I don’t quite see why this is difficult. All that is needed is to attach a set of revisions of the repository to each theorem such that the theorem is valid in the revisions from that set. This set would be always not-empty as a theorem is accepted to the repository only if it checks out with the current version when the author attempts to add it. I think in practice it would almost always contain the current version of the repository. The only way this may not be true is that someone modifies the statement of one of the theorems that the given theorem depends on. From my experience with formalized mathematics this almost never happens. If you need a different statement, you just add one more theorem.

This is one important point.
Because whatever other way you might use it for you can think of (sharing lessons, finding better answer) readability is used everywhere. An author is also a reader even when he contributes….

I guess also the main obstacle is a system that allows incoherent states, within different realm.
The quest for a single large library wont be won and should not be fought.

A few big libraries will eventually emerge from a substrate of incoherence and partial attempts, but I doubt they will never be built upfront.