- Propagated 0x8012A464 / 0x8014FFDC / 0x801502EC -> 138/138 byte-identical, 0 stragglers, 3 new
groups. R22 clean-fleet 140/140; dedup 1872/0; C1 237654/237654.
- FINDING (R14/R31): the §20 propagation cap was gating source-level DE-DUPLICATION, not coverage.
The whole uniquify campaign (Buf+MATRIX+Vec8, 223 files renamed, ~1559 copies stripped, 5
propagations, 4 R22 cycles) moved the fleet by +6 functions / +558 ins / -6 stubs and 0.00pp on
all three headline metrics. The freed cores' members were ALREADY matched individually;
propagation just consolidated them into shared macros. "Unblocked" != "unmatched".
- SESSION ATTRIBUTION: of -837 stubs / +0.3pp instr / +0.23pp fn-count, -831 stubs came from the
FIRST batch (broad lift -> 13 cores). The uniquify campaign contributed -6.
- Roadmap B4 re-labelled: a maintainability item, not a coverage lever. Remaining camps
(Handler/Blk8/V8/Prim/Prim_8016E7C8) are small and now known low-yield — opportunistic only.
- distinct-code sat at EXACTLY 3811442/5634875 = 67.6% at open and close. Fresh cracks are the sole
mover of that number; point the next session there.
MEASUREMENT CORRECTED THE PLAN. The checkpoint called for a "per-camp field-access reconcile";
measuring the camps refutes that: Vec8 = {s32 w[8]} (32B) in 180 files AND {s16 unk0..} (8B) in
139 files; MATRIX 48B/32B/32B; Buf 16B / 0x20+ / DrawEnv. These are DIFFERENT types sharing an
identifier across TUs of the same overlay — reconciling to a canonical layout MERGES them, the
same failure that broke 103 binaries on Prim. The right op is UNIQUIFY: rename the non-majority
camp (byte-neutral — a type name emits no code; TU-local by construction), which makes every camp
single-def and liftable by the existing lift_types rules.
- NEW tools/uniquify_type.py: deterministic camp ordering (file-count desc, then normalized text,
so re-runs assign the same suffixes); majority keeps the name, camp n -> <T>_c<n>; rewrites ONLY
files that DEFINE that camp (a file that merely USES the name gets it elsewhere and is untouched);
\bT\b word boundaries so `Buf` never matches `Buf80153978`.
- VALIDATED on Buf (578/6/1 files): 11 identifiers across 7 files -> 3 camps LIFTABLE -> lifted
(585 local copies stripped) -> R22 140/140 -> blocked core queue 13 -> 11 (0x8012ea90, 0x801749c8
freed). Propagated 0x8012EA90 ×138; 0x801749C8 dropped (straggler in ov_SC07_006).
- YIELD, HONESTLY (P9): ZERO new matched functions. fn-count 88.61% / instr 79.3% / stubs 40281 all
UNCHANGED; dedup 1867->1868, C1 +138. 0x8012EA90's members were ALREADY matched in all 138
overlays — the propagation consolidated duplication into one shared macro (DRY), not coverage.
The value is the PROVEN RECIPE + the queue moving 13->11, not the numbers.
- dedup_propagate (R32): the skip line printed a COUNT and no names, and aggregated three unrelated
causes into n_local — a body skipped merely for a `//` comment (macro-unsafe, 1-line fix) read
identically to one genuinely using an overlay-local type. Now named and split by cause.
- cookbook §64a (uniquify-vs-reconcile + the validated recipe + remaining camps by cost).
With the §20 local-type cap lifted (commit:0863), dedup_propagate --auto-from planned 17
self-contained cores that were previously skipped "not self-contained (local types)".
- BANKED: 12 cores via --auto-from + func_80175308 (propagated separately, 138/138) = 13 ×138.
func_80175308 is the PROVABLE unblock — the core the SESSION-13 checkpoint named as
local-type-blocked. No claim is made that all 17 were unblocked by the lift; measuring that
needs a pre-lift re-scan I did not run (P9).
- 5 correctly DROPPED as cross-overlay stragglers (0x8012A018, 0x80172C50, 0x80173A60,
0x80144090 in ov_SC01_000; 0x801495C4 in ov_SC07_006): h_exact sharing is all-or-nothing and
those overlays' bytes diverge. --recover NOT used (the documented quadratic thrash hazard).
- GATES: R22 clean-fleet 140/140 BYTE-IDENTICAL (2nd full cycle this session); make report green;
dedup-check 1854 -> 1867 validated / 0 failed, C1 236964/236964; 0 NON_MATCHING (G4).
- FLEET: instr 79.0 -> 79.3% | fn-count 88.38 -> 88.61% | INCLUDE_ASM stubs 41112 -> 40281 (-831).
distinct-code stays 67.6% — correct: propagation replicates already-distinct-matched code;
only a fresh crack moves that number.
- STILL BLOCKED: 13 cores "not self-contained (local types)", blocked by exactly the 8 deferred
VARIANT entities (MATRIX 3-def, Buf 3-def, Vec8, Prim, Handler, Blk8, V8, Prim_8016E7C8).
Next lever = a per-camp FIELD-ACCESS RECONCILE, not a lift (lifting them blindly is what broke
103 overlays earlier this session).
Crack wave w9lidyi5b (24 fresh LIVE=138 families): 20 self-assessed MATCH, 3 near.
Banked x1 (R22 clean-fleet 140/140):
- 2 self-contained via plain harvest_verify: func_8012B4B8 (§52b-wall crack),
func_80169228.
- 5 via gate_stage's reconcile ladder (cast_call_sites in the draft's own TU):
func_80175308, func_8012E138, func_80130C08, func_8012A1BC, func_80137178.
Correction (decision-log follow-up): fix_header_decl v1/v2 is FRAGILE for shared
multi-caller decls — rewriting an engine_core.h decl breaks callers that use the
return differently (CC1-FAIL). func_8014CD80 was a lucky single-caller/ignored-return
case. gate_stage's call-site-cast is the right tool for multi-caller plumbing (banked
5 where fix_header_decl broke the build). Remaining ~13 near/deeper-plumbing drafts
staged in .run/drafts-sc07006-fresh/. Propagation x138 next.
Bounded probe (SESSION-13, token-free) that REVERSED the same-session "fresh
reach-138 well is spent" verdict. Target func_8014CD80: 138 live, 0 matched, NO
DEFINE macro, a universal body (only universal callees + param offsets, zero
overlay-local D_* refs), clean MATCH draft from batch-1.
