> "Passing the password through MD5 reduces the complexity to 128 bits, you can't get that back."
Assuming that the new hash is secure (and sha512 is generally agreed to be secure), then, given a specific sha512 hash, the original MD5 hash can only be determined via rainbow tables, which is a Big-O operation. Even though entropy is reduced, it's still a significant work to determine the original MD5 hash (significant in this instance being longer than the heat-death of the Sun, given current extrapolations of computing performance).
Attacks against MD5 are based around knowing the original MD5 hash. In this instance, the original MD5 hash is unknown, so there is no mathematical shortcut to finding a collision.
In this case an attacker isn't looking for a collision (which would mean creating two passwords with the same hash, and what hash that is doesn't matter).
The attacker needs a password with a specific hash, and the best reported attack for that is around 2^128.
Attacks against MD5 are based around knowing the original MD5 hash. In this instance, the original MD5 hash is unknown, so there is no mathematical shortcut to finding a collision.