If you integrate the characteristic function of the rational numbers over any interval, you get zero because rational numbers are very rare.
So they don't exist either?
To be less glib, I don't see Feynmann's argument to be bringing anything new. We already knew that counterexamples, if they existed, would be very rare because we tried looking for them with computers and we couldn't find them. But stuff being rare still doesn't prove anything.
Many of us were fooled by Skewe's number:
https://en.wikipedia.org/wiki/Skewes%27s_number
There's no way to conclude that this exists via brute calculation. It's just inconceivably large and would have eluded any of Feynmann's methods.
So I don’t think you have produced a compelling counterexample yet (though I expect you are right that one exists).
Roughly, it goes as such:
1) the probability of N being a perfect square is proportional to 1/sqrt(N).
2) For any N_0 arbitrarily high, if you integrate from N_0 to infinity the expression (1/sqrt(N) dN), you get infinity.
3) The expression in 2) is the "Feynman equivalent" of the expected number of square numbers above N_0.
So Feynman's nonproof actually turns out to be true, despite it not being a proof in this case as well.