Original 'counts events per KST day' test used UTC times that bucket
identically under both KST and naive UTC slice — would not catch a regression
where kstDate was replaced with ev.ts.slice(0,10). Add an explicit
near-midnight case (2026-05-01T15:30Z = 2026-05-02 00:30 KST) that fails
under naive UTC and passes under correct KST conversion.
6 tests pass (was 5).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