- Blocker was a def-side header decl: engine_core.h DEFINE_func_8014CD0C()
forward-declares it `void func_8014CD80(s32,void*,void*)` while the byte-true
def is `int func_8014CD80(s32,u16*,u16*)`. gate_stage's arity pre-pass is
param-COUNT-only (misses return/ptr-type); §54 reconcile_def_sig fixes the
wrong direction.
- One byte-neutral header edit (void->int, void*->u16*; call site passes u16[3]
arrays + ignores the return -> codegen unchanged) -> harvest_verify banked x1
BYTE-IDENTICAL -> dedup_propagate --addr propagated 138/138 overlays
byte-identical (live 138->0) -> R22 clean-fleet 140/140.
- Fleet 78.7->78.8% instr, 88.22->88.26% fn-count from this one family;
tools-health green (dedup 1851/0).
Quantified market (decision-log 2026-07-23): of the 75 fresh (>=100-live)
families, 46 carry an engine_core.h caller forward-decl, 38 SIMPLIFIED = the
func_8014CD80 pattern -> each a candidate x138 (~+1.5-2.8pp instr). NEXT: build
tools/fix_header_decl.py + a fresh-family wave. func_80165CA0's 0/135 was a
non-universal BODY (different failure mode), not this blocker.
Propagation of the 6 batch-1 x1 banks (§55b: banks committed first in commit:0848,
then targeted propagate as a standalone step):
- dedup_propagate --addr: func_801325B8 -> +3 onboarded-tail siblings
(ov_SC07_007/010/011). func_8014A048/func_801678F0 byte-diverge in the SC07
cluster (kept x1); func_8014FE60/func_80167540 local-type-blocked §20 (x1).
- func_80165CA0: consolidated its h_exact subgroup (dedup group registered, +0
new), then family_sweep --hseq 0/135 — a PER-MEMBER WALL (cf func_80133AB0
0/136). The x135 "fresh family" prize does not exist here.
- Net batch-1 yield ~9 newly-matched functions; fleet 78.6->78.7% instr, distinct
flat; ov_SC07_006 84.6->84.8%. R22 clean-fleet 140/140; tools-health 1850/0.
The finding (R14/R35, decision-log 2026-07-23): nins*reach leverage over-counts —
rank by LIVE-siblings. build_wave_args.py --rank live now ranks by the true lever
and reports the fresh(76)/onboarded-tail(44) split. The reach-138 family well is
largely SPENT via wave+gate; the fresh families are the hard tail (def-side
plumbing/DIFF/per-member walls), not free x138 fuel.
- batch-1 wave: 25 non-jtbl ov_SC06_018 targets (reach 3-14 modal), 13/25 match_one
MATCH; whole-binary banked 8 (4 plain + 4 via gate_stage --src-file), swept 4 families
-> 30 members across 13 overlays. 5 matches deferred (missing-sym/s58/deeper plumbing);
12 nears are permuter fuel (several close=2/3/4).
- KEY: non-jtbl fns in a jr-split file need gate_stage --src-file <the jr TU> (the default-
TU reconcile misses them -- same class as the jtbl --src-file fix).
- BUG FIX (R33): gate_stage's arity-undo snapshotted src/<bin>/*.c BEFORE _gate1 splices
the banks there, so an unbanked draft in the batch triggered a snapshot-restore that
SILENTLY REVERTED the banks (measured: a 9-draft run banked 4, the 5 unbanked reverted
all 4 to INCLUDE_ASM). Fix: restore ONLY src/shared/ (the fleet hazard the snapshot
exists for); the binary's own TU arity edits are local + byte-neutral. GATE_NO_ARITY=1
was the interim workaround. s61's law from below (undo scope must not EXCEED write scope).
- incremental check-all 140/140 (concurrent with batch-2 drafting; full R22 after batch 2).
fleet distinct 67.5->67.6% (+38 unique fns), instr 78.6% steady.
- func_8017D648 (102, reach 82, MODAL/cross-address family): cracked fresh in ov_SC06_018
(+ a byte-neutral s17a-1 callee-cast reconcile of func_8012F14C to the TU's 1-arg
convention), banked x1, then swept 70/82 siblings (85%) via family_sweep --hseq
--source ov_SC06_018 --allow-pins (+3 gate-fail, +8 remap-refused unresolved-immediates).
- 3rd data point for the family-specific finding: across 3 thesis families the fresh-
exemplar sweep rate is 100% (func_801365B8) / 0% (func_80133AB0) / 85% (func_8017D648).
Refined model (-> decision-log): the sweep works for the MAJORITY of families at high
per-family rates, with a genuine per-member-wall minority; cracking is the generator.
- R22 clean-fleet 140/140 byte-identical; tools-health green; 0 NON_MATCHING.
fleet 78.5->78.6% instr / 88.18->88.20% fn (session net 78.4->78.6 instr / 67.1->67.5 distinct).
- binary-aware crack wave (new tools/workflows/wave_binary.js): 8-target calibration
over ov_SC06_018 substantial stubs, 7/8 match_one MATCH
- func_801365B8 (155, reach 133): cracked FRESH in ov_SC06_018, swept 132/132 siblings
via family_sweep --hseq --source ov_SC06_018 --allow-pins -- SESSION-10 refused this
family 0/133 from an ov077 exemplar. THESIS CONFIRMED (fresh exemplar unlocks it).
- func_80133AB0 (137, reach 137): cracked fresh + banked x1 (+ a byte-neutral s17a-1
cast reconcile of banked caller func_801343C4), but the family sweep FAILED 0/136 even
from the fresh exemplar (reverted clean) -- THESIS REFUTED for this family.
- FINDING (R14/R31 -> decision-log): the fresh-exemplar sweep is FAMILY-SPECIFIC, not a
blanket mechanical x137. A fresh crack is necessary but not sufficient; the byte-gate
arbitrates each family (~50% on this 2-family sample -> discount the ~1.5pp estimate).
- tooling (R33): family_sweep --source override now searches matched_members (a fresh
member leaves 'members' after a sig-regen); cdecl._depth0_spans consumes backslash
line-continuations so a raw-draft #define macro no longer trips audit-cdecl.
- R22 clean-fleet 140/140 byte-identical; tools-health green (dedup 1849/0, C1 234615);
0 NON_MATCHING. fleet 78.4->78.5% instr / 67.1->67.5% distinct / 88.14->88.18% fn-count.
The contract requires all three headline metrics to include the main EXE. Since
Phase-27 T10 main was reported as a SEPARATE provisional line, so the headline
silently measured 139 of 140 binaries — and flattered itself by omitting the
LEAST-decompiled one.
