My personal exprience is that very few users are aware of or care about the strict accuracy of the collation sort order and other locale aspects. But they care a lot about index corruption.
So I'd argue that we should not have any breaking changes at all, even in cases where the provider is clearly wrong.
FWIW, using external ICU libraries is a nice solution for users who need strict and up-to-date Unicode support.
Cell phones do often get support for new code points before databases. So databases can end up storing characters before they are aware of the meaning. (Slide 27 in the pgconf.dev talk illustrates a recent timeline of Unicode & phone updates.)