types for func_8015FAAC` (cc1 exit 33), and the `family_sweep --reconcile`
(canon_sig_reconcile) path also banked **0/2470**. Only reading the *actual cc1 error* (not the masked
metric) got the truth.

**The byte-proven root cause + fix.** `src/shared/engine_core.h` forward-declares the member
(`extern void func_8015FAAC(s32 *a0);` — a shared engine fn CALLS it), while `family_remap` copies the
EXEMPLAR's signature (`void *a0`) onto the member's def → `conflicting types` → the member TU never
compiles. (The exemplar `func_80131EEC` has NO engine_core.h decl, so it banks cleanly — that asymmetry is
why the family templates in ov_SC01_077 but not its members.) Fix = **reconcile the member draft's DEF
signature to the shared-header canonical** (`s32 *a0` not `void *a0`) — byte-NEUTRAL (a pointer-type param
diff doesn't change codegen; `(s32)a0` is identical), and the whole-binary gate arbitrates anything else
(G3/P9). Implemented as `family_sweep --fix-def-sig` (`header_sig_map` + `reconcile_def_sig`, 1005 mapped
fns). Result: pool 1 **2331/2470 (94%)**, pool 2 **2470/2496 (99%)** = **4,801 members banked**, one member
hand-verified byte-identical first.

**The generalizable lesson (this is the FOURTH instance of one class).** `family_sweep` must reproduce every
build step the member's own bank requires — the §53 carve, the -O0 flag (Task 1), AND now the member's
CANONICAL DECLARATION when a shared header forward-declares it. The plain sweep's premise ("remapped drafts
are self-contained") is false whenever `engine_core.h` already declares the member with a caller-derived
signature. ~~`--fix-def-sig` should likely be default-on for the h_seq path.~~ And the meta-lesson, hammered
three times in one task: **a masked/intermediate MATCH is a candidate, never a diagnosis — reproduce the
real build and read the real error before naming the cause (R35).**

> **⛔ SUPERSEDED (P30 T4 audit, 2026-07-31) — do NOT act on the struck sentence.** Making
> `--fix-def-sig` default-on was byte-refuted by **T84 / §119**: the flag is a **REPAIR, not a default**.
> It rewrites a member draft's def signature to the shared-header canonical, which is right when the
> draft contradicts a *correct* header and **destructive when the draft is right and the header is
> wrong** — on `0x80161c98` it imposed a signedness-wrong `s32 a1` over the true `u32`, turned a
> byte-correct draft into a 1-instruction DIFF (`slti` vs `sltiu`), and **held 137 members at 0 until
> the flag was DROPPED**. Verified this session: the flag defaults OFF (`action="store_true"`, single
> consumer via `getattr(a,"fix_def_sig",False)`) and **no caller anywhere passes it**. The posture is
> correct; only this recommendation was stale. Left struck-through rather than deleted so the original
> reasoning stays legible (R31) — but a forward-looking "should be default-on" in a doc a fresh session
> reads for direction is a live hazard, not a historical note.

### 2026-07-18 — P29 jtbl 8-align wall: the half-pin was INVERTED (vacuous probes), the fix is a pad-spec filter (§8e)

**Context.** The 4 jtbl giants (`func_80131340`/`80159C84`/`8013C414`/`8013F350`, all match_one MATCH,
~2.6M agent-tokens of preserved drafts) were blocked on ONE tooling gap: banking `func_80131340` into the
shared `_jr_8012ACE0` TU produced a +4 pad at rodata 0xCC → image-wide %lo shift → SHA1 fail. The session-2
checkpoint recorded a half-pin — "cc1 AND maspsx both emit the jtbl `.align 2`; the +4 is a downstream
`as`/`ld_interleave` artifact" — and told the next session to start from there.

**What the evidence actually said (R35, again).** Both preserved probes were VACUOUS: an empty `j $31`
function with NO jump table — the `.align 2` they "showed" was the function-entry `.text` align. The honest
stage-walk (real draft spliced into the real TU, `.run/probe_jtbl/`) inverted every clause: cc1 emits
**`.align 3` before every table**; **maspsx passes it through verbatim** (the famous `maspsx.py:435` "drops
`.align`" is an inventory-only pass — the §8a-pad cookbook claim was false too); `as` bakes the pad
section-relative; and the LINK side was never guilty (`SUBALIGN(2)` + `ALIGN(.,4)` place even 4-mod-8 carve
starts tight — the banked `0xb07dc` carve proves it). Two Explore subagents produced OPPOSITE readings of
maspsx (one read the inventory pass as the output path); the tie was broken by reading the code path myself
plus one byte observable — the clean object's `.rodata` sh_addralign=8, which only a surviving `.align 3`
explains (R34: make oracles argue; R14: settle on bytes).

**The design fork and why the filter won.** The obvious fixes all fail a generality test: blanket align-demote
breaks the main EXE's island (its intra-TU pads are load-bearing); pure isolation fails multi-table functions
whose first table sits at vram ≡4 mod 8 (`.align` is section-relative, so the section-start parity flips every
internal pad); sed/as/ld have no per-occurrence mechanism. The winning shape: **replace each rodata `.align 3`
with the ORIGINAL's exact pad bytes** — derived per span by interval arithmetic from the carve config
(`pad[K] = start[K] − end[K−1]` ∈ {0,4}), emitted as a per-object `JTBL_PADS` make var, applied by a ~50-line
post-maspsx filter with fail-loud drift guards. Parity-independent, per-sibling self-adapting (each overlay's
own addresses), and structurally fleet-neutral (every pre-existing carve is single-table → no var → pipeline
byte-identical). A red-team subagent pre-verified the transform empirically (verbatim 0xE4/pad-at-0xCC vs
filtered 0xE0/tight) and surfaced 7 hardening items, including the LATENT bug that produced the original
failure (tight abutment silently merged into a bytes-impossible span) and a byte-witnessed wrong-TU splice
(`stub_file` first-match returned a stale duplicate stub — the "conflicting types" cascade was never the
draft's fault).

**Hindsight better-path.** The checkpoint's half-pin cost nothing this time because R35 forced re-derivation —
but only because the vacuous probes were LOOKED AT. The transferable rule: **a probe whose output contains no
instance of the probed thing pins nothing** — check that first, before trusting any recorded verdict. And when
a wall involves a multi-stage pipeline, walk it stage-by-stage with one byte observable per stage before
designing anything; the whole design fell out of five observables in under an hour.

---

## 2026-07-21 — Phase 29 Task-13A: the permuter's problem was TARGETING, not a missing transform

**Context + belief going in.** `docs/hindsight-study.md` §7 framed the offline endgame as *mine the
permuter's failures*: capture structured residuals, batch-diagnose them with an LLM, and route each plateau
to **missing-transform** (extend `permuter_weights` — "the highest-value bucket and the whole point"),
**seed-structural** (one LLM seed), or **genuine-wall** (file with an expiry). The implicit premise was that
the permuter is pointed at reachable work and is losing for want of the right mutation. Task 12 had plumbed
the telemetry; Task 13 was to build the classifier and the LLM autopsy on top.

**What the measurement said instead.** Before writing the LLM tier I materialised the corpus the classifier
was supposed to read — and it did not exist: 1 of 6,169 backlog records carried a `residual`, 0 carried
`passes_tried`, because Task-12's telemetry only fills records written after it landed. But 1,752 open
near-misses had their draft and their target .s on disk, so the residual was ~1 s of CPU away per function.
Recomputing all of them (21 s at -j12, through the existing `match_one` path) and classifying deterministically
gave the real distribution: **699 `redraft` · 578 `structural` · 306 `integration` · 75 `permuter` · 2 unknown.**

Of the 972 records the grinder's OWN filter admits, **75 (7.7%) are permuter-shaped.** The daemon has been
spending ~92% of its CPU on residuals a search-closer provably cannot close — 547 structural (a different
load width, an extra instruction, a flipped branch) and 348 drafts that are not the function at all. That is
the byte-grounded explanation of the Phase-22 audit's "7 banks all-time, all in Phase 21, and 0 since," and it
is a *targeting* defect, not a missing transform. Fixed for free: `grinder.candidates()` filters on the
measured bucket (1,303 → 78) and takes its directed profile from the measured class rather than the logged
label — 91% of records have no label, so the directed search had been silently running on gcc defaults.

**The pivot.** §7's ordering is now inverted for the rest of the phase: **do not run an LLM batch autopsy over
the backlog.** The deterministic classifier resolves 96% of it into three non-LLM routes, and the remaining 75
have not yet been permuted *under correct targeting* — so calling any of them a "plateau" today would be
diagnosing a search that never properly ran (the R35 failure mode: a probe from a mis-aimed instrument is not
evidence). Correct order: run the directed permuter over the 75 → collect genuine plateaus with real
`passes_tried` → only then spend the LLM, on what survives.

**Two findings that change other numbers.** (1) `closeness` conflates "one instruction off" with "this draft
is a different function"; 699 records rank as near-misses at closeness up to 278 purely from a length
artefact. They are un-attempted work misfiled as a backlog of hard functions — fresh crack fuel, and a reason
`docs/backlog.md`'s closeness ranking overstates how nearly-done the frontier is. (2) A 12-draft gate probe of
the `integration` bucket banked **1 of 12** (11 PLUMBING), so the 306 prices Task 14's reconcile ladder rather
than promising 306 free banks — stated as a measured conversion, not a projection, precisely because this
phase already over-projected once from a staged count (§57a).

**Hindsight better-path.** The corpus was one command away for months; the reason nobody ran it is that the
backlog's scalar `closeness` *looked* like a diagnosis. The transferable rule: **when a queue is ranked by a
scalar, check what the scalar is measuring on a sample before building anything that consumes the ranking** —
here, 40 % of the queue's "closeness" was a length artefact, and the tool built to consume it (the grinder)
had been quietly wasting 92 % of its work for two phases. Also: fixing the instrument surfaced a genuine
concurrency defect (`masked_diff`'s shared probe file) that had been silently dropping 0.8 % of drafts in
every parallel wave — a crashed self-check is indistinguishable from a failed draft, so it never got reported.

## 2026-07-22 (Phase 29, §61c) — the "clean-invalid jtbl bank" blocker does not exist: two reads of one polluted tree are not a replication

**Context + belief.** Phase 29's session-7 checkpoint named a single blocking finding and gated the entire
jtbl track behind it (cookbook §61c): the carve+isolation path produces a bank that is *incrementally valid
and clean-invalid*. `func_80135A4C` gated BYTE-IDENTICAL through `harvest_verify` every time, and
`make clean && extract-all && check-all` came back **139/140, `[FAIL] ov_SC06_018`, twice, identically.**
The stated implication was that the whole-binary byte-gate — the project's sole arbiter since Phase 12 —
cannot see this class of defect, *because the gate IS the incremental build* (§42b in its worst form). On
that reading, no jtbl core could be banked by anyone, and the 9 preserved cracks were frozen. It was written
up as the single next task with a precise diagnostic recipe: diff the incremental vs clean object set.

**What happened when the diagnosis ran.** It never reached the object diff, because the failure does not
reproduce. Re-applying the bank through the single-function automated path and then measuring:
per-binary clean rebuild BYTE-IDENTICAL; `make clean && extract-all && check-all` **140/140**; an
independent second full-fleet run **140/140**. The path is reproducible from committed config + source.
There is no extraction-order effect and no mid-flow asm.

**Why the original measurement said otherwise.** The failing R22 runs were taken on the tree left by the
*batch* `_jtbl_prep` — the run that ended `6 table-bearing → 1 carved, 4 isolate-FAILED, 1 stale-asm carve
fail`, i.e. a tree carrying the residue of five failed preps (stranded carves, half-applied isolations).
The per-function snapshot-restore that removes exactly that residue landed **after** those runs, in the very
commit that named the blocker (`41d65af73`). The measurement was real; the attribution was to the mechanism
rather than to the tree it ran on. The failing tree is gone, so that stays the best-supported explanation
rather than a byte-proof — but the load-bearing claim (the path is clean-invalid) is byte-refuted twice.

**The pivot.** §61c is retired; jtbl cores bank again, one draft per `harvest_verify` invocation (fault 2 —
isolation repartitions shared source, so a per-function undo is unsound in a batch — is real and stands).
`func_80135A4C` (181 ins) is banked and clean-fleet-verified; its family is 138 members / 24,978 ins ≈ 0.19pp
and the 9 remaining preserved cracks are unfrozen.

**Hindsight better-path.** "Twice, identically" felt like replication and was not: it was two reads of the
*same* contaminated state, which is one observation. A replication has to re-create the state, not re-run the
check — especially when the session that took the reading had, in the same hour, documented the tree as
polluted and then shipped the fix for the pollution. This is R35 turned on ourselves: we are disciplined
about not trusting a *tool* until it is verified, and much less disciplined about not trusting a *tree*.
The cheap guard is procedural and costs one command: **before writing a fault down as a property of a
mechanism, re-apply it from a known-clean tree.** Had that run before the checkpoint was written, the phase
would not have spent its single named next task on a blocker that was already fixed. That makes six
"structural walls" in this project that resolved to our own state or tooling (B2, SC07, pin-crash, the ~3%
-O0 artifact, the grinder targeting, and now this) — the base rate is now high enough that *the first
hypothesis for any new wall should be our own tree or instrument*, not the 1997 compiler.

## 2026-07-22 (Phase 29) — the shared byte-gate compared one binary against another binary's hash, for a month, because a default was truthy

**Context + belief.** `gate_stage` is the project's shared banking spine — the ladder every wave, the grinder,
and every manual harvest run through. Phase 29 had spent two sessions treating its verdicts as measurements:
"the ladder converts 0/10" was used to price Task 14 stages 2-3, and the Task-5 wave's "11/12 match_one MATCH,
the gate banked ZERO" was written up as three named integration walls (§61a).

**What the bytes said.** `main()` did `good_sha=a.good_sha or DEF_SHA`, where `DEF_SHA` is **ov_SC01_077's**
locked hash. Being truthy it beat `run_gate`'s per-binary `good_sha or _check_sha(binary)`, making that lookup
dead code on every CLI invocation. So the gate BUILT `ov_SC06_018` and compared it to `ov_SC01_077`'s SHA.
It cannot match. Every draft came back `"near"` — built, wrong bytes — which is *exactly* what a genuine
codegen residual looks like. **Nothing could ever bank outside ov_SC01_077 from the CLI, since 2026-06-21.**

**Why it survived a month.** The programmatic callers take a different path and were all correct:
grinder/idiom_hunt pass `None` (per-binary lookup); lora_grind/bulk_harvest pass an explicit per-binary sha;
orchestrator only ever gates 077, where DEF_SHA happens to be right. So the tool banked fine for daemons and
never for a human — and the two paths were never compared. This is R34 experienced from the inside: we had
two oracles all along (the CLI verdict and the direct `harvest_verify` verdict) and never made them argue.
The tell was visible and ignored: Task 13B's grinder banked `func_80181F78` in **ov_SC03_014** in the same
week my CLI ladder banked **0/10** on the same tree.

**Blast radius, measured not assumed.** 0 of 6,708 backlog records come from the affected path (worker 2724 /
bulk-harvest 2275 / lora-grind 766 / grinder 522 — all correct). The backlog needs no re-run. The void
verdicts are exactly the manually CLI-gated non-077 functions: the 12 preserved t5wave cracks. Re-run after
the fix, **7 of 12 now bank**, including two giants (478, 673 ins) and `func_801299C8`, which had been filed
as "PLUMBING: prototype declaration".

**The pivot.** Three "findings" are withdrawn: §61a's three integration walls, the 0/10 ladder pricing, and
the "9 compile / 0 bank ⇒ image-level effect" reading. Task 14 stages 2-3 remain unpriced — but now against
a gate that can actually bank.

**Hindsight better-path.** The defect is one truthy default, and the fix is one line — but the *detectable*
signal was a **disagreement between two paths to the same answer**, which existed for weeks. The transferable
rule: **when a tool has both a library entry point and a CLI, gate them against each other, because a
divergence there is invisible to every downstream check** — the byte-gate is a perfect correctness oracle and
a null oracle for "was the right question asked". Concretely: a gate must ASSERT that the SHA it is comparing
against belongs to the binary it just built (R32-style — assert the premise, not just the result). That
assertion would have failed loudly on the first non-077 CLI run in June. This is the fourth wall in one
session, and the fifth this phase, to resolve to our own tooling rather than gcc-2.7.2.

---

## 2026-07-22 (Phase 29, SESSION-11) — the fresh-exemplar sweep is FAMILY-SPECIFIC, not a blanket mechanical ×137

**Context + belief.** SESSION-10 closed on a strong claim: "sweeps only pay when they RIDE a fresh ×1 crack;
the discriminator is the exemplar's PROVENANCE — every success templated from a core freshly banked ×1 today,
every refusal from an ov_SC01_077 exemplar." The plan was to crack ov_SC06_018 exemplars fresh and sweep their
h_seq families (Task-5 measured ~1.59pp of fresh families for this one overlay).

**The experiment.** Two families SESSION-10 had swept 0/137 from an ov077 exemplar — `func_801365B8` (0/133)
and `func_80133AB0` (0/137) — were cracked FRESH in ov_SC06_018 (agent drafts, whole-binary byte-gated ×1),
then swept via `family_sweep --hseq --source ov_SC06_018 --allow-pins` from the fresh exemplar.

**The result (byte-gated, R22 140/140).** SPLIT: `func_801365B8` → **132/132 siblings banked** (thesis
confirmed — a fresh non-ov077 exemplar unlocked the whole family ov077 could not). `func_80133AB0` → **0/136,
reverted clean** (thesis refuted for this family — the members diverge in more than reloc symbols; no fresh
exemplar remaps into them).

**The refined finding.** A fresh ×1 crack is **necessary but not sufficient**. The fresh-exemplar sweep is
**family-specific** — the whole-binary byte-gate arbitrates each family, and on this 2-family thesis sample
the sweep rate was ~50%. The SESSION-10 provenance claim holds DIRECTIONALLY (func_801365B8 swept where ov077
refused) but overstated it as a mechanical ×137. **Consequence:** the ~1.5pp "fresh families" campaign
estimate must be discounted — cracking generates sweep fuel for SOME families, not all, and only the gate
says which. The honest yield model is (families that crack) × (per-family sweep probability), not
(families) × (reach).

**Why the difference (hypothesis, not yet exhaustively byte-proven).** func_801365B8's members differ only in
per-overlay reloc symbols over a uniform pinned regalloc → the fresh pinned exemplar templates cleanly.
func_80133AB0's members carry per-location immediate/codegen divergence the single exemplar does not share.

**Tooling caught in the act (R33 — fix the instrument).** The `family_sweep --source` override only searched
`members`, but a freshly-banked source member moves to `matched_members` after a sig-regen, so the override
silently missed it and would have templated from ov077 (a false refutation). Fixed to search both lists —
without it the thesis test would have "confirmed" SESSION-10's refusal for the wrong reason. Second fix:
`cdecl._depth0_spans` now consumes `\`-continuations, so a raw-draft `#define` macro no longer trips
audit-cdecl (the committed, cpp-expanded source was never affected — the gate was reading scratch).

**Hindsight better-path.** Test the sweep-from-fresh mechanism on ONE family before scoping a 95-target
campaign around it. The calibration cost ~1 wave and delivered a load-bearing correction to the yield model —
exactly the probe-before-scaling discipline (Phase-15/R35). Carry it forward: every fresh core crack this
phase must be followed by a gate-arbitrated sweep attempt, and the sweep's 0/N is data (per-member wall),
not a tooling artifact — provided the tools (`--source`, cdecl) are themselves verified first.

**Addendum (same session, 3rd data point).** `func_8017D648` — a MODAL/cross-address family (reach 82, a
structurally different case than the two per-location families) — cracked fresh in ov_SC06_018 and swept
**70/82 (85%)** (+3 gate-fail, +8 remap-refused "unresolved immediates"). So across three thesis families the
fresh-exemplar sweep rate is **100% · 0% · 85%** — 2 of 3 sweep, and the two that sweep do so at high rates.
The refined model: the fresh-exemplar sweep works for the MAJORITY of families at high per-family rates, with
a genuine per-member-wall minority (func_80133AB0). So the ~1.5pp campaign estimate should be discounted by
the wall-family fraction (~1/3 here) and the per-family remap-refusals, not treated as a flat 50% haircut —
and the only way to know a given family's rate is to crack it fresh and let the gate sweep it. Cracking
remains the generator; the sweep is high-yield but not universal.

