joint-dual-replay-fast: opt-in adapter removing the joint-dual replay's per-output triple scans (1.74x replay stage, identical checks)

[navinor]

joint-dual-replay-fast: opt-in adapter removing the joint-dual replay's per-output triple scans (1.74x on the replay stage, identical checks)

Aaron Wang, with help from opencode. Verifier engineering; no kappa, no certificate, supplier or selected-frontier change. Apache-2.0.

Scope

• Card: replay-runtime (scope key b1dd41940c60d240b47a; chooser OS nonce 24ae43237c3eaed2410bcdf516d854cd; priority revision 3; base 0b1cccf86e2f5132ed286a26ef0a681a495740ea; selected-record SHA-256 10b9e340fd8e354a37a0c4813c90e5de1dd42044b5f6aca4a3fc5a42b685fa32).
• Card selection note (pair-collision). My first-choice drawn card portable-replay (scope key 831cc1a3c8f9ff01b197) matched the fresh claim posted 2026-10-14T11:24Z in Handoffs and requests for independent checks #1166 (completed as Repair locale-dependent reads in test_research_current_status (no new kappa) #1834), so per the collision rule I took my second-choice card. Recorded as a card-selection change in research/joint-dual-replay-fast/TASK.json.
• Object (scope refined). One measured replay bottleneck: the per-output full-triple scans in scripts/experiments/binary_frame_replay.py:replay() on the pinned words certificates/joint-dual-word-{23,25}.json.gz (frozen chain research/joint-dual, source freeze research/joint-dual/SOURCE.json, enforced by joint_dual_compose.check_sources()). This is the independent word check of the PR 25-PR 58 frame-compiler line and the dominant cost of its longest isolated test module (tests/test_joint_reclaim.py). Verification throughput only.

The measured bottleneck

cProfile of replay() on the h=23 word (CPython 3.12, Windows): 2.045 s total, 0.84 s in the per-output full-triple scan expected=sum(1<<i for i,t in enumerate(triples) if common in t and not(set(t)&(set(triple)-{common}))) and tt=[i for i,t in enumerate(triples) if common in t] (1,020,119 generator iterations at that line; 5,336 outputs x 1,771 triples).

What the package does

research/joint-dual-replay-fast/fast_replay.py re-executes the historical replay body with every assertion, both full-basis orientations and the receipt identical, replacing only those scans with containment masks over the same combinations(range(h), 3) enumeration: expected(common, triple) = center_mask[common] & ~pair_mask[common][x] for x in set(triple)-{common}, which is exactly the literal condition. check_word (including PR 64/rfu08's cancelling-scatter control, retained verbatim) and the whole transitions stage are the frozen historical implementations, imported unchanged. Nothing frozen is modified: the historical scripts, certificates and tests/test_joint_reclaim.py keep their bytes and their pins — the opt-in adapter pattern of research/ck-dyadic-replay-fast.

Evidence (all in the package; RESULT.json is the recorded receipt)

• Identical receipts: historical vs adapter replay receipts equal on both pinned words and equal to the recorded receipts in certificates/joint-dual-compiler.json; the delegated transitions stage equals certificates/joint-dual-transitions-{23,25}.json.
• Full cross-check: on all 12,261 outputs of both words the mask algebra equals an independently written literal scan.
• Oracle matrix: the exact twelve mutations of the frozen tests/test_joint_reclaim.py (11 rejections + scatter-reversal acceptance) on both words, both checkers: 24 cases, every accept/reject outcome agrees with the historical module, and the documented rejections reject.
• Failure controls: a corrupted containment table is detected by the cross-check (40/44 mismatching outputs at h=23/25); an unchecked word rejects; -O entry refuses ("Assertions must remain enabled"); source-freeze and word pins re-verified every run.
• Adoption simulation: the unmodified frozen test module runs against the adapter through shim/binary_frame_replay.py and passes all 16 tests (verify.py runs this as a load-bearing control).

