[JarryShaw]

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort) — not run
• make test passes, and a test case covers the change — tests/project only: 397 passed, 1 skipped
• Added a changelog entry — N/A, added centrally

What is the purpose of your pull request?

• ci — workflows or build tooling

---

Closes #1387

Guard (fails the step). The test step now logs every measured process with its parent (COVERAGE_DEBUG=pid,process). Before combine, Report coverage fails and posts no total unless the controller and each of its direct children (the xdist workers) left a .coverage.* file. The set is read from the run's own log, so -n auto needs no hard-coded count. A missing child of a worker is not flagged: the 20 duplicate-hash "skipped" files are such children. If the log names no worker, the check also fails, so it cannot pass while blind. Why fail rather than warn: coverage-comment.yml accepts only four tests values, so a warning marker would need an edit there. With no artifact, it already posts a notice and exits 0.

Mechanism: still not traced. I ran 3 local -n 4 runs (2 pinned to 4 CPUs), and every worker saved each time (47 files). The one candidate I measured is mostly ruled out: xdist SIGKILLs a worker 10 s after gw.exit(), but workers saved 2–3 ms after teardown began, and each save took about 0.3 s. Adding patch = ["_exit"] covers the documented os._exit paths: execnet's shutdown fallback (gateway_base.py:1253) and pytest-timeout's thread method. sigterm = true was not added, because nothing in the chain sends SIGTERM, and coverage.py's docs warn it can cause problems (issue 1310).

Tests. New tests/project/test_coverage_data_guard.py (8 tests) runs the extracted guard on synthetic logs and on a real coverage run log. All 8 pass. Mutation checks: the unpadded child regex, the round-1 guard on a missing log, and a warn-only guard each fail their test. The real-log test caught two parse bugs (New process: cmd: and space-padded pids). The guard also passed on all 3 real local runs. The citation shift (+25 lines) is in workflows.rst/releasing.rst.

[JarryShaw]

Cross-review verdict on 2958401f7: NEEDS CHANGES. One test is flaky (Sonnet reviewed; author Opus).

• The flaky test. The regex in tests/project/test_coverage_data_guard.py:142, ^(\d+)\.\w+: New process: cmd:, does not allow for the leading spaces coverage adds to pids under 5 digits. The fixture at :61 writes {pid:5d}, and the workflow guard's own regex (' *(\d+)') handles them. With child pids 9832 and 9836, re.search returned None and the test died with a TypeError. It will fail on any runner whose pids are below 10000. Fix: ^ *(\d+)\..
• Guard on a real run. It passes a real xdist run: -n 4 on tests/project gives 395 passed, 1 skipped, the controller plus 4 workers, and nothing missing. Workers that run zero tests still save a file. A killed worker fails the guard, which is correct. A log with no workers exits 1.
• patch = ["_exit"]. It is supported from coverage 7.10.0, which matches the CI pin coverage>=7.10. An unknown key raises ConfigError, so a typo fails loudly. The execnet and pytest-timeout paths are UNVERIFIED.
• Logging env. With it set, tests/project gives 395/1 skipped, identical to a run without it. The author's local failures did not reproduce.
• Permissions and docs. No write permission is added. All the shifted line citations in workflows.rst and releasing.rst match.

Notes, not blocking:

• The revert proof fails on the missing heredoc, which is weak evidence.
• A missing coverage-processes.log gives a traceback rather than ::error::.
• A guard failure on the 3.14 leg now blocks Required checks passed. Before this PR it only under-reported coverage.

[JarryShaw]

Round-2 verdict on 8447c8156: GOOD TO GO (Sonnet reviewed; author Opus). All three round-1 findings are fixed.

• Padded pids. Every lookup now goes through the CHILD regex (^ *(\d+)\.). Coverage really does pad the pid, as f"{os.getpid():5d}…" at coverage/debug.py:320 shows. CHILD matches 7, 123 and 12345 with or without padding, while the old regex returned None on the padded line.
• Missing log. A missing coverage-processes.log now gives ::error::, adds the summary line and exits 1, with no traceback.
• Revert proof. This is now strong. The reviewer wrote their own mutants:
  • a warn-only guard fails 4 tests;
  • the old regex errors the padded-pid test;
  • dropping the exists check fails 1 test;
  • dropping or missing fails 2 tests.
• Citations. All 16 shifted unit-tests.yml line citations match the new file, and no stale citation remains anywhere in docs/, .github/ or util/.
• Test runs. The guard module passes 8, and tests/project gives 397 passed, 1 skipped. I re-ran the guard module myself at this head and got 8/0.

Nits, not blocking:

• A log that exists but cannot be read still raises before the ::error:: line. The step still fails, so only the message is worse.
• open(...).readlines() never closes the file handle.

[github-actions]

Coverage: 89.40% (unit tier, Python 3.14, bc7e19685, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	4	93.55%
pcapkit/const	18791	968	2352	869	90.90%
pcapkit/corekit	2018	88	628	24	94.78%
pcapkit/dumpkit	234	2	84	2	98.74%
pcapkit/foundation	2573	99	910	44	94.63%
pcapkit/interface	112	7	40	5	92.11%
pcapkit/protocols	16262	203	4296	165	98.17%
pcapkit/toolkit	539	74	168	3	84.58%
pcapkit/utilities	429	4	122	4	98.55%
pcapkit/vendor	4409	2342	1006	157	43.25%

Per-file detail: the coverage-html artifact of this run.

[JarryShaw]

The red Python 3.x legs are main's, not this diff. test_protocol_make_unit.py::TestMakeRawItems::test_unparsable_raw_item_raises_protocol_error also fails on main at 66950b1f2, where I reproduced it locally. It is a semantic conflict between #1344 and #1352, tracked in #1394, and a fix is in progress. This PR needs a re-run once #1394 lands.