---

## 2026-07-22 (Phase 29, SESSION-11) — the jtbl families are near-misses/walls, not plumbing wins; the post-carve reconcile makes the gate honest

**Context + belief.** SESSION-11's calibration wave drafted 4 jtbl families to match_one MATCH; the plan
(SESSION-11 checkpoint) billed the 3 reach-138 ones (`func_80135EB0`/`func_80135260`/`func_8012AAAC`) as
"≈+0.58pp, drafts done, just bank via the §8e carve." Banking them all failed with `conflicting types` —
looked like a jtbl tooling wall. Drew set /effort max to "fix the tooling once."

**Root cause (diagnosed, not assumed).** The jtbl carve's §8b carried-decl layer (jr_isolate_all's ambient
file-scope decls) conflicts with each draft's own externs. The reconcile chain that fixes this
(`cast_call_sites` + `reconcile_tu`, both `--src-file`-aware) already exists but runs PRE-carve against the
wrong TU — a jtbl fn's real TU is the split file, which doesn't exist until harvest_verify carves. gate_stage
had even deleted its batch jtbl stage noting "harvest_verify owns the splice." So the fix is one hook:
`harvest_verify._jtbl_reconcile` runs the same chain POST-carve against the carved TU (cookbook §62).

**The fix works — and that is exactly how it delivered a NEGATIVE result.** Validated on two functions
(`func_80135260` callee, `func_80191C50` data): both went `conflicting types` → a genuine codegen DIFF. The
plumbing was real and is now dissolved. But dissolving it revealed that all four jtbl drafts have a DEEPER
issue the plumbing hid: `func_80135260`/`func_80191C50` a real `%hi`-sharing regalloc residual (the agents'
reloc-masked match_one MATCH over-claimed it — R14); `func_8012AAAC` a def-side-arity conflict that is ALSO
fleet-shared (engine_core.h) and ALSO still DIFFs after the arity fix (a def-side register-threading wall);
`func_80135EB0` a carve `isolate FAILED`. **So the "+0.58pp from 3 reach-138 jtbl families" is REFUTED** —
they are genuine near-misses/walls needing per-function matching (re-draft/permuter), not cheap plumbing wins.

**Why this is a good outcome, not a wasted phase.** (1) The post-carve reconcile is the durable fix Drew
approved — it BANKS any jtbl family that is plumbing-only-blocked with a true MATCH, and it makes the jtbl
gate HONEST: it now attributes the blocker (plumbing vs codegen) instead of reporting every loose-typed jtbl
fn as an unbankable wall (the §26/§53-class error that manufactured two phases of wrong doctrine). (2) It
corrected an optimistic read: a reloc-masked match_one MATCH is NOT a whole-binary MATCH for a jtbl fn that
references shared symbols — the mask hides both the reloc-resolved codegen AND the carried-decl plumbing
(§58, extended).

**Hindsight better-path.** The calibration wave's jtbl drafts should have been whole-binary-gated (not just
match_one) before the checkpoint billed them as "drafts done, +0.58pp." A match_one MATCH on a jtbl fn is the
weakest MATCH signal we have (two masked layers). For the remaining ov_SC06_018 harvest: gate jtbl drafts
whole-binary early, and expect the reach-138 shared-region jtbl families to be walls (they are the most
loose-typed code in the engine). The cheaper yield is the ~83 non-jtbl targets (no carve, no §8b layer).

**Two §61-class traps re-confirmed** (both in cookbook §62): gate jtbl functions ONE AT A TIME (a mid-batch
isolate-FAIL corrupts the whole batch, `final SHA None`); and `fix_arity` on a fn in engine_core.h edits
fleet-shared state — a `git checkout src/<ov>/` restore misses `src/shared/`, and the per-overlay build stays
byte-identical so nothing flags the leak (caught here by a full `git status` + R22 clean-fleet).

## 2026-07-23 (Phase 29, SESSION-13) — the reach-138 wave on a P27-onboarded overlay is LOW-ROI; the true lever is LIVE-siblings, and the fresh families are per-member walls / def-side plumbing (R14/R35)

**Context + belief.** The SESSION-12 checkpoint's option (b): a fresh-exemplar crack-wave on a *higher-reach*
overlay (Task-5 greedy cover, ov_SC03_015/ov_SC07_006 "each ~+0.3–0.6pp of FRESH families"), billed as a
cleaner path than the drained ov_SC06_018 non-jtbl tail. Belief: ov_SC07_006's **122 draft-now reach-138 WAVE
families** (all cached, zero prefetch) are untouched fresh fuel — one crack + sweep banks ×138 (the SESSION-11
`func_801365B8 ×138` precedent).

**What happened.** `tools/build_wave_args.py` (new) emitted the top-24 reach-138 families **ranked by the fuel
manifest's `nins*reach` leverage**. Wave (`wave_binary.js`, 24 xHigh): 16 self-assessed MATCH, 8 killed by the
Anthropic session usage limit. Byte-gate: **2 via plain harvest_verify + 4 via `gate_stage` reconcile = 6 ×1
banked**; 17 failed as PLUMBING (`conflicting types for func_XXXX/D_XXXX`), 2 CC1-FAIL, 2 DIFF. Propagation:
`dedup_propagate --addr` banked **func_801325B8 → +3 onboarded-tail siblings**; func_8014A048/func_801678F0
**byte-diverge** in the SC07 cluster (kept ×1); func_8014FE60/func_80167540 **local-type-blocked** (§20 cap);
**func_80165CA0 consolidated its h_exact subgroup (+0 new)** and then **swept 0/135** via `family_sweep --hseq`
— a per-member wall like func_80133AB0 (0/136). **Net batch-1 yield ≈ 9 newly-matched functions**; ov_SC07_006
84.6% → 84.8%; fleet +0.1pp instr, ~0 distinct. R22 clean-fleet 140/140 twice (the engine_core.h arity edit
was fleet-safe); tools-health green.

**The finding (R14/R35 — verify the leverage assumption against the bytes BEFORE scaling).** The fuel
manifest's `nins*reach` leverage **badly over-counts**: a reach-138 family already matched in ~135 overlays
yields **+(live siblings)** on a fresh crack, not +138. Re-scoping the 122-fn pool by **actual live-sibling
count** (grep INCLUDE_ASM): **76 "fresh" (≥100 live) vs 44 onboarded-tail (<5 live)**. And the two classes have
OPPOSITE difficulty: the **onboarded-tail** families bank *easily* (they have a matched sibling in ~135
overlays to port verbatim — 5 of the 6 banks) but yield only +few; the **genuine fresh** families are the HARD
tail — batch-1's fresh-138 attempts FAILED as **def-side plumbing** (`conflicting types for func_XXXX`, needs
§54 `--fix-def-sig`, which `gate_stage`'s caller-arity pre-pass does NOT clear), **genuine DIFF** (permuter
fuel), or **per-member walls** (func_80165CA0). A fresh crack does **not** reliably unlock its family.

**Why this is a good outcome, not a wasted batch.** (1) The corrected lever is durable: `build_wave_args.py`
now ranks by `--rank live` and reports the fresh/tail split, so future scoping targets the true fuel and never
again mistakes an onboarded-tail family's inflated `nins*reach` for leverage. (2) It confirms — from a fresh
overlay ov077/ov_SC06_018 never sourced from — that the reach-138 *family well is largely SPENT via wave+gate*
(the Phase-26 "h_seq templatable-families thesis is byte-proven SPENT" finding, now re-confirmed on the SC07
cluster). The remaining reach-138 residual is per-function (permuter + §54), not breadth.

**Hindsight better-path.** (a) Rank by live-count AND *require a matched-sibling-to-adapt* — that combination
is what banks (the tail wins had siblings; the sibling-less fresh families didn't). (b) The onboarded SC07
overlays (006/007/010/011) are a **distinct less-shared ~84% cluster** (1549 vs 1702 distinct-code base), not
merely un-integrated — so a "sweep every matched family into them" pass will hit the same per-member
divergence batch-1 saw (func_8014A048/func_801678F0 diverge; two more local-type-blocked). (c) The genuinely
higher-ROI next move is NOT more ov_SC07_006 wave batches — it is either the per-function grind (permuter on
the DIFFs, §54 `--fix-def-sig` on the def-side-plumbing failures) or a different lever entirely (Task 7's
ROI-gated close arithmetic now has a third low-yield data point: ov_SC06_018 non-jtbl tail ≈0.1pp, this ≈9
functions). **Do NOT close P29 on ROI — the burn-down floor is still undetermined (needs 3 session-close
deltas).**

**UPDATE (same session, R35 — the probe REVERSED this verdict).** Rather than defer the header-decl reconcile
as future work, I ran the bounded probe on `func_8014CD80` (138 live, 0 matched, NO DEFINE macro; a clean MATCH
draft with a **universal body** — only universal callees + param offsets, zero overlay-local `D_*` refs). The
def-side blocker was engine_core.h `DEFINE_func_8014CD0C()` forward-declaring it `void func_8014CD80(s32,void*,
void*)` while the byte-true def is `int func_8014CD80(s32,u16*,u16*)`. **One byte-neutral header edit**
(`void`→`int`, `void*`→`u16*`; the call site passes `u16[3]` arrays and ignores the return, so codegen is
unchanged) → `harvest_verify` banked ×1 BYTE-IDENTICAL → **`dedup_propagate --addr` propagated 138/138 overlays
byte-identical (live 138→0)** → R22 clean-fleet **140/140**. **So the fresh-138 families ARE recoverable ×138 —
the blocker was purely the def-side header decl, NOT a wall.** (func_80165CA0's 0/135 was a non-universal
*body*, a different failure mode; h_exact=1 does not distinguish them — the BODY's universality does.)
**Quantified market:** of the 75 fresh (≥100-live) families, **46 carry an engine_core.h caller forward-decl,
38 SIMPLIFIED (`void`/`void*`) = the func_8014CD80 pattern** — each a candidate ×138 (≈+1.5–2.8pp instr if
half-to-most bank, gated by whether each draft's body is universal + byte-correct). **DECISION: build a
`fix_header_decl` tool** — parse the byte-true def sig (from the banked def or a MATCH draft), rewrite every
engine_core.h forward-decl of that fn to match, then hand off to the existing bank→dedup_propagate→R22 chain.
It edits fleet-shared engine_core.h → INHERITS the §61 snapshot-undo constraint (undo = restore, never an
inverse; validate FLEET-WIDE via R22). Pair it with a fresh-family wave (`--rank live --min-live 100`): the tool
is the INTEGRATION half, the wave supplies the byte-true draft. **This reopens option (b) as the campaign's best
lever — the "low-ROI, spent" read above was measured on the WRONG 24 (leverage-ranked → onboarded-tail) and is
superseded for the FRESH pool. The onboarded-tail read stands.**

## 2026-07-23 (Phase 29, SESSION-13) — the full remaining-work re-derivation + two roadmap-bucket corrections (R31, roadmap delta)

**Context.** While a crack wave ran, re-derived the ENTIRE remaining-work map from `family_hseq.json` +
`asm/nonmatchings/` (main) to answer "largest families, largest unique fns, what's MCP-automatable." This is
the re-derivation the roadmap §0 mandates (numbers rot; consume §2 only through one). It does NOT supersede the
roadmap's PHASE SEQUENCE (families P29 → main+tail P30 → behemoths P31 → flip P32 — CONFIRMED), but it
re-baselines the numbers and corrects two buckets.

**The re-derived baseline (2026-07-23).** Fleet 78.8% instr · 67.6% distinct · 88.26% fn-count. Remaining:
**overlays 40,395 stubs / 2,738,677 ins / 6,472 families** (159 reach-138 · 2,510 reach-2..133 · 3,803
singletons) **+ main EXE 2,002 fns / ~84k ins (~1,048 game-code, 0.7% done).** **CONCENTRATION: top-20
families = 23% of remaining instr, top-100 = 53%** — half the entire remaining project is 100 shared cores,
each ×N. Largest cores: `0x80176734` 371×138=51k · `0x80176218` 327×138 · `0x8013c414` 329×137 (jr) ·
`0x8014d820` 304×138 · `0x80135eb0` 289×138 (jr); ~half the top-20 are jr (jtbl-carve path).

**CORRECTION 1 — B1 pessimism partially REVERSED (cross-ref the SESSION-13 fix_header_decl entry).** B1 =
"substantial h_seq families, mechanical templating byte-proven dead, per-member cracking only." FALSE for the
def-side-blocked slice: `fix_header_decl` recovers them ×138 (func_8014CD80 proved; ~38+ candidates). The
templating-dead verdict holds for the h_seq-adapt path; it does NOT hold once the shared-header decl is fixed.

**CORRECTION 2 — B7 behemoth list is STALE/INCOMPLETE.** B7 lists 5 behemoths topping at `0x8017bf14` (4,763).
It MISSES **`0x80183814` (5,122 ins, ov_SC07_006) — now the largest single function in the game** — and
`0x8017dc1c` (1,518, ov_SC07_006). Cause: the 4 SC07 overlays were P27-onboarded AFTER the roadmap's
2026-07-15 baseline, so their singletons never entered B7's count. **P31 must rebuild B7 from
`family_hseq.json` singletons, not the roadmap's list.** (R14 caveat: verify `0x80183814` is one function, not
a mis-split, before scoping it.)

**MCP-automation finding — the prefetch gap is in the TAIL, not the top.** reach-138 families are **146/159
already cached** (draft-now, no MCP); the 6,027 tail families + main (0/2,002) are uncached. So the roadmap's
P28 "fleet Ghidra-C prefetch" is a **P30 fuel-generator for the tail+main**, NOT a P29 blocker — the top-100
(half the project) is already fuel-ready. The only MCP-dependent link left in the pipeline
(prefetch → wave → fix_header_decl → dedup_propagate) is that tail prefetch.

**Path forward (abiding by the roadmap, refined by concentration).** Stay in P29; order by ×138 byte-weight
top-down; **STOP capping the wave at 150 ins — the GIANT fresh families (150–371 ins) are the biggest single
wins and were being skipped.** Non-jr via wave→fix_header_decl→dedup_propagate; jr-half via the carve path.
Defer main+tail to P30 (the MCP prefetch pays off there); behemoths (incl. 0x80183814) to P31.

## 2026-07-23 (Phase 29, SESSION-13) — CORRECTION: fix_header_decl is fragile for SHARED multi-caller decls; gate_stage's call-site-cast is the right tool (R14/R31)

**The over-claim.** The earlier SESSION-13 entry billed ~38 fresh-138 families as a `fix_header_decl` ×138
market. On the actual crack wave (24 fresh families, 20 MATCH drafts), bulk-applying `fix_header_decl`
(v1 self-def AND v2 --reconcile-externs) **BROKE the build** (`SHA None`, CC1-FAIL across the batch).

**Root cause (byte-proven).** `fix_header_decl` rewrites a decl in `src/shared/engine_core.h`. But that decl
is SHARED by MANY caller macros, each using the fn differently. Changing the return (e.g. `s32`→`void`) breaks
a caller that USES the return (`void value not ignored`); changing a callee's params to match one draft's
loose extern is an ABI change the tool correctly REFUSES — but a bulk pass still corrupts the header.
`func_8014CD80` (the ×138 proof) worked only because it was a LUCKY single-caller / ignored-return case. The
header-rewrite lever is therefore NARROW: it is byte-neutral only when the decl change is compatible with
EVERY caller — a minority of the 38, not all. **Also:** most of these fresh families' def-side conflict is a
PER-OVERLAY-LOCAL forward-decl in the split `.c` (emitted by a matched sibling), which `fix_header_decl`
(src/shared only) never touches.

**The right tool — `gate_stage`'s reconcile ladder.** It casts the CALL SITES in the draft's OWN TU
(`cast_call_sites`) instead of rewriting the shared decl, so it never breaks other callers. On the same wave
it banked **5** (func_80175308/8012E138/80130C08/8012A1BC/80137178) where `fix_header_decl` broke the build;
plain harvest_verify banked 2 self-contained (func_8012B4B8 §52b-wall + func_80169228). **7 of 20 MATCH banked
cleanly; R22 140/140.** The other ~13 are near / deeper-plumbing, staged for a member-adapt/gate_stage pass.

**Doctrine (supersedes the "build fix_header_decl v3" next-step).** Keep `fix_header_decl` for the narrow
single-caller/ignored-return self-def case (it's cheap + proven there). For the fresh-138 integration in
general, the SPINE is `gate_stage` (call-site casts + arity pre-pass, byte-gated), NOT header-decl rewriting.
"Matching is solved; integration is the bottleneck" holds hard here — the 20 bodies matched; the plumbing is
the wall, and the call-site-cast ladder is the way through it, per-family, not a bulk header edit.

## 2026-07-23 (Phase 29, SESSION-13) — the §20 type-lift is a FLEET-WIDE collision-resolution op, not a tool-run (R32/R35; do NOT improvise it)

**Goal.** Unblock the 4 §20-capped fresh-138 cores (func_8012B4B8/80175308/8012E138/8012A1BC — "not
self-contained (local types)") via `build_engine_types` so they propagate ×138 (+552 stubs; roadmap B4).

**Diagnosis (correctly measured, after fixing a broken instrument twice — R35).** The cores reference
fleet-local struct types (`Vec8`, `Mat32`, `Buf`, `MATRIX`) that live at overlay file scope, NOT in
`src/shared/engine_types.h`. `dedup_propagate` conservatively skips any body referencing a non-shared type.
`build_engine_types` lifts such types to the shared header (byte-neutral: `--strip` removes the defs, type
decls emit no code) — the proven §19 Phase-20 lever.

**The blocker — MULTI-DEF COLLISION across the fleet.** Robust scan (`}[ ]*<T>[ ]*;`, NOT the `[^;]*` regex
that silently under-counts multi-field structs — that bug read "Mat32 = 1 copy/SAFE" when it is 138+1, and I
nearly trusted it: R35, twice): **`Mat32`/`Vec8`/`Buf` each have 2 distinct defs across 138 overlays, `MATRIX`
has 3.** The extra defs are drafter-invented simplified variants (`Mat32 = {int w[8]}` from THIS session's
wave vs the canonical `{s32 w0,w4,w8,wC; s16 h10,hpad; s32 t0,t1,t2}`). So a naive `build_engine_types
--source ov_SC07_006 --strip` would push a NON-CANONICAL def into the fleet-shared header and break the 138
overlays holding the canonical local copy. The safe scoped lift (ov07-UNIQUE types only, e.g. the named
`Cam8012E138`) unblocks ~1 core — not worth an R22.

**Why this is not a tail-of-session improvisation (the responsible call).** It is a genuine fleet-wide op: (1)
pick the canonical def PER type (resolve the 2–3-way collisions); (2) reconcile every non-canonical draft
(this session's simplified Mat32/Vec8/Buf + historical variants) to the canonical — byte-checking each (a
32-byte struct COPY is byte-neutral across layouts, but any FIELD access is not); (3) lift canonical →
engine_types.h; (4) `--strip` FLEET-WIDE (138 overlays, no make target exists — Phase-20 did it manually); (5)
R22. Rushing a fleet-shared header edit with colliding type defs is exactly the SESSION-12 corruption class.

**Bounded payoff (cookbook §20).** The lift only helps type-blocked-BUT-otherwise-clean bodies. The DOMINANT
§20 fraction is the DEF-conflict loose-typing wall — byte-proven unrecoverable by text transform, only path is
RE-DRAFTING under the caller-canonical sig. So the type-lift is real but not a fleet-% silver bullet.

**RECOMMENDATION.** Do the type-lift as a DEDICATED operation with: a correct multi-field type-scanner (the
`[^;]*` one is retired), a per-type canonical-def picker + a draft-reconcile pass, fleet-wide strip
orchestration (build the missing `make lift-types` that loops all overlays + `--exclude` the irreconcilable),
and R22. It is high-value (roadmap B4) but must be planned, not improvised. The 4 cores stay ×1 until then.

