I really like your analysis. Ive been thinking about this subject for a week or so. My main thoughts are, what are the negatives of picking a “low” white point value. By low a mean slightly lower. In your example, lets say 15600 instead of 15830. It seems to me at worst, we would be telling HL module and maybe other modules that values above here are clipped when maybe a few pixels are not. The HL algo will still use them ( I think) and will the be ok.
So in your example, using the same simple steps, would you get a significant different image if the white point was set to 15500? what about 15000?
I think we could all agree, too high is a problem since HL module will not try to recover it.
One problem is that the restored region will effectively expand as the cutoff is reduced. Usually when restoring you want the closest non-clipped edges possible to the clipped area to get a good estimate.
A second issue is that the cutoff itself is an estimate. If you treat the clip as a region with no gradients (i.e. flat), then quite often the cutoff will not match that. For example below is the gradient view of one channel of this image, which doesn’t quite match that obtained from a cutoff alone (clipped regions are solid fill):