RESTATED CAVEAT (the old "stale, PROVISIONAL" wording was misleading): main's sig
is Ghidra's (2026-06-14), but function BOUNDARIES derive from the original bytes
and do not change, and matched-vs-stub comes from the LIVE corpus.stubs — so the
numbers do NOT drift. The real limitation is R34: sig_image cannot independently
validate a PS-X EXE's boundaries, so main has no SECOND, DISAGREEING oracle for the
PHANTOM/TRUNCATED class. The sig also excludes the LINKED PsyQ objects, which is
exactly right for a GAME-CODE contract.
EFFECT — the headline DROPS, which is the point:
instr-weighted 78.7% -> 78.4% (10,299,493 / 13,141,652)
distinct-code 67.9% -> 67.1% (3,782,990 / 5,634,875)
A continuity line reports the ex-main figure so pre-2026-07-22 readings stay
comparable, and the binary-count label is corrected to "main + resident + 138
overlays" (it read "resident+139 overlays").
Metrics-only change; no build impact (ov_SC06_018 re-verified BYTE-IDENTICAL).
burndown.py averaged the last 3 INTER-COMMIT deltas. The phase's ROI criterion is
"per-SESSION yield floors out", and historically one snapshot was taken per session
— but I seeded THREE inside this session (9, 9b, 9c). Averaging them drove the mean
to +0.23 and printed "AT THE FLOOR — consider closing P29" for a session that
actually yielded +0.7pp instr (78.0 -> 78.7), comparable to recent sessions.
I nearly closed the phase on it. Same error class as everything else this session:
an instrument answering a DIFFERENT QUESTION than the one asked, where the wrong
answer is indistinguishable from the right one.
- --session-close marks a snapshot as a session boundary; the floor verdict is now
computed ONLY from those (older records predate the flag, so a label containing
"close" counts too). Mid-session snapshots still record for tracking.
- honest output now: "0 SESSION-to-SESSION delta(s) logged — need >=3 for a floor
verdict (1 session-close snapshot on record)".
=> P29 must NOT close on ROI grounds yet: the floor is UNDETERMINED and needs 3
session closes to become computable. The eyeballed "+2.5 -> +1.1 -> +0.6 -> +0.2"
trend is not the instrument's verdict either.
gate_stage._jtbl_prepare carried the SAME config-only undo as harvest_verify's did,
and ate the tree again on the first ladder run: 5 orphan region files, truncated
TUs, `undefined reference to func_80192F64`. That INVALIDATED the run's 0/10, so it
was re-measured rather than reported (R35 — a probe from a broken tool is not
evidence). Tree restored from HEAD and re-verified byte-identical first.
DELETED, not patched (R33 — the best outcome is a deleted stage). It was wrong on
two independent axes:
1. §61b already byte-proved THE CARVE MUST FOLLOW THE SPLICE. A batch pre-pass
carving unspliced functions reports "prepared" and yields a spec that fails
once the body lands — which is why it banked nothing.
2. Its undo snapshotted only config/, while jr_isolate_all rewrites region 0 back
over the ORIGINAL src/<ov>/<nm>.c truncated.
harvest_verify's per-draft prep is the correct mechanism, snapshots the full source
set, and undoes per function. Two implementations of one capability, the outer one
ineffective AND destructive.
THE HONEST RE-MEASUREMENT (clean tree; tree verified clean after):
- 0/10 bank, but 9/10 now COMPILE and land as whole-binary byte-DIFF; 1/10 plumbing.
- match_one close=0 on several (the function's own bytes exact) and rtu_match says
MATCH-in-real-TU for func_80135888 — while func_801299C8's transformed draft does
not compile in its real TU at all. The residual is MIXED, not uniform; at least one
is an IMAGE-level effect rather than the draft or its TU decl context (prime
suspect: jtbl/rodata carve placement). NOT generalized from one data point.
- This PRICES Task 14 stages 2-3 by measurement: the existing ladder converts 0 of
10, so they are not "wire in normalize_self_decls + the type-lift and collect ten
banks" — the projection error §57a already caught once this phase.
- R22 clean-fleet 140/140 BYTE-IDENTICAL with the giant func_8018F694 banked and the
func_80135A4C family swept 138/138
- cookbook §61d (the tree-eating undo in two tools; the constant-label defect; the
re-probe + ladder measurements; the general rule: an undo whose scope is narrower
than its write scope destroys work no byte-gate can see)
- decision-log + CURRENT_PHASE updated (R30/R31)
The session-7 checkpoint gated the entire jtbl track behind one finding: the
carve+isolation path yields a bank that is incrementally valid and clean-invalid
(139/140, [FAIL] ov_SC06_018, "twice, identically"). The prescribed diagnosis
(diff the incremental vs clean object set) never ran, because the failure does
not reproduce.
MEASURED, with the bank applied through the single-function automated path
(harvest_verify --chunk 1 -> [jtbl] carved -> + chunk(1) -> BYTE-IDENTICAL):
per-binary clean (rm asm+build; extract; build) -> BYTE-IDENTICAL cbbc4f44
make clean && extract-all && check-all (run 1) -> 140 passed, 0 failed of 140
make clean && extract-all && check-all (run 2) -> 140 passed, 0 failed of 140
ATTRIBUTION (best-supported; the failing tree is gone): the 139/140 runs were
taken on the tree left by the BATCH _jtbl_prep (6 table-bearing -> 1 carved,
4 isolate-FAILED, 1 stale-asm carve fail) — five failed preps' residue of
stranded carves + half-applied isolations. The per-function snapshot-restore
that removes exactly that residue landed AFTER those runs, in commit:0803, the
same commit that named the blocker.
THE LESSON (R35 on ourselves, -> decision-log): "twice, identically" was not a
replication — two reads of the SAME contaminated state is one observation. A
replication must RE-CREATE the state, not re-run the check. Standing guard:
re-apply a fault from a known-clean tree before writing it down as a property
of the mechanism. Sixth "structural wall" to resolve to our own tree/tooling.
- BANKED: func_80135A4C (181 ins) x1 in ov_SC06_018 — isolated into its own
code subseg + .rodata carve (single-table, no JTBL_PADS; tail3..tail18 renumber)
- §61c faults 1-2 STAND: a stranded carve poisons the overlay; per-function undo
is unsound in a batch -> ONE jtbl draft per harvest_verify invocation.
jr_inventory's 1:1 ownership assertion was right and is unchanged.