## 2026-07-23 (Phase 29, SESSION-13) — UPDATE: the §20 type-lift IS safely executable when SCOPED to clean types (lift_types.py works)

Refines the "do NOT improvise it" entry above. The fleet type-lift is NOT all-or-nothing: `tools/lift_types.py`
(built this session — reuses build_engine_types' brace-aware parser, picks each type's CANONICAL/majority def
fleet-wide, strips all local copies, R22 arbitrates) lets you lift the CLEAN subset safely while deferring the
variant-heavy types. Applied to Mat32 (138 canonical + 1 copy-only variant) + Cam8012E138 (unique): R22
**140/140**, unblocked **func_8012B4B8 + func_8012E138** for ×138 propagation. **The doctrine:** classify each
type first (`lift_types --types … ` dry-run reports distinct-def counts + variant overlays); lift the 1-def and
copy-only-variant types NOW (byte-neutral, R22-verified); DEFER the genuinely fleet-split types (MATRIX 3-def,
Vec8 180/139, Buf 3-def, M8 2-def) to a per-camp reconcile pass — they need field-access reconciliation, not a
blind strip. So: not a monolith, an incremental clean-first lift. The variant reconcile is the remaining hard
part of roadmap B4.

## 2026-07-23 (Phase 29, SESSION-13) — the BROAD 100+-type lift needs collision-vetting + precise strip; the CLEAN 2-type lift stands

Attempted #2 (broaden the clean §20 lift): discovered 102 single-def fleet-local types + added a topological
sort to `lift_types.py` (dependency ordering — a type with a value member of another lifted type must follow
it). Two edge cases blocked the broad lift, both caught by R22 (as designed — nothing committed):
1. **Name collisions.** `actor4c` (struct) vs `Actor4C` (typedef) are the SAME logical type declared under
   case-variant names across overlays — lifting both → `redefinition`/`redeclared as different kind`. A
   case-insensitive name-collision exclude dropped it (100/102), taking R22 138-fail → 1-fail.
2. **-O0 strip precision.** The last straggler (ov_SC01_077's `_o0.c`) hit a link `multiple definition of
   D_801DAA08` — stripping a type def indirectly perturbed a nearby declaration in the -O0 file format (the
   diff showed nothing removed AT D_801DAA08, so it is an indirect/format edge case in the strip span logic).

**Verdict:** the broad lift is 139/140-close but needs (a) full name-collision vetting (case-variant + struct-
tag/typedef aliasing), and (b) a strip that is exact against the -O0 file format. Both are real tool work, not
a tail-of-session push. **The CLEAN 2-type lift (Mat32+Cam8012E138, +276) stands committed** and proved the
lever; the broad lift is a follow-up with the harder strip/vet. `lift_types.py`'s topo-sort is kept (correct +
needed for any future multi-type lift). Doctrine unchanged: classify-first, lift the truly-conflict-free
types, and let R22 arbitrate — it did.

## 2026-07-23 (Phase 29, SESSION-14) — the broad §20 lift LANDS (154 types, R22 140/140); all three carried "blockers" were misdiagnosed

**Context + belief.** SESSION-13 closed the broad lift at "139/140-close, needs (a) full name-collision
vetting (case-variant + struct-tag/typedef aliasing) and (b) a strip that is exact against the -O0 file
format — both real tool work." I opened by treating that as the spec.

**What actually failed.** Fixing the instrument before trusting its readings (R35) changed all three answers:
1. **(a) is not a naming problem.** `actor4c`/`Actor4C` is a single TAGGED TYPEDEF counted twice with
   OVERLAPPING spans — 13 such pairs / 6,142 occurrences fleet-wide, 0 with a standalone tag. The
   case-insensitive exclude that appeared to fix it was a heuristic over a structural fact, and would have
   wrongly dropped the legitimate `Obj`/`obj` and `Vec`/`vec` pairs. `build_engine_types` had solved this
   correctly since Phase 26-A; `lift_types` simply carried its own copy of the model (the R33 failure mode).
2. **(b) is not strip precision.** `ov_SC01_077_o0.c` is the 1 TU of 3,226 that deliberately omits
   `engine_core.h`. Stripping its types deleted them; `multiple definition of D_801DAA08` was three steps
   downstream (undeclared type → parse error → implicit int → tentative definition → link collision). The
   link error named a data symbol that no diff ever touched, which is why the strip-span theory survived.
3. **A third blocker, introduced by me this session.** `--candidates` classifies per ENTITY but emits per
   NAME; passing `Prim` dragged in the deferred VARIANT `typedef Prim`, repointing 103 overlays at the
   header's different layout. Compiled clean; per-binary pre-filter green; **R22 37/140**, and the 103
   failures were exactly the 103 Prim-stripped overlays (set equality).

**The pivot.** Three guards, each at the point of mutation rather than in the selector: one shared
containment model (R33), a pairwise-disjointness assertion (R32), a header-visibility check, and the strip
invariant *"remove a local def only if what becomes visible is textually identical to it."* A selector bug
can no longer reach the source. Byte-grounded: R22 **140/140**, `engine_types.h` +510 lines, 2,958 files.

**Hindsight — the better path.** Two of my own measurements lied before the tools did: an `__attribute__`
regex artifact invented a "defs that also declare an object" class (zero real instances — I nearly built a
`cdecl` vetter for it), and a `while read` loop counted the literal string `check-all:` as an overlay. R14
applies to the three-line script I just wrote, not only to sub-agents. And the pre-filter lesson generalises
§61 one level down: **a pre-filter is evidence only about what it filtered** — `ov_SC01_077` passed the
Prim-broken run too. Pre-filter on a binary that FAILED.

**Deferred, named, not dropped:** 8 VARIANT entities (MATRIX 3-def, Buf 3-def, Vec8, Prim, Handler, Blk8,
V8, Prim_8016E7C8) for the per-camp field-access reconcile — still the remaining hard part of roadmap B4;
14 carried tags; 5 types kept local in the -O0 TU.

## 2026-07-23 (Phase 29, SESSION-14) — "should a fresh decomp do types FIRST?" — no for matching; yes for one cheap naming convention

**Context.** Drew asked whether doing the type work up front would help a lot — e.g. for Vagrant Story or a
fresh game decomp. Worth recording because the intuitive answer ("of course, types make code readable and
matchable") is **byte-refuted by this project three separate times**.

**(a) Types are byte-NEUTRAL for matching.** Phases 16, 17 and 18 each re-confirmed it: gcc's output is
determined by access WIDTH and OFFSET, which we read directly off the MIPS opcode; a struct definition is a
spelling convenience for `*(s16*)(p+0x24)`. Phase 16 spent an entire phase on "recover the actor struct →
matching gets easier" and the byte-gate refused it. SESSION-14 closes the loop from the other direction:
lifting 154 types fleet-wide banked **zero** new matched functions.

**(b) What types gate is SHARING, and that is architecture-specific.** BFM's economics are "match once →
stamp ×138 overlays"; a matched body naming a file-local type cannot enter the shared header, so it cannot
be stamped (§20 cap). That lever exists because 138 overlays run the same engine. A decomp without that
duplication gets far less from types-first. **For VS specifically: unmeasured.** Same compiler and CC0, so
the gcc idioms transfer; whether its structure supports propagation is an open question, not a claim.

**(c) The cheap exception is a NAMING convention, not type recovery.** Seed documented SDK types on day one
(we found THREE contradictory `MATRIX` layouts, and the 578-file majority — `{s32 m[3][3]; s32 t[3]}`, 48B —
is almost certainly WRONG versus the documented PsyQ `{short m[3][3]; long t[3]}`, 32B, which sits in only 71
files; it spread precisely because it never mattered for bytes). Then forbid bare generic type names at
DRAFT time: measured, **7 of the 8 collided names are bare** (MATRIX, Buf, Vec8, Handler, Blk8, V8, Prim),
and the one address-suffixed collision (`Prim_8016E7C8`) differs only in a member's type spelling.

**The transferable lesson** (→ the public "how to AI-decomp" wiki): the type camps were **self-inflicted by
parallel agents** each inventing a local name and layout for the same memory. Enforcing an address-suffixed
naming convention at draft time costs nothing; cleaning it up afterwards cost a session. **Types-first would
not have made one function match sooner; a naming convention would have saved most of SESSION-14.**

## 2026-07-23 (Phase 29, SESSION-14 close) — the §20 propagation cap was gating DE-DUPLICATION, not coverage

**Belief going in** (carried from Phase 19/20 and restated in the SESSION-13 checkpoint): the §20
local-type cap is "the single biggest propagation unlock" — free the capped cores and the fleet %
follows. This session tested it end to end and the bytes say otherwise.

**Measured.** The broad type-lift freed 17 propagatable cores (13 banked ×138) = **−831 stubs**. Then the
uniquify campaign (Buf → MATRIX → Vec8; 3 camps, 223 files renamed, ~1,559 local copies stripped, 5
propagations, 4 full R22 cycles) freed 6 more cores and moved the fleet by: **+6 functions, +558
instructions, −6 stubs, 0.00pp on all three headline metrics.**

**Why.** A core capped by a local type was still MATCHED in every overlay that has it — each overlay's copy
had been banked individually. Propagation replaces those N individual definitions with one shared
`DEFINE_func_*()` macro. That is a source-DRY win (and it shrinks the registry/gate surface), but it banks
no new bytes, because nothing was unmatched. **"Unblocked" and "unmatched" were being conflated.**

**Consequence for the roadmap.** Roadmap B4 ("close the propagation cap") should be re-labelled as a
maintainability item, not a coverage lever. The remaining camps (Handler, Blk8, V8, Prim, Prim_8016E7C8)
are small AND now known low-yield — do them opportunistically, never as the session's main bet. **The only
lever that moved distinct-code today was nothing: it sat at exactly 3,811,442/5,634,875 = 67.6% at open and
at close.** Fresh cracks are the sole mover of the distinct-RE number, and that is where the next session
should point.

**What the campaign IS worth keeping for:** the uniquify recipe + `tools/uniquify_type.py` (the correct
operation for same-name-different-type camps, §64a), the blocked-queue drop 13 → 7, and the R32 fix that
makes `dedup_propagate` name what it skips. Cheap to re-apply later; just not a yield play.

## 2026-07-24 (Phase 29, SESSION-15) — crack-wave efficiency audit: the bottleneck is INTEGRATION, not idioms

**Question (Drew):** waves cost millions of tokens each — how many succeed, and are we missing a new idiom?

**Measured (2 LLM waves, byte-verified):**
- s14: 24 drafted / 20 match_one MATCH / **6 whole-binary banked** / 2.52M tok.
- s15: 24 drafted / 22 match_one MATCH / **6 whole-binary banked** / 2.77M tok.
- Draft success ~**92%**; bank success ~**27%**.

**Are we missing an idiom? NO — verified against the bytes.** Re-ran match_one on 6 s15 NON-banks:
all 6 are MATCH (byte-correct bodies). The drafters find the right idioms (92% byte-correct C). The
functions that don't bank are **byte-correct-but-unintegrated** — the whole-binary build rejects a
declaration/type conflict (def-side sig, data-extern type, unshared struct), NOT a wrong instruction.
The only recurring genuine codegen residual is the phantom-frame schedule class (~2/wave) — minor.

**The waste:** each wave produces ~22 byte-correct functions and banks 6, **stranding ~16 paid-for
correct functions**. 461K tok/bank now; if integration recovered all 22, ~126K tok/bank — a **3.7×
efficiency gain for ZERO additional drafting tokens.** We already bought the correct code; we throw
away 73% of it at the gate.

**Conclusion / next investment:** the bottleneck is INTEGRATION AUTOMATION, not drafting and not idioms.
Pointing more tokens at drafting strands more correct functions. Build a **fleet-safe integration-recovery
pass** (def-side sig reconcile + data-extern reconcile + auto struct-def lift, R22-validated) that runs
after the gate — NOT fix_header_decl (fleet-blind, §63). It ~3.7×'s the yield of every wave, past and
future. Stop waves; invest in integration tooling first, then resume at ~3× efficiency. (Drew stopped the
waves on this instinct — the data confirms it.)

## 2026-07-24 (Phase 29, SESSION-16) — the integration-recovery pass: measured, and the §20 DEF-conflict wall refuted per-overlay

**Belief going in (SESSION-15 audit).** Waves bank ~27% of drafts; the ~73% stranded are "byte-correct
bodies stranded by plumbing", so a fleet-safe integration-recovery pass would **~3.7×** every wave's
yield for zero new drafting tokens. Drew stopped the waves on that finding and this session was
chartered to build the pass.

**What measuring first changed (before any recovery tooling existed).**
1. **The premise was ~2/3 true, not true.** Among the 36 STRANDED drafts, `match_one` says **24 MATCH,
   11 near, 1 ERR**. The "~92% byte-correct" was a whole-wave figure; the residue is **67%**. The 11
   `near` are unfinished drafts — and they are precisely the ones that compile in their real TU and
   DIFF. So the recoverable fuel was ~24, not ~36, and the 3.7× was over-stated at the source.
2. **The blocker was not the class the existing ladder targets.** Running `cast_call_sites` +
   `reconcile_tu` over all 36 clears `callee_decl` 19→3 and `data_decl` 16→**0**, and converts **1 of
   36** to compiling — which then DIFFs. §61d verbatim. The dominant blocker is the **shared-header
   self-decl (21 of 36)**, which no draft-side transform can reach.

**The move that worked, and why it was available.** `fix_header_decl` is off-limits (fleet-blind: 3/3
per-binary then R22 139/140). §20 had concluded the DEF-conflict class is "byte-proven unrecoverable by
text transform" because the shared macro's `extern` is the only declaration in the 137 stub overlays and
"can't be edited per-overlay — it's in the shared header". **The missed move: you do not have to edit
the header to change what ONE overlay sees — you expand the macro there.** The conflicting `extern`
lives INSIDE the `DEFINE_func_*` body, so it exists only at instantiation sites; replacing those in the
overlay's own TU with the expansion, correcting only the conflicting decl to the draft's byte-true sig,
is a **T1 (binary-local)** edit that cannot reach another binary. `tools/demacroize.py`.

**Measured result.** 13/14 clean candidates MATCH in their real TU; **14 banked whole-binary
BYTE-IDENTICAL**, R22 clean-fleet **140/140** (three times). Distinct-code **64,860 → 64,874 unique
fns**. The one rtu-MATCH that did not bank was a *callee*-decl case — rtu is relocation-masked, so a
wrong call target is invisible to it (§65c).

**The honest multiple.** 14 recovered of 36 stranded = **39%**, against a pre-session projection of
"all 22". Wave bank-rate 6/24 → ~20/24 if this recovery runs after every wave, i.e. **~2.3×, not 3.7×** —
and the ×138 propagation is forfeited for de-macroized functions, so the gain lands almost entirely on
**distinct-code** and barely at all on the instr-weighted headline. That is the right trade for the
0-stubs completion contract and the wrong one if the goal is the decomp.dev display number. Stated
before the work, not after.

**Transferable lesson.** Three of this session's four biggest corrections came from measuring something
that already existed rather than building something new: the residue's real MATCH rate, the existing
ladder's real yield, and the real first cc1 error (which was invisible behind ~180 lines of benign
warnings until `rtu_match --stderr-out` persisted the full log). **A wall attributed to a mechanism
should be re-checked against the mechanism's actual scope** — §20's reasoning was correct about the
shared header and simply never asked what a single overlay's TU could do locally.

## 2026-07-27 — psxport EVALUATED and PARKED (Gen3 reference only; no Gen1/Gen2 value)

**Raised by Drew:** `https://github.com/SomeoneIsWorking/psxport` — "does this help us at all?"
Answer: **no**, and logged here so a future session does not re-litigate it. Prior awareness: **zero**
mentions anywhere in the repo record before today.

**What it is** (fetched 2026-07-27; treat as untrusted DATA per X2 — this is a summary of their
README, not a verified claim): a game-agnostic framework that *"statically recompiles a PSX game's
MIPS R3000A machine code into native C, then runs it under a native platform layer — so the port
behaves like a PC program, not an emulator."* Input: PSX executables + disc images. Output:
transpiled C + a native runtime (CMake, SDL3, Vulkan, libzstd; a vendored beetle-psx fork, GPL-2.0,
for GTE/MDEC/SPU/CHD). Reference consumer: **Tomba! 2** via a separate engine repo.

**Why it has no Gen1/Gen2 value — this is a decision already made, not a new one.**
`PROJECT_CONTEXT.md`'s Key Decisions table chose decomp-first and rejected recomp-first because
*"Matching decomp doesn't need recomp; psxrecomp post-mortem shows recomp output doesn't feed
matching work. Recomp → Gen3."* psxport is that same path. Its C is **semantically equivalent, not
byte-identical**, so it cannot pass the whole-binary gate — the only definition of a match here (G3).
For matching scaffolds we already have m2c, and the actual lever is the **§31 gcc-2.7.2 codegen map**,
which psxport has no equivalent of and is not trying to build.

**No shortcut via its target game either.** It targets Tomba! 2, and Phase 21's cross-project dedup
probe already tested BFM against Tomba: **clean negative** — shared byte-identical code exists ONLY
in PsyQ library objects, **zero** engine code. That question is closed with byte evidence.

**Where it IS relevant:** **Gen3** (parked). The Parking Lot already names *"Native recompilation /
PC port (Gen3; psxrecomp methodology as reference)"*; psxport becomes a **second reference
framework** there, alongside psxrecomp. Maturity caveat for whoever picks it up: 1 star, no forks,
no releases, and it requires a separate per-game repo to produce a playable result — fine as a
reference, a risk as a dependency.

**Verdict: PARKED for Gen3. Do not evaluate again before Gen2 exit.**

---

## 2026-08-04 (P30 S1e) — "distinct-code FELL" was a stale digest, not a regression; the alias lever is UNGATED

**Context + belief.** S38 closed with the def-side asm-label alias cracking a 208-conflict class
138/138 — the phase's best-performing lever. Its checkpoint then gated it in bold: *"distinct-code
FELL 89.3 → 89.2 — UNEXPLAINED. Do NOT scale the alias lever until it is resolved. The BYTES are
proven (R22); the ACCOUNTING is not."* The recorded lead was `progress.py:423`'s `SIG` regex booking
`void aF80146A6C(…)` under the alias name — with the honest caveat, written at the time, that a pure
naming artifact would move fn-count and distinct-code *together*, and these had diverged.

**What failed.** Two mechanisms, both mine, both asserted before being derived (R14):
1. *The recorded lead.* Real blindness — but it feeds `classify()`, which computes **fn-count only**.
   Neither weighted metric ever sees a C identifier; they derive from `matched = sig − corpus.stubs`.
2. *"The harvest reverted functions to INCLUDE_ASM."* Attractive because that is **byte-neutral**
   (INCLUDE_ASM pastes the original asm), so R22 would stay 140/140 across genuine coverage loss —
   the R34 blind spot. Refuted by one grep: **483 stub lines removed, 0 added.**

**The pivot.** Stop hypothesizing; prove the arithmetic. Identical sigs (both denominators unchanged)
+ unchanged `tools/` + zero `INCLUDE_ASM` additions ⇒ HEAD's stub set is a strict subset of the
prior commit's ⇒ HEAD's matched set is a superset ⇒ **both numerators are forbidden to fall.** A
reported fall is therefore a statement about the *digest*, not the tree.

**The byte/measurement-grounded why.** Reconstructing each commit's stub set from its own committed
tree (0 unresolved symbols):

| | instr | distinct | unique fns |
|---|---|---|---|
| `10f954627` **true** | 12,394,533 | 5,022,306 | 77,895 |
| `10f954627` *as committed* | 12,402,412 | 5,029,324 | **78,025** |
| `HEAD` true **= committed** | 12,405,402 | 5,025,082 | 77,952 |

True delta: **instr +10,869, distinct +2,776 ins / +57 unique fns — everything rose.** The
`10f954627` digest was **committed stale** (generated from a working tree still holding work that was
reverted before the commit landed; overstated +7,879 ins / +130 unique fns, never regenerated). The
next honest digest was lower than the stale one, so the metric *appeared* to fall.

