My only input is that we have at least 2-3 (depending on whether the latter two are to be considered separate) hierarchies in want of refactoring: Functor/Applicative/Monad, Num and friends, and Monoid. Ideally any solution would "solve the problem" for all of them, but even if it doesn't (and only solves, say, Monad's case), I think it should be a requirement that it at least allows for the others to be solved as well in an incremental fashion afterwards (whether by 'upgrading' the by-then existing feature, or adding a new, orthogonal one). The undesirable scenario would be where you would have to "change the world" all over again a second time to resolve the remaining problems.
Another place where this might help would be with the RandomGen/SplittableGen issue: http://hackage.haskell.org/trac/ghc/ticket/4314 If "design goal 1" is met, then clients would not have to refactor their Random instance to match the new class factoring. -Ryan