Ah, I forgot we didn't do within-line diffs. If we did those, then latin-1 would be less bad choice than ascii+surrogateescape. As it is, either should work in this case (since they just need to tunnel the raw bytes, and aren't being sliced at all).

I agree with most of Terry's comments, but think the case can be made this is a bug worth fixing in 3.3 (it's definitely borderline, but making it feasible to port Mercurial is a pretty big gain for a relatively tiny risk).

3.2 is about to enter security fix only mode though, so dropping that from the list of affected versions - while the fix should apply just fine, it's definitely not appropriate to make a change like this in the final planned maintenance release.