[JarryShaw]

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort)
• make test passes, and a test case covers the change
• 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?

• test — tests only

---

Description of your pull request and other information

On GraalPy, test_finaliser_alone_releases_an_abandoned_handle failed on main's first full GraalPy run (job 118912680486).

GraalPy's gc.collect forces a Java GC, sleeps briefly, and then runs only the weak-reference callbacks that are queued by that point (GcModuleBuiltins.javaCollect, graal-25.4.4). Under load, a weakref.finalize can therefore run after collect returns.

A new collect_until helper keeps collecting for up to FINALISER_GRACE: 10 s on GraalPy, 0 on CPython. A finaliser that never runs still fails. A stand-in finaliser that runs only on the third collection shows the helper waits for it. I could not reproduce the failure locally, so the cause is inferred from GraalPy's source.

The run's nine other FAILED ids were worker crashes, not test failures. Each worker hit --vm.Xmx2g with ExitOnOutOfMemoryError after its heap grew across tests. Each of those tests passes alone on GraalPy, so gating them would be wrong. The fix for the crashes goes to the GraalPy leg separately.

Overlaps #2000 in test_extraction_ownership.py; I'll rebase whichever lands second.

Part of #1963.

[JarryShaw]

GOOD TO GO at c33e6b797. A Sonnet review of 190b48930 checked GraalPy's GcModuleBuiltins.javaCollect at graal-25.4.4: it calls forceFullGC(), sleeps 15 ms, then polls only the reference queue as it stands. That is the race this PR describes.

• FINALISER_GRACE keys on sys.implementation.name, and at 0 on CPython a late finaliser still fails.
• The loop is bounded by time.monotonic().
• The third-collection stand-in shows the helper waits.

c33e6b797 applies the same helper to test_extractor_abandoned_before_eof_still_releases_its_handle, the review's one gap; the diff is that one line, and I checked it myself. The file passes on CPython (16) and on GraalPy 25.4.4 with the leg's heap flags (16).

[github-actions]

Coverage: 91.54% (unit tier, Python 3.15, c33e6b797, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	2	95.16%
pcapkit/const	19597	881	2618	867	92.11%
pcapkit/corekit	2540	71	808	37	96.54%
pcapkit/dumpkit	276	3	100	3	98.40%
pcapkit/foundation	2929	23	1020	25	98.78%
pcapkit/interface	138	4	52	3	96.32%
pcapkit/protocols	17468	222	4600	187	98.13%
pcapkit/toolkit	872	36	298	10	95.21%
pcapkit/utilities	523	2	158	4	99.12%
pcapkit/vendor	4431	1995	1018	167	51.73%

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