It's important to note that the Wikipedia page specifically says that using two hash functions in case one is broken is a valid use. It's also important to note that the assumption of a corresponding collision in B if a collision is found in A is most useful if A and B are not orthogonal but have a common basis. For example, using SHA-1 and MD5 is not smart as they have common elements. If A and B are completely orthogonal, then knowing a collision in one should tell you nothing about whether a corresponding collision exists in the other.
The effective strength of two hashes can never be greater than a hash with an equal number of bits to the two combined, assuming all three hashes are orthogonal and have no known weaknesses. One of the problems with simply adding bits to a hash is that doesn't guarantee it will be any stronger. If the bits are poorly generated, it might even weaken it. This could happen if the extended hash is insufficiently close to random and information is exposed.
If, however, you're already working with hashes that are as long as you can usefully generate them, the combined strength of the hashes (again, assuming there are no common elements) must always exceed a hash of twice the length because the algorithms are no longer safe in the extended form, creating an unnecessary weakness in addition to any weakness that might be inherent in the algorithm anyway.