Fix rootdir conftest fixtures not visible to items collected outside the rootdir

[breidenbach0]

Problem

Fixes#14683.

doctest_namespace injections (and any fixture defined in the rootdir conftest) stopped being available in pytest 9.1 when --rootdir points to a subdirectory and tests/doctests are collected from a parent directory. Doctests then fail with NameError, e.g. the reporter's invocation:

pytest --rootdir src/xclim/testing --config-file=src/xclim/testing/conftest.py \
--doctest-modules src/xclim

Reproduced: 9.0.3 passes, 9.1.1 fails with NameError: name '...' is not defined.

Cause

Bisected to 46478fad5 (#14098, pytest 9.1), which changed conftest fixture visibility from nodeid-based to node-based:

• Before 9.1: a conftest located in the rootdir got an empty baseid (relative to rootdir), which matches every collected item — its fixtures were visible session-wide.
• After 9.1: that conftest is scoped to its own Directory node. When items are collected from outside the rootdir, that Directory is a sibling subtree, not an ancestor of the items, so the fixtures are no longer visible.

Fix

Restore the pre-9.1 behavior by attaching the rootdir conftest to the Session node, in the two places a conftest gets parsed:

1. _flush_pending_conftests_to_session — also flush the initial rootdir conftest (e.g. when it is the --config-file) to the Session.
2. pytest_make_collect_report — when collecting the rootdir Directory, attach its conftest to the Session instead of the Directory.

This is safe:

• Behavior-identical for normal layouts (everything under the rootdir): the Session is an ancestor of all items, so visibility is unchanged.
• Preserves Conftest fixtures leak to sibling directories when testpaths points outside rootdir #14004: only the rootdir conftest is affected; nested conftests still scope to their own Directory, so sibling-leak does not return.
• Fixture overrides still work: a subdir conftest remains more specific than the Session per is_visibility_more_specific, so it still wins.

Tests

• testing/test_conftest.py::test_rootdir_conftest_visible_outside_rootdir
• testing/test_doctest.py::TestDoctestNamespaceFixture::test_namespace_fixture_from_rootdir_when_modules_outside_rootdir (mirrors the reporter's exact --config-file + collect-from-parent pattern)

Both fail on main and pass with this change. Added news fragment changelog/14683.bugfix.rst.

Note on #14635

#14635 (Home Assistant, "fixtures not found when a parent directory appears multiple times in an argument list") is tracked as a duplicate and is being worked on separately by @RonnyPfannschmidt (multiple Directory nodes for the same path). This PR targets the rootdir-conftest visibility facet (#14683) only; making the rootdir conftest Session-visible also makes it more robust against the multi-node scenario, but it is not meant to be the complete fix for #14635.

[breidenbach0]

Heads up on the two red CI jobs — both are a known repo-wide infra issue, unrelated to this change:

• build (ubuntu-pypy3-xdist) fails during dependency installation (before any pytest code runs) while building hypothesis 6.156.6 from source:

  error: the configured PyPy interpreter version (3.10) is lower than PyO3's minimum supported version (3.11)
  

  The latest hypothesis now ships Rust extensions via maturin/pyo3 0.29, and pyo3 0.29 dropped Python 3.10 while the PyPy CI image is on 3.10. Since hypothesis is pinned only as >=3.56, pip pulls the incompatible latest.
• check is just the re-actors/alls-green status gate — it fails solely as a downstream consequence of the pypy3-xdist failure.

For confirmation, another currently-open PR (#14693, unrelated) fails the identical two jobs with the same cause. All other jobs on this PR (macOS/Ubuntu, Python 3.10–3.15, xdist, plugins, doctesting, etc.) are green.

[RonnyPfannschmidt]

reopend to trigger flaky ci/codecov

[breidenbach0]

Hello, I was wondering, if I would be able to take up commit access to pytest-dev. Happy to keep going through the usual review process for my own PRs and to start helping others.

Best regards,
breidenbach0

[RonnyPfannschmidt]

that privilege and responsibility is not something to ask for after one pr merged and one pr in the pipeline

it is something we propose to individuals after a trusting relationship has been built - and not all of those take it up

[breidenbach0]

Ah sorry, I missunderstood the Contributor guidelines. Thanks for clarifying. Looking forward to build that relationship and continue building

[RonnyPfannschmidt]

turns out we hadn't update that guideline since 2016 - so you didn't misunderstood the text as written