ci(timing): the Python 3.15 leg's p90 passes half its 45-minute cap, and the unsplit project ordering name fires the step check #2571

Problem. Run 41303273918 (73add05a3, 2026-10-25 to 2026-11-01) failed two hard checks; artifact: https://github.com/JarryShaw/PyPCAPKit/actions/runs/41303273918/artifacts/13083455648

• H1: Python 3.15 p90 23.2 min, 52% of its 45-minute cap; Run unit tests takes 3.54x its paired legs (W3, n=227).
• H2: Plain unittest ordering (project) step p90 12.7 min, 64% of its 20-minute step cap (W1: median 9.9 to 12.6).
• H2 is an artifact. ci(ordering): shard the project ordering leg and cache the tier guard's workflow parse (#2254) #2258 split this cell on 2026-10-26T18:14Z, so the unsplit name holds only the 47 samples before it; the halves are separate keys (analyse keys on the API job name, nothing pools) at p90 7.7 and 5.9 min. The 2026-11-02 report spans the split as well (19 pushes, step p90 12.9) and will likely fire H2.
• H1 is real. 3.15 runs the whole suite under coverage: modules outside the ALL tier are 56% of its test time, and the shared ones run 1.7x slower than on 3.14. It ran 2 xdist workers on 4 CPUs (-n auto counts physical cores once psutil is installed). Daily medians for 2026-10-30 to 2026-11-01 are 22.5 to 24.7 min (n=2-5), against weekly medians of 19.6 and 20.6; it finishes last on 63% of PR runs.
• Options. H2: (a) wait for the window to clear the split, 2026-11-02T18:14Z; (b) make evaluate skip a key with no sample in the window's last day. H1: (c) -n logical on the CPython unit legs; (d) cache or trim the heaviest modules, as ci(ordering): shard the project ordering leg and cache the tier guard's workflow parse (#2254) #2258 did for the tier guard; (e) add 3.15 to JUNIT_LEGS, so the report names its modules; (f) split the leg (ruleset edit, one more runner); (g) raise the cap.
• Recommendation. H2: (a), no code. H1: (c) and (d) in one ci PR with before and after Run unit tests minutes, plus (e); refresh the 8.1-minute and 1.8x figures in the test job's matrix comment (conventions/process, "CI Timing Review"). Keep the 45-minute cap: the maximum is 25.4 min. (f) adds a job to a queue whose p90 is 34.9 min on PRs.