**Consequences (what changed).**
- **The alias lever is UNGATED** — the blocker was a phantom. It is the phase's cheapest large lever
  and should be scaled (S4 onward), subject only to the §61 small-batch discipline the two real R22
  failures taught.
- `progress.py stub_addrs` no longer swallows `corpus.stubs`' refusal. The old bare `except` turned a
  fail-closed oracle into a guess: byte-witnessed reporting **instr 100.00% / distinct 100.00%** in a
  tree with no `asm/`.
- **`make audit-digest` is new** (in `tools-health`, after `report`): recomputes the three headline
  metrics from the current tree and fails if the committed digest disagrees. Compares **integers, not
  percentages** — the +7,879-instruction staleness printed as "94.4%" both before and after.
- The same swallow was found twice more in the **integration spine** (`cast_call_sites.tu_for`,
  `reconcile_tu.tu_for`), where it silently reconciled drafts against the default `<ov>.c` instead of
  the jr/-O0 split TU — the exact bug `cast_call_sites`' own docstring says it exists to fix. Both now
  propagate. Given the phase's ~24k PLUMBING vs 4,917 DIFF base rate, that class presents as a
  codegen wall.

**Hindsight — the better path.** The three greps (denominators / `tools/` diff / `+INCLUDE_ASM`
count) cost under a minute and settle the question *before* any hypothesis is formed. The general
form, now cookbook §140: **a committed number is a claim about a tree; if it cannot be recomputed
from that tree it is not evidence, and it must never gate a lever.** The deeper repeat is that this
is the *fourth* consecutive phase where a "wall" resolved to our own instruments — and this time the
instrument was the scoreboard itself, which is the one nobody thought to audit because the byte-gate
is green over it by construction (R34).

---

## 2026-08-04 (P30 S39) — Drew's MASTER_REMAINING_FUNCS proposal: adopt the goal, reject the mechanism

**Context.** Drew, mid-session: *"should we build a full list of all funcs in the entire game… scan
every single file and use ghidra to verify all funcs, a total list. search for missing overlays,
missing mains… then check all of our banked funcs and create a third list MASTER_REMAINING_FUNCS =
total − banked. Each new func we bank, we must remove from the second list."* Explicitly flagged as
thinking-ahead, not a work order.

**Assessment (recorded so it is not re-litigated from scratch).**

**Lists 1–3 already exist, DERIVED, and are recomputed on every read:**
- **total** = `.run/sig.*.jsonl` (`sig_image`/rabbitizer over the ORIGINAL bytes, independent of
  splat): 87,459 distinct fns / 13,141,652 instructions across 140 binaries.
- **banked** = never stored — derived from the proven invariant (`INCLUDE_ASM` pastes the original
  asm ⇒ a function not wrapped in it is byte-exact): `matched = sig − stubs`.
- **remaining** = `corpus.stubs(binary)` — 13,493 open stubs, derived from the filesystem.
- **"remove it when we bank it" already happens**: banking IS deleting the `INCLUDE_ASM` line, which
  is the same act as leaving the remaining set. There is no second bookkeeping step to forget.

**Why the MAINTAINED form is the one thing to avoid.** A hand-updated master list is precisely the
shape the Phase-26 audit deleted ~10 of (file allowlists, `REGION_SUB`, `func_`-only regexes). The
canonical failure: `.run/fuel_manifest.json` recorded **130 live stubs when the truth was 30**, hiding
**91.6%** of remaining gain — and nobody noticed, *because a target that is never nominated produces
silence, not an error*. R33 exists for this ("the best outcome is a DELETED SCANNER"). A
MASTER_REMAINING file would drift silently, and always in the flattering direction.

**Where the instinct is RIGHT, and the work is genuinely open — the DENOMINATOR:**
- Phase 27 found **4 hidden SC07 overlays** invisible for a month (code at PAC entry 1); onboarding
  them moved the honest headline 68.9% → 67.0%. We had been grading against an incomplete game.
- **39 type-1 code modules remain un-onboarded** (load-address RE pending; roadmap bucket **T**,
  owned by P31). Until they resolve, **no 100% claim is meaningful** (R34/`disc-completeness.md`).
- **`main` has NO independent boundary oracle**: `sig_image` cannot sign a PS-X EXE, so main's
  function list rests solely on a Ghidra sig 7 weeks stale and missing 757 of 2,002 stubs.

**Caveat on the Ghidra half:** for overlays it is **partly circular** — those programs' boundaries were
seeded FROM splat by `DefineFunctions.java`, so Ghidra would largely confirm splat to itself. The
independent oracle is `sig_image`, and `make audit-corpus` already makes the two argue (that is how the
193 `listCdBuffer` phantom slices became visible). **main is the exception** where Ghidra is all we have.

**VERDICT: adopt the goal, reject the mechanism.** Do NOT build a maintained list. Build a **derived,
coverage-asserted RECONCILIATION** — one command computing total/banked/remaining from the oracles that
**fails when the independent views disagree**. That is the actual gap: "what's left" is answered today by
`corpus.stubs`, `worklist.md`, `backlog.md`, `family_hseq.json`, `fuel_manifest.json` and
`frontier-p30.md` — each individually derived, **never cross-asserted**. P30's T0 had to hand-reconcile
exactly that (family_hseq 29,961 vs progress.py 28,296, an unexplained R32 gap). A gate that refuses to
be green while they disagree catches the next one for free. Same shape as `audit-digest` (S39) and
`audit-binaries` (R36). **Sequencing: P31's opener (it is already the T-bucket phase), or a P30 close
item if the honest denominator is wanted before the next re-baseline.**

---

## 2026-08-05 (P30 S6/S41) — the definitive disc audit: assert a PARTITION, don't extend a list

**Context.** Drew: *"im getting tired of learning there was more code all along, we really need a full
audit that definitively lists ALL code that we need to decomp to complete this game."* Justified —
three separate discoveries in three phases, each a real expansion of the denominator:

| when | what was found | how it had hidden |
|---|---|---|
| P27 | 4 SC07 overlays (136→140) | the extractor globbed `0.4.dec`; their code sits at PAC entry **1** |
| P28 | type-4 row was vacuous for 138 known binaries | `disc_code_sweep` decoded only RAW bytes — **blind to compressed code** |
| P27 | **39 un-onboarded type-1 code modules** | resident-class; each loads at its OWN address, so not mechanically onboardable |

**The diagnosis (this is the part worth keeping).** Not one of these was a wrong answer. Each tool was
**correct about the subset it examined and silent about the rest** — a glob, a decode layer, a
4,096-word window. Extending any single list would have produced the same class of surprise again.
Measured while writing this: the current sweep windows at 4,096 words, so **782 of 1,328 PAC payloads
are only partially classified (~55.5M words never examined)**. Almost certainly data — but *nothing has
checked*, which is exactly the shape of all three findings above.

**The decision: assert a PARTITION over the disc, not a list of code.**

> Every byte on the disc belongs to exactly ONE bucket — onboarded-code / classified-data /
> audio-video / filesystem-metadata / unused — the buckets sum to the disc, and **residue is a
> DEFECT** (R32).

Once a gate enforces that, "more code all along" becomes structurally impossible: a further discovery
would have to come from outside the disc image. This is the same move as `audit-digest` (S1e) and
`audit-binaries` (R36) — the two gates that ended the metric and citizenship surprises — applied to the
denominator itself. **A partition with an asserted residue of zero is a completeness proof; a longer
list is only a longer list.**

**Three layers (tasks #10, #11).**
- **L1 static partition** — walk from the DISC IMAGE, not our configs; every ISO file → `.CD` sub-file
  → PAC entry → **both** raw and decompressed layers; classify WHOLE payloads (no window); emit
  `docs/disc-ledger.md` with per-payload `claimed-by <binary> | UNCLAIMED`; assert the sum.
- **L2 second oracle (R34)** — today's code test is a heuristic (`valid ≥ 0.90` AND `jr $ra ≥ 0.01`);
  a small code payload can fall below 1% `jr` density. Cross-check with `sig_image` boundary carving;
  disagreements become the review queue.
- **L3 runtime census** — static analysis says "looks like code"; only the emulator says "was loaded to
  X and executed". A scripted PCSX-Redux tour logging every load (payload → RAM addr → len) and every
  executed PC range.

**Why L3 is sequenced with the type-1 onboarding rather than after it:** it is the *same run*. The 39
modules are blocked on load addresses that only runtime RE can give (P9 — a build binary needs its
address to byte-verify), and the census needs the same instrumentation. Instrumenting it to log EVERY
load rather than only those modules makes one pass deliver the onboarding data **and** the completeness
proof. Doing them separately would pay for the tour twice.

**Expected direction of the number, stated in advance so it is not read as a regression:** onboarding
the 39 RAISES the denominator and LOWERS the headline %, exactly as the main sig regen did today
(94.5→94.4) and the P27 overlay find did (68.9→67.0). **"100%" is not claimable until the 39 are
onboarded-and-matched or explicitly excluded with a stated reason** — already in the completion
contract, and this makes it enforceable rather than remembered.

**Sequencing (Drew's call):** finish the serial crack queue → L1+L2 (cheap, deterministic, and they
sharpen L3's target list) → L3 + type-1 onboarding. Fold into **P31**, which already owns bucket T.

## 2026-08-06 (P30 S44) — the 78-payload campaign: static addresses dissolve the emulator dependency; "modules" mostly dissolve into overlays

**Context + belief.** `make audit-disc` (S43) enumerated 78 unclaimed code payloads (~3.4 MB). Standing
doctrine (`disc-completeness.md`, from P27): these are "type-1 modules" whose load addresses are "only
knowable by runtime RE" — so onboarding was gated on an emulator session (L3), and the completion
contract carried them as a 39-module backlog.

**What the measurement said (3 read-only agents, byte-verified).** (1) The load addresses are STATIC
for 46 of 78: the EXE's `loadDestPtrTable` + boot literals + two index tables inside the resident +
`resident.c:641` + the SC07 pair's own headers give every MAIN payload and the SC07 pair a derived
address, corroborated by two independent corpus-side voting methods at ~500:1 margins
(`memory-map.md` §S44). (2) The three biggest "modules" are ORDINARY OVERLAYS stored uncompressed
(type 1 = raw overlay, type 4 = LZSS) for the standard 0x80128158 slot — ~75–77% of their functions
h_exact-identical to the onboarded corpus, 802 genuinely novel across all three. (3) The remainder
tiers honestly: 35 small actor modules at two statically-known ping-pong slots; SC07/3+4 at their own
slot; 28 script modules (7 × 4 per-disc builds) + 4 stragglers genuinely runtime-determined.

**The pivot.** L3 shrinks from "the onboarding prerequisite" to a small runtime-confirm pass (28+4
payloads + R34 verification of the static addresses). The campaign inverts: tooling updates → onboard
the big 3 through the EXISTING overlay machinery → dedup-bank the h_exact majority → batch the small
modules — all emulator-free. The "new binary class" tooling burden collapses to: `config/modules.mk`,
a de-ov_'d R36 gate, a vram-derived `family_remap`, glob widenings, and a parameterized
`new_binary.sh`. Full per-tool table: `tooling-audit.md` §S44.

**Why this was missable for 30 phases.** Each prior tool was correct about its subset and silent about
the rest (the audit's founding observation) — and the doctrine layer had the same shape: the P27
"only knowable by runtime RE" sentence was true of the tools that existed then, and nobody re-derived
it after the loader cluster was matched (the tables were sitting in matched C + the resident's own
bytes). A confident negative doctrine is a claim like any other — date it, cite its evidence,
re-measure before letting it gate a campaign (the §146/§147 lesson at doctrine scale).

**Hindsight better path.** When Phase 3 T5 wrote "entries [1]+ are runtime-indexed (no static xref)",
the honest follow-up was a named open question ("WHERE do the indices live?") rather than a doctrine.
The answer was one grep away once the resident was matched in Phase 12.

## 2026-08-06 (P30 S45) — Part II lands: the module fleet onboards emulator-free; the denominator is now partition-complete minus a 34-row parked ledger

**Context/belief.** S44's Part I proved the campaign shape on the big 3; Part II was checkpointed as
"mechanical" — onboard the 35 small MAIN modules + the SC07 pair, dedup, verify, retire the
superseded tools, re-baseline.

**What happened (byte-verified).** All 40 modules (38 MAIN + SC07 pair) built **byte-identical on
their FIRST build** at the §S44 static addresses — zero parked, byte-corroborating the loader table
across all four slots (A/B/boot/SC07). The dedup measure came in exactly as predicted (LOW): 69/1,113
module fns h_exact-match matched corpus code; the scoped `family_sweep --hseq` banked **408 members**
(182 into modules, 226 into the big 3 — families Part I's `--only` scoping missed). R22 183/183;
`audit-disc` UNCLAIMED 75→34 at residue 0.

**What the checkpoint's "mechanical" hid (the session's real work — five instrument findings).**
(1) A module header can carry a function's JUMP TABLE → the hdr carve must be a dot-typed `.rodata`
PAIRED with the c segment (standalone rodata emits cross-object `.L` refs; `bin` links in the data
block). (2) The A4 symbol-window law bit again: `symbols.resident.txt` in the boot trio's stacks
minted a phantom `DsMix` function inside md_MAIN_011. (3) `family_sweep --hseq`'s stub map globbed
`sig.ov_*` only — every module member silently "not-stub" (the I.1d widening class; the tool sat on
the audit's "auto-OK" list). (4) `--bootstrap` sig boundaries GLUE adjacent functions around
jtbl-dispatch code — 24 false TRUNCATED slices; `sig-modules` now seeds from the built ELF's
`func_*` symbols. (5) `corpus.audit` counted jtbl `.word` lines as instructions and `progress.py`
left `INCLUDE_RODATA` symbols unbucketed — both R32 holes the new module layout exposed. Every fix
carries a negative control.

**The pivot/state.** The completion contract's denominator is now **183 onboarded binaries + a
34-row parked-for-L3 ledger** (28 script modules + SC02/9 + MAIN/7/9 + **SC03/53/54/56** — three
rows the S44 exploration never tiered, found by the audit's arithmetic refusing to close). Honest
baseline: **94.0% instr / 95.96% fn-count / 87.6% distinct** on the grown denominator (was 94.4%
instr over 143 binaries — the headline fell because the game grew, the honest direction). The ~931
module stubs + ~1,700 big-3 novel stubs are ordinary crack-wave frontier; L3 shrinks to a bounded
runtime-confirm pass (P31 bucket T).

**Hindsight.** "Mechanical" batches over a NEW binary layout are where instrument blind spots
surface — the five findings above were all invisible until 40 same-shaped binaries went through the
pipeline in one afternoon. The S44 plan's per-tool audit was right to exist and still under-reached
(family_sweep was "auto-OK"; the plan's "zero build refs" for the retirees was wrong for 3 of 7).
The discipline that worked: R37 probe-first (one module before 29), negative controls per fix, and
the R22/audit ladder after every batch.

## 2026-08-07 (P30 S45, part 2) — the L3 tour: a live emulator session run as a measurement campaign

**Context.** The S45 module campaign left a 34-row parked-for-L3 ledger. Drew was available; we
ran the emulator session same-day instead of deferring to P31 — emulator-first was the right
sequencing call because his availability was the scarce input and the ledger was the last
denominator unknown.

**The instrument.** The retail debug menu, summoned by forcing `gameMode=7` in per-frame writes
over the Redux web API (two instrument lessons the hard way: hex `offset` params are silently
parsed as 0 — my first 150 "writes" landed in kernel space and a no-op 200 had "verified" the
write path (R35: a no-op is not a control); and a single write never latches — the game rewrites
the mode every frame, which is WHY the original GameShark code is a constant-write). Drew
transcribed the full AREA/SCENE list by hand (docs/debug-menu-list.txt) — flying blind ended and
10 targeted loads replaced ~150.

**The arc.** One accidental pre-crash capture (SC03/76 during an INN load) was the only positive
for an hour of scene/dialogue/flag probes — until the INN replication cracked the law: city
INTERIORS stream script modules, member k ↔ interior k, AREA selects the chapter. After that, 27
modules fell in ~20 minutes of menu-hopping at four byte-verified slots. MAIN/3 was discovered
FIRST — the very first snapshot showed id 0x39 at the resident slot during the main menu — the
audit's classified-data bucket had hidden a 121 KB module both oracles missed; the id-word census
then proved it was the only such miss.

**What refused to appear.** MAIN/7/9, SC02/9, SC03/53/54/56 — parked with per-state negative
evidence. A crash mid-tour (dynarec at the kernel vector — wild copy from a state-mismatched
scene load) cost nothing: the frozen RAM still held its capture, and the savestate-hub pattern
made later crashes ~20-second events.

**Banked same-session (R30):** 29 onboardings byte-identical on first build; fleet 212; R22
212/212 (after THREE catches on md_MAIN_003 — the A4 resident-symbol leak again, then an
extract-order-sensitive splat boundary that the bytes resolved as a data-sentinel-in-text +
function at +4, now pinned in the curated symbol file); audit-disc UNCLAIMED 34 → 6 at residue 0.

**Hindsight.** (1) A human with a transcribed menu beats an agent guessing scene semantics — the
DEBUG-MENU-LIST was the session's force multiplier. (2) Negative results with evidence are the
product: six payloads now carry "never loads in X/Y/Z" instead of "unknown". (3) The write-API
no-op control was a real R35 miss — verify instruments with a VISIBLE effect. (4) R22 catching
md_MAIN_003 three times in one evening is the rule working exactly as designed.

## 2026-08-13 (P30 S50) — the A-prop conversion gap: the experiment the artifacts had already run

**Context and belief.** S49's closing checkpoint made one thing the session's first job, on the
grounds that it "prices everything else": re-gate the 42 (truly 35) unbanked A-prop drafts ONE PER
TU, to test §170's hypothesis that family-batched cards concentrate members into a single
destination TU and die of the §169 collision. 320 further batched members — roughly 15M tokens of
wave — were explicitly held behind that measurement. The belief was that A-prop's 91%-agent /
57%-gate conversion was an integration-topology problem.

**What failed.** The hypothesis, and the framing that made it look like it needed an experiment.
Three artifacts already in `.run/` answered it: `gate_aprop1.json` records **5-draft single-TU
groups banking 5/5** (batch size is not the discriminator); **11 of the 35** unbanked drafts were
already single-draft groups, i.e. the proposed test had already been run on them; and
`harvest_failed.ov_SC03_107.classified.txt` names the actual failure verbatim — `PLUMBING:
undefined reference to 'D_80181900'`, eleven times. §169's own law ("read the classified file
before theorising about any sweep failure") was written for exactly this and was not applied to the
sibling lane.

**The pivot.** Diagnose from the recorded verdicts first; run the experiment only for what they do
not answer. A 40-line static audit — compare each draft's vram-suffixed symbols against the symbols
the target's own `.s` relocates — classified all 35 in under a second: **24 stale-seed-symbol, 11
genuine DIFF**, zero ambiguity. One probe banked, then 22 of the remaining 23.

**Why (byte-grounded).** A per-location data symbol is the seed's ENVIRONMENT, not its logic.
`match_one` compares instruction encodings and is blind to a relocation's target NAME, so a carried
symbol scores MATCH standalone and fails at link inside the host TU. A-prop's real conversion is
**87% (79/91)**, not 57% — the lane was never the problem, and the 320 held-back members are worth
substantially more than they were priced at.

**Hindsight — the better path.** The cheap deterministic audit should have been part of the wave's
verification step from the start: it needs no build, it is the second oracle (R34) for the one class
`match_one` structurally cannot see, and it would have converted these 24 in S49 instead of leaving
them to be re-measured a session later. Generalized rule: **before designing an experiment to
explain a failure rate, grep the failure verdicts the tools already wrote** — and when a checkpoint
declares a test the top priority, that is a hypothesis with a plan attached, not a finding (R14/R35
applied to my own handoff notes).

