C0: add _pytest/pytest to FREEZEGUN_IGNORE so --durations stops reporting the bogus ~1.7e9s setup for test_api_baseline[convey]; the freeze itself is unchanged, baselines pass identically. C1: test_cancel_commit_race_runs_at_most_once synchronises on a threading.Event from commit() + a join on the real Timer handle instead of a blind 0.15s sleep. Real Timer and 50 iterations preserved, no fake clock; at-most-once/xor assertions kept and strengthened (timer-dead + commit-not-fired now asserted). 7.63s -> 0.03s. C2: callosum stop() now closes the listening socket via an idempotent _close_server_socket() (lock-guarded read-and-null, OSError-swallowing), shared with the accept-loop finally, so a blocked accept() is interrupted on shutdown instead of waiting the 1.0s accept timeout. No double-close, no stop()-originated exception, behaviour otherwise equivalent. Drops the test_cortex_client teardown cluster (~0.8-1.0s x ~8). C3: test_readiness timeout/poll-interval trims (real-timeout behaviour unchanged; assertions verbatim). C4: test_importer_stall_timeout wait args trimmed; elapsed<2.0 assertions unchanged. Real timeout-path tests (test_wait_for_agents_timeout_actual, _missed_event_recovery) untouched. Counts unchanged (4923/11/8). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>