We're using the offline replica feature quite extensively in our application. We're adding many tables to the replica which are only used in read-only mode and are never changed. When returning the changes to an online system the SynchronizeReplica method doesn't allow us to define which tables we really want to return (as opposed to CreateReplica which accepts the layerquery options and ObjectIds to limit the data).
I propose that you add a possibility to define which tables/featureclasses are returned on the sync and the rest of the tables would simply be ignored. This would give us a huge performance boost, because most of our data is simple read-only data.