- UNFROZEN: this family = 138 members / PURE / 24,978 ins ~ +0.19pp (jtbl_family_bank,
§53 carve law); the 9 preserved t5wave cracks (Task 14 stages 2-3, §57 plumbing)
- R22 clean-fleet 140/140 x2; tools-health OK (dedup 1848/0, C1 234481/234481,
cdecl 53189/53189, audit-binaries 140); 0 NON_MATCHING (G4)
- fleet 78.0% instr / 66.5% distinct / 87.95% fn-count
- also: preserve the 4 untracked wave-4 .o0 drafts (R20); killed an orphaned cc1
from the Jul-21 session burning a full core for 13h23m
DIAGNOSED, not assumed. The 12-draft integration probe banked 1/12 and reported the SAME
label for 10 of the 11 failures: `conflicting types for built-in function 'memcpy'` — the
§58 red-herring (a WARNING, from an unrelated TU position). Splicing three top-reach
failures individually and reading real cc1 stderr gave the actual causes:
conflicting types for `func_XXXX' 3/3 <- loose-typing ARITY conflict
redefinition of `struct V8' <- a SECOND class (type-lift), stage 2
A banked shared caller macro in engine_core.h declares the function with FEWER params than
its byte-true definition takes (the original calls K&R-style with fewer args than the callee
reads); a C89 prototype makes that a hard error. tools/fix_arity_callers.py --any-proto
already fixes it and was simply NEVER WIRED into gate_stage's ladder (only family_sweep
carried §57). Now wired as a TU-side pre-pass.
MEASURED: 2 of 7 top integration candidates banked (func_8016EFC8, func_80164418, both
reach-138) vs the 1/12 old-ladder baseline. R22 140/140; tools-health OK (dedup 1848/0).
INCIDENT — this stage BROKE 138/140 AND R22 CAUGHT IT (nothing was ever committed):
pairing `--apply --any-proto` with `--revert` for the unbanked drafts corrupted declarations
fleet-wide. `--revert` rewrites ()->(void), which inverts a PLAIN apply but NOT --any-proto,
so an unbanked fn whose real decl was `extern void func_801708B0(void *a0)` came back as
`(void)` — in engine_core.h (included by all 138 overlays) and 6 sites in ov_SC01_077's own
sources. harvest_verify --binary ov_SC01_077 reported BYTE-IDENTICAL and was RIGHT about that
binary; the other 137 were structurally invisible to it. Repaired to the exact lines.
ROOT CAUSE FIXED: the ladder now snapshots every file the pre-pass touches and undoes by
RESTORE + re-apply-for-the-banked-set-only — exact by construction, cannot invent a signature.
NEW HARD CONSTRAINT (cookbook §61): any ladder stage mutating SHARED state must be undone by
snapshot restore, never an inverse transform, and validated FLEET-WIDE (R22) rather than by
the per-binary gate that authorised it. §55b's propagation law, one level down. The planned
type-lift stage edits engine_types.h and inherits it by default.
ALSO FIXED: the first wiring passed only --drafts (the narrow-param FILTER) without the
required --funcs, so the stage exited `no funcs given` as a SILENT NO-OP and the gate reported
0/6 as though diagnosed. sh() does not raise on non-zero exit -> explicit rc check added.
The hindsight-study §7 taxonomy predicts plateaus decompose into missing-transform (the
"highest-value bucket and the whole point"), seed-structural, and genuine-wall. Run against
real plateaus this class produced NO missing-transforms, and the answer needed no LLM.
MEASURED: `length` probe, 20 targets, 1 win. tail 1/6; partial 0/12.
AUTOPSY (read directly from the bytes, 3 partial plateaus):
- func_8017F0C0 / func_801806C8: target has `sltiu $v0,$v0,1` = gcc's codegen for `!x`/`x==0`;
the drafts wrote `(u32)(D_x ^ 1)` which emits `xori`. No local mutation crosses that.
- func_8017FF90: draft stores to arg0+8, target stores to a GLOBAL. Different function.
=> these are WRONG DRAFTS wearing a small closeness, i.e. seed-structural, not a mutation gap.
THE FIX IS THE OPPOSITE OF "ADD TRANSFORMS" — a tighter ADMISSION rule:
- _drift_route: permuter only when |d|<=2 AND explains=="tail" (the shape that measurably
converts). length pool 339 -> 34; permuter bucket 389 -> 84.
- SIZE-MISMATCH: added a PROPORTIONAL test (|d| >= 0.5*nt). max(2,0.15*nt) is far too
permissive on a tiny target — a 2-ins draft vs a 4-ins target read as a near-miss.
permuter_weights needs NO extension for this class.
Transferable (cookbook §60b): raising a search-closer's yield is at least as often about
refusing it unreachable work as widening its mutation set. Same knife as Task-13A's
targeting fix, one cut finer. Drafter idiom recorded: `sltiu rd,rs,1` => `!x`, never `x^1`.
17 unit tests green; corpus re-collected (1654 rows, closeness cross-check clean).
PROPAGATION (§55b, its own targeted batch): dedup_propagate --addr 0x80141B90 --recover
-> "138 overlays byte-identical after propagation"; 117 remaining stubs -> 0; 1 new
dedup group. This was the ONLY one of the 21 directed-run banks worth propagating.
THE REPRICING (R14 — measure a bucket's VALUE, not just its conversion rate):
the directed run converted 27% (21/77) but moved the fleet ~0.03pp, because h_exact
reach of the 21 is: func_80141B90=138, TEN at reach-1 (nothing to propagate), rest 2-10.
Instruction-weighted, the ENTIRE permuter bucket is worth ~0.36pp at 100% conversion.
The mechanism is validated; the fuel was small. Priced frontier (ins-weighted / 13.08M):
LENGTH-DRIFT |d|<=2 472,178 ~3.6pp (339 fns) <- the real permuter-adjacent lever
integration 419,162 ~3.2pp (305 fns) <- Task 14's ladder
WIDTH 71,593 ~0.55pp (45)
permuter (current) 46,571 ~0.36pp (74)
BRANCH-POLARITY 9,462 ~0.07pp (22)
So WIDTH/BRANCH-POLARITY are NOT worth prioritizing; my earlier "~200 candidates"
framing undersold LENGTH-DRIFT 10x and oversold WIDTH.
