The cases where bisect fails you are, basically, ones where it lands on a merge that does too much - you now have to manually disentangle the side that did too much to find out exactly what interaction caused the regression. But this is on the rarer side because it's rare for an interaction to be what caused the regression, it's more common that it's a change - which will be in a non-merge commit.

The squash merge workflow means every single commit is a merge that does too much. Bisect can't find anything useful for you by bisection anymore, so you have to get lucky about how much the merge did, unenriched by any of the history that you deleted.