Currently, there can be multiple backouts per push, which makes classifications less precise (a classification pointing to a given backout push, could refer to any backout from that push) and so it makes things harder for our algorithm (in https://github.com/mozilla/mozci) to find regressions.

OTOH, doesn't more backout changesets mean bisections would be more likely to hit broken trees?

Glandium's concern is valid. I don't think this is worth pursuing given that and other improvements we could make instead.