For anyone interested in this issue I would suggest taking a look a Neil Fraser's Differential Synchronization.[0] It took me 2 weeks as a solo developer to add real time collaborative editing to our product using that method.
Differential synchronisation works very well indeed for plain text, but as Neil himself has pointed out, it doesn't necessarily work well for tree-structured data (like XML or JSON) because there's no guarantee that the combination of separate syntactic-validity-preserving edits will necessarily be syntactically valid.
Using diffs was our very first approach. It was quite naive (not to say dumb) solution but we also came to a conclusion that it will be problematic for tree-structured data and there were also performance issues (for example when you bolded the whole document).
So instead of trying to fix those issues, we tried a different approach.
The solution to this problem is to add a translation layer from the tree structure to plain text. If you can do that you can work on any arbitrary data structure easily. I'll leave it up to the reader how to figure out how to do it ;)
We've been there (although, in case of OT). When implementing CKEditor 5, we started with a linear model and just a couple type of operations (the usual set). At that time, we weren't able to transform that linear model to a tree model because conflict resolution based on that simple set of operations was too dumb. As I mentioned in the other comment about CRDT vs OT, we'd still consider having a linear + tree model duet, but so far it's purely theoretical discussion whether it has a chance to work.
Nice! It's my understanding that the default method used by Neil requires a fixed polling interval, and thus updates are limited to every 100ms or so. Did your implementation get around that?
[0]: https://neil.fraser.name/writing/sync/