Final S50 state: 307 instances banked, stubs 12,468 -> 12,161, fleet 95.3% instr / 90.0%
distinct / 96.65% fn-count. R22 clean rebuild 4x, check-all 213/213 every time.
- tools/aprop_autodraft.py + tools/draft_prechecks.py: seed body + symbol_map + a MINIMAL
synthesized preamble. The seed's decl layer never travels — that layer is family_sweep's
dominant failure (331 of 458 S49 verdicts). 256 banked at zero agent tokens, against the
~20M the same work would have cost as a wave.
- Macro seeds (567 of 1196 members, all 3737 de-macroize) take the DEFINITION only; the block
stays the decl source. Pasting it whole measured 28% vs inline's 68% — func_8016AB6C's macro
is 1,891 lines of which 108 are the function.
- IMM is a second engine, not a wall: T2a's imm_map_tier1 resolves a per-location LITERAL like
symbol_map resolves a per-location SYMBOL. 131 of 275 IMM members resolve.
- draft_prechecks negative-controlled against ALL 205 banked drafts: zero false positives,
catches 39 of 67 known failures. That control found two bugs in the checks themselves —
C89 `f()` declares UNSPECIFIED parameters (not zero), and a member's own definition read as
a call to itself. Conservative by design: a pre-check that discards good drafts is worse
than one that lets a few builds fail.
- The A-prop pool is now priced exactly: PURE 437/37,376 ins, IMM 275/8,849, STRUCT 238/4,259.
- Cookbook §171a; SETUP rows; CURRENT_PHASE S50 FINAL checkpoint.
- The blocker was carried as "one missing file-scope extern gates 83 PURE members". Both
halves were wrong (R14): corpus.stubs says 4 open members, and D_801ED98C is a DEFINED
const Blk8 whose rodata lives inside the member's own nonmatchings .s — replacing the stub
deletes the data with it. gather_externs can carry an extern DECL, never a DEFINITION,
which is why it reported "no file-scope decl" for a symbol md_SC05_023 defines on line 114.
- Fix: paste typedef + const definition + body per sibling (data bytes verified identical
across md_SC05_024/025/028/029). 4/4 banked.
- aprop_symfix: new `local-only` class — draft-DEFINED identifiers that merely carry a
vram-looking suffix (Blk8_…, S8_…, L_call_…) are not stale symbols. Measured: that is every
non-clean case in the whole wave-7a/7b stored-draft residue, which holds ZERO stale-symbol
recoveries (a clean negative result — the defect was A-prop-specific).
- cookbook index regenerated (tools-health fails closed on a stale index — it caught §171).
- R22 clean rebuild: check-all 213 passed, 0 failed of 213. Stubs 12,445 -> 12,441.
- REFUTES §170's open hypothesis (batched cards concentrate members into one TU ⇒ §169
collision): 5-draft groups banked 5/5; 11 of 35 unbanked drafts were already one-per-TU;
and the two "concentrated" groups banked 12/12 and 10/10 once the real defect was fixed.
- The cause: a per-location data symbol carried out of the seed body unrebased. match_one
compares instruction ENCODINGS and is blind to a relocation's target NAME, so it scores
MATCH standalone and dies at link in the host TU. 24 of 24 concentrated failures, all 1:1
rewritable at one constant vram delta (0x4128).
- tools/aprop_symfix.py: audit + --fix, emits a gate_lane-shaped slate; deterministic and
build-free, so it runs BEFORE the gate. The R34 second oracle for the class match_one
cannot see.
- family_cousins.py --aprop-cards: members now carry sym_map, the explicit {seed -> member}
renames, read from the seed's C BODY (a matched seed has no .s of its own) vs the member's
.s. Two case-mismatch defects fixed while wiring it (sig lowercase vs splat uppercase).
- 23/24 banked. Stubs 12,468 -> 12,445. Fleet 95.2% instr / 89.9% distinct / 96.57% fn.
R22 clean rebuild: check-all 213 passed, 0 failed of 213. dedup 2,043/0.
- A-prop's true conversion is 87% (79/91); the 320 batched members are unblocked.
- Cookbook §171 + §170 struck in place; SETUP row; decision-log (R31).
- NEW family_cousins.py --aprop-cards + tools/wave/aprop_wave.js: lane A (1,700 open fns /
76,419 ins) had NO card type — cousin diffs are empty for h_seq-identical members, so the card
is a positional WORD diff vs the matched sibling, grouped BY FAMILY (one agent, N drafts).
Head cards: 13 families / 433 members, median TWO differing words each.
- calibration 9 batches / 108 members: 98 agent-MATCH (91%, best of any wave) -> 56 BANKED (57%),
~80k tok/banked fn vs 157k (cousin card) vs 400k+ (crack wave). R22 213/213 BYTE-IDENTICAL.
- HONEST GAP (R14): 91% agent -> 57% gate is the worst conversion measured; 14 groups banked 0.
Hypothesis TESTABLE not proven — family batching concentrates members per destination TU, the
§169 collision. Re-gate unbanked ONE PER TU before scaling the remaining 320.
- >=16 head diagnosed: 3 of 4 blockers are plumbing — the --band substantial default hid 5 of 13
families from every prior sweep; one missing file-scope extern (D_801ED98C) gates 56 PURE
members; dedup_extend is macro-only. Only func_8017C294 is a genuine crack.
- fleet 96.56% fn / 95.2% instr / 89.9% distinct; stubs 12,535 -> 12,468; dedup 2,043/0.
- cookbook §170.
- thresholds relaxed to <=6 blocks/<=16 tokens UNION edit-fraction <=0.20: cards 518 -> 721,
MIXED 310 -> 50 skeletons; the 753-ins func_8017BEBC (0.987 sim) became reachable.
- 59 cards -> 48 agent-MATCH (81%) -> 44 BANKED (92% MATCH->bank, 75% end-to-end), 6.9M tok.
- FINDING (the actionable one): 7b's bank rate crushed 7a's because it SPREAD 48 drafts over 35
destination TUs; 7a's failures were per-TU declaration collisions between sibling drafts.
Cookbook §169 updated with the spread law.
- R22 213/213 BYTE-IDENTICAL from clean; fleet 96.55% fn / 95.2% instr / 89.9% distinct;
stubs 12,584 -> 12,535; dedup 2,035/0.
- incidents 3 & 4 recorded: an agent wrote a TRACKED header (guard caught it, prose is not
enforcement); my own gate_lane filtered on the wrong key and printed 'gating 0 drafts' as a
result (R32 silent skip) — fixed with a coverage assertion that refuses to report 0.
- pilot 30 cards -> 25 agent-MATCH (0 refuted) -> 16 banked; 29 instances banked tonight
(89 incl. propagation); 2.7M tokens haiku-tier ~= 30k/banked instance vs a crack wave's ~75k.
- R22 213/213 BYTE-IDENTICAL from clean; fleet 96.53% fn / 95.1% instr / 89.8% distinct;
stubs 12,613 -> 12,584; dedup 2,029/0.
- R14 CORRECTION: a banked cousin usually does NOT propagate (2 of 8; cousins are byte-variant).
The card 'reach' column is cousin fuel, not dedup copies — priced wrong in my earlier framing.
- FINDING: the 9 gate failures are per-TU INTEGRATION (standalone-MATCH, host-TU-rejected),
clustered 5+2 in two binaries — the reconcile-ladder class, not codegen.
- TWO INCIDENTS (mine): an outer timeout tighter than gate_stage's own scaled timeout killed a
healthy 5-bank group mid-write AND orphaned its dedup_propagate child, which kept rewriting
src/ through a git checkout. Killed, inspected, reverted; the same 5 drafts banked 5/5 untimed.
