[squid-protocol]

Part of #5133

What

.test_durations regenerated from the real per-test seconds of main run 38901664312 (ea93351d8, after #5247): the three test-times-part-* artifacts merged, 21549 tests, none unrecorded (the old file lacked 180 and held 13 stale ones).

Predicted per-part sums (least_duration, replayed offline against the real seconds)

split	part 1	part 2	part 3
actual on main (balanced by the 2026-10-12 nightly cache)	906 s / 4.9 min	1278 s / 7.1 min	840 s / 4.8 min
old committed file, replayed	957	1070	999
refreshed file, replayed	1008	1008	1008

Replay used my own re-implementation of least_duration (sort descending, give each test to the lightest group); pytest-split is not installed in the box venv, so the real plugin's tie-breaking was not run. Wall scales by about 0.33 min per 100 s of test seconds, so the refreshed file predicts about 5.6 min per part, not the 7-8 min the brief gave as a target: there is no total left to spread, the suite is 3025 s of test seconds in all.

Ground truth (not fixable by this file)

mainframe-ground-truth.yml splits by explicit file lists, not durations. Summing this file over tests/cobol_mainframe/: part 1 (everything but two files, plus the ledger check) 1071 s, part 2 408 s (test_evidence_components.py has no entries here, so part 2 is understated). least_duration 2-way would be 740 / 740 s, about 7 min each. Matches 10.8 / 3.3 min wall.

Open items (workflow edits, not authorized here)

1. ground-truth: replace the matrix tests: lists with --splits 2 --group N --splitting-algorithm least_duration (and put the ledger step on the lighter part).
2. smoke-test full-suite-part balances by the newest test-durations-* cache (run 38247246921, 2026-10-12), not this file; the file is only the cache-miss fallback and the full-suite-gate input. The next nightly saves a fresh cache; until then the smoke parts stay at 4.9/7.1/4.8.

Sibling gaps

• full-suite-gate OS matrix reads the committed file: gets the refresh.
• Stale nightly durations cache feeding smoke (above).
• ground-truth: not duration-based at all (above).

Gates

pr_gates --fast 6/6 pass; --ratchets 8/8 pass; time_ratchet.py check (reads slow marks from source): 0 over the limit, 7 warned.

Lessons for agents

• time_ratchet.py check --times .test_durations skips the source slow-mark read and false-fails on module-scoped fixture cost; run it with no --times to check the committed file.
• Test-times artifacts are {"seconds": {...}, "slow": ...}, not pytest-split format: take ["seconds"].

🤖 Generated with Claude Code

[squid-protocol]

No description provided.

[squid-protocol]

No description provided.