urandom only implements computational security, whereas random makes an attempt at information theoretic security (albeit an extremely poor one). Computational security is so practically useful that the difference between the two seems moot, but there still is a valid distinction.
To me, XOR using /dev/random should be usable for a true OTP [0] whereas XOR using /dev/urandom has the security of the effective stream cipher.
[0] (malleability disclaimer, lest you think I'm arguing for the practicalities of "OTP" or something)
We can define our measure of entropy such that the output of AES (/SHA1) qualifies, or such that it does not. The way it is used in the random/urandom distinction, it's clearly not meant to include PRNG output.
They should be completely partitioned pools, and collected entropy that goes out random should never be mixed into the urandom pool. Of course it's likely random still wants whitening, likely resulting in the use of a computationally-secure algorithm.
Practically since it's anyone's guess what the random clients expect, it's probably the right thing to `ln -f urandom random`. Bona fide hardware sources can come up with their own device/daemon to administer and/or stir into the kernel P/RNG. I'm just pointing out that there is a worthwhile distinction they were trying for, despite the engineering pitfalls of trying to make it available!
First: while I don't know what you mean by "teasing out differences with respect to various oracles", the fact that you have a kernel CSPRNG that your system relies on means that diversity in CSPRNGs is a bad thing. Simply: alternatives to the LRNG lead to multiple single points of failure.
Entropy is not "scarce". In fact, it's infinite and renewable. To generate a practically unbounded number of random bytes from just 16 bytes of starting state entropy, just use those 16 bytes as an AES key, and use it to encrypt an incrementing counter.