## 2026-08-14 (P30 S50-Max) — the ordered finish of func_8017C294: how a Max-effort failure gets banked

**Context.** Drew set Max and ordered: read the gcc source, finish cracking `func_8017C294`
(NEAR 2/246, ×16 reach, ~90 prior refutations), and document the idioms from the three families
worked. The residual: the target frame carries 32 bytes of never-referenced spill slots the draft
lacks — invisible dead compiler state, not code.

**What the source reading changed.** Five files deep (combine/cse/reload1/caller-save/mips.md +
toplev/function), the session replaced inference with mechanism: the complete list of
never-referenced-slot producers (combine USE-orphans; eager caller-save areas — a discovery, with
`-fcaller-saves` on at -O2; per-hard-reg spill slots), the exact alignment arithmetic, the orphan
rule with its `(set (reg:HI) (subreg (reg:SI)))` rewrite, and two proofs by construction: opacity
that defeats cse equally blinds combine (`num_sign_bit_copies`), and cross-jump cannot delete
slot-bearing code. R35 was applied to my OWN S50 verdict: the universally-quantified
"impossible" claim was re-tested from scratch — including the one test nobody had ever run, the
actual whole-binary gate on the NEAR drafts (verdicts held), a 200-variant randomized structural
sweep (one new vars-moving dimension found: swapped-arm recomputes, cost ~1:1 in real code), the
inline-function axis (collapses the chain — proving the source is textual macros), and the cc1
flag axis (invariant).

**The honest outcome.** Not cracked. The wall is real and now sharply bounded: the missing 32
bytes require structurally different source with coincidentally identical bytes — a haystack
outside systematic derivation. Floor stays NEAR 2. Parked for P32 with a complete siege kit.

**The transferable lesson.** A Max-effort "finish it" on a hard wall should end in one of exactly
two states: the crack, or a mechanism-complete refutation that future work can stand on. The
difference between this wall-verdict and the cheap kind: every claim in it is either a source
citation or a byte-measured probe, the instruments (cc1_dumps.sh, sweep_gen.py) outlive the
attempt, and two NEW reusable decompilation tells (§172a: lhu/lh typing; macro-vs-inline
redundancy) came out of the failure. Failures bank too, if you make them pay rent.

## 2026-08-14 (P31 open) — the re-charter: organize the frontier before grinding it

**Context and belief.** Phase 30 closed the overlays at their measured ceiling (95.3% instr,
213/213 byte-identical, 12,059 stubs left). Roadmap-v2's P31 chartered "Scope-Complete + Main &
Resident" — in practice a per-function agent grind over main (~79.5k weighted ins priced at
~490 tok/ins) plus queue-consumption. The standing belief was that the remaining mass was
organized as well as it could be (h_seq families + the S49 cousin tier) and only grinding remained.

**The pivot (Drew, at the Phase-Start gate).** Do not grind blind. Read the cookbook, the gcc
source, and the PsyQ material deeply; characterize every remaining function by the compiler
behavior that dominates it; group the thousands of "unique" functions into CRACK GROUPS so one
exemplar (or a near-2..6..20) carries a whole group; widen the mechanical tooling to accept more
near-misses, more permutations, more structural/length differences; token efficiency is a
first-class constraint — deterministic zero-token lanes first, agents only for exemplars and
genuinely novel classes. Milestone shape: campaign to ceiling (the P30 pattern). Main fully
included from day one.

**Why (measurement-grounded, from the plan-mode exploration).** (a) The cold tail — 3,238
units / 215k ins, 37% of the non-main remainder — has NO grouping at all below cousins@0.85,
and the only sub-exact similarity metric in the repo is one SequenceMatcher tier; no CFG/frame/
tell features exist anywhere. (b) The LEN wall that refuses every length-drifted member is ONE
LINE (`classify_member` → LEN), and the measured dominant drift class (li-expansion, 25/86
near-pairs) is mechanically resolvable. (c) The biggest classified failure class is PLUMBING
(1,217 distinct fns, ~204 symbols) — declarations, not codegen — plus 75 byte-correct
integration-blocked MATCH drafts: cheap banked wins sitting idle. (d) Main's fuel gap was
already closed (2,001/2,002 Ghidra-C cached) and 33 main stubs have exact-h_seq matched seeds
in the fleet — main is not as barren as the 0.85-tier verdict suggested; nobody had ever
clustered main against itself. (e) The §172b detectors (EXTPAIR/SELECT) were declared but never
implemented, and the 892-record near-miss audit with per-draft opcode-transition histograms was
never joined to anything. Organizing first converts N independent cracks into one crack + N
mechanical/cheap transfers — the same economics that carried every prior tier (h_exact dedup,
h_seq remap, A-prop autodraft).

**Ratifications.** Plan approval (gate 1, 2026-08-14) formally ratified R37 (probe before
costing), R38 (read the recorded failure verdicts before designing an experiment), and R39
(negative-control every new refusal-check against the already-succeeded population) — all three
operated as binding through P30 and are now rules.

**Hindsight — the better path.** This is accelerator A6 ("regroup the residue by structure
before calling it unique") applied one level up: build the ATLAS — features + tiers + evidence
joins + lever labels — the moment a frontier stops being family-shaped, not two phases later.
The next project should build the feature/similarity layer right after its first propagation
engine exists.

## 2026-08-14 (P31 T7) — the LEN-LI mechanical cousin lane: killed by its own probe, correctly

**Context and belief.** The phase plan's Leg B promised a fully-mechanical lane for the LI-ONLY
adapt cards: align a cousin member against its matched seed, recognize the li-cluster length
drift, swap the constant, bank for $0. The alignment engine (`tools/family_align.py`) was built
and negative-controlled (NC-1 verdict-equivalence with `classify_member` 157/157 banked pairs —
after two real classifier fixes the NC itself caught: R-type non-shift sa diffs are STRUCT, and
registers are tested BEFORE the reloc skip; NC-2 imm-engine parity 21/21).

**What the probe said (R37, before any driver was built).** Classifying all 26 live LI-ONLY
cards: **0 mechanical** — STRUCT-ALIGNED 16 (regfields drift ×19) + LEN-STRUCT 10.

**Why (the premise error, named).** A cousin seed is a 0.85-similar DIFFERENT function, not an
h_seq sibling — its register allocation naturally differs everywhere, so a word-level positional
remap between cousins was never viable. §168 law 1 says exactly this ("a cousin is a SEEDED
CRACK, never a family_sweep remap") — the plan's mechanical lane contradicted the measured
cousin law, and the probe re-derived the law for ~$0 instead of ~30 wasted builds.

**What survives.** `family_align` itself — the aligned classifier + the aligned/cluster imm
engine — whose correct consumer is the LEN+N NEAR-MISS pile (T8): a draft vs its OWN target is
the SAME function, where registers agree outside the drift regions and the §172b tells route the
indels. The 26 LI-ONLY cards stay agent cards (correct all along). Also parked for T8: lui-bearing
clusters need a reloc-vs-constant range discriminator (`reloc_indices` conservatively flags every
lui+consumer as an address anchor).

**Hindsight.** The probe order in the plan (build NCs → probe ONE card → then the pile) was
right; what it should ALSO have said is "probe the CLASSIFIER against the pile before building
any driver" — that reordering is what saved the effort here.

---

## 2026-08-24 (P31 S59) — the A-prop 0-bank decomposition: consume the verdicts you already compute

**Decision.** When the free A-prop lane banked 0/117 three passes running, the response was to
decompose the failure population INDIVIDUALLY (four probes, four DIFFERENT causes) rather than
treat "the lane is broken" as one defect — and then to fix each class in the LANE, not as one-off
rescues.

**What the decomposition bought.** 117 staged = 82 already-refuted-by-match_one (staged anyway:
the filter read `status==AGREE` and ignored `shape`) + 19 judged by the WRONG oracle (standalone
compile for a TU-destined draft; 7/27 were byte-perfect in their real TU) + 16 byte-correct
bodies stranded on TU decl walls the pipeline had a tool for but no wiring to. Even the "gate
never saw it" probe artifact decomposed: a triage harness racing itself on an fn-keyed scratch
dir — the third fn-keyed-collision defect found in one session (staging filter, reloc --fix
rowmap, triage dir), which is now a named anti-pattern: **key nothing by bare fn name; overlays
share function names.**

**The general law (cookbook §270).** A mechanical remap is judged four times — instructions,
symbols, TU coexistence, whole-binary bytes — and a lane that stages on fewer than all four burns
a build per missing layer per pass, forever. Result of wiring all four + the decl-layer fixes:
the same population banked 64+ in one session, zero model tokens, and the mechanisms are in the
unattended lane.

**Ceiling honesty (R41).** The lane's own number ("117") was a stage count with a
three-quarters-hopeless numerator. The real denominator is 973 open members; the mechanically
reachable slice is ~120/pass banking at ~50–60%; the remaining ~850 are named residuals (169
STRUCT cracks, 121 type-inference, ~73 IMM tier-2, wrong-family cards) — work for different
tools, not this lane. Saying so beats reporting a flattering fraction of the wrong denominator.

## 2026-08-25 (P31 S61) — the wall is an integration wall: stop re-drafting solved functions, re-judge them

**Context and belief.** Through S58–S60 the campaign optimised the wide wave: band, mix, card
count, sibling-inclusive draws, straggler handling, gate parallelism. Each lever moved yield by
single digits while first-gate conversion slid from 51% (dd) to 1–6% (en/eo/ex/ey/ez), and the
going explanation was population exhaustion of a *codegen* frontier — the gen6+ "wall".

**What the measurement said instead** (`docs/tool-designs/frontier-analysis-s60.md`, a read-only
Fable audit at session end): of the 292 functions refused six or more times, 178 had ALREADY
produced a closeness-0 draft — byte-equal at the object level, refused by the whole-binary gate —
and across the open pool ~571 functions had finished drafting. The fleet kept re-drafting them
(10,049 reject rows over 574 distinct functions), and the gate spent ~3 whole-binary builds per
failing draft confirming failures that were never about the draft's text.

**The pivot.** Build the deterministic lane the analysis asked for instead of touching the wave
again: `tools/integration_resolver.py` treats the ledgers as an INDEX (not a promise — a July
closeness-0 draft can CC1-fail today because the fleet's declarations moved), re-judges every
candidate at the REAL split TU with `rtu_match`, requires `reloc_identity` to agree on symbols
(rtu masks relocation fields, so a wrong symbol name still reads MATCH — R34's disagreeing
oracle), rebases with `aprop_symfix` where only names are wrong, stages, gates on the whole-binary
SHA, commits at once (R42). Refuses main by name, `//@EDIT` drafts, dirty trees, collapsed
registries; a negative control over recently-banked functions must pass N/N before a verdict is
read (its first form picked carve moves as banks and failed 9/12 — fixed before any stock verdict
existed, R35).

**Why, in numbers (denominators, R41).** First pass, 2026-08-25 23:38–23:48: 1,352 nominated
(binary, fn) pairs → 901 already banked, 27 main → **424 judged in 41 s → 245 staged (57.8%) →
63 banked** (net INCLUDE_ASM delta; the commit subject's "72" counted 9 carve moves), zero model
tokens, ~10 minutes wall. The wave lane's best recent gate banked 13 of 222 in a ~30-minute gate.
The falsifier (<5% survive intake) was not close.

**What the refusals then taught, byte by byte.** 182 doubly-verified drafts were still refused.
Probe 1 (`md_SC03_076/func_801EFBB4`, 407 ins): the spliced binary builds and 66 bytes differ at
file offsets 324–2440 — the `.rodata` jump tables the stub's `.s` carried (`jtbl_801EF5AC`,
`jtbl_801EF5E4`) move by one word when the compiler emits them itself: the §8e/§260 table-placement
class, not codegen. Only 43 of the 182 carry rodata at all; 139 (in 16 binaries) were refused with
no rodata coupling — that residual is the next thing to characterise (cookbook §293 records what the
probes find).

**Hindsight — the better path.** The ledgers held the evidence for weeks: `closeness: 0` rows with
"whole-binary gate rejected — CAUSE NOT DETERMINED" as their verdict. A verdict that names no cause
is a measurement nobody can act on; the cheap move — re-judge the stored body at the real TU and
diff the BYTES of the built image against the good one — costs ~90 s per function and should have
been the gate's own failure report from the day the ladder was built. Generalisable: when a
correctness oracle says "no" without saying where, build the second instrument that says where
before tuning anything upstream of it.

## 2026-08-26 (P31 S61, small hours) — the RED fleet: the wave collapse had a third cause nobody was billing

**Context.** After the resolver banked 63 zero-token functions, 182 of its doubly-verified drafts
(rtu-MATCH at the real TU, reloc-AGREE on symbols) were still refused by the whole-binary gate, and
the per-binary refusal counts were all-or-nothing: 23/23, 23/23, 18/18, 13/13 in single binaries.

**What the probes said.** All-or-nothing per binary is not a property of drafts. A clean-tree build
of the fully-refusing binaries: 15 of 214 fleet binaries were baseline-RED at HEAD — seven not even
building (stale JTBL_PADS after S60's evening banks/reverts), the rest byte-shifted (a missing
interleave entry; a wave-committed half carve). They had been red for 4–12 hours because the fleet
R22 sweep was guard-skipped whenever any gate was in flight — i.e., always. Every wave gate and
every resolver gate against them since was a measurement of the BINARY billed to the DRAFTS —
part of the "1–6% conversion" story S58–S60 attributed to population exhaustion.

**The pivot.** Stop diagnosing drafts; audit and repair baselines: per-binary clean audit → the
byte-proven repairs (jtbl_pads_fix for pad drift — itself first un-broken three ways, R40;
insertions-only interleave regeneration from the splat yaml for layout shifts) → a gate-side
BASELINE-RED refusal so a red binary's drafts are never judged (negative-controlled both ways) →
re-open the mislabeled refusals in the resolver ledger for automatic re-judging.

**Casualties of the night, recorded because they are the pattern.** The repair tool's own writer
poisoned the registry into a make parse error that failed every build of every binary for ~9
minutes, voided one wave's re-gate and its own candidate search, and was adopted by a blanket
pre-gate commit (R52 instance #2). Two gater-restart helpers killed themselves via unanchored
pgrep self-match (the S60 hazard, from the other side). Every one of these is the same lesson:
the instrument — including its write path, its restart path, and the baseline it measures against
— is part of the experiment.

**Hindsight.** The S60 close declared "tree clean, all lanes stopped" and was true, while 15
binaries were silently red — "clean" and "green" are different invariants, and only one of them
was checked. A session-close ritual (and any wave post-mortem) should quote the fleet's GREEN
count next to the tree's cleanliness; tonight that number was 197/214 pretending to be 214/214.

---

## 2026-08-31 (P31 S67) — The frontier's largest single class is CARVE PLUMBING, not codegen; and an optimistic probe is how it stayed invisible

**Context.** S66 closed with 530 open functions and a free-wins audit naming four "zero-drafting"
lanes. S67 opened by measuring what is actually on disk: `tools/strand_census.py` finds **193 of the
530 open functions already have a draft**, classified in their real TUs as 37 MATCH · 67 NEAR ·
89 CC1-FAIL (219 draft files; 1,885 wave targets seen, 1,521 already banked, 166 open functions
never drawn at all).

**What the probes said.** Gating the 37 MATCH drafts banked **0 of 13** before I stopped the run —
and the reason was not codegen. Probing the whole jtbl class with the REAL planner instead of the
cheap probe: **159 of the 530 open functions (30%) reference a jump table, and 96 of them cannot be
carved at all** — `build_carve` refuses a plan whose same-subseg `.rodata` carves would be
non-contiguous (one object cannot leave a hole for an uncarved neighbour's table). 75 of those are
non-main, across 38 subsegs. Not a compiler wall, not a declaration wall: carve plumbing.

**Why it was invisible.** `jtbl_carve --probe` called only `island_probe`, which answers *where does
this table live* — necessary, not sufficient. Every one of the 96 probes "carveable". The S66 audit
priced 32 of them as free on exactly that reading, and `ov_SC02_000:func_8017F950` — named on that
free list — refuses. The planner is a PURE function, so the probe can just call it; it now does
(cookbook §322). The generalisable rule: when a cheap probe and an expensive applier disagree about
feasibility, check whether the applier's DECISION half is separable from its MUTATION half — if it
is, the probe must call it, because an optimistic probe does not merely lose opportunities, it
manufactures work plans.

**The pivot.** Stop treating the stranded-draft pool as a gating backlog. Two lanes, in this order:
(1) the **89 CC1-FAIL drafts**, where two of the five levers their errors call for
(`scope_data_externs` §8d and `normalize_self_decls`) exist but are wired ONLY into the family
lanes — a draft written by a wave agent has never seen either, and 43 of the 89 sit on exactly
those two; (2) the **jtbl unblock**, which is the bigger prize but is gated on `jr_isolate_all`
round-tripping, which it currently does not.

**What the attempt at (2) proved, and cost.** ov_SC02_000 took three defects to reach the byte gate
and still fails it: the carried-decl layer emitted `struct sprite8` four times (dedupe added), then
re-emitted a type `engine_types.h` already defines (header-skip added), then compiled and linked
and diverged (`#if` guards do not travel with a carried block — open, cookbook §323). 20 of 35
blocked overlays dry-run clean, and that number means nothing until one round-trips.

**Casualty, recorded because it is the pattern.** I diagnosed the 13 `near` verdicts as my own
mis-invocation and killed a running gate loop on it. Wrong: `--src` is never defaulted,
`harvest_verify` derives each draft's TU, and `match_one_closeness` re-derives the asm subdir per
function — all three documented in comments I had not read. Cost: the last five binaries of that
pass, main's nine included. The correct reading was available in the code before the pkill.

**Hindsight.** Two of this session's three real findings came from distrusting a green-looking
signal: R53 caught a failed build whose stale binary still hashed to the locked SHA (twice), and
§322 caught a probe that had never been asked the blocking question. The one thing I did NOT
distrust in time — my own first diagnosis — is the one that cost work.

## 2026-09-01 (S70) — The postgame reframe: tools must work over the CRACKED corpus, not just the frontier

**Context + belief.** Phase 31's whole apparatus — waves, cards, gates, the residual classifier — was
built to consume the *remaining* work. Every tool's population is "open stubs". The frontier is now
355 real functions (67 main game-code + 288 non-main; main's other 960 open stubs are PsyQ library
code that is not a matching target at all), so tools scoped to the frontier are scoped to a shrinking
and unrepresentative slice.

**What prompted the pivot.** Drew, mid-probe: *"this tool should work for all previously cracked funcs,
not just the remaining work. we are focusing on the postgame now, this project being used for all
future decomps, the tools/cookbook, everything."*

**The measurement underneath it.** The S70 coverage probe found `residual_rules_b` returns UNKNOWN on
**52.8% of real near residuals, and 57% of the cleanest (<=8-diff) band** — the band where a rule is
worth writing. Hand-labeling 4 of 4 sampled UNKNOWNs mapped them to existing cookbook buckets; one
(`WIDTH/lhu!=lh`, a one-word `u16`->`s16` fix) already has its discriminating signature COMPUTED by
`residual_class` and still returns `top=None`. The asymmetry is 1,062 cookbook sections vs ~13 coded
rules, concentrated in the buckets with the most prose (regalloc 127, types/width 93, structs 86,
scheduling 85). So the rules are worth building — but hand-authoring 100 rules from prose is exactly
the slow path this project keeps learning to avoid.

**The pivot.** The ~850 MATCHED functions are a **labeled ground-truth corpus** and were never used as
one. For any banked function we hold both the earlier failed drafts (on disk under `.run/`) and the
known-good final C (in `src/`): the residual is recomputable and the correct label — *the fix that
actually worked* — is derivable from the draft->final diff rather than guessed. That converts rule
authoring from prose-reading into mining known answers, supplies a real precision/recall validation
set, and makes R39 negative control free (a candidate rule must not misfire anywhere in the banked
corpus).

**Why it matters beyond BFM.** A classifier that learns from *a project's own matched corpus* is
portable: any decomp with a byte-gate and a growing `src/` can bootstrap it. That is the postgame
deliverable — the reusable method (tools + cookbook), not this one binary. It also inverts the
economics recorded all phase: the matched corpus GROWS while the frontier shrinks, so a
corpus-trained tool gets stronger exactly as the frontier gets harder.

