I have trouble understanding how guard works. Why does it type check? Isn't mzero capable of returning some m a where a /= ()?
guard :: (MonadPlus m) => Bool -> m ()
guard True = return ()
guard False = mzero
mzero :: m a
Yes, mzero is capable of returning some m a where a /= (). But it's also capable of returning m (). guard uses it in this second case.
It's similar to this:
n :: Int
n = 5
5 can be a Float or Double, but can also be an Int. The compiler chooses the needed interpretation of 5 during type checking.
Similarly, the compiler chooses the right type for mzero in the original example during type checking. More precisely, it sees that a m () is needed, so it chooses that type.
The important bit here is that
mzero :: MonadPlus m => m a
actually means
mzero :: forall m a . MonadPlus m => m a
which states that the caller of mzero gets to choose the actual values for m and a (as long as m is a MonadPlus). Hence, the caller can choose a=() to make things type check. This choice can be made by the user through a type annotation, otherwise the compiler will try to infer the correct choice during type checking.