That's not true, because the RDRAND entropy is mixed in last. So once you're under the assumption that RDRAND is nefarious, the microcode only needs to detect the rdrand-to-xor pattern to make the entire entropy pool predictable (for example: by setting the non-rdrand input to the xor operation to zero it could disable all other entropy sources).

A microcode backdoor capable of reading the existing entropy pool state is going to be a hell of a lot more powerful than a RDRAND backdoor, to the point of making a RDRAND backdoor worthless.