[JarryShaw]

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort)
  • Note: isort -l100 -ppcapkit --check-only tests/_support.py flags the import block; left as is.
• make test passes, and a test case covers the change
  • Note: only tests/foundation/test_extraction_ownership.py was run; make test was not.
• Added a changelog entry under docs/source/changelog/ and regenerated CHANGELOG.md, if the change is user-visible
  • Note: N/A, centralised.

What is the purpose of your pull request?

• fix — corrects a defect

---

Description of your pull request and other information

collect_until ended on the first collection once its grace ran out. PyPy 7.3.19 ran an abandoned extractor's finaliser only after a second collection (observed), so a slow first collection may have left the handle reported as still open (#2481). The cause is unverified; the fix matches the observed two-collection behaviour on PyPy. With a grace, collect_until cannot give up before MIN_COLLECTIONS_WITH_GRACE (2) collections. grace=0 is unchanged, as its existing test pins it.

Regression test: fake clock, each collection costs 2 s against a 1 s grace. It fails with the floor set to 1 and passes at 2.

Verified: tests/foundation/test_extraction_ownership.py 20/20 on CPython 3.14 and PyPy 3.10 (7.3.19). The CI failure did not reproduce in isolation, so the full-suite cause is UNVERIFIED.

Refs #2481

[JarryShaw]

NEEDS CHANGES (Sonnet, 7025040c2). Fix is sound on CPython: the new test fails on the old helper, passes on the new one, and grace=0 is unchanged across callers.

• tests/_support.py collect_until docstring: "after the first collection" and "done() held before the grace ran out" are false on the new path; document the two-collection floor. Rename local collections (~L1206), which shadows collections.abc.
• PR body: "make isort already fails on main" is false (Makefile covers pcapkit, util, examples only). Reword. Changelog box: "N/A, centralised".
• "measured" for two collections cites KTLO pypy3.10: finaliser backstop test fails on main #2481, which records no measurement. Say how it was measured or drop "measured".

[JarryShaw]

NEEDS CHANGES (Sonnet, 1780a9b4f). Shadowing, isort wording and the PyPy-two-collections claim are verified.

• tests/_support.py: "at least MIN_COLLECTIONS_WITH_GRACE collections run" and the #: comment are false when done() holds after one collection. Say "cannot give up before N collections". Same fix in the body and commit message; the FINALISER_GRACE comment's "after the first gc.collect" contradicts the docstring.
• PR body: "make test passes" is ticked with one module run; the changelog box is ticked with no entry. Leave both unticked with a note.
• "a slow first collection reported the handle as still open" is stated as fact; the body calls the cause UNVERIFIED. Hedge it, and use Refs #2481 instead of Closes.

[JarryShaw]

NEEDS CHANGES (Sonnet, e78e54b41). The fix holds on CPython and PyPy; the wording is now true on the code.

• tests/_support.py FINALISER_GRACE block is 12 lines; trim to 8 or fewer. Its "(ci: categorise the tests into tiers and run each tier where it belongs #1958, measured on PyPy 8.0.1)" cites a run that is not this one; drop the measurement claim.
• Changelog note: "no changelog entry added" should read "N/A, centralised".