fix: guard FDCapture.snap() against prematurely closed tmpfile

[RonnyPfannschmidt]

Summary

Fixes #14528 — ValueError: I/O operation on closed file during capture teardown on Windows + Python 3.14.3.

Guards FDCapture.snap() and FDCaptureBinary.snap() against a prematurely closed tmpfile, returning EMPTY_BUFFER and emitting a PytestWarning instead of crashing. The captured data is unrecoverable regardless, so crashing only makes things worse.

Reproduction attempts

The issue was reported on Windows 11 + Python 3.14.3 + pytest 9.0.3 with pytest-asyncio, anyio, and pytest-httpx. Collection short-circuits inside a dot-prefix path (.claude/skills/actionmail/tests/) with import errors, and the capture tmpfile is already closed by the time snap() runs.

We attempted to reproduce on Windows CI with:

• Python 3.14.3 (incremental GC active, gc.get_threshold() = (2000, 10, 0))
• The reporter's exact plugin set (pytest-asyncio==1.3.0, anyio==4.13.0, pytest-httpx==0.36.2)
• An identical dot-prefix tree layout with __init__.py, two test files with sys.path.insert + failed import run, and a fixtures/ sibling

The issue did not reproduce. All EncodedFile.close() calls traced to explicit teardown paths (pytest_keyboard_interrupt → stop_global_capturing → done → tmpfile.close()). No GC-triggered file closure was observed. The trigger appears to depend on additional environmental factors specific to the reporter's machine.

Notably, Python 3.14.5 (which reverts the incremental GC) also showed no issue, consistent with the reporter's observation that the problem is 3.14-specific.

Changes

• src/_pytest/capture.py: Guard snap() in both FDCapture and FDCaptureBinary — check self.tmpfile.closed before seeking, emit PytestWarning, return EMPTY_BUFFER
• testing/test_capture.py: Parametrized regression test using os.pipe() for isolation
• changelog/14528.bugfix.rst: Changelog entry

Test plan

• Regression test TestFDCaptureClosedTmpfile — deterministically closes tmpfile before snap(), asserts EMPTY_BUFFER and PytestWarning
• Existing capture tests pass
• Windows CI with Python 3.14.3 + reporter's plugins — no crash (exit code 2)
• Pre-commit clean

[Copilot]

Pull request overview

Guards FDCapture.snap() and FDCaptureBinary.snap() against the case where the underlying tmpfile has been closed externally (notably by Python 3.14.0–3.14.4's incremental GC finalizing the TemporaryFile before pytest's capture teardown runs). Instead of raising ValueError: I/O operation on closed file during _ensure_unconfigure → stop_global_capturing, snap() now returns EMPTY_BUFFER, preserving the intended exit code (e.g., 5 for no-tests-collected) and avoiding confusing tracebacks in test output.

Changes:

• Add early-return on self.tmpfile.closed in both FDCaptureBinary.snap() and FDCapture.snap().
• Add changelog entry for issue #14528.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/_pytest/capture.py	Returns empty buffer from snap() when tmpfile is already closed, avoiding the seek(0) crash.
changelog/14528.bugfix.rst	New changelog entry describing the fix.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[RonnyPfannschmidt]

running another expeirment first - im of the impression the initial explanation doesnt fit and only coincides