That would be my guess, too.
Data gets scaled down, then sent through the pipe.
Depending on wether that one pixel to create a supernova – and one is enough, I’ve tried – is among the scaled down data will determine the outcome of the preview. But scaling algos that emphasize extreme values usually look really ugly and pixelated, so the choice is not that wrong.
It’s just the tool chain that allows for hidden nasties.
As I see it, the technical aspect is correct while the usability aspect is all wrong. But usability was never really a strong suit of AP, anyway. Probably wasn’t scientific enough. I had discussions about that with him way too often.