Pass teardown exception to pytest_fixture_post_finalizer explicitly

[krishansubudhi]

Reworked per review: the hook now gets a teardown_exception argument instead of relying on sys.exc_info(). Also fixes #14800.

What changed and why

Up to pytest 8.0, pytest_fixture_post_finalizer was called from a finally: in FixtureDef.finish(), so a hook could inspect a failing teardown with sys.exc_info(). In 8.1, 434282e collected teardown errors into a list first. The hook then ran after the errors had already been caught, and sys.exc_info() became (None, None, None). Later, #14114 (719f1ec) made the hook the first registered finalizer so that a raising hook could not leave the fixture half torn down. That kept the hook out of reach of the active exception.

This PR calls the hook from finish() again. It runs in a finally around re-raising the collected teardown error(s), so the error (or the BaseExceptionGroup) is active while the hook runs. The cache reset moves into a nested finally, which keeps the #14114 guarantee: a raising hook still cannot stop cached_result from being cleared or the finalizers from being dropped. The existing cached_result is None early return still limits the hook to one call per teardown (#5848). The special-case lambda finalizer and its assert not self._finalizers in execute() are gone, so the source diff is net zero lines.

One behaviour difference is worth a reviewer's eye. If a fixture's teardown fails and the hook also raises, both errors are still reported, but they are chained ("During handling of the above exception, another exception occurred") instead of grouped in one BaseExceptionGroup. That is how 8.0 behaved. If you would rather keep the group, I can rework it.

test_fixture_post_finalizer_hook_exception from #14114 still passes unchanged.

How it was tested

• New test test_fixture_post_finalizer_sees_teardown_exception in testing/python/fixtures.py. It fails on main and passes with this change. It also checks that sys.exc_info() is empty when the teardown succeeds.
• Full suite on Python 3.11 (pytest -n 8): 4681 passed, 49 skipped, 13 xfailed, 7 xpassed. The one error came from my own -p no:cacheprovider flag, and testing/test_legacypath.py passes when run normally.
• pre-commit run --all-files is clean (ruff, ruff-format, mypy, codespell and the rest).
• Changelog entry changelog/12306.bugfix.rst; added myself to AUTHORS.

This PR was written with an AI coding agent and reviewed by me (Krishan); the commit carries a Co-authored-by trailer as the PR template asks.

[RonnyPfannschmidt]

🤖 Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Thanks for digging through the history here. The timeline in the description is accurate. I'd still like a different design, because this PR repeats the mistake that the earlier changes in this area made.

The hook should not depend on sys.exc_info(), and it should not need a particular call site to see the error. 434282e broke the implicit exc_info contract by moving the call, and #14275 made the hook a finalizer. Both changes were about where the hook runs, and both broke something: #12306, then #14800. This PR moves the call site again so that the implicit state lines up. The next refactor of finish() will break it the same way.

What I'd like instead:

• Add an explicit parameter to pytest_fixture_post_finalizer that carries the teardown error: the single exception, the BaseExceptionGroup, or None. pluggy allows hookimpls to accept a subset of the spec's arguments, so existing plugins keep working.
• Call the hook directly from FixtureDef.finish() after the finalizers have run. Drop the lambda finalizer and the assert not self._finalizers in execute(). This part of your change is good, and it also fixes #14800 (overlaps with #14801).
• An exception raised by the hook itself is a plugin bug and should fail hard, not be reported like a fixture teardown error (see my comment on #14114).
• Docs and the changelog should point users to the new argument. If sys.exc_info() still happens to work, we shouldn't present it as the supported way.

One concrete bug in the current diff: when a parametrized fixture is torn down because the next test needs a different param, execute() calls self.finish(request) with the new request. With this PR the hook receives that request, so the hook for the param=1 instance sees request.param == 2. On main the hook gets the request that set the fixture up. Whatever the final design is, finish() needs to use the setup request.

Please also state in the PR whether the hook should run for a setup that failed before a result was cached. With this PR it doesn't, and #14801 tests that it does.

---

Generated by Claude Code

[krishansubudhi]

Thanks for the thorough review. Agreed on all points. I'll rework it:
explicit teardown-exception arg on the hookspec, call from finish()
using the setup request (good catch on the param bug), let hook errors
propagate, update docs/changelog. On failed setup: I'll make the hook
run there too, consistent with #14801. Happy to coordinate with that PR
if you'd rather land one of them.

[krishansubudhi]

Thanks for the thorough review, reworked in f3b9ce1:

• pytest_fixture_post_finalizer gets a new teardown_exception
  argument (the exception, a BaseExceptionGroup if several finalizers
  failed, or None); nothing relies on sys.exc_info() anymore, and
  docs/changelog point to the argument.
• The hook is called directly from FixtureDef.finish(); the lambda
  finalizer and the assert in execute() are gone.
• finish() now uses the request stored at setup, so the param-switch
  case gets the param=1 request again (regression test added).
• Exceptions raised by the hook propagate instead of being treated as
  fixture teardown errors.
• On your question: yes, the hook now runs (once) for a setup that
  failed before a result was cached, with cached_result being None.
  This also fixes Internal AssertionError on stale _finalizers when a pytest_fixture_setup hookimpl raises during setup of a parametrized argument #14800, so I added a fragment for it; happy to drop
  that if Fix stale finalizers when a pytest_fixture_setup hookimpl raises #14801 lands first.