Law: never wrap a self-timing tool in a tighter cap; kill process GROUPS, not pids.
- cookbook §169 (the lane + the three laws + the threshold sizing table).
Wave 6 added 110 (24 cracks + 23/24 families propagated). Fleet 95.1% instr /
89.7% distinct / 96.51% fn-count. R22 clean-fleet run 9x this session, 213/213
every time. Bank rate across six waves: 67/79/69/68/73/60%.
Records §166a (the destination-TU oracle) and the four-instance pattern it
completes: a tool asserting a conclusion it never reached. A confident wrong
label costs more than a missing one.
Wave 5 added 152 (29 cracks + 28/29 families propagated) — the session's
largest. Fleet 95.0% instr / 89.6% distinct / 96.48% fn-count; stubs
13,345 -> 12,771. R22 clean-fleet run 8x this session, 213/213 every time.
P30's milestone is '>=95% instr fleet, or every remaining overlay stub on a
named ledger'. THE FIRST HALF IS NOW MET — T5 (phase close) is a live option.
Bank rate across five waves: 67% -> 79% -> 69% -> 68% -> 73%.
Wave 4 added 105 (27 cracks + 27/27 families propagated). Fleet 94.9% instr /
89.4% distinct / 96.44% fn-count. R22 clean-fleet run 7x this session, 213/213
every time.
Bank rate now measured four times: 67% -> 79% -> 69% -> 68%. Prior-notes
seeding 10/12 (was 7/9). func_8017C294 — the x16 family, largest item on the
board — is NEAR at 2 ins after three seeded attempts (18 -> 11 -> 2).
Also records the 4th comment-blindness defect and its blast radius (one draft
comment refused a binary's stub oracle, failed 5 later binaries, and left
drafts spliced in src/ so 17 re-gates read a poisoned tree as 0/17), and that
the wave harness now lives in tools/wave/ with its contracts written down.
Wave 3 added 84 (27 cracks + 21 propagated families). Fleet 94.8% instr /
89.2% distinct / 96.41% fn-count. R22 clean-fleet run 6x this session, 213/213
every time. 96 commits.
Bank rate measured three times: 67% -> 79% -> 69%. The dip is the cost curve
(wave 3's tier was 29 Opus-band / 14 jr vs wave 2's 8 / 5, median reach x6 ->
x3-4), not a regression.
Two levers proved out and belong in every future wave: the hardened harness
contract (0 drafts lost vs 21) and prior-notes seeding (7 of 9 previously
failed targets converted, incl. both long-standing NEARs and all three wave-2
gate misses). NEAR is a resumable state, not a write-off.
Session close state. Three parts: stage 0b (91, zero decompilation), wave 1
(26), wave 2 (116). Fleet 94.4% -> 94.7% instr, 88.3% -> 88.9% distinct,
13,345 -> 13,112 stubs. R22 clean-fleet run 5x, 213/213 every time.
The campaign now has a MEASURED rate, twice: 67% (wave 1, all-Opus) then 79%
(wave 2, 20 of 28 Sonnet) of cracks survive the whole-binary gate. The Sonnet
band beating the all-Opus wave is the session's most useful economic finding
and sets wave 3's routing.
Resume order changed on evidence, twice over:
- harden the wave harness FIRST (per-agent dirs, sha1-last verifier, and a
tools/recover_drafts.py built from the transcript-replay method that
recovered 21/21 today);
- then wave 3, sized on 79%, not on the reach-15 prior.
Error ledger grew to 6. The two that matter: I wrote off 21 verified cracks as
lost when the run transcripts held every one of them, and my first two
recovery passes both failed by reading a single tool record instead of
replaying the file's mutation history.
Stage 0b closed (91, zero decompilation) + Stage-1 wave 1 (8 cracks -> 26
instances). Fleet 94.6% instr / 88.8% distinct / 96.36% fn-count.
Resume order changed on measured evidence: FIX THE md_ MODULE LANE FIRST.
16 of the wave's 42 member slots were unreachable for tooling reasons, not
matching reasons — 12 on a carve that assumes raw data lives in
<binary>/data/*.data.s (modules do not), 4 on an uncarried extern
(`D_8011511A' undeclared). Both are named with verbatim errors; probe one of
each before pricing (R37). Precedent: 0b's three repairs banked 91 for ~0
agent tokens; the wave spent 3.36M for 26.
Also recorded: the frontier re-derivation (1,955 zero-crack families /
330,622 templatable ins), the tier-ordering correction (ins-per-crack is flat
across x5-x8, so rank by templatable weight, not by tier), and the §162
harvest with its two in-place cookbook corrections.
Stage 0a's first defect, and the largest single zero-token bank of the session.
family_remap's kind test asked ONE question — is this address a function in the SIBLING'S OWN sig?
— and defaulted to `D_` on "no". But a body calls outside its image constantly: an overlay calls
resident helpers, an md_* module calls the overlay-range engine. Those addresses are absent from
the sibling's sig, so the test fell through and emitted a DATA NAME FOR A FUNCTION —
`D_800183E0`, `D_800D1EBC`, `D_80171A1C`. None exist anywhere in src/ or config/symbols.us.txt,
while `func_80171A1C` alone has 1,061 references. Measured: 611 member-rows across 45 symbols,
the largest named residue class. "Not in MY sig" means "not mine", not "is data".
Fix — three oracles, strongest first, never a blanket fallback:
1. the sibling's own sig (authoritative for its image; this is what preserves the Phase-29 T82
case where a slot is a function in the exemplar and DATA in the member — unioning every sig
would have re-broken the 251 members T82 fixed),
2. the always-linked images via extern_fn_addrs() — resident + main, 2,146 addresses whose ranges
cannot collide with an overlay's,
3. the exemplar reached it by `jal` — a call target is a function BY DEFINITION, which covers an
external address neither sig claims (0x80171A1C from an md_* module, 112 rows).
Only a non-call reloc no oracle claims still falls to `D_`.
Result: BANKED 205 member-matches, failures 670 -> 575, derived net = report = 205.
R22 clean-fleet 213 passed / 0 failed of 213.
Fleet 94.4% instr / 88.3% distinct / 96.27 -> 96.33% fn-count; stubs 13,563 -> 13,345.
THE RESIDUE HAS CROSSED OVER: DIFF is now the LARGEST class at 143 of 575 — real byte divergence
outranks plumbing for the first time this session (undefined-ref 611 -> ~8, PLUMBING-other
231 -> 81). The ~5:1 plumbing:DIFF ratio that justified "tooling beats volume" has inverted in this
queue, exactly as the frontier analysis predicted: the declaration-axis vein was one-time.
The new head class is `conflicting types for func_80175414` (27) — the same addresses this fix
started naming correctly, now surfacing the NEXT layer (the symbol resolves; its declared signature
disagrees). That is the conform axis, not the remap axis.
Note for anyone auditing this class: rtu_match MASKS HI16/LO16, so a wrong %hi/%lo symbol still
reports MATCH (the T82 comment records `MATCH (10 ins)` on a member the fleet gate refused). This
defect is invisible to the per-function tool by construction — only the whole-binary gate sees it.
W1b — the 3 targets whose agents died on API rate limiting, retried with cookbook §160 in the
prompt: func_801EFBF4 (reach 12), func_801EFDC8 (12), func_8018CC40 (10, jr). 3/3 confirmed by an
independent verifier, all banked, R22 clean-fleet 213 passed / 0 failed of 213.
func_8018CC40 failed the first gate with `too many arguments to function func_80178970` — which its
own crack agent had PREDICTED in its report, naming the §17a-1 remedy. Dropped the draft's
empty-paren externs and cast 6 call sites instead; banked. Read the agent's integration notes
before diagnosing a gate failure — it has already seen the TU.
