[leonardoaraujosantos]

Summary

Two tests failed intermittently on unrelated PRs (#113, #115, #116), and one of them failed main's own Windows build at db9e171e. Both are test timing problems, not product bugs.

disposed_binary_request_retains_worker_ownership_until_actual_exit (and its sibling check)

The binary worker returns its path permit before its activity Scope drops. Dropping the scope is what removes the activity from the registry, and it happens on the worker's own thread. The test asserted available_permits() == 1 and then read activities.is_empty() once, so a loaded runner could still see the entry. The test now waits up to 5 s for the registry to empty (activities_drained), at both places it checked. The ownership guarantees under test are unchanged.

reviewed_recovery_busy_confirmation_requires_interrupt_ack_and_a_new_review (Windows)

The test starts a background cyber exec … GATE and gave it 10 s to reach the scripted model. On a loaded Windows runner, starting exec and its first request can take longer. That failed main's Windows build, while the same test passed on other runs (e.g. #117's).

• The deadline is now 60 s, the bound other Windows paths use.
• The exec's stderr is now captured. If exec exits before reaching the model, the test fails with its exit status and stderr instead of a bare "GATE never reached provider".

Tests

• Both tests pass locally.
• cargo nextest run --workspace --profile ci on top of the current main (with Re-send failed Turns on the model fallback chain #115–Drive cyber exec over stdin with --input-format stream-json #117 merged): 5770 passed, 0 failed.
• clippy -D warnings.

Not addressed: a one-off Windows failure in mcp_oauth::a_refresh_timeout_fires_during_a_blocked_write_which_keeps_the_lock seen once on #115; it passed on rerun.