I’ve used it to fix regressions, most recently in a register allocator for a compiler. There’s pretty much no chance I would’ve found that particular bug otherwise; it was caused by an innocuous change (one of those “this shouldn’t matter” things) clashing badly with an incorrect assumption baked into a completely different part of the allocator.

I had seen the same effect from an unrelated bug on a different program. When I added a new test and saw the same effect, I had a “didn’t I fix this already?” moment. When I saw that the previous fix was still there, I checked if an older version of the allocator exhibited the same bug on the new test, and it did not. Bisecting found the offending change relatively quickly and further conventional testing exposed the incorrect assumption.