Software is too complex to document all intended behaviors. Any meaningful library will be too complicated for that. Software is hard and semver seems like it makes it a little more manageable, but practically it's not a lot better than single numbers because software is too complex to fulfill semver's guarantees.
If any program that used the previous behavior is still correct, then it's a non-breaking change. Despite what you say, this is a very common thing. Adding a new function, for example, usually fits here. But that is always verified orm the point of view of documented behavior, never looking for actual behavior.
If people are relying on undocumented behavior, they will have problems upgrading their libraries. They should be expecting it, if they are not, it's their problem not the library maintainer's. Wether to depend on undocumented behavior is a decision made by the library users at their discretion by weighting their opotions, it shouldn't impact upstream.