Tooling: 69× faster mixed-point Circuit.verify supports with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, merged CrocSwap/integer-mult-bounds#1984 sped up mixed-point scalar_check digests but left the general Circuit.verify path on scripts/experiments/mixed_point_circuit.py on the historical set-intersection genexpr

sum(1<<i for i,source in enumerate(inputs) if len(set(source)&set(target))==r)

Profiling on build(30,3,3,1) shows that genexpr dominates Circuit.verify wall time. This PR replaces it with an equivalent closed-form point-incidence mask: precompute vertex[p], then for each r-subset U of the target keep inputs that contain every point of U and avoid every other target point. Every output support still matches the historical checker, verify still returns True, and support admission uses require so it stays enforced under -O. Only the source pin for scripts/experiments/mixed_point_circuit.py changes inside the two certificates that bind that file.

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3 (Linux-6.12.94+-x86_64-with-glibc2.39), five alternating fresh-process pairs timing only Circuit.verify on the fixed fixture set build(20,3,3,{0,1}), build(26,3,3,{0,1}), build(30,3,3,1) (construction excluded) gave median wall time 8.331 → 0.121 seconds (69.03×). Timing is host-descriptive and is not a correctness gate. Every verify receipt matches the baseline checker (True return and identical output-support fingerprints).

Package: research/mixed-point-verify-masks/.

Validation

python3.12 -I -B research/mixed-point-verify-masks/verify.py
python3.12 -I -O -B research/mixed-point-verify-masks/verify.py
python3.12 -B scripts/research_package.py --package research/mixed-point-verify-masks
python3.12 -I -B research/mixed-point-verify-masks/benchmark.py --rounds 5

Expected:

PASS_MIXED_POINT_VERIFY_MASKS verify_cases=11 mask_cases=3994 certs=2 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/mixed-point-verify-masks outcome=completed
PASS_IDENTICAL_MIXED_POINT_VERIFY_MASK_RECEIPTS

Default verify is read-only. Mask nonequivalence mutant, corrupted-support ValueError, and temp-copy digest corruption controls reject. 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 objects or delivery): CrocSwap/integer-mult-bounds#756, merged CrocSwap/integer-mult-bounds#1984 (scalar_check digests on the same file; this only rewrites general Circuit.verify), CrocSwap/integer-mult-bounds#2001 (ExclusionCircuit / DisjointTensor support masks; same style, different file/relation), CrocSwap/integer-mult-bounds#1861 (open; same parity-side.json, different pin keys), CrocSwap/integer-mult-bounds#2117 (open; same parity-side.json, different pin key disjoint_tensor.py). No competing open PR on mixed-point Circuit.verify at refresh.

Douglas R. Colkitt authored the mixed-point circuit. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.