Cookbook §161a-c (index 469 sections):
§161a case 0: break; is LOAD-BEARING when a jump table is indexed from zero. The natural
case 1..5 makes gcc-2.7.2 pick minval=1, emit `addiu $v1,-1`, and shift every table index —
58 of 77 mismatched on a byte-perfect body. Tell: the table's FIRST entry points at the
function's own end address. Family-wide (10 members).
§161b aliasing a parameter into a local can force a SECOND callee-saved register (+8 frame,
+3 ins) even when uses are mutually exclusive. Suspect it before reaching for register pins.
§161c loose-prototype engine helpers: don't fight the TU's (void) decl, cast at the call site.
G2 — THE MAIN EXPERIMENT. family_hseq excludes main as "structurally barren — zero h_exact
overlap". True and irrelevant: an h_exact claim guarding an h_seq tool. There is not even a
sig-main target — main had never been signed for this pipeline. Signed it (2,002 fns, seeded from
splat boundaries via corpus.stubs rather than --bootstrap, which glues functions around jtbl
dispatch and would have corrupted the hashes under test).
Result: main is ~85% singleton work, not 100%.
internal h_seq families (>=2): 207 families / 748 fns / 11,537 ins (13.7%)
shapes shared with the fleet: 161 fns / 1,346 ins (1.6%)
genuine x1 remainder: ~71,034 ins (84.6%)
IMMEDIATELY ACTIONABLE: 44 classes / 151 main functions / 1,239 ins already have a matched exemplar
in the fleet — free propagation, invisible only because main is not in the map.
Long-term: 748 of main's 2,002 functions (37%) are templatable once one exemplar per family is
cracked, which refutes "2,002 independent cracks" as the planning assumption for the 79k-ins tail.
OPEN, deliberately not done unilaterally: adding a sig-main target and dropping main's exclusion
from family_hseq.load() changes a fleet-shared oracle every targeting tool reads. Needs Drew's call.
func_801EDC18 (md_SC05_023) is the largest multiplier remaining — 57 members. The wave agent
abandoned it at "closeness 6" with class SIZE-MISMATCH [redraft]. It was two lines from correct.
THE CODE (cookbook §160a): the target copies 8 bytes with lwl/lwr + swl/swr — gcc-2.7.2's
emit_block_move for a type with ALIGNMENT 1. The draft used a u32 copy (aligned lw/sw), which is
wrong by construction. `typedef struct { char c[8]; } Blk8; buffer = D_801ED98C;` reproduces it.
Six spellings were tried in parallel; two independent agents converged on the same one.
THE INSTRUMENT (§160b) — this is the part worth more than the function. The target .s bundles a
leading `.section .rodata` block (D_801ED98C as two .word) ahead of .text. Those lines carry the
same `/* off vaddr HEX */` shape as instructions, so masked_diff.insns_from_s counted them as TARGET
instructions, while insns_from_object (objdump -j .text) can never emit them. A byte-perfect draft
therefore read `mine=26, target=28, 26 mismatched` — every position shifted by a constant +2 — and
got classified as needing a redraft. 116 of 12,583 .s files in the corpus have this shape, one at
-29 instructions. Every one of them would report a false wall to any agent that tried it.
Fixed: insns_from_s tracks .section and counts only .text. Full-corpus control: 12,467 unchanged,
116 corrected, 0 regressions. Same artifact class as §129a (post-carve jtbl inflation).
THE OWNERSHIP LAW (§160c) — my own error, corrected by the gate. Four sites declare
`extern short D_801ED98C;` and nothing in src/ defines it, so I shipped an extern-only draft. The
gate refuted it: `undefined reference`. The .s block the draft REPLACED was the definition. The
variant emitting `const Blk8 D_801ED98C = {{...}}` banks clean. Never infer ownership from externs.
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
ALSO BANKED — the wave's idiom harvest, which had been sitting unwritten in workflow transcripts
(R16/R30 debt): §160d the ASYMMETRIC INDEX RELOAD (a just-stored narrow field read twice emits
reuse-then-reload; the C is deliberately asymmetric), §160e a stack-layout scheduling rule now
byte-proven on a SECOND independent function (promoting it from coincidence to rule), §160f the
address-only global store via array decl, §160g sibling-search keyed on the CALLEE SET as step 0 of
every wave prompt (one grep turned a 126-instruction crack into a copy-edit).
Cookbook index regenerated: 468 sections.
The 10 exemplars from W1 flipped modal -> matched in the regenerated map, so family_sweep could
template them. 9 non-jr families swept: BANKED 140 member-matches / 32 failed across 50 overlays.
Derived net = report = 140 (no untracked carve files this time, so the two counts agree).
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
Fleet 94.3 -> 94.4% instr / 88.2 -> 88.3% distinct / 96.22 -> 96.27% fn-count; stubs 13,713 -> 13,563.
WAVE ONE, FULLY ACCOUNTED: 10 agent cracks + 140 templated members = 150 functions for 1.36M
tokens (~9k tokens/function). Still owed from this wave: 73 member-slots in 2 NEAR families,
34 in 3 rate-limited targets, 9 in the jr family (routes to jtbl_family_bank, §53).
TWO MEASUREMENTS THAT CORRECT MY OWN FORECASTS (R14):
1. Conversion was 81%, not the 58% I projected from this morning's propagation run. Today's
plumbing fixes (alias-drop, cpp-derived TU type map, group-level draft-vs-draft aliasing) are
paying off in a population they were not tuned for.
2. The effective multiplier was 15x, not the 2-3.5x I predicted. That estimate used the MEAN
family size across the whole zero-crack pool (3.55); this wave deliberately targeted the TOP of
the reach distribution, where families run 10-28 members. Ordering waves by reach is what
produced the difference — the mean was the wrong statistic for a wave that selects on the tail.
The regen step is load-bearing and now byte-proven twice: a fresh crack reads as `modal` until sigs
+ family_hseq are rebuilt, and family_sweep templates only from `matched`. Skipping it sweeps a
stale map and the multiplier evaporates (the Phase-26 finding, whose surviving qualifier is that
remap works BEHIND a fresh crack).
The one-line fix committed ahead of this run (CdFileLoc_80128C98 aliasing CdFileLoc) cleared the
largest remaining propagation-sweep class. Re-sweep: 138 member-matches banked, failures 875 -> 737,
`conflicting types for cdFileLocTable` gone entirely (136 -> 0).
Derived net (138 INCLUDE_ASM removed, 0 re-added) equals the report's 138 — they agree.
R22 clean-fleet: check-all 213 passed / 0 failed of 213.
Fleet 94.2 -> 94.3% instr / 87.9 -> 88.1% distinct / 96.15 -> 96.21% fn-count; stubs 13,780.
Residue reclassified — no symbol dominates any more: 227 PLUMBING-other, 125 DIFF (real byte
divergence, 17%), 93 CC1-FAIL(no-diagnostic), 26 memcpy, then a tail of small data-symbol
conflicts (D_80114F24 12, D_800AE620 11, D_800183E0 9, D_80126B58 6, D_80078EB4 6).
CC1-FAIL rose 77 -> 93 and that is NOT a regression: members that previously died earlier on the
cdFileLocTable conflict now reach a different compile error. Those 93 are hard gcc errors whose
text the sweep's classifier discards because it greps for `error:`, which gcc-2.7.2 never emits on
hard errors. That classifier is now the highest-value instrument fix left — three times today a
no-diagnostic verdict concealed something cheap.
scope_data_externs §8d drops the draft's decl of any symbol the TU already declares at file scope.
