I wouldn't have thought I'd dare to intervene in a MFDB-vs.-MFDB thread... but as this is mostly about dynamic range I'd throw in the 2 (euro) cents of a cheapskate only shooting with a goodol'300D (yeah, the original DRebel ; some results in the link below in my signature).
First, the theoretic definition of sensor DR as difference between saturation and noise floor is theoretically correct... But the difference between theory and practice is greater in practice than in theory, isnt' it?
Where perception kicks in is in the noise structure : with the same measured Signal to Noise Ratio, a perfect gaussian noise on one side will be far more acceptable, photographically speaking, than the same noise
only made of banding (that fugly orthogonal pattern in noise, very evident in some cases with my camera).
I'd really like to stress that one :
banding noise shall also be measured in a way or another, besides noise. That can also be applied to other structures of noise, as eg the "noise blotching" (color noise splattered on a few pixels) of some P&S sensors (though that is a bit farther from the topic).
Second, I'd say there is yet another factor in real-life DR, especially when dealing with large ranges above 10 stops :
lens veiling (sometimes called flare - I'm not talking about the colored artifacts showed with a point light source near or in the field, but about the general contrast reduction caused by light scattering in the lens or sensor chamber).
This parameter can be offset by shooting the same low-contrast target in a dimmer environement with + and - exposure correction, but in a real world image, the whole thing is about to capture shadows and highlights in the same shot, so veiling can also be there.
In a first time, lens veiling could improve the ability to record high DR scenes, as it levels up the shadows (it was a darkroom trick in the form of preexposure if I remember well).
However, with deeper shadows, it doesn't only level up the tonality but decreases contrast, to the point that initial texture is drowned in a grey veil, and so it could also limit DR.
I've encountered this phenomenon while playing with modern raw files (see link below) : sometimes while pulling up the shadows, it only brings an uniform grey
before bringing up objectionable noise.
Third, about marketing claims vs. facts : the "playing with raw files" was actually
a game on a french forum (link in french with MANY side discussions), where someone (some here may know
Thierry Legault?) took some raw files of various origins and zeroed the N least significant bits, effectively converting them into 14bits to 8bits raws.
Bottom line : no reasonably visible difference past 10 bits, no pixel-peeping significant difference past 12 bits, even with MFDB files (MFDB pages 7-8-9, note that the first test on page 1 had a 2bit bias). There is just plain noise in the last bits,
as said elsewhere. However, I suspect now that in some cases the veiling might limit the DR too much.