I have seen the arguments for both and neither are convincing.
Ask yourself "why do semiconductor manufacturers go to the expense of creating a precision 1.024V reference?"
Does that help?
Not really. No. It's a hardware issue. We know that over the range of possible input voltage, the 1024 possible fixed output levels looks like a staircase if drawn on a graph. There are several different ways that the graph can be drawn, sliding it the right and left a little.
Furthermore, with 1024 output levels, there are only 1023 vertical steps between them. If 0 == 0V and 1023 == 5V, then the height of each step is ( 5V - 0V ) / 1023.
If 0 == 0V and 1 == 5 mV, then at what input voltage does the transition from an output count of 0 to an output count of 1 occur ?
at 0.0000001 V ? at 2.5 mV ? at 4.9999 mV ?
and at what input voltage does the transition from 1022 to 1023 ( the highest possible value ) occur ?
If I draw a diagonal line on my graph which touches each step at the outer edge of each tread, or the inner edge of each tread, which of those lines intersects the origin ?
The answers to these questions would show whether a denominator of 1023. or 1024. is appropriate. I don't see how your precision voltage reference is relevant, at all.