Does the replication key need to be in the schema for a stream? Or can you use an inline mapping (stream_map) to define a replication key?

This is an interesting question - I presume if you need a mapping to generate this, then the source isn't aware of the replication key. In which case, surely you won't be able to request 'new records' from the tap using this key value? Or is it some kind of compound key that the tap can partially use to filter? Or are you doing this purely to limit the amount of data loaded into your target?