Tooling: 6.47× faster production R420 mutants geometry cache with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, production research/e8-bit-r420/code/mutants.py still rebuilt full geometry for every gate-only mutation control. The opt-in adapter research/r420-geometry-cache/ already proved geometry reuse with identical mutation receipts. This PR wires that equivalent cache into the production harness used by verify.sh: gate-only controls keep identical geometry fields; Word.__init__/labels() are cached; full()/interface() always rerun. research/e8-bit-r420/MANIFEST.sha256 binds the new digest.

No certificate κ pins change. The pairwise-era mutants.py digest in research/r420-geometry-cache/SOURCES.json is moved to provenance_inputs so research-packages CI does not fail working-source drift. Leave #2263 xcheck.rref alone (this tip does not edit xcheck.py). Both PRs update MANIFEST.sha256 for different file digests — expect a trivial manifest merge if both land.

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3, five alternating fresh-process pairs gave median mutants-harness wall time 9.602 → 1.485 s (6.47×). Timing is host-descriptive and is not a correctness gate. Mutation stdout digest stays 8c2738b3258e…; JSON receipts match the baseline harness.

Package: research/r420-geometry-cache-production/.
Tip: feb0b96b5c471be0a8c72e6ecb8f9e1f74b48991.

Validation

python3.12 -I -B research/r420-geometry-cache-production/verify.py
python3.12 -I -O -B research/r420-geometry-cache-production/verify.py
python3.12 -B scripts/research_package.py --package research/r420-geometry-cache-production
python3.12 -I -B research/r420-geometry-cache-production/benchmark.py --rounds 5

Expected:

PASS_R420_GEOMETRY_CACHE_PRODUCTION controls=10 corruption=subprocess_failure mutants_sha=8c2738b3258e certs=0 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/r420-geometry-cache-production outcome=completed
PASS_IDENTICAL_R420_GEOMETRY_CACHE_PRODUCTION_RECEIPTS

Default verify is read-only. Temp-copy stale-full foil rejects via subprocess failure. Package gates under -O still launch assertion-enabled children. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated harness change.

Overlap and credit

Closest priors (different delivery): research/r420-geometry-cache/ (opt-in geometry reuse; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2257 / #2263 (same production-promotion + provenance pattern). No competing open PR on production mutants.py geometry reuse at refresh.

Credit the authors of research/r420-geometry-cache/ and eumemic (PR543 harness/checker). Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

[Mister-Meeseeks]

Maintainer review of feb0b96b5c471be0a8c72e6ecb8f9e1f74b48991: the cache correctly keys geometry and reruns full()/interface() on a fresh Word; the normal production receipt comparison passes. There is an integration regression to address before merging.

research/r420-rref-performance/adapter.py:bind_sources() validates every original execution input at import, including research/e8-bit-r420/code/mutants.py. Its manifest still requires e2fd3429..., while this PR installs bc6b0106..., so this previously retained historical replay fails immediately. Updating/demoting the geometry-cache package's pin alone does not cover this consumer, and a consumed baseline must not simply become unbound provenance.

Please preserve the original baseline bytes for historical A/B replays, route their executable inputs to those immutable copies (or an explicit pinned historical replay), inventory other executable consumers, and demonstrate affected registered replays still work. Coordinate with #2263 because both modify the same production manifest and R420 historical inputs. The production logic itself looks promising; this hold concerns reproducibility and integration, not the claimed mathematical result or host-specific benchmark ratio.

Supplemental exact-head pin evidence will be in research/community-efficient1300-audit/combined-r420-regression.json at the next checkpoint.