This feels far better than the current model used in most languages. If you've ever had struggles creating an uberjar you know this pain.
E.g. repo managers like Maven central/etc. should use binary API analysis to reject any jar upload that has breaking changes.
My only hesitation is that, AFAIK, semantic import versioning has never been tried at scale, so having to constantly bump imports from "com.foo.v1" to "com.foo.v2", and deal with "app1 wants to pass com.foo.v1 objects to app2, but it expects com.foo.v2 objects" might introduce more pain than expected.
Granted, right now app1/app2 are blithely passing around "com.foo" objects that may/may not be compatible, but if it's an 80/20 thing, or 99/1 thing, and most of the time you get lucky and it works, perhaps that's good enough.
But would be great to have go be the first community to try this at scale and see how it goes. I like it.