When inferring roles it is Much More Kosher to work on the source
type, as written by the user, rather than the representation type as
computed by GHC.  Error messages may be better and, more subtly, the
representation type is the result of a pretty complicated calculation
and I'm worried about accidental cycles.