With lower(), we can expect we'll get the right transformation of string A each time. If instead, we store string A, and then store string B as A.lower() and copy it... A.lower() will always be A.lower, but it's much easier for someone to come along, screw with the database, and change B.
They need to store the verbatim username in order to know how to display the username in the UI.
They need to store the canonical username in order to efficiently know whether a given canonical username is in use.
But - when the issue here is the question of the reliability of the implementation of the canonicalisation function, having it done once in python, and then again by PG is going to be a huge issue.
Well, I still think that it is better to have two (hopefully) correct fields in the database, rather than only one. (Consistency of the two fields can be checked once in a while).