It keys on the SYMBOL, but a §37 asm-label ALIAS binds a DIFFERENT C identifier to that symbol:
the TU declares `D_801851BC`, it does NOT declare `tbl_D_80187044`. Dropping the alias left the
body referencing an undeclared name, which cc1 reports with no `error:` prefix — so the sweep
classified all 132 siblings as CC1-FAIL(no-diagnostic), i.e. as a codegen wall.
The bitter part: the alias exists PRECISELY BECAUSE the TU declares that symbol with a conflicting
type (a `void (*[])(void)` dispatch table vs this function's 20-byte-stride view). The drop rule
fired on exactly the declarations written to survive it. Why 1 of 2 died was fully determined:
tbl_D_80187048's symbol is not in the TU, so it demoted normally.
Fix: is_asm_alias() — an alias is demoted into the body, never dropped (the identifiers differ, so
it cannot collide with the TU's decl). Control-tested 6 ways incl. self-labels and plain externs.
Measured: func_80132018 3/135 -> 135/135; full re-sweep +16 more. Total +148 members.
R22 clean-fleet 213 passed / 0 failed of 213. tools-health OK, dedup-check 1949/0.
Fleet 96.11 -> 96.15% fn-count, 87.8 -> 87.9% distinct; stubs 14,120 -> 13,972 = -148 (2nd oracle).
CORRECTION TO MY OWN CLAIM (R14): after the probe I said the 58% aggregate was concealing a broad
problem. The re-sweep refuted it — only 16 more banks fleet-wide. The alias class really was one
family; the first read ("outlier") was right and the correction was wrong.
875 sweep failures classified: 231 PLUMBING-other, 141 DIFF (real divergence, only 16%),
136 `conflicting types for cdFileLocTable` (ONE symbol — biggest single class left),
77 CC1-FAIL(no-diagnostic), 26 memcpy, 12 D_80114F24, 11 D_800AE620, 9 D_800183E0.
STILL UNFIXED, and the most dangerous instrument left: the sweep's failure classifier greps for
`error:`, which gcc-2.7.2 never emits on hard errors. Every hard error therefore reads
CC1-FAIL(no-diagnostic). That is how a missing declaration looked like a codegen wall across 132
functions. rtu_match was fixed for this at T0(b); this classifier was not.
family_sweep --hseq --band all -j 8 over every matched-exemplar family: 553 families /
203 overlays / 1,419 banked / 1,023 failed (58%). R22 clean-fleet 213 passed / 0 failed of 213.
tools-health OK, dedup-check 1949 validated / 0 failed.
Fleet: 93.9 -> 94.2% instr / 87.2 -> 87.8% distinct / 95.72 -> 96.11% fn-count.
Second oracle (R34): INCLUDE_ASM stubs 15,542 -> 14,120 = -1,422, equal to the diff-derived net
(1,451 removed - 29 re-added = 1,422 = 1,419 sweep + 3 probe). Three independent counts agree.
B -> C -> P IS ONE CHAIN, NOT THREE WINS. 1,102 of the 1,422 landed in ov_SC02_037 (409),
ov_SC03_107 (364), ov_MAIN_012 (329) — the three newly-onboarded binaries from C, which had never
been wired into the shared-body ecosystem, so every matched exemplar was unreachable from them.
B fixed the declarations, C wired the include, P poured through the opening. A repeat sweep will
NOT pay like this; the opening was one-time.
S47 total: 1,481 functions banked with zero agent drafting, all from removing plumbing.
Two findings recorded, neither fixed (deliberate, costed):
- --band defaults to `substantial`: the first probe returned a confident {"families": 0,
"banked": 0} on a real 135-member `mid` family. Always pass --band all.
- The alias-gather defect: probe on 0x80132018 banked 3/135, all 132 failures classified
CC1-FAIL(no-diagnostic) because gcc-2.7.2 emits no `error:` prefix. Real error is
`tbl_D_80187044' undeclared` — the exemplar declares TWO §37 asm-label aliases and uses both,
family_remap carried one. T7-S1's "gather" class. Measured as an OUTLIER (aggregate 58%),
which is why the sweep ran before the fix.
Refused by design, all named: 50 jr families / 183 member-slots (§53 interlock — it printed its
own coverage and reason), 264 STRUCT, 112 unresolved immediates, 3 not-stub.
Regenerated after the S47-B/C banks (family_hseq.py + report): docs/family-hseq.md,
docs/progress.fleet.md, docs/backlog.md. Numbers only — no analysis change.
Frontier at this HEAD (overlays only): 7,085 families / 14,508 instances / 752,073 ins.
with siblings (>=2): 2,429 fams / 9,852 members / 467,634 ins (62.2%)
- matched exemplar (propagate, ~0 tok): 460 fams / 2,861 members / 125,630 ins
- zero-crack (crack 1 -> templates to N): 1,969 fams / 6,991 members / 342,004 ins
singletons: 4,656 fams / 4,656 members / 284,439 ins (37.8%)
- matched exemplar: 143 / 6,603 ins - zero-crack (pays x1): 4,513 / 277,836 ins
Structural: the x138 era is over — 3 fleet-wide families remain and ALL 3 already have matched
exemplars, so no fleet-wide CRACK is left, only propagation. 82% of remaining code now sits in
the two worst cost profiles (x2-9 zero-crack 45.5%, singleton zero-crack 36.9%).
- BANKED: 11 functions at 400-952 ins from the cascade (func_8017D898 952, func_8017CE58 733,
func_801902EC 673, func_8018C2D8 673, func_8018A8D4, func_8017C6F4, func_800CBB38,
func_800CF3A4, +3). check-all 213/213 from a clean tree. 6 near = jr/switch (§53 separate
banking step), 1 failed. The cascade agents wrote 6 new cookbook sections incl. §158.
⚠️ tools-health UNVERIFIED at commit (stale cookbook index fixed, confirming re-run
interrupted) — run it first next session. check-all is the byte oracle and it is green.
- WASTE PREVENTION (Drew: "prevent this from ever happening again, however you need to"):
* tools/validate_targets.py (NEW) — names 5 defect classes (NO-ASM / MID-BODY /
OUT-OF-RANGE / ALREADY-DONE / NO-BOUNDARY), exits non-zero.
* WIRED INTO wave_snapshot so it fails closed — every wave passes through there for its .s
files, so no path from target list to spawned agents bypasses validation. Negative-control:
a 3-target bad list is refused with the exact mid-body offset (+72 bytes of 100).
* The cascade `done()` predicate now short-circuits on SKIPPED as well as MATCH. It tested
only MATCH, so a non-existent target fell Sonnet -> Opus -> Fable and three agents each
proved the same phantom absent: ~29 invalid targets x 3 tiers = 87 of 119 agents, ~9.7M
tokens. A tier that cannot act must END the pipeline, not escalate emptiness.
* docs/accelerators.md A9, including that wave_snapshot's own R32 assertion REFUSED that list
(24 of 57 found) and was routed around — the one instrument warning that was right and ignored.
- B RE-SCOPED (S46-10) and deliberately NOT done: the extend blocker is INTRA-HEADER, not
target-side. engine_core.h declares memcpy FOUR incompatible ways across its DEFINE_ macros;
two in one TU collide. NOT a safe cleanup — the in-tree note at ov_MAIN_012.c:14333 records
that `extern memcpy` disables gcc's builtin and turns an inlined block-move into a CALL, so the
declaration CHANGES CODEGEN. Probe one macro in one binary and byte-gate before any sweep.
