I wrote a similar tool to this many years ago that used JATS as an internal document model, using contentEditable, and I can relate to the pain you refer to. Getting the semantics right in cases where you split or join text that has varying types of nested context was hard.

I like what you did here in terms of the cursor that points one way or the other: it tells you if you're in "normal" mode, or for example, in a bold region. I don't think I've seen that before so, if you made it up, that itself is a cool invention.