Tooling: 3.09× faster production frame-pair enumeration via eight-bit neighborhoods with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, production research/complex-frame-pair-decoders/verify.py:frame_pair_solutions still rebuilt partner lists with nested list comprehensions on every mask. woklerx’s opt-in adapter research/frame-pair-enumeration-fast-seed93/ (incorporated #1099) already proved an eight-bit neighborhood enumerator with identical PR968 D/E/H receipts. This PR wires that equivalent rewrite into the production verifier.

No certificate κ pins change. Same-tip provenance hygiene demotes previous live usage:input pins on this verifier to provenance (and drops them from live VERIFY.json inputs) in:
complex-puncture-depth-ceiling, complex-general-normal-boundary, p104-term-shear-seed8be7, p104-two-copy-drops, e8-connectivity-incidence-boundary, selected-p104-port-adapter, and p112-batch-centre-floor. Live pairwise model pins retarget to vendored baseline.py via auxiliary e8_model (not a historical files[] pin) for selected-p104-port-adapter, p104-nullpair-compensated-shear-seed318bd5, and p104-term-shear-seed8be7. Legacy WT-hash loaders p104-asynchronous-carrier-bound, root-pair-first-drop-seed48, complex-decoder-field-extension-seed44, and e8-factor-linear-overcomplete47 likewise load that baseline instead of the rewritten production verify.py. 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 gave median D/E/H wall time 13.642 → 4.421 s (3.09×). Timing is host-descriptive and is not a correctness gate. Receipts stay identical (frames=1260, decoders=7168, term_ranks=57344).

Package: research/frame-pair-enumeration-production/.
Tip: 462593231536f771050da8f44390b3e4eba95783.

Validation

python3.12 -I -B research/frame-pair-enumeration-production/verify.py
python3.12 -I -O -B research/frame-pair-enumeration-production/verify.py
python3.12 -B scripts/research_package.py --package research/frame-pair-enumeration-production
python3.12 -B scripts/research_package.py --package research/complex-puncture-depth-ceiling --check-only
python3.12 -B scripts/research_package.py --package research/selected-p104-port-adapter --check-only
python3.12 -B scripts/research_package.py --package research/p104-nullpair-compensated-shear-seed318bd5 --check-only
python3.12 -B scripts/research_package.py --package research/p104-term-shear-seed8be7 --check-only
python3.12 -I -B research/p104-nullpair-compensated-shear-seed318bd5/verify.py
python3.12 -I -B research/p104-asynchronous-carrier-bound/verify.py
python3.12 -I -B research/root-pair-first-drop-seed48/verify.py
python3.12 -I -B research/root-pair-first-drop-seed48/check_sources.py
python3.12 -I -B research/complex-decoder-field-extension-seed44/verify.py
python3.12 -I -B research/e8-factor-linear-overcomplete47/verify.py
python3.12 -I -B research/p104-term-shear-seed8be7/verify.py
python3.12 -B -m unittest tests.test_frame_pair_enumeration_production
taskset -c 0 python3.12 -I -B research/frame-pair-enumeration-production/benchmark.py --rounds 5

Expected:

PASS_FRAME_PAIR_ENUMERATION_PRODUCTION frames=1260 decoders=7168 term_ranks=57344 corruption=subprocess_failure certs=0 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/frame-pair-enumeration-production outcome=completed
PASS_PACKAGE_METADATA_AND_PINS
PASS_P104_NULLPAIR_COMPENSATED_SHEAR
PASS_P104_ASYNCHRONOUS_BOUND
PASS_ROOT_PAIR_FIRST_DROP_S48
PASS_COMPLEX_DECODER_FIELD_EXTENSION_S44
PASS_FACTOR_LINEAR_OVERCOMPLETE
PASS_P104_TERM_SHEAR_FINITE_SCREEN
PASS_IDENTICAL_FRAME_PAIR_ENUMERATION_PRODUCTION_RECEIPTS

Default verify is read-only. Temp-copy omission foil rejects via subprocess failure. Package gates under -O still compile checkers with optimize=0. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated verifier change.

Overlap and credit

Closest priors (different delivery): woklerx’s research/frame-pair-enumeration-fast-seed93/ / #1099 (opt-in neighborhood enumerator; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2348 / #2357 (same production-promotion + provenance pattern). No competing open PR on production frame_pair_solutions at refresh.

Credit themilkman216 and woklerx. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

[Mister-Meeseeks]

Maintainer integration review at 462593231536f771050da8f44390b3e4eba95783: the remaining downstream pin migration needs to include at least two more consumers before this production rewrite can land.

I executed research/p25-root-pair-bound-seed632639/verify.py:check_pins against its complete pinned inputs. It passes with the original PR968 source. Replacing only research/complex-frame-pair-decoders/verify.py with this PR's production bytes fails with:

ValueError: working input SHA drift: research/complex-frame-pair-decoders/verify.py

That consumer is not changed in this PR. The newly integrated research/cycle47-shared-frame-correction also imports the same model at runtime and lists the original bytes as a live input in SOURCES.json / VERIFY.json. This second dependency arrived after your base, so please include it when refreshing the integration.

Please preserve the historical model for these consumers (or explicitly update their live binding and replay the affected checks), keeping their expected receipt/source metadata consistent. Merely demoting a runtime dependency to provenance would not address reproducibility. A complete manifest-and-loader scan should catch any further consumers.

This is an integration hold, not a rejection of the neighborhood-mask optimization or its numerical results. The optimization's full review remains pending. Exact reproduction evidence will be retained in the efficient1300 batch audit; themilkman216, woklerx and your production-wiring contribution remain credited.