SQLite stream retention cutoff on a whole second deletes same-second rows early #464

The SQLite retention sweep (cleanup_expired_stream_records in crates/storage-sqlite/src/stream.rs) deletes rows with created_at < cutoff as a text comparison. created_at carries a millisecond fraction (2026-10-23T14:19:07.123Z). The cutoff is format_timestamp(now - 24h), RFC 3339 from the time crate, which renders a whole second with no fraction (2026-10-23T14:19:07Z). When the cutoff lands on an exact whole second, every row in that same second (...07.123Z) sorts before it, because . sorts before Z, and is deleted up to a second early.

Reaching it needs now_utc() to return a zero sub-second on the sweep's read, so it fires rarely. The fix is to render the cutoff in the same millisecond shape as the column (sqlite_util::format_epoch_millis), with a unit test for a cutoff on a whole second.

Found while working on #463, which did not introduce or widen it.