Just had another look at the third report and was struck by how hard the PF appears to be working the roll axis right after the disconnect. See page 29..
http://www.bea.aero/docspa/2009/f-cp...90601e3.en.pdf
In the 30-35 seconds following the AP disconnect he made 17 more or less alternate roll inputs with the aircraft rolling 10 degrees first one way then the other. Was he simply too busy or too focused on roll to notice he was climbing ?