The ghc-devs discussion gave me some thoughts relevant to this differential. I put them as a comment in Trac #10598.

Never mind, I had a misunderstanding about how superclass dictionaries combine with newtype deriving (i.e. pretty much not at all).

Hm. What about subclasses of Typeable, particularly Data? E.g. does the following compile:

Great! I'm not going to understand all of this code, but a couple things I noticed on browsing through it:

I suspect type variables as heads cannot be so easily dismissed. Consider