Fix Windows lockfile contention race

[MuteJester]

What changed

• close the atomically-created lockfile handle before entering the protected section; the lockfile itself remains the ownership sentinel
• retry once when Windows reports PermissionError but the holder deletes the file before the existence check
• preserve prompt propagation for persistent genuine permission failures
• add a regression test for the disappearing-holder race

Root cause

Windows can report a sharing violation instead of FileExistsError while the owner keeps the lockfile handle open. The owner may then close and unlink the file before the waiting thread checks exists(), causing valid contention to be misclassified as a genuine permission failure.

Validation

• 100 runs with six competing threads passed via a direct lock-module stress harness
• git diff --check
• local pytest is unavailable because the host test environment lacks PyTorch; Windows CI is the platform-authoritative regression check