CDC snapshot-to-stream handoff

You're bootstrapping CDC on a hot 3 TB Postgres table that takes ~6 hours to snapshot, while it receives ~5k writes/sec the whole time. You must produce a downstream copy that is eventually consistent with the source with NO gap and NO duplicate-induced incorrectness, but you can't lock the table for 6 hours and you can't afford to miss writes that happen during the snapshot. Design the snapshot-to-stream handoff so the consumer ends up with exactly the right final state, and explain how a mid-snapshot failure is recovered.