The test fast-forwards simulated time from February to April by having the live
real-time handler fire a scheduled event that advances a ManualTimeProvider each
step. That advance is therefore bounded by wall clock, but the test waited only
15s and then asserted on the collected events regardless of whether the advance
had completed. Under load the wait timed out around late March, so the final
week's scheduled event (2024-03-25) had not fired yet and the collection assert
failed with 6 vs 7 events.

Give the advance a generous budget and assert it actually completed, so the test
reliably observes the full timeline (and a genuine hang fails with a clear
message instead of a misleading count mismatch).

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>