PaulS:
I change the reference voltage BEFORE doing a conversion, so that's not a problem.
So, the conversion is done AFTER the reference voltage change, which the data sheet says may produce incorrect output.
So, what is the problem when this actually does happen?
The problem is that without the delay, the A/D results are too low. I read the voltage source (an LM-35 temperature sensor) and figured what the approximate A/D result should be. It was way low. After adding the delay, the result got closer and closer to correct. At around 80 uSec it stopped "improving" and 100 uSec is as far as I can reliably go without crashing the stack.