As hashing functions aren't injective, composing them leads to a reduction of range (ie, a smaller set of possible output values). As such, (bcrypt . MD5) would almost certainly be a weaker hash than bcrypt alone. Might not be enough to make a difference, but I'd consult a cryptologist before betting on that.
Yeah, it would be weaker against a collision attack. But that shouldn't be a concern as the purpose of the password hashing function is to protect against a preimage attack: finding the password given its hash. I can't see how the composition could be more vulnerable to a preimage attack than its parts. (Provided that the inner function's output is evenly distributed.)
