Tooling: 1.88× faster production joint-dual replay via containment masks with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, scripts/experiments/binary_frame_replay.py:replay still paid an O(outputs × triples) literal scan for each paid output when checking expected source symbols. Aaron Wang’s opt-in adapter research/joint-dual-replay-fast/ (incorporated #1880) already proved exact containment masks with identical receipts. This PR wires that equivalent check into the production source used by make joint-dual-verify.

Only the source pin for scripts/experiments/binary_frame_replay.py changes inside certificates/joint-dual-kappa.json, certificates/skip-frame-kappa.json, and research/joint-dual/SOURCE.json (plus the consequent source_manifest_sha256 / skip-frame certificate digest). Compiler replay bytes and every assertion / check_word / full-basis orientation receipt stay identical.

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 both pinned words gave median wall time 5.873 → 3.120 seconds (1.88×). Timing is host-descriptive and is not a correctness gate. Every replay receipt matches the baseline checker fieldwise. The measured path is both-words replay() (the historical checker used by make joint-dual-verify), not full make wall time including C++/transitions/unittest.

Package: research/joint-dual-replay-production/.
Tip: 8fdb992c7d40239f8a5b7602fa6f4636f3d97961.

Validation

python3.12 -I -B research/joint-dual-replay-production/verify.py
python3.12 -I -O -B research/joint-dual-replay-production/verify.py
python3.12 -B scripts/research_package.py --package research/joint-dual-replay-production
python3.12 -I -B research/joint-dual-replay-production/benchmark.py --rounds 5

Expected:

PASS_JOINT_DUAL_REPLAY_PRODUCTION words=[23, 25] outputs=12261 receipts=2 no_new_kappa
PASS_JOINT_DUAL_REPLAY_PRODUCTION_O_POLICY refused_by_assert_gate=2 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/joint-dual-replay-production outcome=completed
PASS_IDENTICAL_JOINT_DUAL_REPLAY_PRODUCTION_RECEIPTS

Default verify is read-only. Historical checks use assert and deliberately refuse to load under -O (same production policy as before). 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): Aaron Wang’s research/joint-dual-replay-fast/ / #1880 (opt-in containment masks; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2194 / #2210 / #2231 (same production-promotion pattern on other objects). No competing open PR on production binary_frame_replay.replay at refresh.

Chafik Boukhalfa authored the joint-dual / binary-frame replay verifier path. Aaron Wang authored the containment-mask algorithm. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.