I read on diyaudio about someone trying to blind test speaker drivers, and they had to cancel because all they learned was that they couldn’t consistently hear ANY difference between ANY of the drivers…
That's the wrong approach, at least from a scientific point of view. Assuming that the experiment is flawed when its outcome contradicts what is expected is definitely not the right way to do it.
If you have an hypothesis, devise an experiment to test it, perform the experiment and the results contradict the hypothesis... the conclusion is not to cancel the experiment, it's to try to figure out what's going on.
I'm not arguing that the experiment might be flawed, it can definitely be the case, but it can also be that the experiment is well designed and one of the assumptions in the hypothesis is simply wrong.
My favourite analogy: Look at a bowl with 100 M&M’s. Now look at a bowl with 99 M&M’s. Can you tell which is which? Probably not. Now eat 100 M&M’s, and compare with eating 99 M&M’s. Can you tell the difference? Probably not. Now find 100 5-year olds and tell them they can each have one M&M. You should now have a very noticable auditory response letting you know if you had 100 or only 99… you changed the test conditions, so now the difference matters.
If you cannot tell which is which, the difference is irrelevant from the point of view of "fidelity". I'm not sure how the analogy works with the "splitting" compared to an audio sample, you suggest comparing e.g. different frequency components? Different frequency components can lead to the same perceived composite audio sample, especially given that human hearing has different responses to different frequency inputs.