The whole point of doing it randomly is that even if somebody figures out the salt for one user, they would still need to figure it out individually for every other user too. Using a "system", like you're doing, significantly decreases the effort required to figure out the salt.
I think it was in WWII where the Americans would intercept the German broadcasts, which were encrypted by Enigma. But because the Germans routinely broadcast the same sorts of information in the same format every day - weather reports, for example, it gave the US a known set of data of limited scope which significantly improved their chances of breaking the codes. What you are doing, especially because you're broadcasting it on the internet, is essentially the same thing.
So from my un-trained-in-cryptography eyes, the username approach is solid.
Previously I've used a hash of site-wide random string - password - user specific random string. For my latest project I've moved to phpass which is an implementation of bcrypt. My understanding is that in terms of security:
bcrypt > stretched salted hash > hashed salted hash > salted hash > hash > plaintext