It's a self inflicted situation by dependees. Tools in widespread use, JupyterLab for instance, has users tossing dependencies onto the pile, offering next to no guidance on dependency version control, as if expecting them to have to ponder such issues is unreasonable.
Well, if that's how you're going to behave, foregoing entirely obvious and solved engineering problems, you get to glue all the pieces back together when reality asserts itself. I don't imagine for one minute numpy folks didn't know what this would look like, and they did it anyway. Good on them. Breaking changes are necessary. That's unqualified. Necessary. If you can't afford a big blowup in your work then manage your work.