NEW: permuter_weights._LENGTH profile (perm_temp_for_expr/perm_expand_expr are the only
passes that change instruction COUNT; the reorder/decl-order levers that dominate the
regalloc+schedule profiles cannot, so they are down-weighted here) + residual_class
._drift_route (|d|<=2 -> permuter/`length`, larger stays structural — same class,
opposite tool) + classify() accepts a PROFILE NAME directly (the measured profile beats
re-parsing a free-text label). 17 unit tests green.
grinder: --profile filter (probe ONE residual class's conversion) + a PERSISTENT attempt
ledger. `tried` was in-process only, so every fresh --once run re-permuted the previous
run's losers — the permuter is deterministic given (base.c, target.o), so that CPU can
never produce a new win. Measured: a 20-target probe drew 19 already-tried targets.
Keyed by draft_sig so an improved draft legitimately re-opens the function.
First run of the Task-13A-targeted grinder (--once --batch 8 --permute-secs 90):
targeting: ON — 1665 classified; only bucket=permuter is admitted
targeting skipped 1601 non-permuter candidates
{redraft 707, structural 583, integration 305, unclassified 4, unknown 2}
permuter WON func_80181F78 @ ov_SC03_014 (close was 1) [~6 min]
BANKED 2 (both whole-binary byte-gated, R22 clean-fleet 140/140):
- func_80181F78 (8 ins) — classified DELAY-SLOT / schedule
- func_80141B90 (29 ins) — classified IMM-VALUE / cse
Both were classified bucket=permuter by residual_class BEFORE the run: the classifier
predicted a search-closer could reach them, and one did. First end-to-end validation of
the targeting thesis.
THREE latent defects, all pre-existing, all unreachable while the grinder banked nothing
since Phase 21 — the fix made it win and every one fired at once (cookbook §60a):
1. gate_stage commit path crashed on src=None. `src` is DELIBERATELY never defaulted
(Phase 26-A: a default silently PINS the gate to the main .c), but the commit did
`git add src …` unconditionally -> every caller that omits it (grinder, orchestrator,
idiom_hunt) crashes THE MOMENT IT BANKS. Now `git add -u src/`, which also retires the
`src/ov_*/*.c` filename glob that once omitted 4 R22-verified banks from a commit.
2. _xform ladder dirs (-cn/-cast/-rc/-uni) ACCUMULATE across runs: they held 34 stale
drafts, so the gate processed 34 when the grinder submitted 1, and banked one function
it was never asked to try — which would have been committed under a message naming a
different one. Nothing wrong entered the tree (G3/P9 held; the gate banks only
byte-identical output) but report and work had diverged. Now cleared per run.
3. grinder called gate_stage with the default propagate=True = `dedup_propagate
--auto-from`, the §55b fleet-wide path that timed out at 3600s and left 90/140 overlays
broken — and being INSIDE the gate it takes the banks down with it. Now propagate=False;
banks commit cheap, propagation is its own targeted --addr batch (§55b law).
R22: make clean && extract-all && check-all -> 140 passed, 0 failed of 140.
tools-health OK (dedup 1847/0, C1 234343/234343); 0 NON_MATCHING (G4).
The autopsy (hindsight-study §7) assumed the permuter loses for want of a mutation.
Measured over the whole open backlog, it loses because it is aimed at work a
search-closer provably cannot close.
- NEW tools/residual_class.py: decide a near-miss's class FROM THE BYTES. Decodes each
mismatching MIPS word -> (op-skeleton, register-fields, immediate); drift FIRST (one
inserted insn inflates `closeness` by the tail length), then consistent-injective
register map -> REGALLOC-PERM (§31 S11/RC-3), same-multiset-reorder -> SCHEDULE-REORDER,
DELAY-SLOT, WIDTH/BRANCH-POLARITY/STRENGTH/ADDRESSING/IMM-OFFSET/IMM-VALUE. Every class
routes to a BUCKET = which tool the failure wants. Uncovered opcode -> UNKNOWN, COUNTED
(R32). 16 synthetic unit tests (test_residual_class.py).
- NEW tools/autopsy.py: `collect` materialises the corpus Task-12's telemetry never filled
(1 of 6,169 records had a residual) by recompiling every open draft through the EXISTING
match_one path (R33) — 1,752 drafts in 21s at -j12. `report` -> docs/autopsy.md.
- NEW corpus.o0_sources()/is_o0(): the opt-level oracle DERIVED from the Makefile's own -O0
rules, coverage-asserted. Scoring an -O0 target at -O2 makes the residual 100% artefact
(the trap this phase hit four times).
- R34 cross-check baked in: residual_class's closeness vs masked_diff.structured_diff's,
asserted per row; 1,673/1,673 agree, 0 classifier errors.
FINDING: of the 972 records the grinder's own filter admits, only 75 (7.7%) are
permuter-shaped; 547 are structural and 348 are drafts that are not the function at all.
~92% of the daemon's CPU went where it could not win — the byte-grounded explanation of
"7 banks all-time, all Phase 21, 0 since" (Phase-22 audit). grinder.candidates() now
filters on the measured bucket (1,303 -> 78) and takes its directed profile from the
measured class, not the logged label (91% carry none -> it ran on gcc defaults).
Degrades to undirected if uncollected and says so; --no-targeting A/Bs it.
Two measured corollaries (R14, not projections):
- 699 records rank as near-misses at closeness up to 278 purely from a length artefact:
un-attempted work misfiled as a backlog of hard functions -> new `redraft` bucket.
- a 12-draft gate probe of the `integration` bucket banked 1/12 (11 PLUMBING), so the 306
prices Task 14's reconcile ladder rather than promising free banks. func_80167714
(104 ins, reach-134) banked x1, un-propagated by design (§55b).
Two defects fixed forward:
- masked_diff._common_typedefs() used ONE shared probe path, so parallel match_one
processes clobbered each other: 14 of 1,752 drafts lost in a single 12-way run (0.8%),
silently, in every parallel wave ever run. Now per-PID.
- gate_stage.match_one_closeness never passed --o0 -> phantom residuals for every -O0
function, written straight into the backlog this autopsy reads.
R22 clean-fleet: check-all 140 passed, 0 failed of 140; tools-health OK (dedup 1847/0,
C1 234343/234343); 0 NON_MATCHING (G4). Flywheel captured in-session (R30/R31):
cookbook §60, decision-log entry, SETUP.md inventory.
tools/burndown.py snapshots the 3 fleet metrics + remaining substantial frontier
each run (docs/burndown.jsonl, git-HEAD-keyed), computes velocity vs the previous
commit-snapshot, and flags the yield floor (mean instr-pp over last 3 snapshots <
threshold => 'consider closing P29'). Seeded with the session-4-close baseline.
