We really don't want to unify (a b) with (Eq a => ty).
The ever-ingenious Oerjan discovered this problem;
see comment:101 in Trac #9858.

See Note [Decomposing fat arrow c=>t] in Type.hs