Can you sketch your levels over my cartoon? I'll pretty them up an update (if we'll all agree)
Sure! Although I'm not sure what my sketch can really add - and apologies if I'm misunderstanding what you're asking me to sketch. This image circles your diagram's label showing that the undithered signal covers 13 bits/80dB of range. It also draws arrow lines and labels the peak level and noise floor of that undithered signal: approx. 93dB peak, approx 13dB floor = 80dB as the label says.
Not to be pedantic, but 13 bits isn't actually quite 80dB. To my knowledge it comes out to 78.26dB (6.02dB per bit, AFAIK - happy to be corrected if I'm wrong). Conversely, my prior comment that this is 16dB less than 16-bit redbook was also slightly off: it's really 18.06dB less (3x6.02)
At any rate, for everyone else reading, I think it's worth re-stating and re-clarifying that your diagram, per your earlier comment above, is meant to represent what happens with an "MQA CD," in other words an MQA file in a container format that only has 16 bits' worth of depth, yes?
In that event, and if your diagram is correct, the 3 bottom bits are not simply "dither," because they cannot contain the randomized and/or noise-shaped info dither usually contains - some portion of those 3 bits are needed to contain MQA flags and codec information.
What I'm not clear on is whether or not, in this example, those 3 bits also contain the lossy-encoded ultrasonics we can see in the diagram: There's a bump-out in the red box, from 22.05k to 44.1k (because this would be an 88.2k MQA file not a 96k one), and that bump-out includes ultrasonics only up to about 55-60dB below peak.
My understanding has been that when it comes to MQA CDs or any other MQA file in a 16-bit container, those ultrasonics are simply not encoded - they are discarded because there is nowhere to put them. My understanding has been that only a 24-bit MQA file can accommodate those ultrasonics. That in turn makes MQA CDs completely useless as they sacrifice the bottom 3 bits only for the purposes of containing a flag telling the DAC to use a particular reconstruction filter.
On the other hand, if I am mistaken and MQA CDs/16-bit MQA files put some or all of those ultrasonics in those lower 3 bits, then that makes it even worse because it only adds to the non-random quality of the noise in those lower 3 bits, further degrading the effective noise floor/bit depth.
Regardless, this diagram is a nice illustration of MQA's underlying conviction: MQA is based on the claim that it is worth reducing the unmolested bit depth by 3 bits/18dB in order to preserve low-level ultrasonics. Personally I don't see that as a worthwhile tradeoff.