Before/after wall time (same fixture; timings are measurements, never a correctness gate)

Paired alternating runs, CPython 3.12, Windows (BENCHMARK.json):

Fixture	Historical median	Adapter median	Speedup
replay() on both pinned words (5 pairs)	2.233 s	1.285 s	1.74x
frozen tests/test_joint_reclaim.py, unmodified (3 pairs)	41.47 s	37.72 s	1.10x

The unit fixture clears the stated target of at least 25% (42% reduction). The suite fixture improves only 10% because most of that module is JSON deep-copies, early rejection paths and the unchanged transitions stage; no coverage was dropped to improve either number.

Retained checks: check_word, every replay assertion, both full-basis orientations, the unchanged transitions stage, and the frozen test module as oracle. Invalidation: containment tables are rebuilt from the word on every call — no persistent cache, no staleness; freeze/pin bindings are re-checked each run.

Reproduce

python3.12 -B scripts/research_package.py --package research/joint-dual-replay-fast --check-only
python3.12 scripts/research_package.py --package research/joint-dual-replay-fast
python3.12 -I -B research/joint-dual-replay-fast/benchmark.py

Expected: PASS_PACKAGE_METADATA_AND_PINS, PASS_JOINT_DUAL_REPLAY_FAST_EQUIVALENCE then PASS_REGISTERED_RESEARCH_PACKAGE research/joint-dual-replay-fast outcome=completed, and paired timings in BENCHMARK.json (values vary by machine). The documented gate python3 -B -m unittest discover -s tests -p 'test_research_*.py' passes 42 tests at base and after this change (this PR adds only research/ content).

Limitations / remaining obligations

• Opt-in only: the repository-wide suite wall time benefits only if a maintainer later adopts the adapter, which would require superseding the research/joint-dual/SOURCE.json freeze. This PR does not do that.
• Domain is exactly the two frozen words at h in {23, 25}; the mask identity is unit-tested on synthetic shapes up to h=12 but no other word is claimed. Not checked: the C++ profile stage of verify_joint_dual.py, make frontier1094-verify and any cold replay of the selected construction (not affected by this change).
• No mathematical claim of any kind.

Overlap

Entry and submission-time scripts/find_overlap.py refreshes (complete pagination: 519/537 open PRs, 489/521 discussion records, 2026-10-14T12:06Z and 13:05Z) found no work on this object. Distinct from research/ck-dyadic-replay-fast (CK dyadic row kernels), research/git-provenance-stream (research_package git batching), research/omission-lanes-seed92, research/rational-containment-performance (different replays) and discussion 1547's gzip full-payload adapter (different artifact).

Credit

• Chafik Boukhalfa's frame-compiler line (PR 25 by Rohan Arun (rohanarun, with OpenAI Codex assistance) through PR 58 by Chafik Boukhalfa (chafreaky, with OpenAI Codex assistance); scripts/experiments/pin_joint_dual_sources.py, research/joint-dual) owns the frozen replay, transitions stage, words and source freeze; PR 64 (rfu08) is credited in check_word for the cancelling-scatter counterexample, retained verbatim.
• Workflow precedent, credited not modified: research/ck-dyadic-replay-fast (opt-in adapter pattern; paired benchmark receipts) and research/git-provenance-stream (timings as measurements).
• Task selection, profiling, adapter, controls and write-up by Aaron Wang with help from opencode; substantial AI assistance is disclosed in TASK.json. No independent human review claimed.

Claim: #1166 (comment) (released with this PR).

[Mister-Meeseeks]

Batch review at f4e150e4222ce4b4418e66874c6363c43987c8d1.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Opt-in containment-mask adapter preserves both frozen full-basis replay orientations and receipts; full historical16-test suite,24 mutation cases, exhaustive small shapes and assertion-disabled entry rejection pass. Missing historical PR58 fixture restored from exact head. Benchmark speedup remains author-measured, not selected-supplier adoption.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.