fix: don't crash when an XDG dirs variable holds only separators

[darrenhuai]

Two things, the second found while writing tests for the first.

site_*_dir crashes when an XDG dirs variable holds no paths

site_data_dir, site_config_dir and site_applications_dir raise IndexError on Unix and macOS when $XDG_DATA_DIRS or $XDG_CONFIG_DIRS is set to something non-empty that contains no actual paths:

XDG_DATA_DIRS=':'    -> IndexError: list index out of range
XDG_DATA_DIRS='::'   -> IndexError: list index out of range
XDG_DATA_DIRS=' : '  -> IndexError: list index out of range

The lookup splits on os.pathsep and drops blank entries, which leaves an empty list, and then the caller does dirs[0].

What makes it look unintended rather than merely undefined is that a whitespace-only value already falls back to the platform defaults — the outer .strip() makes it falsy, so it takes the super() branch. ":" is just as empty in every sense that matters but takes the other branch and crashes.

It's easy to produce one by accident from a shell:

export XDG_DATA_DIRS="$SOME_UNSET:$ALSO_UNSET"

That's ":", and the caller gets list index out of range from inside platformdirs with nothing pointing at the environment as the cause.

I pulled the split into _xdg_dir_list() so the three call sites share it, and moved the emptiness check to after filtering instead of before. Unset, empty, whitespace-only and separators-only now all mean the same thing.

user_desktop_dir was missing from the shared test property list

While adding cases for the above I noticed user_desktop_dir is the only one of the 27 public *_dir properties absent from the PROPS tuple in tests/conftest.py, and from the copy in __main__.py that test_props_same_as_test checks it against.

That tuple backs the func and func_path fixtures, so the omission quietly kept desktop out of every test parametrised on them — test_windows, test_macos, test_android, the XDG cases in test_unix, the appdirs comparison, and test_no_ctypes.

test_no_ctypes is the awkward one. #519 was user_desktop_dir raising ValueError on Windows builds without ctypes, which is the exact failure that test exists to catch — it just never ran against that property. Including it would have caught the bug before 4.11.0 went out.

Nothing fails once it's added, so there's no second bug behind it; it only closes the gap. python -m platformdirs prints the desktop directory now too.

Happy to split these into separate PRs if you'd rather keep them apart.