Fix Storage.path clobber at import time (#6061)

[evnchn]

Motivation

Fixes zauberzeug#6061. PR zauberzeug#5960 introduced a module-level Storage.path = None assignment in nicegui/testing/general_fixtures.py intended as a "not configured yet" sentinel for pytest. But that line executes on plain import nicegui.testing — not just under pytest — so any production app whose import graph reaches the testing module has its Storage.path clobbered to None.

Concretely, NiceGUI's own docs site does this: website/documentation/content/screen_documentation.py:2 does from nicegui.testing import Screen, which transitively imports general_fixtures, which sets Storage.path = None. The next app.storage.user access then crashes in Storage._create_persistent_dict at nicegui/storage.py:98:

TypeError: unsupported operand type(s) for /: 'NoneType' and 'str'

Empirical repro (pre-fix):

>>> from nicegui.storage import Storage
>>> Storage.path
PosixPath('/.../.nicegui')
>>> from nicegui.testing import Screen
>>> Storage.path
None

Implementation

Stop mutating Storage.path at module-level. The tempdir creation now happens only inside pytest_configure, and its cleanup is registered via atexit.register(shutil.rmtree, ..., ignore_errors=True) — the same pattern already used in screen_plugin.py for DOWNLOAD_DIR. This lets us drop pytest_unconfigure entirely (and its three re-exports in plugin.py, user_plugin.py, screen_plugin.py).

Net change in general_fixtures.py: +1 import, -1 module-level assignment, -1 hook function, -1 sentinel variable (after design iteration; see below).

A regression test is added in tests/test_lazy_imports.py (subprocess pattern matches existing test_module_access_does_not_import_others):

def test_importing_nicegui_testing_does_not_clobber_storage_path():
    result = subprocess.run(['python3', '-c', dedent('''\
        from nicegui.storage import Storage
        orig = Storage.path
        from nicegui.testing import Screen, User  # noqa: F401
        assert Storage.path == orig, ...
    ''')], capture_output=True, text=True, timeout=30, check=False)
    assert result.returncode == 0, result.stderr

Verified the test fails on pre-fix code (stash/unstash) and passes after — not vacuous.

Design discussion: why atexit instead of a `_storage_configured` boolean + `pytest_unconfigure`?

The first commit on this branch used a separate module-level _storage_configured: bool as the sentinel and kept a symmetric pytest_unconfigure hook for cleanup. The second commit replaces that with atexit.register and drops the hook.

Alternatives considered for replacing the broken Storage.path is None sentinel:

1. Module-level _storage_configured boolean — clean, but adds a new mutable module-level variable.
2. Capture _INITIAL_STORAGE_PATH = Storage.path at import, compare in pytest_configure — still a new variable, and fragile if a user has set Storage.path themselves before pytest runs.
3. Class attribute on Storage — leaks pytest concerns into production class. Rejected.
4. pytest.StashKey — proper pytest idiom but still introduces a module-level StashKey instance, same effective scope as a bool.
5. Function attribute (pytest_configure._called) — poor discoverability. Rejected.
6. Drop the guard, use atexit.register for cleanup — zero new module-level variables, matches existing screen_plugin.py pattern. Chosen.

Trade-off vs the boolean + pytest_unconfigure approach:

• For one-shot pytest invocations (CI, normal dev runs): identical behavior.
• For in-process repeat usage (IDE "re-run tests", pytest --looponfail, some pytest-xdist setups): cleanup is deferred to process exit instead of pytest session end. Leaked tempdirs accumulate in the system temp dir until the Python process dies. The directories are small (empty or a few JSON files) and OS-GC'd eventually.

Why the dual-plugin re-entry guard wasn't necessary:
The previous "is already configured" check guarded against pytest_plugins = ['nicegui.testing.plugin', 'nicegui.testing.user_plugin'] both being loaded, which would call _general_pytest_configure twice. NiceGUI's own tests/conftest.py loads only nicegui.testing.plugin (which already re-exports user-plugin fixtures), and the dual-load is redundant by design — it would cause fixture redefinition warnings. Without the guard, the worst case is one extra tempfile.mkdtemp call whose result is immediately shadowed; the orphaned dir is still rmtree'd at exit via its own atexit registration.

I'm happy to revert to the boolean approach if you prefer the deterministic per-session cleanup.

Validation

• pytest tests/test_storage.py tests/test_lazy_imports.py tests/test_main_file_marker.py tests/test_user_simulation.py → 102 passed / 2 xfailed (pre-existing).
• ruff check, pylint, mypy on changed files → clean.

Progress

• The PR title is a short phrase starting with a verb like "Add ...", "Fix ...", "Update ...", "Remove ...", etc.
• The implementation is complete.
• This PR does not address a security issue.
• Pytests have been added (regression test in tests/test_lazy_imports.py).
• Documentation has been added/updated or is not necessary.
• No breaking changes to the public API.

---

Draft against evnchn/nicegui:main for review/finetune before upstreaming to zauberzeug/nicegui. Two commits are currently on the branch; squash on merge if preferred.