The article doesn't explain what is actually going on to produce the second chart, but here is a guess. Assume people have a true ability x, an estimated ability y which is x + some unbiased noise, and a test score z which is also x + some unbiased noise. If you take a sample and collect the lowest quartile of test scores z, then the average of the estimated abilities y in that group is higher than z because the y are produced centered at x rather than centered at z. By collecting the lowest quartile of test scores, you didn't just get the dumbest people, you also got the people who happened to perform badly on the test that day. These people may be estimating their ability accurately, in which case the estimate would be higher than their randomly bad performance that day.

This seems like too obvious a mistake to not have been noticed for this long though.