For the predicted rating formula, the average rating of user u, should we take all the items average? Or only the ones that have ratings in common with user x.

In the example given, if user u is Meg, do we calculate the average rating for Meg by: (2+4)/2   or   (2+4+3+5)/4      or (2+4+3)/3

r_u average is the one you use to calculate corresponding W_x,u

In case of user Meg, the  r_u average = (2+4)/2.