Why does this recommendation require a RBG3 construction based on some physical entropy source to use a DRBG that is XOR'd or re-seeded by the entropy source to produce full entropy?
Remember that RGB3 is designed for the ultraparanoid types, who ask if 1,000,000 random bits, want each of the $2^{1000000}$ possible outputs to be approximately equiprobable [1]. It should be obvious that this goal would be unachievable for any rng with a fixed state much less than 1,000,000 bits.
Hence, to output an arbitrary number of 'truly random' bits, we need some additional entropy inputs as we go along. NIST has opted for two different options to provide this additional entropy input; xor'ing the raw entropy input with the DRBG output, and periodic reseeding. The first option obviously achieves this goal (xor'ing a truly random string with anything not correlated to it produces a truly random string); the second one would appear to (if we make some plausible but unproven assumptions on the DRBG).
[1]: My opinion: don't ask me what they want to do with those random bits - almost all the uses of random bits use them as inputs to some computational complexity crypto, which would void any security we would get from informational theoretical randomness.
