The idea is that you have two messages, m1 and m2, or m1 and m1' if you prefer, and you vary bits in both until you get a collision. You need some area of m1 and m2 that doesn't matter for the application, so that you can change those bits and find a collision. Since all bits of m1 are supplied when entering the password, you have no ability to modify it without getting the user to change his/her password.
If you could collide any arbitrary m1 as it's given to you, then attacks like fake certs with signed MD5 hashes could create the fake cert after submitting it to the CA and getting the signed cert back, rather than before.
Also, the collision process requires knowledge of m1 so you can see the intermediate hash states. If you know m1, the password/passphrase, why are you trying to find a new m1' that hashes to the same value rather than using the pass you already know?
An attack of concern for using MD5 as a password hashing step would be a first preimage attack. [2]
[1] https://www.google.com/search?q=md5+collision+block+birthday... (first link at present is http://www.win.tue.nl/hashclash/SingleBlock/ )
Let's say that "foo" and "bar" are two distinct passwords that have the same MD5 hash. Then bcrypt(md5("foo")) == bcrypt(md5("bar")), regardless of how bcrypt("foo") compares to bcrypt("bar"). By pre-hashing with MD5, you have added possible collisions that weren't there previously, and those collisions remain regardless of how many more hashes you pile on top.
It seems to me that the most obvious problem is that you get two chances at colliding — once with MD5 and once with bcrypt. But bcrypt is not known to be especially vulnerable to collision attacks, so this setup is probably not noticeably worse than MD5 alone. But that's just looking at probabilities — I ain't no fancy crypto expert or nothin', so there might be much more subtle vulnerabilities than the added chance of collision.
Yeah, that's all I'm saying. I was answering a question about being "fundamental worse," and fundamentally, there are now two sources of potential collisions instead of one. In theory, that's twice as insecure! However, the practical effect is unlikely to rise above absolute nil anytime soon.
As a practical matter, we can basically say that never happens. Certainly not for passwords that are user selected and not designed to collide. And since the hash itself is hidden by bcrypt, the attacker won't know md5("foo") even if they were inclined to find a "bar" with the same hash.
Cryptographic pseudorandom number generators also collide and produce cycles. MD5 hashes of arbitrary ASCII strings are reasonably modeled as random numbers, and the concern you cite is unmeaningful.
An upper limit on string length suddenly makes some sense.