the MPTC Dilemma (please solve)

[Claus Reinke]

continuing the list of odd cases in type class handling, here is a
small example where overlap resolution is not taken into account
when looking at FDs. 

context: both hugs and ghc resolve instance overlaps in favour 
of the most specific declaration.

so the following works in both ghc (darcs, 25022006) and 
hugs (minhugs 20051031):

    {- both ghc and hugs accept without 3rd par and FD
       neither accepts with 3rd par and FD -}

    data T = T deriving Show
    data F = F deriving Show

    class    TEQ a b {- tBool | a b -> tBool -} where teq :: a -> b -> Bool
    instance TEQ a a {- T                    -} where teq _ _ = True
    instance TEQ a b {- F                    -} where teq _ _ = False

    test = print (teq True 'c', teq True False)

and both print "(False,True)", so best-fit overlap resolution entirely 
determines which instance to choose! 

now, if we uncomment the third class parameter, they both complain 
(as expected, though about different things).
 
however, if we also uncomment the functional dependency, to fix 
the ambiguity wrt the 3rd parameter, both complain that this FD 
is in conflict/inconsistent with the instances!

as far as i understand it, the potential inconsistency should have 
been eliminated by the best-fit overlap resolution, so I submit this
is a bug (and unlike the earlier example I submitted to hugs-bugs,
this fails with both hugs and ghc; and it is less open to alternative
interpretations, I hope).

cheers,
claus