Tooling: accelerate exact quadratic-affine witness replay

[Curtis-D]

The historical quadratic-affine audit constructs its largest minor through pairwise Hermitian products. This optional adapter computes the same complete rows with carry-safe integer lanes. Tooling improvement; no new kappa or new negative result.

Three alternating fresh-process pairs on CPython 3.14.2/macOS arm64 measured median complete-audit wall time 73.411 s → 10.959 s (6.70x). All six certificates match the frozen original byte for byte. This includes imports, source admission, all twelve witnesses, rational span controls, certificate comparison and output; it is not a whole-repository or selected-frontier speedup. Raw samples and the overlapping preliminary probe are disclosed separately.

Only research/quadratic-affine-packed/ is added. The original checker, default commands, selected frontier and upstream/ remain unchanged. The README supplies the signed-lane, zero-detection and compaction arguments, source pins, provenance and limits.

Reproduce and validation

python3 -I -B research/quadratic-affine-packed/verify.py

Expected final line: PASS_QUADRATIC_AFFINE_DIFFERENTIAL all_frozen_rows_equal=true cases=12.

• 14,662 full rows / 161,333,572 declared matrix entries directly compared against the original pairwise calculation, with unchanged ordered binary elimination, digests, rational span checks and complete certificate.
• 19 test groups pass: 12 packed-arithmetic tests, three source/CLI guard tests and four unchanged original tests. A literal integer-component oracle covers all 125 three-coordinate phase vectors, all 60 small Gaussian labels, seeded cases at every dimension 1–32, signed extremes and lane boundaries.
• Meaningful controls detect missing conjugation, ignored imaginary parts, omitted diagonal entries, adjacent-lane borrowing, dependent rows, source/fixture/certificate corruption, disabled assertions and reused output. The omitted-diagonal example retains the same rank but changes its digest.
• Isolated minimal replay from an unrelated directory and a path with spaces passes, preserving every copied file. Selected-record/source-inventory verification and staged whitespace validation pass.
• Full make verify, selected cold supplier replay, Lean and remote Linux CI were not run or claimed for this optional adapter.

The existing negative result concerns only twelve specified full-orthogonality matrices. Other supported matrices, subfamilies and compilers remain outside it. All inherited all-size, realization, hosting, precision/tape and analytic obligations retain their scope.

Pins, assignment and credit

Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60 (still main at final fetch). Historical source: d55bcd62ceb8f322bcf311a5a66e94c3ac7724f0; upstream: adc7f1241b42e322a6451854ab7e4b4c146bf78a. SOURCES.json pins all 37 predecessor files consumed by the replay. At this base the selected conditional kappa remains #626/#543's 0.001635527380375392.

OS-random seed 92, verification-engineering track. Three candidates and scope are recorded in SELECTION.json; there was no pinned coordination issue, so the claim is local. Final overlap review covered 565 open PRs through #746, then fresh arrivals through #748. PR154's unchanged quadratic-affine hotspot is credited; the newer geometry, Gram, owner-mask, cyclic-inverse and streaming tools concern distinct kernels. No pending code is imported.

Credit Douglas R. Colkitt for the original families, witnesses and audit, and eumemic/PR154 for the profiling lead. Prepared by Curtis Dransfield (@Curtis-D) with substantial OpenAI Codex assistance, including a separate scoped AI arithmetic reviewer and test author. No independent human review is claimed. Apache-2.0; predecessor attribution remains intact.

[Mister-Meeseeks]

100-PR batch review at 3603d8d9184995ae1cc67c0f5e01f02f3a1d634e.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Opt-in packed quadratic-affine verifier; must preserve all exact phase/tableau/witness checks, no numerical approximation.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

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