Tooling: 3.57x faster complete T74 native replay with unchanged exact coverage

[tekuuu]

The complete PR2751 T74 native replay spends most of its time compiling exact Clifford states. This opt-in runner integrates Charles Bethin's PR662 gate tables and Ali Romman's PR1029 sparse evaluator into the two Clifford-using stages while retaining the unchanged independent warp checks and every original native control.

Tooling improvement; no new κ. Three alternating pairs on the same base word measured a 3.57× complete-native speedup, median 306.7s → 85.9s (72% less wall time). All eleven native artifacts matched in every pair after only documented elapsed-time and verified gzip-container normalization; phase recipes were byte-identical. This is a measured native-verification improvement, not a whole-project or multiplication-runtime claim.

Validation and scope:

• Fresh complete base and extra provider replays; the extra runs from a Git-free copy in a path with spaces.
• All 84 inherited negative controls retained (46 local, 25 wrapper, 13 recipe), including the independent all 512-column warp checks.
• 1,488 dense/sparse exact-state comparisons plus 18 new boundary controls for altered source, fixture, word, recipe, receipt, optimized Python and output reuse.
• Exact source closure, both fixtures, complete six-sample benchmark and reference digests are in research/t74-native-fast/. No saved native result is reused. The full supplier/hosting/outer admission remains separate.

Reproduce with Python 3.12 and numpy 2.2.6:

python3.12 scripts/research_package.py --package research/t74-native-fast

The scientific base is 2d3af1ba7e28b74248b47381768e5b116078ba1c; the branch was synchronized to canonical main 936fac15b1b1b59dc5285c713a77f9f3c07b2906 before submission. Native sources and both actual input words are pinned through PR2751 4cfbb0cc8314d7fb4e8e754545a49a42eb590a69 and PR2679. Relevant sources, priorities and selected record were unchanged by the refresh.

This contributes current T74 integration and complete-pipeline evidence, not a new sparse algorithm. PR1041's older U8 runner and Victor Li's related PR2329 multicold production integration are credited. PR2226's borrowing operators, the Rohan Arun/icekylinx/DreamingOfClouds native lineage, Philippe Van Velzen's source snapshot and the complete PR2751 predecessor notices are retained. Frozen sources and the selected frontier are unchanged.

Coordination: discussion #2755. Final overlap search covered all 710 open PRs, the earlier complete discussion snapshot and a fresh 40-discussion fallback; limits are recorded in FRESHNESS.json.

Prepared by Tekalign (@tekuuu) with substantial OpenAI Codex assistance. Apache-2.0. Author-checked evidence; independent human review remains welcome.

Submission-time follow-up: PR2766 (e6df920cd0110e4b629e2ad84eddcb539001e174, David Arias) and PR2767 (e8b771ba132fe632796b28c1742a8d096c04519f, Jeremy Kothe) appeared during final verification with conditional κ claims around0.001717256973467. PR2766 explicitly distinguishes this native-acceleration scope. Those newer constructions were not audited in this tooling round: our measured timings and equality receipts remain tied to the two exact PR2751 words, and do not assert successor-word performance.