The first example also reveals the crux of the issue. The actual problem is that the merging intent is not mathematically solveable in a satisfying manner.

E.g. let's say two people edit the body element depending on the content of the Edit the best solution¹ could be to take both contents (in which order?), to just take one, to take neither or to merge them semantically into a new third thing.

Meaning there are two obvious ways of going for a better merging solution:

1. Retain the ambiguity (if there is one) and present it to the users and let them decide

2. Use an LLM to guess the intent of both edits and task it to resolve the ambiguity if resolveable

¹: with best I mean the the solution real humans merging text from two pieces of paper would chose