A CRDT which models edits of arbitrary text with essentially unbounded revision history (required for merging) almost certainly can only grow (monotonically?) in size. Maybe you can be smart about compression, but I don't think it can ever shrink. Someone smarter than me can probably formally prove this with an argument about entropy while considering a series of particularly pathological edits.

But you're right that I'm wrong when I say "[all] CRDTs are generally append only structures." I meant to say "text-revision CRDTs are generally append only data structures."