Don't trust a zero Windows file ID in samefile_nofollow

[RonnyPfannschmidt]

Fixes #14864.

On Windows, when a collection argument's path does not compare equal to a collected node's path, Session.collect() falls back to samefile_nofollow() to account for 8.3 short paths (#11895). That fallback uses os.path.samestat(), i.e. it compares (st_dev, st_ino).

On Windows st_ino is the file ID, and support for file IDs is file-system specific — a file system that has none reports 0 for every file, the same convention the neighbouring BY_HANDLE_FILE_INFORMATION fields use for unsupported values. The reporter's drive is an sshfs (WinFsp) mount, where the chain is:

• SFTP v3's ATTRS has no inode field, so sshfs never fills st_ino (the identifier does not appear in sshfs.c at all).
• On Linux this is invisible because libfuse's set_stat() overwrites st_ino with its own node id unless use_ino is given. WinFsp's FUSE layer has no such fallback: it discards use_ino and assigns FileInfo->IndexNumber = stbuf.st_ino verbatim.
• FILE_ID_INFO.FileId and nFileIndexHigh/Low are therefore both 0, and CPython passes that straight through to st_ino.

So os.path.samestat() reports every pair of files on the volume as the same file, every node matched, nothing got pruned, and passing a single file collected the whole suite. This is the long-standing python/cpython#78116 (samefile() should not use zero-valued st_dev and st_ino), which also affects Google Drive File Stream and WebDAV volumes.

Fixed by treating a zero file ID as "unknown" in the Windows version of samefile_nofollow(), so the caller falls back to plain path comparison — which is what every non-Windows platform does anyway, since the samestat() fallback is win32-only. The short-path case from #11895 is local NTFS with real file IDs and keeps working; 8.3 short names are an NTFS/FAT feature that the affected mounts do not generate in the first place.

Note that _is_same() in the same module has the same weakness, but there a bogus True only makes the import-mismatch check miss a mismatch rather than break collection, so I left it alone.

🤖 Generated with Claude Code

[bluetech]

LGTM!

[bluetech]

I would put this in a code comment, not in the docstring (it's an implementation detail IMO).