**Hindsight / better path.** The signal was available much earlier. The byte-gate has been a perfect
correctness oracle and a null coverage oracle since Phase 5 (R34), and every banked function since
has carried its own answer. We measured tools against the frontier for ~20 phases because that is
where the *work* was, never noticing that the *answers* were accumulating on the other side. The
generalizable lesson: **when a project accumulates verified outcomes, that archive is training data
for its own tooling — scope a tool to the answers, not only to the open questions.**

---

## 2026-09-02 (P31 S71) — The §406 sweep was priced by SHAPE and refuted by RESIDUAL; the frontier's real lane is integration

**Context and belief.** S70 closed with a fresh-session checkpoint whose first instruction was
unambiguous: "**START HERE — THE §406 PROLOGUE-WEAVE SWEEP. This is the single biggest measured lever
on the board and it needs no agents.** 134 of 1,237 open stubs (11%) carry the shape; the residual is
ALWAYS the `sw $ra` slot; twelve alternative variants are already measured inert." The mechanism was
real and hard-won — traced in cc1's own `.i.sched2` dump, byte-proven on the banked exemplar
`ov_SC02_005/func_8017F898`. The belief was that one scripted edit over a known class would bank
dozens of functions for zero agent tokens, which is the project thesis in its purest form.

**What failed.** Both halves of "134 of 1,237" were wrong, and the error was in the counting, not the
compiler work.

* `corpus.stubs` counts main's **960 PsyQ LINKED library stubs** as open. The real frontier is **210**
  (main 64 + non-main 146) — the checkpoint even records that partition three paragraphs later, and
  the census still did not apply it. Class census 134 -> 77 after the filter.
* The census predicate matched a **shape in the target** (`sw $s0` / `move $s0,$a0` / `sw $ra` within
  24 lines) which is symmetric: it cannot tell a target that keeps `sw $ra` early from one that has
  already sunk it. `main/func_8002EED8` is the second kind, where this lever pushes backwards.

Re-derived from the mine-vs-target disagreement (the direction is two fields of a residual we already
emit), the lever's addressable set is **15 of 210**, and applying it there produced **0 MATCH / 14
applied**. The one member close enough for the lever to be decisive (closeness 8) got **worse, 8 ->
91**: the clobber re-schedules the whole block, it is not a free nudge.

**The pivot.** The same baseline pass — score every real-frontier stub's best stored draft, once —
found **64 of 210 (30.5%) already at standalone `match_one` closeness 0**, across 33 binaries, and
**all 210 have at least one draft on disk**. So the frontier's largest lane is not codegen at all: it
is §376 integration (the TU rejecting a signature the body already gets right). The session redirected
from scripting the lever to gating those 64 and running the `fix_arity_callers` / `cast_self_callers`
/ `--sync-decls` chain on whatever the gate refuses. This is the `matching-is-solved-integration-is-
the-bottleneck` memory arriving as a measurement instead of an impression.

**The grounded why.** A shape census answers *how many functions look like this*; only a residual
answers *how many are broken like this*, and a sweep's yield is bounded by the second number. The 15
WEAVE-SUNK drafts have baseline closeness 8, 85, 103, 104, 112, 146, 158, 158, 158, 200, 231, 302,
362, 400, 502 — the `sw $ra` slot is a symptom inside bodies that are wrong for a dozen other reasons,
and a scheduling lever can only ever close the LAST diff.

**Hindsight / better path.** R37 (probe before costing) was followed and it worked — five members were
probed before anything was built, and the probe is what exposed the two counting errors. What R37 did
NOT force is the step before it: **the checkpoint priced the class when it discovered the mechanism,
in the same breath as the exemplar's win, and that price was never re-derived against the frontier it
would be spent on.** The generalizable rule for the endgame: *a class discovered by cracking one
member must be priced by the residuals of the others before it is written down as a lever* — one
`--json` field on scores we were already running would have said "15, not 134" on the night it was
claimed. Cost of learning it here: about one hour of deterministic compute and no agent tokens, which
is exactly what a probe-first rule is supposed to buy.

---

## 2026-09-02 (P31 S71) — I gated `main` with a tool documented as unable to gate it, and only R22 caught it

**Context and belief.** The session's integration lane was running well: `parallel_gate` in isolated
worktrees had banked cleanly across 30-odd overlays all night. When 15 of the 64 standalone-match
bodies turned out to be `main`'s, I put them through the same tool. It reported **11 banked**, the
merge committed them, and every signal I was watching — worker exit codes, the bank oracle (a stub
disappeared), the summary line — agreed.

**What failed.** The R22 clean-fleet verify returned **212/213**. `main` did not compile from clean
(two `conflicting types` errors). Reconciling both declarations made it build — and it was **still not
byte-identical**. Re-gated one function at a time against a clean tree: **11 of 11 REJECTED.** The
commit was reverted and `main` was verified byte-identical again before anything else proceeded.

**The rule already existed, three files away.** `ox_campaign.gate_main_batch`'s docstring:
*"main is gated by ONE CLEAN REBUILD of the whole EXE, never incrementally … main's extract rewrites
the linker script, so an incremental main gate returns a FALSE DIFF. Measured P31 S58: wave `ab` drew
105 main cards and banked 0 of them."* `parallel_gate`'s worker **is** `gate_stage`, so it inherits
that constraint exactly. I had read that docstring earlier the same session, while looking at
something else.

**Why the failure direction was worse than the one on record.** S58 recorded the false-DIFF direction:
competent drafts thrown away, loud and wasteful. This was the false-PASS direction: wrong bytes
committed, reading green until the next clean fleet check. R53 names the mechanism — *a failed build
leaves the previous object on disk, so a SHA1 check downstream of it reads green* — and R53 was
written for a different tool and never applied here.

**The pivot.** Fixed as a **refusal in the wrapper**, not a note in the callee: `parallel_gate` now
returns REFUSED for `binary == 'main'` and names `tools/gate_main.py`. The main lane was then reopened
properly the next morning and banked 5 (4 after the source-truth correction below), with a bisect
isolating the one bad draft in 7 rebuilds.

**Hindsight / better path.** Three things would each have caught it earlier, in increasing order of
generality: (a) run R22 **before** committing a gate against a binary the lane has not gated before,
not at session close; (b) `gate_main`'s own bank count was also derived rather than measured — it
printed "BANKED 5 of 6" when 4 had applied, because `len(good)` is *what we decided to keep*, not
*what was substituted* — so **count from the source in every gating tool**; (c) the general rule this
session kept re-teaching: **a tool that wraps another tool inherits its refusals**, and the place to
encode that is a refusal in the wrapper. Every constraint documented on `gate_stage` binds
`parallel_gate`, `harvest_verify`, and anything else that shells it.

**Cost of learning it here:** one bad commit, ~40 minutes of revert-and-bisect, and an inflated bank
count I had already reported to Drew and had to correct. Cheap only because R22 exists and was run.

---

## 2026-09-02 (P31 S71) — the drafting pool ran dry, and the lever was an exclude list nobody re-probed

**Context and belief.** With ~40 agents landing at near-100% MATCH, the working assumption was that
drafting capacity was the constraint and the campaign would continue as draw → draft → gate until the
frontier was gone.

**What failed.** Wave 3 drew **1 target** and reported *"0 left in pool"*. Measured at that moment:
174 open, of which `main` 64, and of the 110 non-main — **41 drafted this session, 68 on the exclude
list, 2 proven walls, ZERO genuinely undrawn**. More agents would have had nothing to work on.

**The pivot.** The 68 excluded functions were excluded because the TOOLING could not carve them —
`jtbl_carve` refused their plans with *"subseg would host NON-CONTIGUOUS `.rodata` carves"*. But
tooling had changed **that same session**: `jr_isolate_all` had been fixed twice (file-local `static`
placement, and §323's `__attribute__`-blind regex). Re-probing all 68 found **17 now reporting `tail`
— a standard §8a carve**. Every one already had drafts on disk; scoring them put **10 at closeness 0
for zero drafting**, and the gate banked 9 — four of them in **57 seconds**.

**The grounded why.** An exclude list is a snapshot of *what the tooling could not do at the moment it
was written*. It is treated thereafter as a property of the FUNCTIONS. Nothing in the pipeline
re-examines it, so every tool improvement leaves behind a population that is now tractable and still
marked impossible — invisible, because the draw filters it out before anything measures it.

**Hindsight / better path.** **Re-probe the exclude list after every tool fix, as part of the fix.**
The probe is deterministic, costs no agents, and here it was worth more than the entire drafting lane
at that moment. Generalised: *any list that records a tool's limitation must be regenerated when the
tool changes, or it silently becomes a list of work you have decided not to do.* The same reasoning
applies to `.run/S71_walls_found.txt` — a wall proven against today's compiler knowledge is not a wall
forever, and each entry should carry the refutation list that would have to be beaten.

---

## S72 (2026-09-02) — "11 PROVEN gate-rejects" were one missing carve, and the gate could not have told us

**Context + belief.** S71 closed with `main` as the centre of gravity (59 frontier functions but ~45%
of all remaining instructions) and a hard note in the checkpoint: *"11 main functions score
`match_one` closeness 0 and are PROVEN gate-rejects (re-gated one at a time) … §376 in its purest
form — do not re-slate without a TU-level fix."* The belief was that main's remaining difficulty was
declaration plumbing on top of a hard codegen tail, and the next session's job was the §376 chain.

**What failed.** The verdict was produced by an **ad-hoc script** (`.run/S71_main_bisect.py`), not by
`gate_main.py`. It lifted each body plus a heuristic preamble into the green `src/800.c` and rebuilt.
It never ran the project's own pre-check — which, run afterwards on the same 11 drafts, names **6 of
them** as declaration conflicts against declarations the TU already carries. Four of those never
compiled at all (`conflicting types`), and one of the six (`func_80031988`) is a **false** conflict in
`gate_main`'s type comparison (`struct Ent30D80 *` vs its own typedef `Ent30D80 *`). And of the 11,
**all 11 are switch functions** — a fact nobody had measured, because the gate's entire output is two
SHA1s.

**The pivot.** Build the missing instrument first (`tools/main_diff_locate.py`: attribute a red image
to symbols via the linker map), then let it speak. Every one of the four drafts that compiled came
back with the SAME shape: **1–4 bytes differing inside the drafted function, ~5,200 bytes across ~332
symbols outside it, one uniform positive delta (+28/+52/+76/+84), first moved symbol always
`jtbl_80072A4C`.** That is not codegen. That is a duplicated jump table growing the image.

**The grounded why.** `config/splat.us.exe.yaml` has carried exactly ONE `.rodata` carve since **Phase
7** — LZSS's `jtbl_80072A38`. Every other main jump table stayed raw in the tail data, so a drafted
switch double-emits its table. Extending the carve to the contiguous game-jtbl span
`0x80072A38–0x80072C70` is byte-neutral with no draft substituted (probed first, R37), and with
`jtbl_rodata_pads --derive` taught main's file-offset base, three of the eleven banked
**byte-identical in 14 seconds**.

**Hindsight / better path.** Two rules, both cheap and both skipped:

1. **A gate that can only say "different" will eventually be believed to have said "wrong".** The
   diagnostic was not missing by accident — `gate_main`'s R40 baseline control *rebuilds the tree
   green immediately after a failure*, destroying the red image and its map every single time. The
   control was right and its ORDER was wrong, and that ordering bug is the whole reason 11 functions
   sat parked. Any oracle whose verdict routes work must preserve the artifact the routing needs.
2. **A verdict produced by a bespoke harness is a verdict about that harness.** The project already
   had `gate_main.py` with a decl pre-check, a bisect, an R40 control and a no-op-substitution guard;
   the one-off script had none of them and its output was written into the checkpoint as ground truth
   for the next session. When the real tool exists, a "quick" reimplementation is not quicker.

Generalised (and it is the same shape as S71's exclude-list finding): *a recorded impossibility should
name the instrument that produced it, so the next session knows what to re-probe when the instrument
changes.*

### S72 addendum — the decision to split `src/800.c`, and the estimate I got wrong

**The question Drew asked** was whether the split was worth doing and at what effort, framed as
*"if the devs did it in 1998 let's do it also."*

**The premise needed correcting before it could be leaned on.** The jtbl spans prove there were **at
least** three TU boundaries in that address range — tables pack tight within a compilation unit and
are separated by other data across units. They do **not** prove the devs' files were exactly these
three: a TU containing no `switch` emits no jump table and is completely invisible to this signal.
What we recover is a LOWER BOUND on the original structure, not a reconstruction of it. Happily this
changed nothing about the plan — we need exactly enough objects to give each span its own contiguous
`.rodata` run, three is the minimum that works, and splitting anywhere else would have been
speculation. Evidence-driven and engineering-optimal coincided.

**The estimate I got wrong, and how.** I told Drew the split would be expensive and quoted it as
*measured, not guessed*: "26,543 lines with 2,318 scattered `extern` lines and 175 typedefs — the
exact shape `split_src_region.py` was blocked on for overlays." The measurement was real and it
measured **the wrong quantity**. 2,318 is the TOTAL number of externs; what the split actually costs
is how many declarations are used OUTSIDE the region that declares them, and that is **57 of 1,247
(4.6%)** — 19 typedefs, one definition each, zero shape conflicts, zero file-local statics. The split
took one afternoon and was byte-identical on the first clean build after the typedefs moved.

**The grounded why.** This is the denominator discipline (R41) applied to an *effort* estimate rather
than a cost or yield figure. A number with no denominator attached — "2,318 externs" — reads as
authority because it is precise and true. The question was never "how many declarations are there",
it was "how many are shared", and nothing in the first measurement was pointed at that. **Quoting a
real measurement of the wrong quantity is more dangerous than admitting you have not measured**,
because it forecloses the cheap probe: had Drew accepted the estimate, the 39% of main behind this
would have been deferred to a later phase on the strength of a number I never should have quoted.

**Hindsight / better path.** Before quoting an effort estimate, state the quantity the estimate is a
function of and check that you measured THAT. Here one grep — declarations used outside their region
— was 20 minutes and would have replaced "expensive, needs a shared header, not a naive partition"
with "57 crossing names, mostly typedefs, one afternoon."

## S75 (2026-09-02) — Seven walls, seven instruments; and the decision NOT to tidy 12,000 functions

### Context and belief going in

The S74 checkpoint handed forward a frontier of ~73 functions and a named list of blockers, chief
among them "`reconcile_tu` manufactures declaration conflicts — the one unfixed defect that is
actively costing banks". The working belief, inherited across several phases, was that the remaining
work was a **hard tail of genuine gcc-2.7.2 codegen walls** with some tooling friction around it.
`SaveLoadRoutine` (1,165 ins) had been carried as §434, an unbreakable wall, since the phase opened.

### What actually happened

Seven separate "walls" were run to ground. **All seven were instrument defects.** Not one was the
compiler. The list is in the S75 checkpoint (§442–§447); the shape that matters is that four of them
were *tools reporting something TRUE about a world that was not the one they were asked about*:

* `reconcile_tu` rewrote legal C into illegal C on a **false premise about cc1 written in its own
  docstring** ("a decl BELOW still conflicts" — true at file scope, false at block scope).
* `jtbl_carve` reserved one word too many whenever a function had **more than one `sltiu`**, because
  `sltiu` is also how gcc emits an unsigned range check — so its guard disabled itself on exactly
  the functions that needed it, quietly, across the whole corpus.
* `harvest_verify`'s `overlays.mk` snapshot matched **the first of two blocks** and silently
  half-restored, leaving a binary unbuildable while `git status src/` showed nothing wrong.
* `main_diff_locate.classify()`'s `TABLE REJECT` verdict was **unreachable by construction on main**
  (it keyed on the string `(.rodata)`; main's tables live in `.data` objects) — so a carve failure
  was labelled a declaration failure and the §376 chain was run at it twice, addressing 5% of the
  evidence.

### The grounded why

**A class that cannot fire is worse than a class that does not exist.** It converts "I don't know"
into confident, specific, wrong advice, and that advice then consumes sessions. `SaveLoadRoutine` is
the exact cost: 1,165 instructions — 9.2% of everything left in the project — sat behind a verdict
string that named the wrong subsystem, for a whole phase, while its body was byte-identical the
entire time. Nothing in the pipeline compared the recommendation against where the bytes actually
were (94.9% jump tables, 5.1% code).

The corollary, which is the session's reusable law: **when a verdict names a subsystem, check that
subsystem owns the majority of the bytes before acting on it.**

### The strategic decision: leave the ~12,000

Gating `ov_SC01_005` stalled a gate for 30+ minutes. Cause: `dedup_propagate --auto-from <bin>`
sweeps the WHOLE binary, not the function just banked, and that overlay held **557
matched-but-never-shared functions**. The fleet census came to **~2,073 distinct functions /
~12,116 sweep items across 174 of 217 binaries**, every one already matched.

Origin: the July 2026 mechanical family sweeps (`c993029f0` +16,512 members, `025cc03f6` +17,975
member-matches) bank a proven body as a **private copy per overlay** and register no dedup group.
That was a deliberate throughput trade — it moved the fleet 66%→71% in one commit — and it left
tidying behind that nobody has done since.

**Drew's decision: do not convert them.** The precedent our own cookbook records is that *sotn writes
duplicate functions explicitly*; under that playbook the 12,000 is the normal end state, not debt.
The sharing machinery is a BFM-specific optimisation for having 211 overlays rather than a handful.
So: gate with `--no-propagate` from here, propagate only deliberately for a high-reach new match.
This removes a 30-minute stall from the critical path of every gate permanently, for zero cost to
completion — the backlog is orthogonal to the percentage.

*Caveat recorded at decision time:* the sotn claim rests on **one parenthetical in our own cookbook**,
not on sotn's repository. It is good enough to act on for an optimisation we can reverse at any time;
verify against sotn-decomp before it becomes doctrine.

### Hindsight / better path

Two of the session's own instruments lied before they were trusted — the frontier classifier reported
38 phantom free twin-remaps (it counted a stale duplicate registry and two *prototype builds* as peer
binaries) and 32 phantom never-drafted functions (its draft scan globbed `.run/S7*` and missed five
other draft directories). Both were caught **only** by checking a case whose answer was already
known, and both were the same defect being hunted elsewhere in the session.

The better path is not "be more careful". It is: **every scan ships with the denominator it claims to
cover, and gets tested against one known-true case before its number is quoted to anyone.** That is
cheap, it is mechanical, and on this evidence it is the single highest-yield habit in the project.

---

## S76 (2026-09-03) — EVERY WALL EXAMINED WAS THE INSTRUMENT. FIVE DEFECTS, ONE SHAPE.

### Context and belief going in

S75 closed with "every codegen wall examined — nine of them — was an instrument defect. Not one was
the compiler." S76 treated that as a finished lesson and planned ordinary work: verify the fleet,
make the main gate durable and parallel, then bank the near-term frontier. The belief was that the
tooling was now sound and the remaining cost was drafting.

That was wrong in the specific way S75 warned about, and the session found **five more** instrument
defects — each one hiding work, each one previously reported as a property of the code.

### What was found, in the order it surfaced

| # | defect | what it asserted | what was true |
|---|---|---|---|
| 1 | verbatim drafts accepted by `gate_main` | "9 functions banked, byte-identical" | the drafts were the targets' own asm; `progress.py` moved by **zero** |
| 2 | same gap in `harvest_verify` | (silent) | the module/overlay half of the same wave reached the tree through a gate with no guard |
| 3 | same body offered by `api_agent.prior_draft` | "a previous attempt left this body behind" | it was raw asm; with both gates fixed, agents STILL resubmitted it because the pack supplied it |
| 4 | `match_one` + `rtu_match` modelling `maspsx + as -O1` | a §182/§188 "IMMOVABLE epilogue wall" | those four TUs build through `reorder_passthrough + as -O2`; the wall does not exist |
| 5 | `draw_waves --main` | `main: refusing 49 LINKED subseg(s)` | **main was never iterated at all** — no `src/main/` directory, so it was never in `bins` |

### The measurements that settle each one

* **#1–3.** One census: **1,099 of 704,375** `.c` files in the draft store are verbatim-asm bodies
  under ordinary `<fn>.c` names. Negative control on the detector: **0 false positives across
  45,898** drafts carrying both a real C definition and an inline `__asm__`.
* **#4.** Same draft, `func_8005ECC0`: closeness **5 / 36 ins** under the old oracle, closeness
  **2 / 35 ins** under the real build path. Seven of eleven main agents in one wave produced correct
  C, saw the phantom tail, consulted `oracle_reorder.py` — which said *"file IMMOVABLE, stop
  grinding, no C-level work can ever close it"* — and each submitted a §265 verbatim body instead.
  Rescoring old drafts under the corrected oracle recovered **4 functions for zero agent tokens**.
* **#5.** `0 -> 55` main stubs reaching the pool. Every mixed draw in the project's history
  contributed nothing from main.

