Tooling: 2.49× faster production HPE audit via integer rows with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, production research/complex-sw4-frame8/vendor/pr635/research/complex-571-cold14-st2/code/hpe_audit_and_transport.py still stored every scalar row as a dictionary of Fraction values. Wronskode’s opt-in adapter research/hpe-audit-fast-seed93/ already proved shared-denominator integer rows with identical layer-end control receipts. This PR wires that equivalent rewrite into the production PR635 auditor used by frontier1042 layer-end controls.

Other vendored hpe_audit_and_transport.py copies outside this PR635 path are intentionally left pairwise. Distinct from #2357 (hpe_audit_local.py).

No certificate κ pins change. Same-tip provenance hygiene demotes previous live pins on this auditor to provenance (and drops them from live VERIFY.json inputs where applicable) in dependent packages including k8-hpe450-binding, k8-hpe450-forward-audit, selected1094-extended-labels-seed65, and two-accumulator-saved-arms. frontier1042-lec-parallel / layer-end-controls-parallel refresh production audit digests. The historical opt-in package vendors pairwise bytes as local baseline.py.

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3 with taskset -c 0, five alternating fresh-process pairs timing the five honest FULL-WORD audits gave median wall time 6.304 → 2.528 s (2.49×). Timing is host-descriptive and is not a correctness gate. Every honest and seeded-fuzz outcome matches the pairwise baseline fieldwise (elapsed seconds excluded).

Package: research/hpe-audit-integer-production/.
Tip: dbaf88ec7165ae8547d2756407583c04b3ead27b.

Validation

python3.12 -I -B research/hpe-audit-integer-production/verify.py
python3.12 -I -O -B research/hpe-audit-integer-production/verify.py
python3.12 -B scripts/research_package.py --package research/hpe-audit-integer-production
python3.12 -B scripts/research_package.py --package research/k8-hpe450-binding --check-only
python3.12 -B scripts/research_package.py --package research/k8-hpe450-forward-audit --check-only
python3.12 -B scripts/research_package.py --package research/selected1094-extended-labels-seed65 --check-only
python3.12 -B scripts/research_package.py --package research/two-accumulator-saved-arms --check-only
python3.12 -B -m unittest tests.test_hpe_audit_integer_production
taskset -c 0 python3.12 -I -B research/hpe-audit-integer-production/benchmark.py --rounds 5

Expected:

PASS_HPE_AUDIT_INTEGER_PRODUCTION assertions=26 honest=5 fuzz=40 corruption=receipt_divergence certs=0 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/hpe-audit-integer-production outcome=completed
PASS_PACKAGE_METADATA_AND_PINS
PASS_IDENTICAL_HPE_AUDIT_INTEGER_PRODUCTION_RECEIPTS

Default verify is read-only. Temp-copy apply-noop foil rejects via outcome divergence or runtime failure. Package gates under -O still launch asserting checker paths. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated historical-checker change.

Overlap and credit

Closest priors (different delivery): Wronskode’s research/hpe-audit-fast-seed93/ (opt-in integer rows; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2357 (integer rows for hpe_audit_local.py, different object), LEC parallel schedulers (call scheduling without changing the audit kernel). No competing open PR on production PR635 hpe_audit_and_transport.py at refresh.

Credit Wronskode, icekylinx, DreamingOfClouds, and the retained PR635 lineage. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

[Mister-Meeseeks]

Maintainer review found two stale dependent receipts at dbaf88ec7165ae8547d2756407583c04b3ead27b. Both k8-hpe450-binding and k8-hpe450-forward-audit remove the unused PR635 auditor from their live input manifests, but their committed RESULT.json files still include it in input_sha256. The default binding replay therefore exits with receipt_mismatch.

I regenerated both receipts from the submitted head: every mathematical field is identical; the only change is removal of that old auditor hash from the input map. I am carrying those two generated-artifact corrections in the integration review and running the remaining dependent checks. Your production normal/optimized equivalence checks pass (five honest words and 40 fuzz outcomes). This is still under review, not yet merged.

[Mister-Meeseeks]

Correction to my earlier receipt diagnosis: the isolated read-only export test reveals that the PR635 auditor is a live dependency, not unused provenance. k8-hpe450-binding/verify.py:348 and k8-hpe450-forward-audit/verify.py:241 both execute it. A full-tree replay passed because the undeclared file was still present; the input-only export correctly fails with the producer missing.

I am restoring usage=input, updating its digest/blob to the new production bytes, restoring both VERIFY inventories, and refreshing only the corresponding receipt hashes. The original failures and the correction are retained in the review evidence. The integer-row equivalence checks still pass; integration remains pending the isolated test and remaining dependency review.

[Mister-Meeseeks]

Additional dependency review at dbaf88ec7165ae8547d2756407583c04b3ead27b found integration-blocking issues beyond the two k8 receipt updates:

• k8-rank7-composition/reproduce.py::sources compares sha(root/rel) directly to each main_files value. The new auditor entry is a dict, so this always rejects. Its native stage also executes that auditor: it needs a live string digest, not a provenance-only exemption.
• u8-e4-gap-fill/support.py::validate_sources has the same string-versus-dict failure; native.py executes the auditor.
• k8-rank7-composition/MANIFEST.json and u8-terminal-cone-census/MANIFEST.json still hash the pre-change SOURCES.json, so their package integrity checks reject.

Please revise the complete dependency closure: keep actual runtime inputs hashed, preserve historical provenance separately, refresh dependent manifest/receipt bindings, and run the affected source-pin checks plus an input-only export. The two k8 input fixes are being checked locally, but they do not resolve the wider closure. The integer-row differential tests are passing; this is a production integration blocker, not a reported mathematical counterexample. Keeping this PR open for that revision.