I think the best option at the moment is to break out type classes in their> own packages. That's what I did with hashable. Indeed! I greatly believe in this mantra now. Really, my point was only this banal one -- packages with only interfaces in them have no dependencies and are much less likely than implementation packages to break when the GHC version changes.
Aside: The problem with collections is that we don't have the programming language means to do this well yet (although soon!). The issue is that we want to declare a type class where the context of the methods depends on the instance e.g. class MapLike m where type Ctx :: Context -- Can't do this today! insert Ctx => k -> v -> m -> m Java et all cheats in their container hierarchy by doing unsafe casts (i.e. they never solved this problem)!
Ah, interesting. Is there a proposal to do this? While we need to avoid an infinite regress of generalization and abstraction (type programming => kind programming?, etc) -- it does seem like class contexts on types and type class instances themselves would be nice to have *some* control over. (In the above message, for example I was having trouble due to not being able to hide instances on import.)
I would hope that we could get all the important interfaces into the Haskell Platform eventually (and have all packages there use them).
+1! Cheers, -Ryan