Flaky Windows handle count in loop_drop_terminates_live_children_and_releases_their_handles (83 vs 82) #26

loop_drop_terminates_live_children_and_releases_their_handles (crates/turnloop-contract/tests/windows_lifetimes.rs, added in #13) failed intermittently on windows-2025 default mode in PR #24's run 34941544032. PR #24 only changes release.yml and release.py:

assertion `left == right` failed: live-child drop leaked a native handle
  left: 83
 right: 82

It passed on #13, #16, #22 and #25. The test compares GetProcessHandleCount against an exact baseline after dropping a loop with live children.

Likely causes to rule out:

• Not a leak: a handle opened by the OS or runtime asynchronously during the window, for example a thread-pool wait registration still being torn down after the child's wait callback, or a job object/process handle released slightly later.
• A real leak: a handle leaked on a rare ordering (child exit racing loop drop).

Expected: find which handle it is (for example with NtQueryObject/handle-type enumeration in the failure message), then fix the leak or wait for asynchronous release deterministically. Don't loosen the exact comparison.