node analytics coverage: follow-ups to #494 (late_evicted_observations semantics, fast-path separator test) #504

Follow-ups from the round-2 review of #494 (see "Review — CS-pve-agent3 PR#494" on head ef687ff0) and from the staging check.

1. coverage.late_evicted_observations is misleading. It counts every observation after the evicted tx's FirstSeen + 1 ns, so normal second-observer hearings a few seconds later count too. On staging it rose by about 11,000 per hour (13,919 → 24,999 in 50 minutes). Count only observations that are genuinely late, for example more than 1 h after FirstSeen, and document the threshold. Alternatively, rename the field to say what it counts. No UI uses it today.
2. The fast-path separator is not pinned. Mutant M7 (dropping the raw[13] != ':' check in parseCanonicalUTCSeconds) survives: 2026-10-16T12X30:00Z is accepted, while time.Parse rejects it. Add a parity test that covers every separator position.
3. Nit. Revert the unrelated gofmt re-alignment in types.go and store.go, or leave it, but do not mix it with logic changes.

The staging check of the main fix passed: coverage.from tracked eviction smoothly over an hour, with a 48 h window and no jumps across 115 eviction ticks.