Take the best of 15 timed pairs in the tapscript scaling ratio tests (issue #2613)

[fametrano]

This pull request cites ISS 2613 and does not close it. Nobody has measured the failure rate at a load of 20 or more, where the issue saw the ratio reach 12.0. The issue stays open for that measurement.

The two ratio tests in tests/script_engine/tapscript_scaling_test.py now share one helper, _ratio. It times 15 interleaved small and large runs instead of 5, and takes the best of each. The limit stays at 10, against 16 for a quadratic implementation. Load only adds time, and it is likelier to hit the longer run. A ratio of 12 from a quiet value of about 4 therefore needs every large run stretched. The best of 15 is far less likely to be stretched than the best of 5.

Measured at loads 2 to 11: the ratios were about 4, with extremes 2.3 and 6.0 and no failure. A reviewer planted a quadratic signature-check cost, and the test failed at 13.7. A CPU-time clock was rejected, because Windows' 15.6 ms granularity is coarser than the small run.

Gates on bd9e0ae, whose tree is this sha's (f2b5fa5 changes the message only), run by the writer under the gate lock:

• uv run pytest: exit 0, coverage 100%
• uv run pre-commit run --all-files: exit 0
• sphinx-build -n -W -b html docs/source docs/build/html: exit 0

The local review cleared f2b5fa5, the branch's only commit on b71525a. No rebase was needed.

🤖 Generated with Claude Code

[claude]

ACK f2b5fa5

No blocking or non-blocking findings. One question is left inline, about the 300 s timeout now covering 15 timed pairs. It does not hold the ack.

[pmazzocchi]

A fresh local review cleared f2b5fa5. Nothing blocks it. Four points:

1. Mutation cost. .github/mutation/engine.toml:31 and sig_hash.toml:43 run all of tests/script_engine for every mutant, under a 90-minute budget. On this head CI took 19.6 s and 21.3 s for the two ratio tests; 5-pair branches took 3.3–7.0 s. So each session reaches fewer mutants. One option is to deselect the ratio tests in those two test-commands. That is a design choice, so it is @fametrano's to make. This is an estimate from CI logs, not a measured session.
2. Commit message: the 2.3–6.0 extremes are given "at load 5" there, "at loads 2 to 11" in the body, and a load of 13 in the thread. Make them agree in the squash.
3. Commit message: "over 50 runs" is a count. It belongs in the body, not in what lands.
4. Body: ISS 2613 is not a link.

[pmazzocchi]

Claimed: a /btclib-pr session (pmazzocchi) is rebasing this onto main (bbb0555), running the gates and taking it to the decision. The branch holds no other change.

[claude]

ACK 42cc95d

• non-blocking, tests/script_engine/tapscript_scaling_test.py:151,161: the ratio tests now make 15 timed pairs instead of 5, but their timeout is still _BOUND * 5 (300 s). That leaves about 16 s for a linear large run, down from about 48 s. The single-run tests allow 60 s. So under load the timeout can now fail a linear implementation. Known from the decorators and the loop count in _ratio.

[pmazzocchi]

Approves head 42cc95d.

• Relies on a fresh local reviewer's CLEARED f2b5fa5b6. The rebase onto bbb0555 is a range-diff =, so the clearance carries over. Gates rerun on 42cc95d: pre-commit, pytest (100% coverage) and sphinx -W all exit 0.
• Bot ACK: #2623 (review)
• Every check passed or was skipped on this head.
• Not checked: the Sign-off script was not run locally (CI's lint / Sign-off passed). The failure rate at load 20 or more is still unmeasured, as the body says.