If you go all the way and make it sound with variance annotations, banning asserts & under specified types, etc. then you end up with a language that sucks to use in practice.
If you go to the other extreme, you don’t have static types at all.
If different parts of your code are typed to varying degrees, or you use a lot of unsound types/inferences, maybe there’s a way to assign a confidence score to every type to indicate how accurate you think it is. Maybe there’s a clever way to statically compute that score, or you can instrument some % of prod traffic at runtime to observe types with lower confidence (Node has this, or see MonkeyType or Reticulated Python). Then take that info and feed it into Closure compiler.