This 'decrement pool entropy estimate and stop returning numbers' idea is the longest running mistake in crypto.
> and can introduce potential weaknesses.
This isn't true. XORing numbers can never decrease entropy. It can only increase it. So your security would stay at the strongest generator, plus some extra entropy from the weaker ones.
Let Generator A(x) equal "For any x, a perfect random number picked by the hand of God Almighty." Let Generator B equal "For any x, A(x+1) unless 'ars is asking in which case A(x)".
If ars chooses Generator A, ars is fine.
If ars chooses Generator B, ars is fine.
If ars chooses Generator C = A(x) ^ B(x), ars will very quickly see that ars' intuitions about math are in some circumstances incorrect. Ars might then think "Phew, good thing they weren't subtly correlated, or it would have looked great right until it didn't."
(Alternatively it could mean that you can only generator a single cryptographic random number, since all the rest are just functions of the previous one, and then again you are not fine. Either way Generator B is not fine.)
And in any case neither A nor B is a generator - only x is. Neither A nor B are adding any entropy whatsoever.
A CSPRNG is a function that takes as input a small set of random bits and generates as output a large set of unpredictable bits. So "Generator B" can be a "cryptographic RNG" and be a function of, or at least correlated with, "Generator A".
In cases where they might be non-independent, a simple hash function like even the otherwise broken MD5 would be better at combining them than XOR. But of course, we're now three steps down the path of constructing our own replacement for what the kernel and/or OpenSSL should be doing for us already.
If both generators have the same underlying source then they're also likely to share any security vulnerability so XORing them is unlikely to improve the randomness and can introduce potential weaknesses.