- C (dedup_extend over the 129) stays blocked on B. Full context for both in the checkpoint.
Drew: "make it more multi-threaded... I still see my cpu idle for far too long."
Measured, fixed, and regression-tested against the S46-3 bank as a KNOWN ANSWER.
- THE MEASUREMENT: 31s saturated (33 makes/48 cc1/load 27) then ~25s with ONE build alive
while 31 cores idled, repeating. Causes: ex.map starts in list order so the giants land
last, and apply/restore is single-threaded.
- gate_all -> gate_failures: return EVERY failure the sweep already computed (~138 rounds -> 1).
- Longest-first gate scheduling; results re-sorted into `changed` order so the verdict stays
bit-identical to the serial loop's.
- PER-OVERLAY INDEPENDENT SEARCH, IN PROCESSES. My first cut used threads and the box refuted
it: 0-4 builds alive at load 3, because the work is regex over 15k-line files and 138
"parallel" searches all queued on the GIL. Same logic in a ProcessPoolExecutor: 14-29 builds,
load 34.75, search phase ~100s. Safe because the shared header is written ONCE by the parent
and each overlay owns its own .c files + build/<bin>/. Seeded with one in-process search
first — a pool submitted at once gives every worker an empty suspect list and makes all 138
pay a full bisection. place_in_overlay extracted to module level so the worker and the
in-process apply cannot drift (R33); compiles_standalone's fixed t.c is per-call now.
- THE REGRESSION (the point, not the stopwatch): revert src/+config to pre-bank, re-run the
identical command -> 29 functions (same), 141 overlays byte-identical, 682s vs ~1440s, and
285 exclusions vs ~350 => +62 MORE member instances (249,161). The old prefix-based
necessity probe was OVER-EXCLUDING (charging 4 fns to 9 overlays that did not all need
them); the per-overlay shrink minimises per overlay. The faster path is also more correct —
a timing comparison would never have shown it. R22 213/213 + tools-health green.
- STILL SERIAL, now the actual wall-clock (neither is a build): ~3min setup before the first
gate (registered_addrs() yaml-parsing a 1949-group/249k-instance registry + 213 sig loads)
and ~2.5min of sequential reconcile_caller_extern after the search.
- Captured as defaults: docs/accelerators.md A8 + memory fleet-tool-parallelism-defaults.
cookbook index regenerated (my §155c append left it stale — the gate caught it, exit 1).
- cookbook 156: a FAILED draft can poison the fleet. gate_stage's arity pre-pass writes
the shared engine_core.h before the gate; a rejected draft's caller-signature edit
survived and broke 141/213 binaries. Byte-gate held (fail-closed). The trap: a broken
tree makes every later gate report 'near' -- two batches of verdicts were void, not
evidence. Standing practice: GATE_NO_ARITY=1, assert 'git status --porcelain
src/shared config' empty after every batch, recover by revert+replay (deterministic).
- cookbook 157: the cheap-tier size cliff, measured over two controlled waves.
Haiku 4-27 ins 86% (~44k tok/match); >=50 ins 20% (~177k, 4x worse). The documented
'<=50' band was optimistic. Agent honesty 63/63 claims true across 100 drafters.
- tools/wave_snapshot.py: immutable sha1-manifested per-wave .s copy, so a running wave
can no longer block R22's 'make clean'. Coverage-asserting (exit 2 on a missing target),
negative-control proven.
- docs/concurrency-design.md (Fable5): the lane contract, the false-bank correctness
argument, and the finding that a worktree verify certifies the COMMIT -- strictly
stronger than our main-tree R22, which also compiles untracked strays.
- checkpoint p7.
- the gate DECODED from matched C (func_8012832C case 0x300E -> func_80128998 -> streaming
API with &cdFileLocTable[144]) -> scene arithmetic named the 1ST-BOSS arena -> ONE targeted
load captured it at 0x801E4C60
- RETRO-VERIFIED: Phase-3's dumps/ram_castle.bin (2026-06-14) holds it at the SAME address,
same 6,764-B exact prefix — R10 two independent datapoints two months apart;
bossHp_SteamKnight (0x801E4398) lives inside this module's image
- onboarded md_SC02_009 (id 0x3E, TLO 0x4): BYTE-IDENTICAL first build; fleet 213;
R22 213/213; tools-health OK; audit-disc UNCLAIMED 6 -> 5, residue 0
- the last 5 (MAIN/7, MAIN/9, SC03/53/54/56) reclassified emulator->STATIC-RE targets with
decoded leads (memory-map §S45 p3); loc-id map appended to docs/debug-menu-list.txt
- negatives banked: pause menu, memory-box prompt, new-game intro, high/low game, Minku
spawn (slot-A actor 0x15 = md_MAIN_015 candidate naming)
- THE TOUR (Drew driving the retail debug menu; mode-7 hammer over the Redux web API):
all 28 script modules captured live at four byte-verified per-chapter slots
(SC03/73-79 @0x801EF468 ch2-period, SC03/132-138 @0x801E25E8 ch3, SC04/24-30
@0x801E7B28, SC05/23-29 @0x801ED988); the routing law: debug-menu AREA selects the
chapter, each CITY interior streams its own module (member k <-> interior k).
md_MAIN_011/DISELECT byte-proven 24,236/24,240 in RAM; slots A/B/boot R34-verified live.
- MAIN/3 DISCOVERED: the main-menu module (id 0x39, 121,884 B), mis-bucketed as data by
BOTH audit oracles; live byte-proven @0x800CEDF8 (42,632-B exact prefix); onboarded.
- 29 onboardings BYTE-IDENTICAL on first build -> fleet 212; R22 212/212 after three
md_MAIN_003 catches: the A4 DsMix leak; an extract-order-sensitive splat boundary
(bytes: a 1-word data sentinel in .text + fn at +4 -> pinned in symbols file);
corpus.stubs now treats D_*/jtbl_* INCLUDE_ASM as blob includes (mirrors progress.py)
- module-id census (offline, disc-wide): 77 id-law code payloads, 0 further misses;
SC03/55 = confirmed DATA. audit-disc: UNCLAIMED 34 -> 6, residue 0 — the 6 carry
byte-checked negative evidence; next tier = the CD-read tracer
- docs: memory-map §S45 (slots + routing + debug-menu ops), disc-completeness S45
addendum, decision-log R31 entry, docs/debug-menu-list.txt (Drew's transcription)
- .run/s45 evidence allowlisted (tour logs/scripts/rosters); 104 ram dumps LOCAL-ONLY
- new baseline: 93.8% instr / 95.68% fn / 87.2% distinct over 212
- R22 clean-fleet 143/143; fleet 96.13% fn / 94.4% instr (honest grown denominator; pre-expansion
line 95.00% on 140 kept for continuity) / 88.3% distinct. audit-binaries OK over 143.
- make audit-disc: UNCLAIMED 78 -> 75 payloads (3,564,021 -> 2,038,104 B), residue 0 — the three
claims flipped automatically via check.sha, exactly as the ledger was designed.
- S44 session total: 5,126 member-functions banked into the 3 new overlays; the ~2,051 remaining
stubs are the new frontier, visible to every tool via citizenship (no separate ledger rows —
recorded as a deviation from plan I.2e, redundant by construction).
- Part II handoff live in the plan file + the S44 checkpoint block.
- R22 CLEAN-FLEET: 140 passed, 0 failed of 140. Fleet 12,502,519/13,160,961 = 94.997% instr
(+18,146 instructions this session, 23 functions). 393 instructions from the 95.000% bar.
