ci: bound unittest-leg reimport growth and add stall diagnostics (#1052)

[JarryShaw]

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort) -- test_isort_clean passes; pylint/mypy cover pcapkit/ only, untouched
• make test passes, and a test case covers the change -- new modules 11 + 3 passed (fail on main); project leg 466 OK; make test not run
• Changelog entry: N/A — centralised in docs(changelog): shared 1.5.0 changelog — long-lived, merges last (#610, #616, #617, #618, #620) #657

What is the purpose of your pull request?

• ci — workflows or build tooling

---

Description of your pull request and other information

Part of #1052 (A's faulthandler half, B, D). No workflow edits.

A. run_unittest_leg.py arms faulthandler.dump_traceback_later (--stall-dump, or $PCAPKIT_UNITTEST_STALL_DUMP; default 1080s, 0 off), cancelled on exit.

B, as a collect rather than a restore. Any sys.modules restore hides #981. On its reproduction (32bcfba15^, test_http_unit + test_base_class_contract): no restore gives 1 F + 3 E; restoring per module or per test gives 0 + 0; gc.collect() after each test keeps 1 F + 3 E. --gc-every N (default 10). protocols/internet measured locally (446 tests, all OK):

run	wall	peak RSS
main, and this branch without the collect (×3)	420 / 502 / 432 s	1008 / 1650 / 781 MiB
--gc-every 10 (default)	420 s	491 MiB
--gc-every 1	500 s	320 MiB

The 1650 MiB run reproduced the slow tail: test_mh_unit tests at up to 7.3 s instead of 1.2-1.8 s.

D. pytest-timeout in test; timeout = 900, timeout_method = "thread". The slowest test measured was 21 s, and 15× that is 315 s. signal is out because its pending alarm fails TimeLimitTests.test_nothing_is_re_armed_when_nothing_was_pending. Every pytest job already installs .[test…]. On its own, thread killed every test runner at 13% (e8db64a9c). pytest-timeout keeps each item's Timer alive, and each Timer holds on to the tbtrim excepthooks that one import pcapkit installed, which pins that generation. tests/conftest.py now drops the timer, and tests/const under -n 4 peaks at 1979 MiB (9179 MiB and killed before the fix, 1939 MiB without the plugin).

[JarryShaw]

Cross-review verdict on 87c229eda: GOOD TO GO (ran on Sonnet; author Opus)

• Root cause reproduced. test_const_registry_protocol.py with pytest-timeout 2.4.0 in thread mode:

  Run	Result	Peak RSS
  With the hook	passed	281 MiB
  Hook neutralised	MemoryError after 66 tests	7237 MiB
  Plugin off	passed	283 MiB

  pytest_timeout_set_timer stores item.cancel_timeout and never clears it, and pytest_timeout_cancel_timer is a real hookspec.

• Timeouts still fire. A 5 s sleep under timeout = 2 fails with a stack dump. Earlier timers are cancelled.

• test: three modules pass alone but fail together, and pytest reports it green #981 detection survives --gc-every 0, 1 and 10. It still reports 1 failure and 3 errors on the 32bcfba15^ repro.

• Signal mode is correctly excluded. It fails test_nothing_is_re_armed_when_nothing_was_pending with 900 != 0.

• New tests fail when the change is reverted: all 14 of them, under nine separate mutations.

• CI: the Python 3.10–3.14 legs take 14m51s to 18m55s.

Non-blocking nits:

• PCAPKIT_UNITTEST_STALL_DUMP set to nan or a negative value silently disables the dump, and inf raises OverflowError.
• The "just under a 20-minute step cap" comment is stale: that step has no cap of its own; only the job's 45-minute cap applies.
• Under xdist, a hung test crashes its worker. It is still reported by name, but without the stack dump.