If we did try to collapse the record set classes a bit, we could handle this in a sort of before_insert step so that we don't need to supply a full implementation of import_batch to do it.

I also wonder if the export half of this is necessary (converting sgids to gids) or could we implicitly trust the signing on the way in while converting to new sgids? Effectively re-signing from an unknown key to the valid one. Since the transfer is effectively trusting that aspect of the input data anyway, it could be one less step. (Not sure if it makes much difference in practice or not though.)

I opted to go with a generic RecordSet and specific record sets. That kept the interface simple.

For the conversion. Yes, I could skip the SGID > GID > SGID dance, though this feels nicer.