Tooling: 2.63× faster production hosted digest via fragment reuse with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, production p111-tuned audit_hosted_k.py still hashed every mapped instruction with a fresh json.dumps. Matthieu Pascal’s opt-in adapter research/bit-hosted-digest-fastpath/ (incorporated #1874) already proved JSON-fragment reuse with identical hosted-audit receipts. This PR wires that equivalent digest into the production auditor used by adapt9p-parameterized PR589 hosted audits.

Path: research/e8-bit-p111-tuned/vendor/pr510/e8-r480-nine-stage/vendor/e8-r715-eleven-stage-whole/code/audit_hosted_k.py.

No certificate κ pins change. Same-tip provenance hygiene:

• research/bit589-width7-windows-seed72/, research/selected-bit-hosting-audit/, and research/r439-abc-hosted-invoice/ demote audit_hosted_k.py to provenance (pairwise-era digest) and drop it from live VERIFY.json inputs — a naive sha refresh against an old base_commit is invalid for research-packages pin checks.
• research/bit-hosted-digest-fastpath/ vendors pairwise-era bytes as local baseline.py and compares against that file (production path is provenance-only).

Tooling improvement; no new kappa. Selected frontier untouched. Adapt9p’s five hosted-auditor anchors are preserved.

Result

On Linux / CPython 3.12.3 with taskset -c 0, five alternating fresh-process pairs gave median complete parameterized PR589 hosted-audit wall time 34.880 → 13.268 s (2.63×). Timing is host-descriptive and is not a correctness gate. Emitted-program digest stays 96c7ad2e… (7,843,392 events; 16 mutants).

Package: research/bit-hosted-digest-production/.
Tip: b5c3b567f13b18881161cf6c445e4ef251870558.

Validation

python3.12 -I -B research/bit-hosted-digest-production/verify.py
python3.12 -I -O -B research/bit-hosted-digest-production/verify.py
python3.12 -B scripts/research_package.py --package research/bit-hosted-digest-production
python3.12 -B scripts/research_package.py --package research/bit-hosted-digest-fastpath --check-only
python3.12 -B scripts/research_package.py --package research/bit589-width7-windows-seed72 --check-only
python3.12 -B scripts/research_package.py --package research/selected-bit-hosting-audit --check-only
python3.12 -B scripts/research_package.py --package research/r439-abc-hosted-invoice --check-only
taskset -c 0 python3.12 -I -B research/bit-hosted-digest-production/benchmark.py --rounds 5

Expected:

PASS_BIT_HOSTED_DIGEST_PRODUCTION events=7843392 … corruption=audit_assertion certs=1 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/bit-hosted-digest-production outcome=completed
PASS_PACKAGE_METADATA_AND_PINS
PASS_IDENTICAL_BIT_HOSTED_DIGEST_PRODUCTION_RECEIPTS

Default verify is read-only. Temp-copy digest foil rejects via audit assertion. Package gates under -O still launch asserting children. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated auditor change.

Overlap and credit

Closest priors (different delivery): Matthieu Pascal’s research/bit-hosted-digest-fastpath/ / #1874 (opt-in fragment digest; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2257 / #2263 / #2276 / #2320 / #2329 (same production-promotion + provenance pattern). No competing open PR on this production audit_hosted_k.py at refresh.

Credit Matthieu Pascal, DreamingOfClouds, eumemic, DaysSky, Chafik Boukhalfa and the retained hyperplane599 archive. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

[Mister-Meeseeks]

Maintainer review at b5c3b567f13b18881161cf6c445e4ef251870558: the r439-abc-hosted-invoice manifest change needs revision before integration.

audit_hosted_k.py is still a live computational input, not just provenance. verify.py invokes e8-bit-p111-tuned/code/adapt9p.py with the vendored code directory; that adapter reads code/audit_hosted_k.py to produce the hosted auditor executed later. I replayed the exact parent adapter and instrumented port adapter at this head. The read inventory contains:

research/e8-bit-p111-tuned/vendor/pr510/e8-r480-nine-stage/vendor/e8-r715-eleven-stage-whole/code/audit_hosted_k.py

It is absent from the declared input set after this PR demotes it. Consequently the full verifier's consumed <= set(inputs) gate will reject it even if the preceding native stages succeed. This is an actual adapter replay, not a full native-pipeline replay or a claim that the digest optimization is mathematically incorrect.

Please retain a live binding for the auditor actually consumed: either preserve a pinned historical baseline through the adapter chain, or explicitly bind and replay the updated implementation with the affected complete receipts and rejection controls. Do not remove the consumed-input coverage gate. The other manifest changes and production equivalence still need their complete review.