I've seen this happen with Babel a few times. They've released bugfix releases which fixed a slightly incorrect implementation of a spec, which ended up breaking loads of peoples builds (this was before npm/yarn had a decent way to lock dependencies). Turns out, compilers have a really high bar for what can be considered a non-breaking change.

Generally though, I think for end user software, a "breaking" change is one that requires the user to relearn something, i.e it breaks an existing mental model.