Speed complete QQ16 replay with exact integer rows

[ywz65]

The complete frozen QQ16 caller repeatedly constructs Fractions while executing sparse scalar rows. This opt-in adapter shares one positive integer denominator per row, preserving every literal event, original guard and all 712 input columns. The original source reconstruction, scalar formulas, native/ordered checks and corruption controls still execute unchanged.

Three alternating fresh-process pairs on native Windows / CPython 3.12.7 gave median 26.82 s reference versus 20.63 s integer rows (1.30x; 23.1% less elapsed time). These descriptive timings include the same complete-caller instrumentation in both arms. All five original output files and stdout are byte-identical; all coefficients returned by all nine scalar invocations, including five corrupted words, have identical complete-map digests. The slower third reference run is retained in BENCHMARK.json.

The package proves the shared-denominator invariant, handles exact odd denominators and sparse zero-key/alias semantics, and includes 768 exact prefix comparisons, matching original error guards, changed-map/receipt controls and refusal of all 29 individually changed runtime sources. No mutable-input or acceptance-verdict cache is introduced. Source modules execute from authenticated bytes; the original Git-based archive read remains intact.

Sources: Ali Romman's #1969 at a6e6ad3936f67b8f63b5b68c5e75d540610a65e7, with its retained gauge44/T16 and QQ16 proof/verifier lineage. Construction base 6bd33bc0b87a01e9c9d8b1c01576f3ddbcaee850, priority revision 5. Presubmit main 21ec98bfd2e9f27810a8718d95d9b5d19a5df7d7 has all 37 relevant input/provenance/guidance/runner blobs unchanged. Latest scoped REST PR/comment checks found no competing scalar replay claim; complete Discussion coverage remains explicitly cached at 17:08 UTC. Canonical claim.

Validation on native CPython 3.12.7:

python3.12 scripts/research_package.py --package research/qq16-integer-row-replay
python3.12 -I -O -B research/qq16-integer-row-replay/benchmark.py --pairs 3 --out BENCHMARK.json

The registered primary and all three dependency/domain test groups pass. Independent proof/code review and a separate native CPython 3.12.7 optimized primary plus all three test groups passed against publication commit 18f0843079dd03f509a6724328ceba2fec325022; all ten committed package files are hash-bound in the retained review receipt.

This is opt-in verification engineering, not a new construction or supplier. No frozen file, selected result, hosting/movement/precision obligation or complete supplier invoice changes. Runtime still requires Git and the historical source commit; no universal speed/memory or Git-free full-caller claim is made. Prepared by ywz65 with substantial OpenAI Codex assistance; predecessor attribution and licenses are retained.