Scoped negative result: bounded butterfly search misses the PR681 factory cost

[RossRichardson]

Review continuation completed. Current head 679ac7b760ef2e01262039b906c154012548be81 integrates canonical main 38746ad3... while preserving the original scientific files and receipts. The complete reviewed-head command passes in 80.545 s; the new historical-export command passes in 74.671 s, including all 78,728 recipes and full losing-witness comparisons. These are verification runtimes, not new benchmark claims. A separately audited 29-blob export and eight controls preserve exact historical inputs, active assertions and an unchanged checkout. The focused full-history workflow supplies CI; live status is separate in the follow-up comment.

Scoped negative result for the historical PR622/PR681 eight-register factory map. An exact enumeration of 3,072 output-butterfly recipes and 75,656 weighted-input/output recipes finds minima of 95 and 93 real-component primitives. Both miss the accepted PR681 + PR696 incumbent cost of 81. All 78,728 labeled recipes pass the complete dirty-input matrix check; recipes can duplicate circuits.

The package retains the full declared domains, deterministic rules, cost/map and word-stream hashes, and closest losing witness. Independent dyadic replay checks all four matrix orientations and both primitive lowerings. Fresh reconstruction of all fourteen raw factory windows confirms the closest witness costs 124,440/124,456 primitives rather than 123,768/123,784: 672 extra per orientation. The witness also fails the stronger cleanup-envelope domination criterion, which is recorded explicitly. It is not proposed for installation. The per-macro adapter allowance is also repriced: 320,000 → 323,584; the retained local PR696 bound worsens by 92,948,660,896. This is a local arithmetic recount, not a whole-owner/global invoice run.

No new κ, global optimality, or impossibility theorem. Exact base: 91a2094f2e022930069e8cfdc810a3ea1b81cd09. Seed 21 selected the complex track. The delta is research/complex-factory-butterfly/ and a focused workflow. The 29 historical predecessor/context files are exported from their immutable Git base and checked against their original hashes. PR799 is historical benchmark context; this contribution changes neither the live selected result nor upstream/.

Reproduce offline with Python 3.11+, the standard library and a Git checkout containing the full historical source history:

python3 -I -B research/complex-factory-butterfly/pinned_replay.py

Expected: PASS_FACTORY_BUTTERFLY, 14 blocks, gate counts 34/42, and −672 primitive savings in each orientation. This command replays the entire declared domain and full raw scalar programs, rather than checking only a stored winner.

Original validation includes full normal replay, optimized minimal-source relocation without Git/cached outputs, ten focused tests, twelve mathematical/physical adverse controls, four packaging controls, selected-record integrity and staged whitespace checks. See VALIDATION.md. No full repository suite, selected cold replay, remote CI or Lean is claimed. All-size/compiler, realization, hosting, precision/tape and analytic assumptions remain inherited.

The final overlap check found CubeJerry's concurrent #922 at efd5643c0016b6c6dac26e9ed7dee8a75295ff2a. It proposes noncommuting three-register contractions of the same incumbent. This contribution adds distinct finite-family negative coverage; it does not claim that incumbent is optimal, import #922, or independently audit its witness. #802 corroborates the incumbent separately. A subsequent refresh also found #924 at e6f3cb8ca813dfc54c77738c9acd74a536b607fd, concerning coupled scales and a fixed-topology gauge certificate. That distinct result is not imported or audited; its cost discussion prompted our explicit adapter-allowance recount from pinned sources and our fresh replay.

Credit Alluri (#681), Sagar Thakrar (#696), Jeremy Kothe (#622), Rohan Guliani (#802), and the retained supplier/compiler lineage and notices. Prepared by Ross Richardson (@RossRichardson) with substantial OpenAI Codex assistance, including scoped AI implementations and review. Apache-2.0; no independent human review claimed. No pinned sprint issue existed; the local claim records the result.

[Mister-Meeseeks]

100-PR batch review at 4e2925b076050f8035b054f95416029a63af64fa.

This remains open after scoped triage. Next review step: The complete butterfly/elimination recipe replay reached the 180-second budget. Finish its 78,728-recipe coverage and full losing-witness comparisons; do not infer exclusion from incomplete search.

This is a specific continuation point, not a rejection or a request to redo already recorded checks. Completed local evidence is retained separately from outstanding whole-head obligations.

Batch review, attribution and evidence. Thank you for the contribution.

[RossRichardson]

Completed the requested full replay and pushed current-main compatibility at 679ac7b760ef2e01262039b906c154012548be81.

The exact reviewed head 4e2925b076050f8035b054f95416029a63af64fa passes in 80.545 s with a clean checkout before/after. The integrated pinned-export command passes in 74.671 s, also without changing the checkout:

python3.12 -I -O -B research/complex-factory-butterfly/pinned_replay.py

The export authenticates all 29 original Git blobs/SHA256 pins and calls the unchanged verifier with assertions enabled even when its caller uses -O. It excludes stale bytecode and inherited optimization/import settings. This preserves the historical manifest, producer/checker, search streams, candidate and frozen receipts despite three current-main source-path/hash changes. Eight export controls pass, including corrupt/missing blobs, path/symlink/cache exclusion, active assertions and checkout mutation detection. A focused full-history Linux workflow now runs the same complete gate.

Both full runs cover all 78,728 labeled recipes, zero failed-pivot skips, exact dirty maps, complete stream hashes, all fourteen physical windows, both full native/emitted scalar orientations, losing-witness comparisons and adverse controls. The finite minima stay 95/46 and 93/42 against 81/34. The closest witness still costs 672 extra primitives per orientation and fails the stronger sufficient envelope check at cleanup: endpoint growth increased. No universal exclusion or proposed installation follows. A separate AI-assisted mathematical/code audit found no substantive issue; no independent human review is claimed.

Current-head log SHA256: 0cef455449cdbc6285ac822b294e28751bffee4d2f34d742911ecbf4e1b47228.

Delta and limitations. All 62 hosted checks were queued at this exact head when checked; no hosted success is inferred. Maintainer incorporation remains separate. Original predecessor attribution and Apache-2.0 notices are retained; Ross Richardson with substantial OpenAI Codex assistance. No new kappa, whole-owner invoice, all-size supplier or formal-kernel proof is claimed.