ci: cut queue pressure and diagnose slow legs #1052

CI is queue-bound. Peak concurrency sits at exactly 20 jobs (5 on macOS), which matches GitHub Free's per-account cap. One PR push starts about 69 jobs, about 61 of them in unit-tests.yml. Every timeout in the last two weeks was a throughput problem, not a hang:

• the plain-unittest protocols leg slowing 10–15× in the tail of test_mh_unit.py (purge_modules without a module-table restore);
• apt-get update stalls on 2026-10-01;
• the old 30-minute cap.

A merge queue was probed and rejected for this user-owned repo (422 Invalid rule 'merge_queue').

Approved changes:

• A. Run the unittest leg with --verbose, and add faulthandler.dump_traceback_later to it.
• B. util/run_unittest_leg.py snapshots and restores sys.modules around each test.
• C. A step timeout of about 20 min on the unittest run step; apt capped at about 5 min, with one retry.
• D. pytest-timeout for the pytest jobs.
• E. Fold the 36 engine jobs into about 6, with no loss of coverage.
• F. Create Release skips its full-suite gate when nothing is publishable.
• G. Docs-only PRs run only tests/project, the changelog check and the Pages build. required-checks accepts skipped legs only in that case.
• H. unittest-ordering runs on main pushes and on schedule only.

Kept on purpose: the full run on main pushes. Merges land trees that no PR run tested, so that run is the only test of what lands.