fix(genesis-writer): keep sub-second precision on every emitted timestamp

[rickyrombo]

time.RFC3339 has no fractional-second component, so every timestamp the writer emits was rounded down to the whole second. Measured against a full-snapshot replay compared to its source:

table	sub-second in source	in migrated DB
subscriptions	1,022,554	0
playlist_tracks	1,014,903	0
saves	380,604	0
reposts	361,803	0
follows	104,614	0
track_downloads	78,032 of 78,032	0
comments	7,530	0
tracks.release_date	26,129	truncated

same row:  src  2024-12-01 12:30:13.244308
           etl  2024-12-01 12:30:13

Why created_at isn't cosmetic

The indexer replays each migrated row as of its created_at (migrationBlockTime), so the timestamp drives replay semantics, not just display. Parity also keys track_downloads on created_at — which is why every sampled row of that table reported as missing rather than mismatched. The row counts match exactly (78,032 both sides); only the key failed to join.

Safety

• parseMigrationTimestamp tries RFC3339Nano before RFC3339
• Go's parser accepts a fractional second against a layout that lacks one, so event_create.go's stricter time.Parse(time.RFC3339, …) still succeeds
• A value with no fractional part formats identically under both layouts, so rows without microseconds are untouched

Note on the guard

TestAllEmittedTimestampsAreUTC scans for formatting sites and fails if it finds none — it caught this rename immediately. Its pattern now matches both layouts, so reintroducing the truncating form still trips the .UTC() half.

This is a different failure from the one #519 fixed. That emitted Postgres text format, which matched no accepted layout, so the indexer fell back to block time. This parses cleanly and quietly rounds.

🤖 Generated with Claude Code