a simplified, very brief explanation presenting more correct info can be found here.
I looked back again at Giorgio's post from a couple of years ago, and at the textbook link above, and I didn't really see a lot of fundamental disagreements.
I've referred others to the same link you posted above to understand the effect of quantitization error, as it does a very good job of breaking down the story behind the 1/2 LSB signal that can represent the noise that comes from it. The text loses me a bit when it jumps to the conclusion that a listener could "hear" this error in quiet passages or at the end of echo trails. Even without dither, someone would need to have the volume set to "death by audio" to quote a previous poster, in order to hear the raw error, AND be listening to a file with essentially zero noise floor or "hiss" to begin with. If you're working with any kind of analog tape hiss or less-than-perfect microphone preamp, the noise will be there, and it will do a fair amount of dithering by itself just being there. This is not to say dither is bad - it's free and can't hurt - just that it's not going to make the difference between artistry and suffering, as the text suggests.
People often demonstration quant-noise by synthesizing low-level test tones, then downconverting them from 24 bits to a lower number, but this is really going to have limited application to the real world, as these signals will have no noise, and will allow/force the listener to crank the volume to the point where what we're really doing is examining error around a 3-6 bit magnified range.