Windows: consume_wake_signal drops wakes on transient sharing violation + de-flake race test

[danielhertz1999-bit]

Problem

Two issues surfaced by tests/test_wake_handler.py::test_consume_atomic_no_race failing on macOS CI with assert 0 >= 1.

1. A real Windows bug in consume_wake_signal(). It does an unlink-to-consume inside a bare handler:

try:
    self._wake_signal_path.unlink()
except FileNotFoundError:
    return False
except OSError:          # <-- swallows PermissionError too
    return False
return True

On Windows, unlink() raises PermissionError (sharing violation, WinError 5/32) when a concurrent writer or reader momentarily holds the signal file open — Python's open() there doesn't request FILE_SHARE_DELETE. The bare except OSError swallows that transient failure as "no signal," dropping a real wake and leaving the daemon stranded in HIBERNATION with an unserved socket. On POSIX this never happens (you can unlink an open file).

2. The test itself is flaky. consumed_truthy_count >= 1 had no synchronization: the consumer ran a fixed 200 unlink iterations and could finish before any writer's first replace landed (adverse thread scheduling on a loaded runner), reading consumed=0. That's the assert 0 >= 1 seen on macOS — independent of the Windows bug, and unfixed by a Windows-only change.

Fix

• consume_wake_signal: retry the unlink briefly on a Windows PermissionError (mirrors daemon_state._atomic_replace). On POSIX a PermissionError is a genuine permission fault, so it still returns False on the first failure — no behavior change off Windows.
• De-flake test_consume_atomic_no_race: writers produce continuously and the consumer polls until it consumes (bounded by a deadline) instead of a fixed count. Keeps the atomicity intent (errors == [], no crash under concurrent write+consume) while making the >=1 liveness check deterministic on any scheduler. The writer harness also retries os.replace on a Windows PermissionError, mirroring production.
• Two deterministic regression tests for the retry (simulated via monkeypatch, so they run on every platform): a transient Windows PermissionError is retried and consumed; a POSIX PermissionError is not retried.

Verification

wake_handler suite green; test_consume_atomic_no_race looped 10× locally on Windows with zero flakes. The two new deterministic tests fail against the pre-fix source (the retry path doesn't exist), confirming they gate the fix. macOS behavior is exercised by this PR's CI.

[CodeAbra]

Thanks. v2.0 (Rust rewrite) already has consume_wake_signal, so the wake-signal path carried over, I need to check whether the transient sharing-violation retry you added is in there or still missing. If it's missing it goes into v2.0.3 with the other Windows fixes. Keeping it open till I've confirmed. Nice de-flake on the race test too.