Let me give a simple example to show you why. Imagine you are using your 5D2 to image a chess board from a large distance. Each chess square takes up exactly one pixel. Now let's say you use a TC 2x. Each square of the chess board now covers four pixels. With four pixels you can collect four times as many photo-electrons for each square, meaning that the S/N (and essentially DR) improves by a factor of two per chess square, or per solid angle.
I wonder if your simple example isn't mixing a metaphor (or in this case, mixing theory with practical reality). I understand the theory of 'adding pixels' with photon signal being additive while noise combines in quadrature, so the four pixels have twice the S/N over that spatial resolution. But doesn't that assume invariant illumination across the original pixel or the four 'added' pixels? In your example, the photons from that one chess square are spread over the area of four pixels (which is why the effective aperture with a 2x TC is 2 stops narrower). We're not creating a real superpixel, merely spreading out the light with diminished signal at each pixel, while each pixel still has the same read noise. So although the theory predicts double the S/N, that assumes 4 times as many photons input and twice the noise - and in your example, there aren't four times as many photons to go around. Or am I missing something?
I cannot explain why DxOMark does not find a significant difference in DR between 5D2 and 7D. I didn't find a description of how they measure DR in detail, so I don't know what their numbers mean. It also stands in stark contrast to what is found on Clarkvision, which is much closer to my expectation. Perhaps DxOMark doesn't measure DR per pixel, but per sensor area or something. That would make sense. That a large 5D2 pixel would have the same DR as a small 7D pixel definitely does not make sense, so they must mean something different.
I don't see any major discrepancy between DxOMark and Clarkvision, provided you carefully interpret Roger Clark's excellent and informative site. If you look at Figure 4, the 5DII has a 14.7-stop DR, and the 7D has a 13.2-stop DR - a 1.5-stop DR advantage for the 5DII. That's fine, but Figure 4 is based on Table 2, the sensor specifications - i.e., it's modeled/calculated data, not real, empirical
data. Some of my colleagues have a word for that sort of data, and while that's a bit harsh (even though I won't repeat it here), there's some truth to the idea. Skip down to Figure 5b, which is measured
DR for a few cameras. The 7D is not on the plot, but it's cousin the 50D is (the theoretical data in Figure 4 place it at ~13.4-stops of DR, close to the 7D and still a 1.3-stop advantage for the 5DII). From the empirical data, you can see that the 5DII's measured DR at ISO 100 is ~11.3-stops, and the 50D's measured DR is ~11 stops. The stated explanation for the differences between Figs 4 and 5b, "The dynamic range is often limited by the A/D converter and other electronics in the system, illustrated by the measured data falling below the model at lower ISOs
," is not really satisfactory, since both cameras are falling below 12-bits, yet they both have 14-bit A/D converters. Looking at Figure 5b, you can appreciate at low ISOs, there's very little difference between the 5DII and the 50D in actual, tested DR.
Furthermore, it's worth noting that the measured DR's in Figure 5b are much more in line with DxOMark's data - which is consistent with the fact that DxO is empirically testing the cameras and lenses (although it would be nice if they provided some details about the testing procedures!).