Reads progress.fleet.md + family-hseq.md (regen those first for a true reading).
- func_8016CBC0's 137 members were NOT self-decl-blocked (§57 survey: 0 divergent self-decls) but blocked
on the exemplar's 6 LOCAL typedefs (Rec_/Blk_/SVEC_/CLR_/Poly_/Mtx8_8016CBC0). Lifted all 6 verbatim into
src/shared/engine_types.h (fleet-visible via engine_core.h) + removed the locals from the exemplar TU.
No fleet name-collision; exemplar rebuilds byte-identical d19c9580 (the type-lift is byte-neutral).
- family_sweep --hseq --only 0x8016CBC0 --no-preclassify --normalize-self-decls -> 137/137 banked, 0 failed
(the decl-normalize stage a correct no-op here — 0 self-decls). 0 func_8016CBC0 stubs remain fleet-wide.
- R22 clean-fleet 140/140 byte-identical; tools-health OK (dedup 1846/0, C1 234205/234205); 0 NON_MATCHING
linked (G4). fleet 74.9->75.1% instr / 60.1->60.6% distinct / 86.96->87.00% fn-count.
- R14 lesson: two families that both looked like a 'stuck 137-family' needed DIFFERENT levers
(decl-normalize §57 vs type-lift) — route by the real cc1 error, not the shape. session-total +270 members.
- targeted dedup_propagate --addr per core (NOT --auto-from), --recover for stragglers:
0x8014ADE0 -> 138 overlays byte-identical
0x801325B8 -> 134 (ov_SC07_011 byte-diverges -> auto-excluded, kept x1 — what --recover is for)
0x801387B8 -> 138 overlays byte-identical
= ~410 member-instances; 3 new dedup groups (1840 -> 1843), C1 coverage 233795/233795.
- 2 of the 5 banked cores (func_8014E284, func_80137DD4) stay ×1: "not self-contained (local types)"
-> blocked on the build_engine_types type-lift (the §19/§20 propagation cap). Carried.
- R22 clean-fleet 140/140 BYTE-IDENTICAL; audit-binaries OK; dedup 1843/0; 0 NON_MATCHING (G4).
Fleet instr 71.0 -> 71.4% / fn-count 86.30 -> 86.42% / distinct-code 53.3%.
- SELF-CORRECTION (R14/R35), now fixed in cookbook §55c + CURRENT_PHASE: my earlier claim that this
propagate "needs ~2h+" was WRONG. That timing was taken while the tree still carried the partial
damage of a killed --auto-from (90/140 overlays broken), so every member-gate was failing/retrying.
On a HEALTHY tree a targeted --addr propagate is ~233s/core (all 3 = ~27 min) — ~20x faster. Only
--auto-from is genuinely fleet-slow. A timing taken on a broken tree measures the breakage, not the
tool — recover the tree FIRST, then measure.
- cookbook §55: the wave's new byte-proven levers (§49-variant birthing-boost suppression via
reg_n_sets 1->2; sched1 birthing/LUID + "cc1 -dL" movable introspection; switch-tree vs jtbl
CASE_VALUES_THRESHOLD=5; block-scope-extern beats *(T*)&sym) + the GATE-ORCHESTRATION law
(--no-propagate per group then ONE targeted --addr; commit banks BEFORE propagating; a reverted src
needs a re-extract; gate_stage's default harvest_verified.txt accumulates -> phantom banks).
Acting on Task-1's verdict (the legacy-PURE "~3%" is tooling, not a wall): the -O2
type-lift arm. family_sweep --hseq --no-preclassify --band all --allow-pins templates
each matched exemplar's C onto its unbanked same-family members and whole-binary
byte-gates every one (the sole arbiter, G3/P9). The pin-crash wall being dissolved
(Phase-27/28 _carry_macros) let --allow-pins retry the pinned exemplars.
- BANKED 3,407 member-matches across 136 overlays (INCLUDE_ASM stubs -> matched C).
- R22 clean-fleet verify: 136/137 modified overlays byte-identical from a clean rebuild;
the 1 FAIL (ov_SC07_010) was the SIGTERM mid-gate partial -> reverted, byte-identical.
- make check-all: 140 passed / 140 BYTE-IDENTICAL; dedup-check 1840/0; audit-binaries OK;
audit-cdecl green; 0 NON_MATCHING in any default build (G4).
- Fleet: instr-weighted 68.9 -> 70.2% (+1.3pp) . distinct-code 49.5 -> 51.9% (+2.4pp) .
fn-count 83.94 -> 84.73% (+0.79pp).
- Process lesson (CURRENT_PHASE.md): --band all sweeps are too long for one background
pass (got SIGTERM'd); future Arm B runs go band-bounded + committed-per-batch (resumable).
Residual FAILED members (pin/drift/plumbing) + the -O0 Arm A carve remain.
- RE-SCOPED at plan time (R35): the roadmap's swing number rested on a broken-tool probe. B2's
"structural families bank ~0%" (0/8, which reshaped 2 phases of strategy) was a MISSING CARVE ->
8/8, then 102/115 (88.7%). The ~0% doctrine has NO surviving post-fix evidence.
- T0 img_path derive-not-guess -> un-hid a 230,612-ins SC07 pool doubly hidden (P27 onboarded 4
overlays, never regenerated the map; the hardcode would have called every member "LEN").
- T3-A: the SC07 pool is h_exact + UNWIRED, not h_seq. T4 dedup_extend (NEW) wired 6174/6457 (95.6%)
-> groups 134->138 binaries, C1 coverage +6174. NOT member_adapt (the number said build nothing).
- T5: resident 21->14 (90.34%, 7 banked via an Ultracode wave) + honest dossier for the 14; fixed
progress.py (#if 0 + len()-sum) and match_one's FAKE isolation (found by an agent mid-wave).
- T2 purged the poisoned grinder blacklist (8/22 matched anyway). T6 killed the --chunk 1 double-build.
- T7: make audit-binaries (the R36 citizenship gate, negative-control-proven) + fixed disc_code_sweep's
BLINDNESS to compressed code (the type-4 row was vacuous for 138 known binaries) + the worklist key-bug.
- T3b: the legacy h_seq swing number = ~3% AS-TOOLED, CLASSIFIED (274 DIFF / 37 PLUMBING), ceiling
UNKNOWN -- the 274 DIFF is byte-PURE members whose remapped bodies don't reproduce (the same
tooling-vs-wall ambiguity that resolved to TOOLING twice this phase). P29 disambiguates before scaling.
