When commits from two or more parents have to interact in order to produce a bug, I don't think there is a bisect algorithm that can actually identify that bug. There isn't enough data to automatically merge prefixes of the merge history from respective parents in order to test them against each other. Also, the search space in this case becomes the product of the number of commits from the LCA to each parent. So I really don't understand how we messed up.

I'm just glad we have "--rebase-merges" now because when you have two git servers that are actually important enough not to rebase their history, you do need merges.