- REDO of the S43-9 retraction, done correctly through harvest_verify (splice/build/keep-iff-
byte-identical/revert) instead of hand-building. 5/5 banked, each re-verified three ways:
image SHA == locked SHA, stub gone, real definition present.
ov_SC03_101/func_801814F8 · ov_SC03_104/func_80184934 · ov_SC04_003/func_8017E4F4 ·
ov_SC04_005/func_80181054 · ov_SC04_007/func_8017FF08
- THE DEFECT THIS PROVES: family_sweep --hseq reported this family 0/5 with
"PLUMBING: parse error before 'unsigned'" — but the remapped drafts are byte-CORRECT. The only
`unsigned` in the draft is INSIDE A COMMENT, so a gate-pipeline transform is eating a `/*` opener
and turning comment text into code. Per-transform runs on the draft alone all preserve it, so it
needs the gate's real invocation (--src-file) to reproduce. NOT YET PINNED — and it is silently
costing banks in every sweep it touches. Next: run the three transforms with --src-file and diff.
- Workaround that banked them: carry the exemplar's typedefs by hand (the family_remap _carry_macros
gap, §146/§152) and gate directly, bypassing the sweep's recovery ladder.
- Also killed a self-inflicted infinite poll: an `until ! pgrep -f "permuter_ils.py <fn>"` loop whose
pattern matched its OWN bash command line, so the condition could never go false (spun 2h30m).
Same family as the day's other defects: a check that cannot return the answer that ends it.
- ✅ R22 CLEAN-FLEET: make clean && extract-all && check-all -> 140 passed, 0 failed of 140.
Discharges the [R22 PENDING] caveats on commit:1486 (the 0xECC family x12) and commit:1487
(func_8018D98C). All 18 of today's banks are confirmed, not incremental artifacts (§130).
- FLEET: 96.63% fn-count / 94.99% instr-weighted (12,501,204/13,160,961) / 89.4% distinct-code.
Session +16,831 instructions, 18 functions. P30's 95% instr bar is 1,708 instructions away
(18,539 at session open). NOTE the report line rounds to "95.0%" — the bar is NOT yet met.
- THE 5th WAVE AGENT: func_8017CE58 is TWO bodies at one address (246 in SC02_000/003, 734 in
SC03_092). The 246 body is byte-identical to func_8017C294 — THE FUNCTION §147 WAS WRITTEN FROM —
so one draft covers 4 instances, and it went 12 (with a recorded "stop searching" verdict) -> 2.
- §147 CORRECTED IN PLACE (H5: original text preserved, correction appended):
* A "stratum 3, unreachable from C" is REFUTED — there is NO stratum 3. The frame is declared
locals then reload spill slots in pseudo-regno order; the mystery 0x108 slot is an ordinary
spill on a loop.c-created pseudo, reachable by writing the loop as an INDEX loop (a pointer
walk puts it at the bottom). Prior drafts faked it with volatile pEnd + dead[7]. (121 -> 54)
* B the unreferenced slots are combine-orphaned sign-extension intermediates (combine.c:10839),
not "?: on memory" frame cost.
* E the qty_compare tie IS breakable — §148-C's zero-emission ref slider. (30 -> 25)
* D applied properly (drop volatile out + the $24 pin, let a1 spill) remains: 54 -> 30.
- CONSEQUENCE: func_8017C294's 15 siblings were parked "until stratum 3 is explained" — that hold
is VOID. Both near-misses logged to the ledger with their measured closeness, not forced (P9).
- PROCESS LESSON in §147: a confident NEGATIVE verdict is a claim like any other — date it, name
its evidence, and re-measure it before letting it park work (same shape as §146).
- func_8017EF68 MATCH 969/969, re-verified by me, gated: ov_SC06_000 byte-identical at da4a26ff.
- MECHANISM (from cc1's own -dR trace, not inferred): the r3000 machine description gives the
memory unit load-ready-cost 2 / store 1, so blockage(load,store)=2 — a LOAD CAN NEVER BE PICKED
IN THE TICK IMMEDIATELY AFTER A STORE PICK. sched2 therefore always wedges one ready ALU insn
between the lw and the sh, and the target's zero-wedge order is UNREACHABLE BY ANY STATEMENT
ORDER. That is why ~20 documented hand variants AND the repaired permuter both floored at 2.
The draft's own §49 sched1-LUID story was incomplete — real but secondary.
- THE LEVER (cookbook §151, "the ghost wedge"): a zero-emission tied in/out asm
`__asm__("" : "=r"(v) : "0"(v), "r"(rival));` — 0 bytes, but a schedulable insn that absorbs the
blocked tick, and it sets reg_n_sets(v)=2 which also kills sched1's birthing boost (one
instrument, both passes). Two measured fallouts: rival-read in the same asm (22->12), then a
second re-tie on a HIGH-REF host to restore allocno live-length parity (each in-loop insn is +1
live length for every loop-spanning allocno; a trio of invariant addresses sat exactly on
allocno_compare's integer-floor boundary). Host choice empirical: pkt=MATCH, ot=705, double=10.
- ✅ R22 CLEAN-FLEET: make clean && extract-all && check-all -> 140 passed, 0 failed of 140.
This DISCHARGES the [R22 PENDING] caveat on commit:1484 — all five banks are confirmed, not
incremental-build artifacts (§130).
- FLEET: 96.63% fn-count / 94.9% instr-weighted (12,489,130/13,160,961) / 89.2% distinct-code;
0 NON_MATCHING (G4); dedup 1919 groups. Session +4,757 ins from 2 cracks x 5 binaries.
Distance to P30's 95% instr bar: 13,782 ins (was 18,539 at session start).
Answering "did we do S4?" honestly: NO, not properly. The earlier pass re-gated only the NEWEST
stored draft per draft-exemplar head (8 banked of 35). S6 then proved that is sampling, not scanning
-- its giant's match was the 9th of 31 drafts, and my first pass had reported "closeness 40".
Redone with EVERY stored draft run through match_one, over the 39 draft-exemplar heads + Drew's
named large-function list (38 targets, 33 with drafts on disk):
14 of 33 targets MATCH from a stored draft (some had 51-57 drafts each)
-> 6 banked first pass, +1 after recover_giant = 7 banked
-> including func_8018057C (897 ins), which was on the "needs an agent" list
The 14 came overwhelmingly from ov_SC01_077 -- exactly the heads where only the newest draft had
been tried. The winning drafts sit in .run/_a10_sample-cn-cast-rc/, .run/drafts-wave-cn-cast/,
.run/drafts-wave-cn/, .run/ab-exp/opus-cn/, .run/backlog_drafts/ -- i.e. spread across many
historical pipelines, which is precisely why "newest" is the wrong selector.
7 still open after recovery (5 near, 2 failed) -- integration classes, drafts kept in .run/s41/rec/.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12483035 -> 12484373 instr; distinct +1,338 / +7 uniq; fn-count +7. instr-weighted 94.9%.
audit-digest OK. 0 NON_MATCHING (G4).
STILL OPEN from S4: the 263x5 cluster (0x80182fd4 exemplar) sweeps 0/5 with `parse error before
'unsigned'` in the spliced draft -- NOT the missing-type class, undiagnosed, do not assume codegen.
And the 2 resident stubs with gate-rejected match_one-MATCH drafts remain untouched.
THE RULE (cookbook §146, now paid for twice): SCAN every stored draft, never sample. A head with 57
drafts has 57 chances, and the pipelines that produced them differ in ways that matter.