### The shape they share, and why it is not "be more careful"

All five are the same failure: **a tool made a TRUE statement about a scope narrower than the reader
believed, in language that sounded like a statement about the code.** "Banked." "IMMOVABLE."
"refusing 49 LINKED subsegs." Not one was silent; four of the five were *loud and reassuring*, which
is worse, because a confident message is the thing nobody audits.

Three of them were also **the same defect behind more than one door** — the verbatim body had to be
refused in `gate_main`, in `harvest_verify`, AND in the pack that supplied it, and the reorder path
had to be fixed in `match_one` AND `rtu_match`. Fixing the consumers is not the same as fixing the
supply, and a fix made in one of two paths is a fix in neither (the §442/S74 sibling-provisioner
lesson, now with two more instances).

### Cost, and the hindsight path

The direct waste is small — one wave's worth of verbatim fallbacks, four functions that were already
solved. The real cost is **historical**: a large part of the recorded "wall" history in the 800c3
cluster is instrument error, and several functions currently parked in `src/` as deliberate §265
verbatim banks were parked on a false premise. Agents in S76y are now finding this on their own
("the six journal attempts... modeled the wrong assembler"), which means the journal — our
past-attempt fuel, measured at 38/39 MATCH when it works — has been feeding forward false walls.

The better path, and it is mechanical rather than a virtue: **a tool that reports a population must
report what it excluded from it, and a claim about the toolchain has the same shelf life as the
toolchain.** The reorder island landed 2026-09-01; the docs asserting it was unreachable were not
swept, and four tools kept modelling the retired path. A build change must sweep every doc and every
oracle that asserts what the build cannot do — that sweep is now the missing step in the wave
playbook, not a habit to remember.

---

## S77 (2026-09-03) — the frontier stopped being a drafting problem, and three contract items moved

**Context and belief going in.** Phase 31 T10 had been a drafting campaign: draw a wave, crack
functions, bank them. The working belief was that the remaining frontier is *hard functions*, and
that more/better drafting is the lever.

**What the session measured instead.** A 30-workflow burst over main's entire drawable frontier
returned **9 banks and 21 NEARs — and nearly every NEAR named its own mechanism**, down to the gcc
pass and often the source file and line, with measured negative controls. `func_80032A74` reached
**1 of 422**. The frontier has changed character: it is no longer "we don't know why", it is "we know
exactly why and C cannot express it". **A NEAR whose note cites a pass and a file:line is a §474
wall-proof candidate, not a redraft** — costing another wave against that pool would buy
already-answered questions at full price.

**The pivot.** Effort moved off drafting and onto the three things drafting can never deliver: the
contract's second oracle for main, the SDK residue, and the `-O0` build gap. All three are
deterministic, zero-token, and were blocking the completion claim rather than the completion.

**What that immediately exposed — the session's real theme.** Twelve instrument defects, all one
shape: *a tool asserting about a DRAFT what was true only of the HARNESS*. Ratified as **R61**. The
expensive ones were not subtle-looking; they were confident:

* `gate_main`'s clash pre-check compared **block-scope** externs against file-scope spellings,
  making it stricter than cc1 — refusing **566 instructions of byte-correct work** (§481).
* `psyq_identify`, the placement map the whole library-linking pipeline consumes, parsed
  `objdump`'s **rendering**, which elides repeated words. It read 520 words for a 526-word object
  and printed **"not linked by EXE"** for objects that are linked. **+25 objects / 3,877
  instructions** were invisible, not excluded (§485).
* A Phase-8 exclusion said four PsyQ objects were unlinkable for "scattered `.bss`". True — and
  **three of the four are not blocked by it**: `SYS.o`'s two bases have DISJOINT offset ranges so
  the section splits, and two of the objects have no `.bss` at all (§484).

**Three of the twelve were mine, made the same day**, which is the part worth keeping: a cookbook
claim refuted by its own tool's negative control (§479, corrected TWICE in one session), an R48
collision bug in a brand-new tool, and a shell wrapper reading the wrong exit code so I reported a
red gate as still running. Each was caught by checking against a case whose answer was already known.

**The hindsight "better path".** Every one of these was cheaply detectable at any point in the last
twenty phases by an assertion the tool could have made about itself: compare the parsed word count
against the section size; compare the number of drops against the number of judgements; compare a
static blocker class against what the gate actually does to the draft. **R32 said "assert your
coverage" in Phase 26; the tools that broke here were the ones that never got that treatment.** The
lesson is not "be careful" — it is that a derived number must carry its own denominator, and the
five minutes to add that check is repaid the first time the tool is wrong and sounds right.

**Also recorded:** a yield table is evidence; a story about WHY the yield looks that way is a
hypothesis and needs its own negative control before it enters the cookbook, because the next session
will act on it (§479's three versions).

## S78 (2026-09-04) — the completion sprint is chartered on a census, and main's "wall" band turned out to be Sony's controller library

**Context and belief going in.** Drew asked the direct question: what actually remains, are the
waves done, is the rest tooling? The working belief from S77 was that main's residual stubs were
compiler-wall work (§332 "%lo in a delay slot", §474 wall-proofs) — genuine codegen the pinned cc1
cannot reproduce from C.

**What was measured.** Fleet: 51 stubs / 7,710 ins of 13.5 M (0.06%), every one classified by its
true blocker (`frontier_classify`); 180 verbatim bodies of which only ~9 are Square's or the CRT's
own asm; 5,827 ins of located-but-unlinked SDK objects sitting in subsegs labelled "game code"; five
unclaimed disc payloads. No class needs an agent wave. Then the provenance probe: the psx loader's
per-version PsyQ signature sets (never used before) place the `800c3` band byte-exact as **LIBPAD
4.2.1 + LIBAPI 4.2** — twelve of main's twenty-nine stubs, including all four §332 walls, are Sony's
DualShock library assembled in reorder mode, not game code. PsyQ 4.0 (the archive we link from) has
no LIBPAD; 4.6/4.7 differ except one object. The 4.2/4.2.1x `Ps` stamps identified in Phase 1 were
these two libraries all along.

**The pivot.** The finish is ordered around provenance and plumbing, not cracking: name the band
(#12, done: 46 names, Ghidra-mirrored), wire every placed-but-unwired SDK object (#3/#4 — the
residue the build now prints), hunt a 4.2.1/4.3 LIBPAD.LIB (#13) and otherwise finish the band as C
under the reorder island with real names (#5), then the mechanical/plumbing/near classes, the ~22
genuine redrafts as single journal-noted agents, the verbatim end-state, and the PhaseEnd.

**What the first task exposed.** Main's LINKED build had been RED at HEAD since the S77
`psyq_identify` fix (§485): newly-located in-gap objects merged libgte's 22 stub blocks into 3.
It read green at the gates because worktree gates have no `.run/obj40` and take the stub fallback —
the dual "with AND without SDK objects" invariant was only ever verified by hand. Fixed by wiring
stub↔objects by subseg range (+ exact tiling) and by redefining a library object's exported symbol
to the curated name (`firstfile`/`firstfile2`, which also caught a Phase-21 xdedup mislabel at
0x800430B8). And a rename hazard: verbatim `__asm__` bodies spell `\tfunc_X`, invisible to a `\b`
regex and to the string-masking linter — the linter now scans asm bodies (negative-controlled).

**Hindsight.** The signature JSONs were on disk since Phase 1; one afternoon with them in Phase 8
would have named the band, explained the 4.2 stamps, and kept the §332 wall verdicts from ever being
written. General form (now cookbook §487): a wall inside bytes no archive you hold can place is a
provenance question before it is a compiler question.

### S78 addendum — main was 91.8% game-code all along; the instrument said 59.8%

Wiring thirteen "game code" subsegs as LINKED (task #3) moved `MAIN game-code weighted` from 59.8%
to **56.1%** with no game-code change — the tell of a broken denominator (R35). `progress.py`'s
comment asserted the Ghidra sig excluded LINKED objects; measured, it carried every one of them
(~31,000 ins), and their `INCLUDE_ASM` stub records counted as unmatched game code. With the
exclusion derived live (Makefile stub lists → yaml ranges), main reads **91.8% (44,562 / 48,537)** and
the remainder equals the open-stub instruction sum to the instruction. Hindsight: a metric whose
denominator is a snapshot will drift the first time the thing it snapshots changes; derive it (R33),
and check it against a case whose answer you already know (the 28 stubs' size).

### S79 addendum (2026-09-04, task #4) — the "scattered-`.bss`" wall class is closed, and the probe that measured it was too strict

**Belief.** From Phase 8 to P31 S77 three SDK objects were excluded from the LINKED build as "scattered
`.bss` commons — no single NOLOAD base reproduces them": SYS.o (3,109 ins, kept as 56 hand-matched
Sony functions + 62 verbatim frags in `src/800c.c`), VM_F.o (237, hand-matched as game code in
`sgap_6`), GS_001.o (384, hand-matched as game code in `gsgap3`). S77's probe reframed two of them as
"disjoint ranges → splittable" and confirmed GS_001 as the genuine wall ("5 interleaved bases").

**What happened.** Task #4 built the split as a link-time ELF rewrite (`psyq_bss_split.py`, §489)
instead of a curated-dir artifact, and modelled the section as RUNS of one base in offset order with
cuts snapped to symbol starts — because the original linker scattered *symbols*, not offset ranges.
Under that model GS_001 is six symbol-aligned pieces, and the five cut symbols recover by name from
the other libgs objects at exactly the piece bases. All three link byte-identical; main is `143dbb89`
with and without the SDK objects; 235 placed objects across nine curated dirs pass through with zero
refusals.

**Why the probe was wrong.** It grouped references BY BASE and asked whether the per-base offset ranges
were disjoint. A common that the linker placed between two others (PSDBASEY at +0x38 sits between
PSDBASEX at +0x28 and CLIP2 at +0x30 in the packed section, but in the game X and Y are adjacent)
makes two ranges interleave while every run is still single-base. The right unit was the run; the
right tie-breaker was the symbol table.

**Hindsight.** The worklist's own italic note from Phase 8 said "escalate to a Max general fix (split
each object's `.bss` into per-common NOLOAD sections)". That fix is ~400 lines and one afternoon; it
waited twenty-three phases because three exclusions never looked worth a general mechanism, and the
probe's stricter test then ratified one of them as a wall. General form: when a tool says "no single
X reproduces it", the next question is "can X be partitioned", and the partition should follow the
structure the ORIGINAL producer used (here: symbols), not the structure the measurement happened to
group by.

### S79 addendum 2 (2026-09-04, tasks #13/#5) — the band's twelve "walls" were a library version away; the §332 wall verdicts are closed

**Belief (S68 → S78).** Twelve of main's open stubs sat in `800c3`/`800c2`, four of them curated as §332
"%lo-in-a-delay-slot" compiler walls, one as a "no jump table" wall, all excluded from waves as facts about
gcc. S78 named the band (libpad 4.2.1 + libapi 4.2, §487) but expected to finish it as C under the
reorder island because no archive we held could link it; the S78 brief budgeted a bounded hunt (#13) with
three leads and a fallback to C.

**What happened.** The first lead — archive.org — held a 383 KB 7z of the Runtime Library 4.2 with SCE's
February-1998 libpad 4.2.1 patch inside. Every one of the 46 objects it places in main links byte-identical
(cookbook §490). Task #5 then linked the whole band and the apicard region's C objects (FIRST, PAD, PATCH,
CHCLRPAD) in one pass: four TUs deleted, the `REORDER_TUS` island empty, main byte-identical with and
without the SDK dirs, fleet 213/213.

**Why the walls were wrong as WORK while right as FACTS.** §332 is a true statement about cc1 + `as -O2`
reorder output: no C source reproduces a `%lo` in a delay slot through our pipeline. But Sony never
compiled that C through our pipeline; they shipped the object. A wall verdict answers "can we match this
function?"; it never asks "is this function ours to match?" — and that provenance question, once asked
(§487), was answered by a 383 KB download. The exclude audit had the same blind spot in code: a pinned
`# WALL` outranked LINKED, so `PopMatrix`/`PushMatrix` sat as walls for eleven sessions while living in
libgte3, linked since Phase 8. Fixed: LINKED dominates.

**Hindsight.** The order of questions for any stubborn function in a mixed binary is provenance → archive
→ link → (only then) compiler. We ran it backwards for the band: wall_sweep in S68/S77, names in S78,
archive in S79. The §487 sentence — "a wall inside bytes no archive you hold can place is a provenance
question before it is a compiler question" — now has its second half: **and a provenance answer is an
archive to go and find, not a label to match under.**

### S80 addendum (2026-09-05, tasks #9/#10) — a "bank" that was the assembly, and the permuter that could never see a pin

**Context.** The completion sprint's task #9 (one agent per open function) ended S79 mid-flight; S80 aggregated the
eleven agents that outlived that session from their transcripts (`tools/agent_verdicts.py`), banked the two MATCHes
(`main:func_8001EFE0` 468, `ov_SC02_027:func_80180B3C` 297) and ledgered the eight NEARs. Task #10 was meant to be a
formality — ratify five hand-asm routines, decompile one 45-instruction GAME-C verbatim.

**What the instruments said, and what was true.** (1) The S79 #8 permuter verdicts on pinned seeds ("8 cycles, unchanged")
were ONE cycle each: `permuter_ils` re-copied the decoded waypoint (raw pins) into base.c and every later cycle was a
parser refusal; `hide_asm` only knew the `__asm__` spelling; and `defines_fn` could not see a K&R definition — the R39
control over 5,311 stored drafts found 436 K&R backlog drafts that our own coverage check had refused for four phases.
No permuter verdict on a pinned or K&R seed dated before S80 is a measurement. (2) S79 #7's `md_MAIN_020:func_800CB17C`
"raw splice bank" had spliced the function's ASSEMBLY (the ledger's best_draft was the asm) as a verbatim `__asm__`
body — a P9 fake bank the byte gate cannot see (it IS the bytes) and `progress.py` counted. `verbatim_check --strict`
caught it in task #10, one session later. (3) The one GAME-C verbatim's 37 failed drafts were a declaration: the TU
said `extern void f(void)` for a function it only ever takes the address of.

**Decisions.** Fix the instrument before re-measuring (R35): the permuter fixes shipped with an R39 control and a
positive control on the S79 refusal, then the 8-seed sweep re-ran (no score-0; `func_80039DEC` 9→2, `func_80023BF0`
18→11; a "1" on `func_8017DF28` was a divergent store rewrite and was REFUSED as a closeness — R14). The ≤3 residuals with
a mechanism citation and an ILS null are WALL candidates in `config/wave_exclude.txt`, distinct from §474-PROVED. The
verbatim "bank" was reverted to a stub (byte-neutral) and decompiled properly; both def-side declaration walls fixed as
byte-neutral TU plumbing committed BEFORE the gate (the S77 law). `parallel_gate` now exits 2 on banked-but-not-merged.

**Cost.** Two fake or phantom results survived a session each because the instrument that would have caught them
(tools-health's strict verbatim check; a re-read of the permuter's diff) was not in the per-bank loop. ~0 drafting tokens
were spent on #10's two functions: the C existed (or was seven calls); the cost was the declarations.

**Hindsight.** Run `verbatim_check --strict` in the per-bank close, not only at tools-health; never let a ledger's
`best_draft` be an `__asm__` body (R48's cousin: a draft keyed by name with no "is this C?" check); and read a permuter
waypoint's diff before recording its score.

## P32 S81 (2026-09-05) — three zero-token banks and the parked five onboarded: every blocker was an instrument or a mis-keyed ledger

### Context and belief going in
Phase 32 opened on `docs/frontier-p32.md`: 21 stubs, four of them "B-CARVE" (a jump-table carve the tooling refused), and
five disc payloads parked as static-RE targets whose loaders "are individual + condition-gated". The plan costed the carve
class as tooling work, the md_MAIN_034 row as a Sonnet redraft (its best draft was "wrong-sized"), and the parked five as a
static-RE session that might need the Ghidra MCP.

### What actually happened
* **resident `func_800D128C`** — the stored S71 draft was byte-correct all along. Three instruments stood in front of it,
  each an overlay-only assumption the resident (the fleet's one `common.h`-only, `--pre`-sandwich binary) exposed:
  the isolation carrier dropped a typedef whose name engine_types.h also defines (§496); the carve tool regenerated
  `JTBL_INTERLEAVE` without `--pre hdr.rodata.o`, the extract refused, and the gate linked a STALE script and booked
  the draft as DIFF because `harvest_verify` never read the extract's exit code (§498); `interleave_check` read the
  `--pre` line as n=0. Fixed all three; banked.
* **ov_SC02_017 `func_80186C64`** — the "rename needed" refusal was the carrier keying a bodiless
  `typedef struct Tag Alias;` by the TAG (§497); with the alias key the split was CLEAN with no source change, and the
  d=2 twin remapped in one `family_remap` call plus four TU spellings the real-TU check named (§376).
* **md_MAIN_034 `func_800CB00C`** — the census's best draft was a DIFFERENT function's file under the same bare name
  (R48); the journal (R38) named the real one, `rtu_match` said MATCH 123/123, and the S72 resolver had gated only the
  wrong file three times. The S68 "compiler wall" pin labelled that wrong draft. Banked; pin dropped.
* **The parked five** — a static header probe (`payload_base_evidence.py`, §499) pinned MAIN/7 and MAIN/9 by their
  self-calls (9/9, 6/6 on their own function starts) and SC03/53/54 by their fn-ptr tables (0x801EF468); SC03/56's
  vote base was a shared-engine coincidence caught by asking whether its targets were overlay functions. All five
  onboarded on the first candidate; fleet 213 → 218; the parked ledger is empty.

### The measurements that settle each one
Byte-identical builds: resident `8e17e02f…`, ov_SC02_017 `c0253499…`, md_MAIN_034 `46153c06…`; R22 213/213 after each
batch. Instrument controls: carrier provided-types (overlay set identical, 1,197 names; resident set lacks `CdFileLoc`);
`_merge_pre` 4 shapes; evidence tool 7/7 banked modules re-derive their bases. **The first-build byte check is base-
lenient:** MAIN/7 at +8 builds byte-identical, at +0x1000 fails the link — so the §S44 corroboration never proved an
address, only its class.

### The shape they share
Six defects, one shape: a tool written for the overlay class met the resident/module class and stayed silent (a
dropped typedef, a dropped `--pre`, an ignored exit code, an n=0 parse, a tag/alias namespace collapse) — and a ledger
keyed by bare name pointed a redraft at the wrong file. R40 (exonerate the instrument) and R38 (read the recorded
verdicts) converted a planned agent redraft and a planned static-RE session into zero-token banks. R39 (negative-control
a new checker) caught the evidence tool's own two false scorers before they shipped.

### Cost, and the hindsight path
~0 drafting tokens for three banks and five onboardings; the cost was the diagnosis. Sooner: the `frontier-p32.md`
routes should have carried each row's JOURNAL verdict (R38) next to its best-draft path — two of its three "draft"
routes were already MATCH; and the §S45 p6 "onboard at 0x801EF468 and let the first build decide" step should have run
in P30 — with the finding that the first build decides less than it seemed to.

