fix(pytest): stop reporting pytest's own finalizer as a phantom teardown

[Mohanad49]

Description

Closes #918.

From pytest 9.1, FixtureDef.execute attaches an internal finalizer to every fixture. pytest_fixture_setup wrapped it along with the real ones, so every fixture produced an extra afterStage:

fixture_with_finalizer::teardown    status=passed    <- real
fixture_with_finalizer::<lambda>    status=None      <- pytest's own
simple_fixture::1                   status=passed    <- real
simple_fixture::<lambda>            status=None      <- pytest's own

pytest never calls that finalizer through Allure's wrapper, so nothing ever closes the stage and it is written out with no status — which the report renders as unknown, once per fixture, on every run.

Confirmed as FixtureDef.execute.<locals>.<lambda>, owned by _pytest.fixtures:

[probe] simple_fixture
   name='<lambda>'  module='_pytest.fixtures'  qualname=FixtureDef.execute.<locals>.<lambda>
   name='?'         module='functools'         (the yield-fixture teardown)
[probe] fixture_with_finalizer
   name='<lambda>'  module='_pytest.fixtures'  qualname=FixtureDef.execute.<locals>.<lambda>
   name='teardown'  module='test_phantom'      (the author's finalizer)

The fix

Skip finalizers owned by _pytest.

The check is on the owning module, not on the name. Filtering <lambda> would have been shorter and wrong: a lambda passed to request.addfinalizer by a test author is a real teardown and has to keep being reported. The functools.partial that implements yield-fixture teardown is likewise untouched, so simple_fixture::1 still appears.

Testing

tests/allure_pytest — 273 passed, 2 xfailed (271 before, plus the two added here). ruff clean.

The two tests do different jobs, deliberately:

Test	Without the fix	With the fix
test_pytest_internal_finalizer_is_not_reported	fails	passes
test_user_lambda_finalizer_is_still_reported	passes	passes

The first is the regression test — I checked it fails on unpatched main, otherwise it would be asserting nothing. The second passes either way on purpose: it guards the fix against becoming too aggressive, and would start failing if someone later narrowed the filter to the name.

I ran the existing suite on unpatched main too — 271 passed there as well, so nothing currently covers this, which is presumably how it shipped.

Verified against pytest 9.1.1 / Python 3.11. The filter is a no-op on older pytest, which has no such finalizer.

[CLAassistant]

All committers have signed the CLA.