tombstones are sorta the default answer here (i.e. at simplest, you keep all data forever so you can merge correctly, but you hide anything where you've seen a tombstone after it).

but "makes sense" and ways to optimize that can change massively with context. e.g. for a chat app, as soon as you see "deleted message X", you can reasonably drop X and all past and future changes to X because they won't be shown by anyone (don't even need to sync them). if you do that with "deleted chars 87..93" in a text editor, past-edits that you receive in the future might affect the behavior (it might add chars before those, changing what that range means), so you can't simply forget those chars (e.g. an easy option is to replay all events that occur after an event syncs, but that means retaining all events forever). the semantics you choose and what you do with the data affect your outcomes a lot.

tbh this is one of the reasons I like the idea of a WASM-defined algorithm. no one algorithm will be "best" for all data, and the storage/computation/transmission savings can be extreme.