## P32 S82 (2026-09-05) — the T3 wave: 31 drafters, 20 MATCH, a coordinator that died at bank 9, and eleven deliverables swept by one agent's tidy-up

### Context and belief going in
T2c had left 54 stubs (7 pinned walls, 11 old near/far rows, 36 never-drafted stubs in the five freshly onboarded
modules). The plan's T3 was "one bounded crack pass, Opus agents, one per function" — the S80 one-agent-per-function
shape, expected 1–4 banks from the 11 old rows. The belief: the old rows were the genuine hard tail (S79 had banked 3 of
11 on the Opus tier); the new-module stubs were cheap Haiku work; the coordinator could process results one at a time
as they arrived.

### What actually happened
* **Yield far above the estimate:** 20 MATCH / 9 NEAR / 2 FAIL of 31 — Opus 7/16 MATCH with the other 9 NEAR at exact
  length and zero walls; Sonnet 4/4; Haiku 9/11 with both FAILs not compiler walls. Six of the eleven "hard" old rows
  MATCHed (func_800D06E8, func_80039B20, func_80038698, func_80023BF0, func_80015B6C, func_8002FDE8), and their closers
  were all things the journals had mis-labelled: an alias flag called a scheduler tie, a variable SCOPE called a
  self-coalesce wall, a hand-written mask pair that is the libgpu P_TAG bitfield, an array-vs-scalar extern spelling
  documented two functions away in the same TU, a frame-slot hack that itself caused the "wall" it was meant to fix.
* **The coordinator overflowed** ("Prompt is too long") at 01:07 MDT, four minutes after its ninth bank; 22 completions
  arrived into a dead session. The launch HAD been checkpointed (queue file + recovery route in `LAUNCHED.md`), so the
  successor session recovered every verdict with `agent_verdicts.py`.
* **Eleven Opus deliverables were missing.** One agent, tidying its own scratch in the SHARED `.run/P32/t3/opus/`, ran
  `find … -maxdepth 1 -type f ! -name <mine> -exec mv {} _scratch/` and moved every sibling's file — two MATCHes among
  them (604 + 120 ins). Found by reading the transcripts' commands; the files were moved, not deleted. A transcript-replay
  tool (`agent_drafts_restore.py`) was written as the fallback and rebuilt 26 of 30 exactly (the 4 edited via shell
  after their last Write were stale — the on-disk `_scratch/` copies were the truth).

### The measurements that settle each one
Every unbanked draft re-verified by the successor with `rtu_match` in its real TU: the 10 MATCHes are MATCH, the 9 NEARs
reproduce the agents' closeness to the instruction (2 / 6 / 15 / 17 / 27 / 35 / 46 / 49 / 137). The 10 banks of the
producing session each built byte-identical (main `143dbb89…` via `gate_main` ×3; resident `8e17e02f…`; md_SC03_053
`c0848f30…` ×4; md_SC03_054 `06bd73df…`; md_MAIN_007 `2ff702b6…`); no fleet R22 has run since — the resume order starts
with one. Census 54 → 44 stubs / 5,313 ins.

### The shape they share
Two process defects with one root: the S80 shape was designed for 11 agents and run at 31. A shared scratch dir is a
shared blast radius (R48 — bare-name files, one dir), and a coordinator that ingests 2–4 KB of prose per result cannot
survive 31 results in a session already eight hours old. Neither is a model failure; both are harness shape (R40). And
the matching finding repeats S79/S80's: **not one of the six "hard" rows that fell was a compiler wall** — each was a
mis-read mechanism, and in two cases (func_8001BC6C's five converged agents at 28; func_8002FDE8's four attempts at 35)
a multi-agent consensus on a MECHANISM had been mistaken for evidence about the BODY.

### Cost, and the hindsight path
~16 Opus + 4 Sonnet + 11 Haiku agents (~30–70 min each for Opus) for 20 MATCH (2,111 ins) and 9 exact-length NEAR seeds
with cited mechanisms; the recovery cost ~1 hour of a Max session and one new tool. Sooner: (1) the JSON-only final
message and per-function work dirs should have been in the S80 shape from its first run; (2) start a 30-agent wave from
a FRESH session, not at the end of a T0–T2c day; (3) packs should carry the same-TU neighbours' DECLARATIONS of the
shared globals (the func_8002FDE8 fix was two functions away); (4) a "converged plateau" in the journal is a reason to
re-read the body against a neighbour, not a reason to route to the permuter.

## P32 S83 (2026-09-05) — the crack pass closed on 39 banks, the walls got their final ledger, and every remaining "wall" was re-probed in TU context without touching src/

### Context and belief
The 10:01 checkpoint handed S83 a census of 44 stubs: 10 verified-MATCH rows awaiting the gate, 17 Haiku rows never launched
(the 20-agent cap), 9 NEAR seeds and 7 pinned walls. The belief going in: the MATCH rows were free banks, the Haiku rows were
~50% yield on the ≤25-ins band, the walls were settled.

### What happened, measured
- **The MATCH rows were free only per draft.** Eleven of twelve same-TU drafts were `rtu_match` MATCH alone and the batch
  failed twice: a `s16`/`u16` spelling of one global across two drafts, a §304 rodata block only the stub `.s` owned
  (invisible to a compile-only oracle — it fails at LINK), and a prototype a sibling bank introduced between an agent's
  verification and the splice. The BUILD is the batch verdict; a per-TU bank chain (`bank.sh`) made it deterministic.
- **The Haiku band went 17/17 first pass** (13–25 ins, ~50k tokens each, 46–126 s) — the S82 9/11 was an under-estimate
  because the two S82 FAILs were plumbing, not model failures. 28 banks in one session, then a 29th from the permuter.
- **The permuter's "1" was two sound levers plus one wrong-width mutation.** Subtracting the unsound hunk left a leaf MATCH;
  the lever was the §47 live-length slider (an early BIRTH of a pseudo), re-spelled well-defined. A waypoint carrying a
  semantic mutation is a seed, not a rejection (R63 both ways).
- **The `func_800CF3E8` second look corrected a mechanism and refuted a lever** (cse.c `find_best_addr`, not `canon_reg`;
  the alias lever 5/5 inert; a new pinned-pointer launder frees the load but lands it 4 slots late) — 245k tokens for a
  closeness that did not move, and a citation that is now right.
- **T4: all seven pinned walls reproduce their residual in TU context** — and the three that were CC1 FAILs were
  declaration plumbing, re-probed in a SANDBOX TU copy (`.run/P32/t4/tu/`) rather than by byte-neutral commits to `src/`.
  1 PROVED (§474) + 6 CANDIDATE, citations current. Nothing changed; everything is now measured where it will be judged.

### The shape, and the hindsight path
Three sessions running (S81, S82, S83) the finding is the same: the last stubs fall to PLUMBING, not codegen — declaration
environments, rodata ownership, TU spellings — and the instruments that judge them (a per-draft compile-only oracle; a
`find src -name '*.c'` that admits a live probe) are the walls' co-authors. Sooner: (1) `rtu_match --batch` from the first
multi-draft TU; (2) the §304 sentence in every drafter brief from the first module wave; (3) a sandbox TU for any CC1-FAIL
re-probe — it costs a `mkdir` and two symlinks and spends no commit; (4) treat a permuter waypoint's diff as a lever list.
What P32 leaves for P33: 15 functions (7 walls with citations, 8 near-misses with cost), every module at 100% C except the
ones those rows sit in, the fleet byte-identical 218/218 on every sweep of the day.

## P32 S83 T4b (2026-09-05) — the "final 15" directive: hand-crack first, then one Fable agent per row — 9 of 12 completed rows banked, six of them former pinned walls

### Context and belief
At the T4 close the ledger said: 7 pinned walls (1 PROVED, 6 CANDIDATE with gcc citations and 3–5 bounded attempts each, permuter
null, an Opus second look on one) and 8 near-misses (2…137 rows off at exact length). The belief, held since S79 and re-affirmed at
T4 with TU-context re-probes: the candidates were real compiler walls; the phase would close with 15 rows "ledgered with cost".
Drew's directive changed the finish line: nothing but the original hand-asm and the PsyQ objects may remain; hand-crack each row
and name its blocker before spawning agents; Fable agents permitted (later clarified: permitted, not required).

### What happened, measured
- **The hand pass (Fable 5.1 as coordinator, ≤3 probes per row, 22 spellings): 0 banks.** It refined every blocker to a named
  mechanism and wrote the untried lever per row (`.run/P32/t4b/<fn>/NOTES.md`) — and was WRONG about the lever on the two rows it
  measured hardest (func_800391D4's explicit promotion "18, worse" was a cascade; func_80032A74's caller-save area was refuted by
  the agent with citations). Its value was the mechanism map handed to the agents, not the probes.
- **One Fable agent per row, ~2 h budget, pass dumps + gcc source in the brief:** of the 12 rows completed at this writing, **9
  MATCH** (banked byte-identical, one commit each, pins dropped) and **3 NEAR** with the residual attributed to a pass:
  func_80032A74 1 (a ghost pseudo — near-proved), func_80185810 35 → 13, func_80039308 17 → 4. Three rows are in flight.
  Six of the nine were the T4 "CANDIDATE walls" — every one fell to a mechanism the pin's citation had mis-attributed
  (loop.c → cse quantity; cse → sched1 birthing boost; argument position → allocno priority; alias basin → three passes).
  The one "PROVED" wall (§474, −O0) is still in flight.
- **Cost (R41):** ≈3.9M subagent tokens for 12 rows (≈325k/row; 123k–655k), three usage-limit outages that killed every
  agent mid-run (resumed with context intact via SendMessage each time; deliverables written early survived). For comparison,
  the same 12 rows had absorbed the S79 Sonnet/Opus passes, the S82 Opus wave, the S83 Opus second look (245k tokens, 27 → 27)
  and ~16k permuter/sweep compiles with 0 banks.
- **What every crack had in common:** the agent READ the pass dump the prior attempts had not (`-dS` ready lists with the
  `7f000001` birthing boost, `.loop` desirability lines, `-dl` quantity priorities, `.greg` dispositions, `-dR` hazard walks)
  and attributed the residual to a specific pass and line of the 2.7.2 source before touching a lever; then the lever was
  usually one zero-byte dial (a second live set, a hard-reg copy, a launder, two more pads) or a libgpu-idiom spelling.
- **Instrument defects met on the way:** the bank helper called with an empty function list built the unchanged tree and
  exited 0 (two premature "banked" ledger messages — corrected in the next commit; R43 hardening); gate_main's
  `--assert-baseline --allow-dirty` restores the working tree before building (its GREEN measured the committed tree);
  the helper derived the rtu `--split` from the binary name and refused `_jr_` TUs (SPLIT override).

### The shape, and the hindsight path
The T4 wall ledger was "instrument-exonerated" but not "pass-attributed": each pin cited a mechanism, and six of seven were
the wrong pass. The distinguishing discipline of the agents that cracked them was not model size alone but a budget that
allowed reading three dumps and the source before the first probe — the hand pass, at three probes per row, could not afford it
and mis-read two cascades. Sooner: (1) a wall verdict must name the PASS and the dump line that proves it, never a mechanism
alone (R40 → "pass-attributed"); (2) the §172 "canonicalization wall" and §474 should have been re-tested by a dump-reading agent
before being cited as proofs for years of pins; (3) write deliverables early — every agent lost to a usage limit had to be
resumed, and the one that had written its draft first (func_800391D4) banked from the dead run; (4) Fable on the
compiler-internals class is cheaper than another Opus wave on the same rows (9/12 vs 0/12), and Opus is the right tier once a
sibling's recipe exists (§501-H → the md_MAIN_007 pair). No second round this session (Drew); the three NEAR rows carry their
next lever in the backlog.

## P32 S85 (2026-09-06) — T4b hand pass row (d): the last overlay stub falls to a sibling the atlas had scored 0.55

### Context and belief
Four rows were left after the S83 Fable pass; S84 settled three (one PROVED, one BANKED by a sibling port, one PLATEAU). Row (d)
`ov_SC03_105:func_80185810` (489 ins, DIFF 13) carried a report that had read its residual's mechanism from the dumps — the
unboosted 2nd set behind a needed fence — and judged every honest fix blocked by combine. The checkpoint's step 0 was the §501-N
sibling search; the twin oracle, the family maps and the atlas's own knn entry for the row all said "nothing".

### What happened
A grep for the idiom's constant (`'0x200) << 2'`, the libgpu getTPage chain) across `src/` listed `ov_SC02_027:func_80180B3C`
— which the atlas had recorded only as a reverse 0.55 "weak cousin" — and its objdump window was the target's rows 362–386
instruction for instruction. Porting its window spelling matched in the real TU on the first draft, then with every pin removed:
zero pins, zero fences, zero asm dials, BANKED `cdd9a2cb8`, ov_SC03_105 100% C. A 12-variant census then measured which spelling
elements carry the window (the fresh single-set masks, the branch polarity, the early unpinned shift) and refuted the S83 guess
that a hard-register pin was what kept the `andi` alive (§501-P).

### Why (measurement-grounded)
The similarity metrics score REGISTER-RENAMED same-shape functions as strangers (base registers s0/s2/s1 vs t0/t1/t2 change most
of the instruction words); the idiom's literal constants do not change. The draft's "blocked" verdict was correct about the
mechanism and wrong about the cause: the block's other 2-set births (the `uu -= …` pair, the polarity) were what made the
fence necessary.

### Hindsight path
Step 0 of every hand pass and every wave card for a packet-building function: grep `src/` for the idiom constants and objdump the
hits' windows against the target BEFORE the first dial — the whole crack was one grep, one port and 25 minutes, after ~1M tokens
of dial work on the draft. Remaining open: the two main rows with their verdicts (PROVED 1, PLATEAU 4), carried into the PhaseEnd.

## P32 S85 (2026-09-06) — T4c: the last two "final verdicts" banked; the frontier is empty

### Context and belief
After the T4b close the phase held 2 stubs, both in main, both carrying verdicts written by producer censuses: `func_80032A74`
PROVED at 1 (every stack-slot producer refuted), `func_80039308` PLATEAU at 4 (the phantom slot + a hoist order). Drew:
"let's crack these before closing the phase."

### What happened
The §501-N step 0 found the shape cousins in main's own TUs but the cracks were compiler mechanisms the censuses lacked:
combine's self-update bookkeeping gap mints a no-traffic frame slot (§501-Q — one chain, first probe, `f9a90affb`), and a
hoisted invariant read three times needs three inline temps merged by loop.c plus a `u16` accumulator so cse leaves the arms
on the hoisted register (§501-R — 56 variants, then MATCH with zero pins, `ffb1949a1`). Every pin the S79–S84 drafts carried
came off byte-identical. Census: 0 stubs.

### Why (measurement-grounded)
Both verdicts were "instrument-exonerated and pass-attributed" (R40, §501) yet wrong, because a census is only as complete
as its list of producers: `i2dest_in_i2src` was not in the list, and "a named variable can hoist" was assumed where loop.c
refuses it. The reproducer battery (1-second compiles of 5-line functions) settled in minutes what the real function could
not in hours: the mechanism first, the site second.

### Hindsight path
Build the reproducer battery for every species BEFORE the census (accelerators (15)), read the allocation order before any
register lever (accelerators (16)), and treat "PROVED" as "proved against this list" — a wall verdict should name the list.

## P33 S86 (2026-09-06) — the public flip is IN PLACE with the full history, not a fresh mirror
- Context / belief: roadmap v2 §3 P33 and gen2-roadmap Phase 14 planned a curated, freshly-created public mirror
  (copy-in, no history rewrite, the private master untouched). The S86 history audit (every blob on every ref) found the
  ROM-derived and proprietary content confined to five path sets — the EXE (two historical paths), 28 RAM dumps, the
  Ghidra project (verified to embed the EXE's bytes under Ghidra's page XOR mask), Sony's PsyQ SDK, and a 271 MB
  session-transcript archive holding ~260k lines of game disassembly — plus two redistributed third-party binaries; no
  secrets anywhere; the disc, `asm/`, `assets/`, `expected/`, `build/` never committed.
- Dead-end (if any): none tried; the mirror model was rejected on the owner's stated priority — "maintain the commit
  history, timeline, commit amount" — which a fresh mirror discards.
- Pivot: rewrite the history with git-filter-repo (every commit, date, message and order kept; hashes change; one
  all-purged commit drops), scrub the ~700 cited old hashes from every historical blob and message into inert
  `commit:NNNN` tokens with a committed map and a tip commit that resolves them at HEAD, map the two personal e-mail
  identities to the GitHub noreply address, push the current history to a private archive repo first, force-push the
  rewrite over `Druthulu/BFM-decomp`, and gate the visibility flip on GitHub no longer serving the old hashes (Support
  purge, probed by script; delete-and-recreate under the same name is the fallback).
- Why: hashes are not secrets by obscurity here — our own docs publish them, so old objects reachable by hash on GitHub
  would be reachable by anyone; a content-preserving rewrite keeps everything the story and retrospective are mined
  from (dates, messages, the historical `docs/progress*.md`); the fixed-point argument (a new hash written into a doc
  changes every descendant hash) is why tokens live in history and real hashes only at the tip.
- Hindsight / for the wiki: decide the public/private boundary on day one and keep ROM-derived bytes out of git from
  the first commit even while private (R1's relaxation bought convenience and cost a full-history rewrite); cite commits
  by date + subject in long-lived docs, never by hash alone, if a rewrite is ever conceivable.

## P33 S86 (2026-09-06) — main's game-code denominator was Ghidra's, and Ghidra under-counted by 3,616 instructions
- Context / belief: the published "MAIN game-code weighted 41,534 / 41,534" came from a Ghidra sig whose function
  boundaries are flow-derived; P31 S79 had corrected one instance (+22 on FUN_80023bf0).
- Pivot: `make sig-main` now derives every game-code function's boundary from the BUILD (link-map `.text` sections ×
  each object's `nm` symbols; tiling asserted), Ghidra-free and regenerable on a public clone.
- Why: the derived sig tiles the game-code text exactly (45,150 words) and is never shorter than Ghidra's; Ghidra left
  3,628 words of real game code owned by no function (unresolved switch tails, 52 two-to-four-instruction thunks, a
  jump-table target labelled as a function). All still 100% — the denominator moved, not the verdict.
- Hindsight / for the wiki: a metric's denominator should come from the artifact you control (the build), not from the
  analysis tool; when two instruments disagree by a systematic offset, the "corrected" instance you found is usually one
  of a class.

## P33 S87 (2026-09-06) — the purge set leaves the INDEX before the history rewrite, as its own commit (C3)
- Context / belief: the rewrite (git-filter-repo, Block C) removes the ROM-derived paths from every historical commit;
  the natural reading was that the tip's removal is just the last instance of that and needs no separate step.
- Pivot: a separate preparatory commit first — `git rm --cached` of the whole purge set (`tools/public_rewrite/
  purge_set.txt`), files kept on disk and already gitignored — so the rewrite starts from a tip whose tree equals the
  working tree, and only THEN the bundle, the archive mirror, the bare clone and the filter.
- Why (measurement-grounded): (1) the rewrite becomes purely content-preserving — `verify_rewrite`'s rule that any added
  path fails, and A5's recorded run (which cannot see the index) stays valid for the tip, since a `--cached` removal changes
  no tracked-content byte; (2) the first-push gate (`tools/audit_public.py`) is exercised on a real state: it named exactly
  the purge set (255 offender rows) before this commit and passes after it; (3) the ignore rules were proven with
  `git check-ignore --no-index` on every path BEFORE the paths became untracked — a gap there means a blanket `git add -A`
  re-adds ROM bytes into the public history; (4) the build was proven not to need any of it (main byte-identical with
  `tools/psyq/` and both SDK object dirs moved aside — B3's fresh-clone proof still had `tools/psyq/` in the clone);
  (5) a census showed nothing project-authored under a purged directory (the SDK checksums file had already moved out in B4)
  — anything left there vanishes from ALL history, not just the tip.
- Hindsight / for the wiki: the data-loss hazard flips at this commit, not at the flip — ignored-but-present directories
  are one `git clean -x` from deletion, so the guard (CLAUDE.md fail-safe line, the bundle, the archive repo, the text
  export) belongs in the same commit; and a mailmap or a runbook that names the personal addresses would itself need
  scrubbing, so the mailmap is scratch and the runbook refers to it by path.