Let's discuss a little more. Suppose you have a scatter graph with the X-axis being the parent's IQ and the Y-axis being the offspring's IQ, and suppose that the correlation is less than 1 and greater than -1. Actually you can use any variables' correlations, not just IQ's. If you choose the parent's IQ as a predictor of offspring's IQ, there will be a regression towards the mean. BUT, if you choose the offspring's IQ as a predictor of the parent's IQ, the parent's IQ will also show a regression to the mean. Again this applies to a static data set, not to generational differences.
Let's try a little reductio ad absurdam argument here. If regression to the mean occurs only to the offspring, then selective breeding cannot occur. Clearly this is false. Therefore, the premise is false.
I would love to have this discussion directly with Stephan, as I would like to disabuse him of his false understanding of this concept.