When catch-up phase is in progress and a parallel db write is executed it is possible that the write is lost as it is not neither caught by snapshot nor by stremaing pahse as it can get filtered.
To reproduce check io.debezium.connector.postgresql.PostgresConnectorIT.shouldResumeStreamingFromSlotPositionForCustomSnapshotWal2Json() and execute it for decoderbufs. The LSN of the parallel write gets filtered due to LSN comparison. If the LSN filtering is disabled in the code the record gets in.
This is problem only for custom snapshotters with shouldStreamEventsStartingFromSnapshot() set to false.