I'm suggesting that you are inviting selection bias. That does not provide any grounds to question the median aptitude of the population. It just says that anything you infer from your sample data is most-likely skewed.

Now, a skew is discriminantory (by definition it is bias). Whether or not it is <sexist> is simply a question of distribution of the mechanism responsible.

Is a program of 3.5 years of self-study (lets assume, in social isolation) an equally attractive proposition? Probably not, given that women consider doing anything in isolation to somewhat socially demeaning. We know that women, for example, self-select mates based on quite the opposite: high social standing. A long self-enforced period of social isolation is thus <rationally> counter-productive from the perspective of evolutionary biology. So, if we use this as a gating function, yes we will get biased results.

Fairness is a question of expected value. A fair bet is one where the ratio of cost/benefit is equal. So, in this case where there is an asymmetric cost, the result is uneven fairness (in this sense).

Whether or not it is right or wrong depends upon your criterion for value. By one standard - maximizing the potential peformance - it would be inefficient. This is beacaus you have mean XY and below mean XX performance represented (~regardless of proportionality -- as XX is biased down in both quality and quantity). Whether or not this inefficiency is wrong or acceptable is on the whole, just re-phrasing whether or not gender bias is (or is not) wrong.

That is beyond the scope of my commentary here.