Tooling: 1.71x faster exhaustive data-corner replay with identical receipts

[just-mn]

Change and mathematical scope

The historical research/copied-fixed/data_corners.cpp checker repeatedly scans the same eligible columns for each lower row. This opt-in runner caches that list once per pivot, preserving update order, all existing assertions, non-natural pivot columns, and the complete output receipt. Only research/data-corner-performance/ is added; existing sources, default targets, certificates and upstream/ are unchanged. No new κ, construction, or theorem claim.

Three alternating original/optimized pairs on Apple clang 21.0.0 / macOS arm64 gave median wall times 169.198 s → 98.954 s (1.710×; 41.5% less time). All six fresh processes covered all 4,073,300 triple pairs and produced byte-identical stdout and stderr. The receipt retains 191,444,880 nonzero-pivot checks, 1,409,360,440 ordered-zero checks, 1,771 product records, and the same ten modular failures. Raw samples, hashes, source pins and the loop-equivalence argument are included. These timings concern this historical finite checker only.

Validation

python3 -B research/data-corner-performance/replay.py --benchmark 3 --output /tmp/data-corner-result.json
python3 -B research/data-corner-performance/test_replay.py -v
python3 -O -B research/data-corner-performance/test_replay.py -v

Use a fresh output path. Expected replay marker: PASS_DATA_CORNER_REPLAY pairs=4073300 complete_receipts_equal=true. Seven test groups pass in both Python modes, including independent determinant checks, complete-matrix comparisons, the left-of-pivot mutation, reordered-update positive control, every failure position, corrupted receipts, source drift, and rejection of NDEBUG.

Recorded base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. Both execution input blobs remain identical at refreshed upstream f567884aeb8851b6e794af1184119a3bfbe1e631. Published files were fetched back and verified byte-for-byte against the tested local package.

Not rerun: separate historical rational recovery, whole-repository make verify, selected supplier cold replay, Lean, or remote CI. The rational recovery is unchanged and excluded from the timing. All inherited all-size/compiler/common-label/hosting/residual/precision/tape/analytic assumptions remain.

OS-random seed 86 selected verification engineering; CLAIM.json records the three candidate targets. The final overlap sweep covered 693 open PRs through #902, with no matching optimization found. PR #154 is credited as the profiling lead; its code is not imported.

Prepared for Denis (@just-mn) with substantial OpenAI Codex assistance in implementation, experiments, documentation and a separate scoped AI review; no independent human review is claimed. NOTICE.md preserves Rohan Arun / PR39, James Chang / PR34, Douglas Colkitt, eumemic / PR154, and the broader predecessor credits.