- THREE SELF-INFLICTED DEFECTS fixed forward: the registry yaml.safe_dump (H5, destroyed 47 comments +
1832 hex fields, invisible to every byte-gate), two DIFF mis-reports (R14), the match_one shared scratch.
- rule R36 (a newly-discovered binary is not real until every consumer knows it; enforced by
audit-binaries). cookbook §53. R22 140/140 throughout; tools-health OK; dedup 1840/0.
Ultracode wave: 16 isolated drafters over the resident's non-jtbl stubs (the 5 jtbl deferred —
they need the rodata-island carve, §53). Drafts only; the whole-binary byte-gate arbitrated after.
- BANKED 7/16, byte-gated: func_800CEFD0(77) func_800D0D7C(45) func_800D1B80(22) func_800D1E28(37)
func_800D1FC8(62) func_800D29F8(172) func_800D2D10(39).
Resident REAL 122 -> 129, stubs 21 -> 14, byte-ident 124/145 (85.52%) -> 131/145 (90.34%).
FLEET instr 9017152 -> 9017606 (+454 ins). R22 make clean && extract-all && check-all ->
140 passed, 0 failed of 140 (the first R22 was killed by a terminal crash and RE-RUN, not assumed).
Ground truth on 14 agrees 3 ways: source grep, splat-emitted stub .s count, progress.py.
- §52b's LAW, MEASURED AGAIN INDEPENDENTLY: the agents self-reported 11 match_one MATCH; the
whole-binary gate banked 7 (64%). All 4 blocked MATCHes died on `conflicting types`
(D_8010EDEC / D_80115110 / func_800D1984 / cdFileLocTable) — the loose-typing def-side wall, NOT
codegen. gate_stage's recovery banked 0/5 on them. A match_one MATCH is a CANDIDATE (G3/P9).
- FIX — match_one's isolation was FAKE, and its own docstring was the false spec. It promises
"Fully isolated (own temp dir) so many run in PARALLEL with no shared build -- a real asm-differ
loop for an agent to iterate against", while `--work` defaulted to the SHARED '.run/match': every
concurrent caller compiled into the same t.c/t.o. FOUND BY AN AGENT MID-WAVE, the only way it can
be found — it read another agent's function out of its own scratch ("found another agent's
func_800D2650 in my t.c") and reported it. Every other agent steered by a loop that could hand it
someone else's compile: a CONFIDENT WRONG verdict, worse than a crash. Default is now a private
.run/match/<fn>.<pid>; the default IS the promise. (Some agents had already worked around it by
passing --work themselves.) The byte-gate was never at risk — it is the sole arbiter — but the
iteration loop the agents steer by absolutely was.
- The 14 remaining: 5 PLUMBING (loose-typing) + 4 DIFF (genuine codegen: func_800D2650 close=4,
func_800CFAD0 close=5, func_800D0E30 close=12, func_800D27DC close=48) + 5 jtbl deferred.
Dossier next (T5b) — the agents' per-function residual analyses are the durable asset (R30).
R35: fix the instrument before planting a flag on its denominator.
- #if 0 BLINDNESS: classify() knew `#ifdef NON_MATCHING` (:425) but not `#if 0`, so a dead
analysis body was read as a live definition AND its real INCLUDE_ASM stub counted separately —
the SAME function in BOTH `real` and `stubs`. Live case: resident.c:868-925 wraps a full
void func_800D00E4(s32){...} in #if 0 (its jtbl dossier) and re-declares the stub at :926.
Now the block is skipped entirely: dead code is neither matched nor stubbed.
- THE len()-SUM (the dual defect): `placed` was a set union, so it caught a function in NO bucket
— but `matchable` SUMMED len()s, so a function in TWO buckets counted twice and nothing
complained. matchable/byteident are now SET unions, plus a new OVER-coverage assertion that
fails loudly if any fn lands in multiple buckets. R32 means both directions: nothing missing,
nothing double-counted.
- NEGATIVE CONTROL (the fix must change an answer the old tool gave):
resident REAL 123 -> 122 | matchable 146 -> 145 | 85.62% -> 85.52% | func_800D00E4 no longer
double-counted. FLEET instr 68.9% UNCHANGED (no #if 0 in the overlays) — the fix is scoped.
- FINDING (logged for T7's audit-binaries, does NOT block the flag-plant): the two INDEPENDENT
oracles now agree exactly at 144 — corpus (21 stubs + 123 matched, derived from the tree) and
sig_image (the 2nd oracle) — with EMPTY set difference both ways. progress.py still reports 145
because it counts func_800CEDFC and func_800D33E0, which are DEFINED in resident.c but absent
from sig_image. 0x800CEDFC is the resident's vram base +4 (the first function, code starts at
file offset 0x4 after the leading data word), yet make audit-corpus reports 0 PHANTOM +
0 TRUNCATED. Either sig_image has a boundary blind spot or those defs are not image functions.
progress.py's text-scanning classify() is exactly the re-parsing R33 says should be DERIVED from
corpus instead — a real refactor, logged not rushed.
The flag-plant claim is unaffected: it rests on corpus.stubs('resident') == 21 (tree-derived,
verified 5 ways), not on the contested denominator.
Completes T4 and corrects two defects I introduced, both landed in commit:0649.
- WIRED: 006 1543/1614 · 007 1544/1615 · 010 1544/1614 · 011 1543/1614 = 6174/6457 = 95.6%,
~0 agent tokens. Stubs/overlay ~2400 -> 831/984/898/825. Fleet instr 67.0 -> 68.9%,
fn-count 82.16 -> 83.94%. dedup-check 1840 validated / 0 failed; groups now read
"138 members [138 binaries]" (was 134); C1 coverage 227211 -> 233385 = exactly +6174.
R22 make clean && extract-all && check-all -> 140 passed, 0 failed of 140 at every stage.
- FIX#1 — I DESTROYED THE REGISTRY'S DOCUMENTATION, AND EVERY GATE CALLED IT GREEN (H5).
The first cut wrote config/dedup.us.yaml with yaml.safe_dump, round-tripping the whole file:
47 comment lines -> 0 (including the curated Phase-11 header explaining WHY the share is
source-level) and 1832 `vram: 0x80162FF4` -> `vram: 2148937716` (PyYAML parses YAML-1.1 hex to
int; dumps int as decimal). 25,948 lines rewritten. It passed dedup-check 1840/0 AND check-all
140/140 because _addr() accepts both forms: THE DATA WAS CORRECT AND THE DOCUMENT WAS RUINED.
