Paredit give you multiple ways to achieve one thing. Maybe it is fine for real-time collaborative editing but for a VC-type setup it seems to me that the challenge would be in determining the correct list of operations for a given diff so that the results are unsurprising. Consider how often something simple like standard diffs are wrong (where a new function adds the end to one function and the beginning to the following function as the functions all have a similar ending)

Have you looked at tree edit? I thought it could do this, to an extent.

That all said, I am pretty bullish against this idea, at large. For many reasons, not least that I can't remember that many times it makes this mistake. I noticed it when it happens, but it doesn't actually happen that often.