For example, what would happen if two people change the same paragraph in two different instances at roughly the same time?

This is already a problem for existing distributed versioning systems (like git), and in those the merge conflicts can only happen when the users explicitly request them. With activitypub the merge conflicts would happen in response to asynchronous events and to make things worse, different instances might see different events. How would you surface the conflict to the users so they can solve it? Do you send an email to the user which was a fraction of second late saying their edit got rejected? Do you reject both of them and keep the old content? Do you overwrite the first edit silently?

These are already hard UX problems on centralized wikis, and the technical aspects of a distributed system would make them worse. So much worse that I would say it's orders of magnitude harder to implement than a link aggregator.