borjam, you did a nice job of clarifying the issues, separating frequency response (as a function of sampling rate) vs. dynamic range (as a function of word length / bit depth). Many people confuse these two issues and/or lump them together as "resolution"--a term which is perfect for misuse, since it sounds as if it must mean something definite, but has no agreed-upon meaning in fact.
I have two notes which aren't disagreements, but I hope clarifications of what you wrote. One is that in a properly dithered system there are no "stairsteps". That issue goes away totally, completely, 100%. It isn't relative to the strength of the signal, or the number of bits per sample; it "isn't merely really dead, [it]'s really most sincerely dead." Anyone who still thinks in terms of "more bits = smaller stairsteps" (e.g. 80% or more of "audiophile" opinion, still, after all these years) should find themselves a different concept, because that one has fallen and it can't get up. You are exactly right to equate bit depth/word length with the noise floor and nothing else.
The other thing is that there's some strange hand-waving going on in this thread about low-level signals needing to have the same s/n ratio as high-level signals. That makes no sense at all, as a few people have pointed out--but the hand-waving continues.
Any real-world noise floor shifts in level on a moment-to-moment basis. That's what noise is. But it's imperative that the shifts be random, and not correlated with the signal levels. Otherwise the noise tends to merge audibly with the signal, becoming a form of distortion known as modulation noise (or the phenomenon in general may be called "noise modulation" because the noise is being modulated by the signal). Whatever you call it, it doesn't sound good. It's sometimes called "granular noise" or described as "gritty" or "dirty" or "sandpapery" sounding, especially at the lowest signal levels, e.g. the way that the last moments of reverberant sound die down in an otherwise quiet hall.
(Somewhere I think I still have a Columbia CD of a piano concerto in which you can hear a "frying" sound each time a note decays into silence if you turn the volume up. Someone evidently didn't adjust the A/D converters in their PCM-1610 before the session, or maybe they left the dither switched off, or maybe both.)
We could talk about the issues of floating point representation near the noise floor, but I think that would leave a lot of audio-only people out. Summary is that it's not worth arguing about--as long as you don't inadvertently create modulation of the noise floor by the desired signal.
But given the marketing-type claims that I've seen so far, it's not clear to me whether this principle is being observed or not. If it is, I wouldn't expect to see anyone touting the advantages of potentially having more bits in the mantissa (the magnitude portion of the floating point representation) at and around the level of the noise floor, because that just doesn't matter; the roundoff error is below audible significance in relation to that noise floor, and rescaling the exponent just to get more bits into the mantissa in such cases is an illusory pursuit at best.