The two functions the roadmap has carried as PERMANENT WALLS since Phase 24 are matched in all 138
overlays. Neither needed a siege. Both matched from drafts ALREADY ON DISK.
func_80178004 165 ins x 138 = 22,770 Phase 26: Fable5, ~477k tokens, "intrinsic 3-integer
regalloc wall". THREE stored drafts report match_one
MATCH today; one banked first try, no new work.
func_801412A8 198 ins x 138 = 27,324 close=29/110 since Phase 24. Matched from 1 of 31 stored
drafts + the §37/§124 alias.
WHY func_801412A8 LOOKED INTRINSIC (worth understanding — match_one is structurally blind to it):
the TU declares `extern int func_801412A8(int,int,int,int,int,int)` and its callers USE the return
(`param_1 = func_801412A8(...)`), while the byte-true definition is
`Prim_1412A8 *(Prim_1412A8 *, int, int, int, u16, u16)`. Narrow params cannot agree with an `int`
prototype and the no-prototype escape is illegal once a param promotes, so NEITHER side can move --
and the resulting byte difference is in the CALLERS, which match_one never compiles. The §37/§124
def-side asm-label alias decouples them: the TU decl keeps governing the call sites (codegen
untouched), the definition keeps its byte-true signature.
THEN PROPAGATION RETURNED 0/137 TWICE, both times a missing TYPE, not codegen:
family_remap's `_carry_macros` carries file-scope #defines but (a) NOT typedefs, and (b) is NOT
TRANSITIVE -- it brought addPrim_1412A8 and stopped, though that macro calls setaddr/getaddr and
getaddr casts to PTag_1412A8. Lifted Env_1412A8 / PTag_1412A8 / Prim_1412A8 + OT/getaddr/setaddr
into src/shared/engine_types.h (inside the include guard) -> 137/137, 0 failed.
MY ERROR, CAUGHT BY THE GATE: I lifted the typedefs but did not STRIP them from ov_SC01_077.c, so
they were declared twice and gcc-2.7.2 rejects a repeated typedef even when identical -- the lesson
already recorded at the foot of engine_types.h. R22 came back 139/140 with [FAIL] ov_SC01_077 (the
exemplar's own overlay). Stripped, re-verified, 140/140. A proper lift strips the source;
build_engine_types --strip does both and I did it by hand.
Also a measurement error worth recording: I checked whether the draft defined Prim_1412A8 with a
plain `grep -c` -- which matches inside `addPrim_1412A8` -- and briefly concluded the carry worked.
Substring false positive; the same shape as reading a `return` as a declaration.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12432941 -> 12483035 instr (+50,094 -- EXACTLY the two giants x138); fn-count +276;
instr-weighted 94.5% -> 94.8%. audit-digest OK. 0 NON_MATCHING (G4).
THE RULE THIS BUYS: re-measure a wall before respecting it, and SCAN every stored draft rather than
sampling (my first pass checked 8 of 31 and reported "closeness 40" for a function whose MATCH was
in the 9th). Four minutes of re-measurement was worth 50,094 instructions.
Propagation behind every crack, same session (the multiplier the waves exist for). 19 newly-banked
exemplars from waves 1+2, all in the family_sweep lane (0 has_mid_jr):
87 candidate members / 10,212 ins -> 61 BANKED / 26 failed across 39 overlays
The 26 that did not bank are the known plumbing shapes, not codegen: 20 CC1-FAIL + 5 callee
`conflicting types` (func_8017EFA0 x3, func_8012B23C x2) -- the same classes the S40 recovery ladder
already has levers for (§17a-1 no-proto + call-site cast; recover_giant block-scoping). Left open
deliberately rather than force-banked (P9); they are the cheapest fuel on the board next session.
TOOLING GAP RECORDED: the sweep's classifier writes "CC1-FAIL: make: *** Error 33" WITHOUT the actual
cc1 message, so 20 of 26 failures carry no actionable reason. Diagnosing one currently requires
manually splicing the draft into its TU and rebuilding (done twice this session). The classifier
should capture cc1 stderr the way harvest_verify already does -- worth fixing before the next big
sweep, or every CC1-FAIL costs a manual reproduction.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12425854 -> 12432941 instr (+7,087); distinct +6,145 / +51 uniq; fn-count +61.
instr-weighted back to 94.5% ON THE HONEST (post-main-regen) denominator of 13,160,961.
audit-digest OK. 0 NON_MATCHING (G4).
Two ultracode waves over the open-only h_norm clusters (the pool nobody had ever aimed a wave at),
pool VERIFIED from the sigs first (R14).
wave 1 8 targets 8/8 match_one 5/8 gate first pass -> 8/8 after recovery
wave 2 16 targets 16/16 match_one 14/16 gate first pass -> 16/16 after recovery
THE HEADLINE IS NOT 24/24 -- IT IS THAT NOT ONE FAILURE WAS CODEGEN. All six first-pass gate
failures were TU-integration plumbing, each with an already-documented lever:
func_801802EC redefinition of morph_lerp strip the §77 PROBE LAYER (the draft carries types +
a static inline so match_one can compile standalone;
the real TU already defines them -- scaffolding is
not part of the bank)
func_8018B238 conflicting types D_80115158 recover_giant: draft declared it file-scope as a
struct array, TU declares u8[] BLOCK-scope inside
other functions -> block-scope the draft's externs
func_8017EF54 conflicting types (SELF) §37/§124 def-side asm-label alias (TU declares
void f(void) for no-arg callers; byte-true def takes
s32 in $a0; no-proto escape illegal once a param
promotes)
func_80183D78 conflicting types (callee) recover_giant
func_8017F278 conflicting types func_80146C3C §17a-1: the fleet canonical is the NO-PROTOTYPE
form + the intended signature applied AT THE CALL
SITE; a concrete prototype collides with it
(wave-2's 14 first-pass banks needed nothing -- the wave-1 lessons were folded into the prompt)
=> the gate number measures INTEGRATION, not matching. Run the recovery ladder before recording a
wave's yield or the metrics under-report the drafters and send the next wave hunting walls that are
not there. docs/wave-metrics.md S40-1.
POOL VERIFICATION (R14, and it cut both ways): the frontier report's cluster pool MEASURED
1,677 clusters / 5,795 fns / 319,755 ins at a 3.68x multiplier vs its claimed 1,689 / 5,956 /
326,261 at 2.7x -- within 2-4%, and the multiplier is BETTER than claimed. The SAME document's whale
claim was 3/4 wrong. Verify each claim separately; do not accept or reject a source wholesale.
ALSO: 24/24 members propagated from wave 1's 5 banked exemplars (0 failed) -- the same machinery
that returned 0/39 before this session's cast_call_sites fix.
NEW IDIOMS, distilled in-session (R16/R30):
§144 the LITERAL'S SPELLING picks the immediate encoding (`cnt + 0xff` vs `cnt - 1`: mod-256
identical, both one addiu, but gcc emits 0x00FF vs 0xFFFF from the source text)
§145a combine_givs ANCHOR RULE -- the address-giv group anchors on the LAST address-giv in SOURCE
order (record_giv prepends, combine_givs takes the head); store order decides the base and a
wrong choice spawns a third induction register
§145b a bare `p = r;` is a COMBINE BARRIER (can_combine_p/use_crosses_set_p) -- it preserves a
pointer-bump addiu that combine would otherwise fold into every MEM offset
§145c chained assignment `a=b=c=0` emits stores RIGHT-TO-LEFT
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12420375 -> 12425854 instr (+5,479); distinct +5,479 / +43 uniq; fn-count +43.
audit-digest OK. 0 NON_MATCHING (G4). Cost: 3.73M subagent tokens across 24 agents, 0 errors.
