conftest.py fixtures silently missing for files collected after a file argument from an ancestor directory (pytest 9.1.0+, regression from #14004)

[ChrisKotsis]

Environment

pytest 9.1.1, Python 3.12.9, Linux (WSL2). Not reproducible on pytest 9.0.2. Reproduces with --import-mode=prepend (default) and
a package layout (tests/__init__.py); the conftest lives in the package.

Minimal reproduction

repo/
  test_x.py                 # any test
  tests/__init__.py
  tests/conftest.py         # @pytest.fixture(autouse=True) def guard(): raise RuntimeError("guard ran")
  tests/a.py                # def test_a(): pass
  tests/b.py                # def test_b(): pass

pytest tests/a.py test_x.py tests/b.py

Expected: test_a and test_b both error with "guard ran" (the autouse fixture applies to
everything under tests/). Observed on 9.1.1: test_a errors, test_b PASSES — the conftest's
fixtures are not in test_b's fixture closure at all. pytest tests/a.py tests/b.py test_x.py
and pytest test_x.py tests/a.py tests/b.py behave as expected. pytest 9.0.x behaves as expected
in every order.

Cause (read from 9.1.1 source)

Since #14004 conftest fixtures are parsed against the Directory NODE that collects the
conftest's directory: FixtureManager.pytest_plugin_registered stores the conftest in
_pending_conftests[dir], and pytest_make_collect_report pops it on that Directory node's
first collect report, registering fixtures into _node_autousenames[node] (keyed by node
object). _getautousenames walks node.listchain() and looks up each parent node object.

Session.collect() re-collects an ancestor directory WITHOUT the collection cache when an
argument is a file directly in it (handle_dupes = not (len(matchparts) == 1 and matchparts[0].is_file()) — "files given directly multiple times on the command line should
not be deduplicated"). That fresh collection creates NEW child Package/Dir node objects
and overwrites _collection_cache[ancestor]. A later argument inside the package is then
collected under the new Package object; its first collect report finds _pending_conftests
already popped, so the new node never receives the conftest's fixtures, and nothing warns.

The legacy _nodeid_autousenames fallback would have caught this (both nodes share the nodeid
tests), but it is only populated by the deprecated parsefactories(obj, nodeid) path.

Impact

Any autouse guard in a package conftest (isolation, monkeypatching of live resources) is
silently absent for the affected files. We found it because such a guard protects a real
mailbox and the fence test for the guard started draining it.

Suggested fix

Either key conftest registration by directory path (re-register on every Directory node whose
path matches a registered conftest, not only the first), or make _getautousenames /
_matchfactories also consult the nodeid-keyed table for Directory nodes, or reuse the cached
child collectors when re-collecting for the non-deduplicated file case.

Workaround we use (for reference)

Root conftest regroups positional arguments deepest-directory-first at pytest_sessionstart
(the trigger cannot occur) and refuses the run in pytest_collection_modifyitems when any
item's closure lacks an ancestor conftest's autouse fixtures.

[Muhammad-AbdullahGhani]

I tested the minimal reproduction on the latest main branch (pytest 9.2.0.dev), and both test_a and test_b correctly fail during setup with RuntimeError: guard ran as expected:

tests/a.py::test_a ERROR                                                 [ 33%]
test_x.py::test_x PASSED                                                 [ 66%]
tests/b.py::test_b ERROR                                                 [100%]

It looks like this was already resolved on main by commit fc8f56b (#14635 - "fix: deduplicate Directory nodes on re-collection to preserve fixture identity").

Could this be verified and considered for a backport to 9.1.x?