There is not always regression to the mean downward. Keep in mind that all these analyses are on group data. Any individual data point can move up or down, depending on the sources of error in the respective administrations. If you consider the discussion on error, you'll note that he mentions at least three different kinds of error, any of which could contribute to either underestimates or overestimates of "true" scores.
Also, regression to the mean in the left hand tail means that extremely low scores tend to regress upward, toward the mean. Although apparently not quite as much as very high scores, if one consults the available research on that.