THE BUG. tools/cast_call_sites.py classifies a declaration line with
^([ \t]*)(extern\s+)?([A-Za-z_][\w \t\*]*?)\b([A-Za-z_]\w*)\s*\(([^;{]*)\)\s*;
Feed it a return statement and `return` is a perfectly good identifier where a type is expected:
return func_8012CB64((s32)out, -0xC0, 0x40, -0x60, 0);
^^^^^^ captured as the return TYPE, func_8012CB64 as the DECLARED NAME
so the "rewrite this decl to canonical" path REPLACED the statement with
`extern s32 func_8012CB64(s32,s32,s32,s32,s32);`, DELETING the return. In C89 a declaration after a
statement is a parse error, so the damage surfaced as a bare syntax error in the DRAFT -- reading as
the draft's fault, not the tool's. 9 of 9 staged members of family 0x801848dc lost their return.
fix: a keyword guard (a declaration's type-specifier can never begin with a statement keyword)
family_sweep --hseq --band all over 5 families: 0/39 -> 18/39 banked (only the guard changed)
⚠️ AND THE TRAP INSIDE THE FIX: the obvious R33 move is "route it through cdecl". CHECKED, and it is
WRONG -- cdecl.parse() is a DECLARATOR-GRAMMAR parser that assumes it was handed a declaration; it
reports `return func_X(...);` as declaring func_X and `if (f(a));` as declaring `if`.
Statement-vs-declaration is a question cdecl does not answer. Routing there would have been a silent
non-fix that looked principled. §134's law still holds for line-SHAPE masking; this is a different
question.
BLAST RADIUS (measured, not assumed -- R14): cast_call_sites is in gate_stage's DEFAULT pipeline
(canon_resident_calls -> cast_call_sites -> sig_unify -> harvest_verify) and has been since Phase 20.
Of 44,833 stored drafts, 318 (0.7%) carry a `return f(...);` line this mis-reads, across 67 callees
(func_8014F468 x41, func_8014F6F4 x37, func_8014F74C x32, ratan2 x25). Every one, every time it
passed the gate pipeline, lost its return and failed as PLUMBING. Part of the historical plumbing
tail is this bug.
ALSO IN THIS COMMIT
- S5 CALIBRATION WAVE (8 agents, ultracode, 1.31M tokens). Pool VERIFIED FIRST (R14 -- Fable's whale
claim was 3/4 wrong): measured 1,677 clusters / 5,795 fns / 319,755 ins at a 3.68x multiplier vs
its claimed 1,689 / 5,956 / 326,261 at 2.7x -- its numbers hold, and the multiplier is BETTER.
Result: 8/8 match_one MATCH (close=0), and 5/8 banked whole-binary -- the §52b/§61 gap is
integration, not codegen. Banked: func_801822E0 func_8017EC98 func_801851A8 func_80189A34
func_80188E10 (693 ins x1 before propagation). Not banked: func_8018B238 (FAILED),
func_8017EF54 + func_801802EC (NEAR) -- drafts kept in .run/wave-s40/ for recovery.
- 18 member-banks from the re-run sweep (the cross-address free-h_exact pool: h_exact-identical at
DIFFERENT addresses, which dedup_propagate correctly refuses since it assumes position-locking --
family_sweep is the right lane).
- cookbook §143 (this bug + the cdecl trap + the blast radius); index regenerated.
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet 12419169 -> 12420375 instr; distinct +1,526 / +5 uniq; fn-count +23. audit-digest OK.
0 NON_MATCHING (G4).
NEW IDIOM FROM THE WAVE, not yet folded into §31 (agent was told to write only its draft): a byte
counter must be spelled `cnt + 0xff`, NOT `cnt - 1`. Both are mod-256 identical and both compile to
one addiu, but gcc-2.7.2 picks the immediate encoding from the SOURCE SPELLING (0xFFFF vs 0x00FF).
Also flagged: .run/ghidra_c/func_8017EF54.c is a stale decompile of the WRONG function.
main's sig was dated 2026-06-14 and covered only 1,525 of its 2,002 INCLUDE_ASM stubs: 477
stubs were INVISIBLE to the weighted metrics, so main was being graded against two-thirds of
itself and the fleet denominator was understated.
tools/ghidra_mcp_stop.sh (R23 - clean stop, "Save succeeded", releases the .rep lock)
make sig-refresh BINARY=main -> DumpFunctionSignatures.java, read-only, -noanalysis
signatures 1,729 -> 2,205 (+476)
stubs covered 1,525 -> 2,001 of 2,002 (1 still uncovered, named below)
main .text ins 60,201 -> 79,510 (+19,309)
METRICS RE-BASELINED (a DENOMINATOR CORRECTION, not a regression -- the Phase-27 precedent,
where onboarding 4 hidden overlays honestly moved 68.9% -> 67.0%):
instr-weighted 94.5% -> 94.4% (12,419,169 / 13,160,961; numerator UNCHANGED)
distinct-code 89.3% -> 89.0% (5,029,454 / 5,654,184; numerator UNCHANGED)
fn-count 96.51% (unchanged -- it never read the sig)
MAIN game-code 0.72% -> 0.55% (436 / 79,510)
Both numerators are identical across the change; only the denominator grew. audit-digest OK.
CAVEAT UNCHANGED (R34): main still has no INDEPENDENT boundary oracle -- sig_image cannot sign a
PS-X EXE (header offset, interleaved islands), so main's function list rests on Ghidra alone and
audit-corpus cannot cross-check it. That is the open item in docs/second-oracle.md and part of
task #9's denominator work.
NOTE the sig itself is gitignored/regenerable, so this commit lands only the digests; a fresh
clone reproduces it with `make sig-refresh BINARY=main` (Ghidra required, MCP stopped).
MCP IS NOW DOWN (I stopped it for the project lock). The SessionStart hook restarts it next
session; run /mcp before any Ghidra work (R29).
Measured the h_exact free pool from the bytes rather than trusting the frontier report's
numbers (R14 — its whale claim was 3/4 wrong: it said the whale was open in all four SC07
overlays; three were already banked and I closed the fourth earlier this session).
MEASURED: 215 open function-instances / 8,763 instructions are byte-identical (h_exact,
including reloc payloads) to an already-matched function. ONE class is 86% of that pool:
func_801758FC — 55 ins, same address in all 138 overlays, matched in ov_SC01_000 only,
OPEN in the other 137 => 7,535 instructions.
h_exact means identical INCLUDING jal/lui/%lo reloc immediates, so the matched body compiles
byte-identically at every member with NO remap (dedup_extend's correctness argument, §14).
dedup_propagate --addr authored it once as DEFINE_func_801758FC() in engine_core.h and
instantiated it at all 137 open sites in address order.
[ OK ] 138 overlays byte-identical after propagation; 1 new group in config/dedup.us.yaml
VERIFIED: make clean && make extract-all && make check-all -> 140 passed, 0 failed of 140.
Fleet instr 12411467 -> 12419002 = +7,535 EXACTLY; fn-count +137; instr-weighted crosses to
94.5%. distinct-code unchanged BY DESIGN -- the class was already matched in ov_SC01_000, so
the 137 add fleet instructions but no new DISTINCT function. audit-digest OK. 0 NON_MATCHING.
Note this function had been sitting in the stored-draft backlog for ov_SC06_030 and
ov_SC07_010 and re-gated "no" earlier tonight -- because gating a DRAFT is the wrong move for
an h_exact class. The right move is propagating the already-MATCHED body. Same function, two
routes, and only one of them is free.
Remaining free pool after this: 78 instances / 1,228 ins across 32 classes.
Drew T