Fix module fixtures with --doctest-modules

[minbang930]

Summary

• Fix fixture registration when --doctest-modules collects a Python module before the normal module collector.
• Key parsed fixture holders by their visibility anchor, so normal tests get module-scoped FixtureDefs without relaxing node-based scoping.
• Add a regression test, changelog entry, and AUTHORS entry.

Closes #14533.

Test plan

• PYTHONPATH=src python /tmp/pytest-14533-repro/run_checkout_pytest.py -q testing/test_doctest.py::TestDoctests::test_module_fixture_available_to_normal_test_with_doctestmodules testing/test_doctest.py::TestDoctests::test_doctestmodule_with_fixtures testing/test_doctest.py::TestDoctests::test_doctestmodule_three_tests testing/test_doctest.py::TestDoctestAutoUseFixtures — 37 passed in 0.75s
• PYTHONPATH=src python /tmp/pytest-14533-repro/run_checkout_pytest.py -q testing/test_conftest.py::test_conftest_fixture_scoping_with_testpaths_outside_rootdir testing/test_conftest.py::test_conftest_fixture_from_ancestor_above_rootdir testing/deprecated_test.py::TestFixtureNodeidDeprecations — 6 passed in 0.14s
• PYTHONPATH=src python /tmp/pytest-14533-repro/run_checkout_pytest.py -q testing/python/fixtures.py::TestFixtureManagerParseFactories::test_parsefactories_conftest testing/python/fixtures.py::TestFixtureManagerParseFactories::test_parsefactories_conftest_and_module_and_class testing/python/fixtures.py::TestFixtureManagerParseFactories::test_parsefactories_relative_node_ids — 3 passed in 0.10s
• git diff --check
• python -m compileall -q src/_pytest/fixtures.py testing/test_doctest.py
• python -m ruff check src/_pytest/fixtures.py testing/test_doctest.py
• python -m ruff format --check src/_pytest/fixtures.py testing/test_doctest.py

AI assistance

OpenAI Codex helped reproduce the bug, draft the regression test, and prepare the initial patch. I reviewed the diff, checked the fixture scoping risk, added the changelog and AUTHORS entries, reran the targeted tests, and take responsibility for the change.

• Include new tests or update existing tests when applicable.
• Allow maintainers to push and squash when merging my commits.
• Add closes #14533 to the PR description and commit message.
• Credit the AI agent in a Co-authored-by commit trailer.
• Create a changelog file in changelog/.
• Add myself to AUTHORS in alphabetical order.

[RonnyPfannschmidt]

this is a good start and a fix

please expand the docs on the test

@bluetech a later major followup could make dedicated fs nodes part of the collection tree and specific nodes could decide if they ancor fixtures in the file node or the specific node

this first idea needs more iteration

[RonnyPfannschmidt]

please add a mark that links the issue and add a note on the regression on the docstring

[minbang930]

Thanks, updated the regression test docs for the DoctestModule/Module fixture case. CI is green now.

[bluetech]

If we key _holderobjseen by (obj, node), under which conditions will it be effective? I don't think parsefactories needs to be idempotent. So if this is the solution we go with, I prefer to just remove _holderobjseen.

The PR should document in the changelog the fact that higher-scoped autouse fixtures will now execute twice when using --doctest-modules, and how to avoid it if that's a problem (move said fixtures to conftest.py). We should also add a test for this.

If we impart to users the mental model that --doctest-modules effectively means that each python file is processed twice, I think the fixture behavior would make sense to them.

@bluetech a later major followup could make dedicated fs nodes part of the collection tree and specific nodes could decide if they ancor fixtures in the file node or the specific node

Interesting idea, but regarding the "specific nodes could decide if they ancor fixtures in the file node or the specific node" part:

• If e.g. Module and DoctestModule decide differently, then fixtures will still be registered twice, so no good.
• If they both choose to anchor at the FS node, it seems sort of ugly to me for the first one to "take over" the FS node and rely on a _holderobjseen-like mechanism to avoid duplication (ugly because it breaks encapsulation, which is why I'd really like to get rid of _holderobjseen).
• So presumably the way would be for the FS node itself to do the parsefactories, and remove it from Module and DoctestModule. But parsefactories requires an obj which I think is not appropriate for a pure "FS" node to know about.

[minbang930]

Updated per feedback: removed _holderobjseen, documented the higher-scoped autouse fixture caveat, and added a regression test for it.

CI is green now.

[RonnyPfannschmidt]

@bluetech hence the need for iteration - theres need for certain cooperation to make things work nice

technically i would actually go as far as considering doctestmodules a collector that intently is actually a child of the module, that it is at the same level as the module its "contained in" is the real issue for me

[bluetech]

technically i would actually go as far as considering doctestmodules a collector that intently is actually a child of the module, that it is at the same level as the module its "contained in" is the real issue for me

One problem with this is that it makes the DoctestModule dependent on the Module, but I think it's a nice property that they're independent, e.g. I can run just the doctests without running the python tests.

Another problem is that module is not the only relevant scope, there is also class scope. I think if module-level fixtures are shared then class-level fixtures are expected to be shared as well, and Module -> DoctestModule -> DoctestItem wouldn't work for that. Scratch that, doctest currently doesn't support class scope so this wouldn't be a regression.

[bluetech]

Updated the PR:

• Reworked commit message and changelog entry
• Updated docs a tiny bit
• Improved the tests a bit
• Added a couple of semi-related small commits