>Berkson's Paradox seems to rely on the selection criteria being a combination of the two traits in question

100% correct. For traits x and y, selecting for datapoints in the region x + y > z will always yield a spurious negative correlation for sufficiently uncorrelated data, since the boundary of the inequality x + y > z is a negatively sloping line.

>But in TFA, surely the "high performance" selection filter applies only to the adult performance level?

Doesn't seem that way. Reading the full paper [0], they say:

   In sports, several predictor effects on early junior performance and on later senior world-class performance are not only different but are opposite. [...] The different pattern of predictor effects observed among adult world-class athletes is also evident in other domains. For example, Nobel laureates in the sciences had slower progress in terms of publication impact during their early years than Nobel nominees. Similarly, senior world top-3 chess players had slower performance progress during their early years than 4th-to 10th-ranked senior players, and fewer world top-3 than 4th- to 10th-ranked senior chess players earned the grandmaster title of the International Chess Federation (FIDE) by age 14.

[0] https://www.kechuang.org/reader/pdf/web/viewer?file=%2Fr%2F3...