"on one hand, it will give more levels at each subject brightness level, and these extra gradations will then be preserved when one does compensating EC downwards in 16-bit mode,"
Please explain how the extra gradations will be preserved when you do compensating EC downwards. There must be something I'm missing here.
OK Don, here is my reasoning, let me know what you think.
In brief, it is due to the extra accuracy allowed by the extra four significant bits used in 16-bit editing.
At great length,
a) Suppose that a one stop subject brightness range down in the shadows is placed by traditional exposure metering at -8 stops (8 below the maximum brightness handled by the A/D converter) and so 12-bit D/A conversion give you only eight levels over that one stop.
Increasing the exposure by two stops puts it at -6 and gives you 32 levels over the same one stop range of subject brightness.
c) Converting to 16 bit linear increases the number of numerical levels 16-fold, so those 32 values are scattered over a range of 512 (15 as yet unused levels have been added in between each level output by the D/A converter.)
d) If you then correct downwards by the same two stops, dividing each of these value by four, the range is compressed to 128 levels; still a gap of four between each recorded level: no distinctions of level are lost. In other words, each full stop of downward EC discards one bit at the least significant end of the 16-bit number, but since originally the last four bits were meaningless zeros anyway, you can go down four stops before arithmetic round-off loses any read information from the A/D converter. EC by other than whole stop amounts will make the numbers more complicated, but the rounding error from the 16-bit arithmetic will not be as much as the A/D discretisation error until the drop is four stops (a factor of 16) or more.
e) Gamma compression increases the number of levels in each f-stop range in the shadows, so the situation is even better if EC is applied after gamma.
Another point: as Don has mentioned in a different way, if you simply EC back down and do a straight print, these deep shadows will be off the bottom of the scale of any print anyway; the only point to making this effort to accurately record these shadow levels is to keep them somewhat above their natural level, by contrast reduction or whatever. So if you are wanting to print those shadow details, they should not in the end be EC'd all the way back to where they came from.
(I sometimes teach numerical computing, so I hope I have got at least this stuff about rounding, bits and errors right, but I could more easily be wrong about the many other aspects of this situation, and in particular I make no claim to any expertise about the ultimate aesthetic value of the different methods.)