Fixed forward (R6, no history rewrite): restored from commit:0649~1 and re-applied the 6174
memberships via a surgical text edit (add_members_surgical). Verified: 1545 insertions / 1545
deletions, 0 non-`binaries:` lines changed, 47 comments + 1908 hex fields intact, and the
rebuilt fleet is byte-identical to the destructive version (140/140).
THE LESSON: every oracle this project owns measures BYTES, so a formatting-destructive write is
invisible to all of them by construction. R34 says the byte-gate is a null COVERAGE oracle; this
is the same hole one layer out — it is a null DOCUMENT oracle too.
- FIX#2 — I MIS-REPORTED THE DIFFs, TWICE (R14).
(a) commit:0649 claims ov_SC07_006's 71 non-banks were "ALL PLUMBING, ZERO DIFF". FALSE — I read
head -6 of the classified file and generalized. It has the same 4 DIFFs as the others.
(b) I then built the jr guard assuming those 4 were the §53 jr class BECAUSE ov_SC01_077 hosts
them in _jr_8017A4AC.c / _jr_80182268.c. has_mid_jr is FALSE for all four (33-52 ins, no
jump table): they merely live in a carved jr-REGION split, which sweeps in every function in
its address range. HOSTING FILE != FUNCTION CLASS.
The guard is KEPT (preventive, §53-correct, currently skips 0 — no jr fn is in the extendable
set) with its docstring corrected to record what it is NOT. The 12 DIFFs (0.19%) are UNDIAGNOSED
and logged, correctly left as stubs by the gate — not dressed in a story.
- The 283 non-banks: 271 PLUMBING (the loose-typing conflict class + the whale, whose body lives
in src/shared/func_80144B9C.h so no DEFINE macro exists to expand) + 12 DIFF. Existing tools
cover the plumbing (cast_call_sites / canon_sig_reconcile / reconcile_tu).
The 4 SC07 overlays P27 onboarded were byte-clean but NOT citizens: their .c included only
common.h (never ../shared/engine_core.h), so no shared body could reach them, and they
appeared in ZERO dedup groups (1689 groups read "134 binaries", never 138). Each sat at ~80
matched / ~2400 stubs while its siblings were ~2150 matched.
- NEW tools/dedup_extend.py — the missing mode. dedup_propagate is built for CRACK -> AUTHOR
MACRO -> INSTANTIATE: --auto-from scans INLINE DEFS (planned only 11 here; the ~1600 shared
bodies are ALREADY DEFINE_func_* macros in engine_core.h) and --addr dies "no source overlay
has it matched" because no overlay holds an inline def. Extending an existing MACRO-BACKED
group to a newly-onboarded binary is a different operation and nothing implemented it.
- SAFETY (explicit — this feeds the byte-gate): h_exact is the SHA1 of RAW INSTRUCTION BYTES, so
two instances sharing one are identical INCLUDING their jal/lui/%lo reloc immediates — same
callees, same data addresses, same symbols. The body that compiles byte-identically at one
member does so at the other with NO remap. (Exactly why dup_report calls h_exact "guaranteed
byte-match" and h_norm "candidate-only".) A bug here can only FAIL TO BANK, never falsely bank.
- REUSE, DON'T REBUILD (R33): owns only the set computation + the registry edit. The splice and
the gate are harvest_verify verbatim (it already derives each stub's home TU from the corpus
oracle, chunks + bisects, reverts on failure). h_exact members are byte-identical by
construction -> the happy path is ~1 build per binary, not one per function.
- RESULT ov_SC07_006: 1543 / 1614 banked = 95.6%, ~0 agent tokens. Stubs 2374 -> 831.
The 71 non-banks are ALL PLUMBING, ZERO DIFF, in two named classes with existing tools:
* func_80144B9C "undefined reference" — the whale's body lives in src/shared/func_80144B9C.h
(the -O0 shared header), not engine_core.h, so no DEFINE macro exists to expand.
* "conflicting types for D_800A5E60 / func_8012C750 / func_8012C0EC" — the loose-typing
conflict class (cast_call_sites / canon_sig_reconcile / reconcile_tu already exist for it).
- GATES: R22 make clean && extract-all && check-all -> 140 passed, 0 failed of 140, 0 FAIL lines.
dedup-check 1840 validated / 0 failed; groups now read "135 members [135 binaries]" (was 134);
C1 coverage 227211 -> 228754 = exactly +1543. The second oracle accepts the extension.
- Mechanism had been proven by hand first (probe-before-investing): +include + ONE stub ->
DEFINE_func_80128158() -> ov_SC07_006 built 7ca772be BYTE-IDENTICAL, then reverted.
The family the roadmap recorded as 0/8 ("~0%, structural families do not template" — the
number that rewrote P29's arithmetic to "(cores cracked) x (reach)") banks at 88.7% when
swept with the carve its own exemplar required. ~0 agent tokens.
- SWEEP: jtbl_family_bank.py over the remaining 107 members ->
{'BANKED': 94, 'gate-fail': 7, 'remap-refuse': 6}. Family total 8 (T1) + 94 = 102/115.
R22: make clean && extract-all && check-all -> 140 passed, 0 failed of 140, 0 FAIL lines.
- FLEET (measured, make report): instr-weighted 67.0 -> 67.7% (+0.7pp, +97,104 ins);
distinct-code 47.8 -> 49.4% (+1.6pp); fn-count 82.16 -> 82.19%. 102 x 952 = 97,104 =
the exact measured instruction delta — the arithmetic reconciles to the byte.
- THE 13-MEMBER TAIL is the predicted shape, and both halves are data for T3:
* 6 remap-refuse = EXACTLY the family's 6 IMM members (cls_counts PURE 109 / IMM 6).
imm_map_tier1 REFUSED rather than guessed: "unresolved immediates: [(512,
'asm-ambiguous')]" — 512 also occurs at a non-differing position, so a blind swap could
corrupt it. This is the concrete shape of T3's IMM stratum.
* 7 gate-fail = genuine byte-DIFFs, correctly rejected. Verified to leave NO residue
(all 7: split_file=none, cfg_refs=0) — no false-bank risk.
- HYGIENE: the 7 "git checkout ... did not match any file" errors are benign (revert of a
never-tracked path). Verified 0 untracked splits belong to a non-banked member; 91 new
splits + 3 banked into existing splits = 94.
- SCOPE (P9, unchanged): still n=1 family, and jr is the rarest class (3/163 matched-exemplar
families). This demonstrates the mechanism at family scale; it does NOT give a rate for the
PURE/IMM mass (98% of the population). T3 measures the swing number.
Drew T