I have no reason to doubt your assertion, that DxO creates a better output from 5D2 (or from Canon) raw files than ACR. All I said is, that this situation is not suitable to prove that. If I wanted to compare the converters, I would define some outcomes: how bright certain parts should be, how much can be annihilated by black point; then making the best conversion with both converters, once with and once w/o noise reduction. However, I personally am not interested now in the outcome; my interest was in the raw data (how noisy it is in reality).
I just checked again and black point is set to 0. I also used 1.2 as exposure compensation and the amount of noise is almost not altered in the output. Noise reduction is really set to 0, both for luminance and for chroma noise.
I think however that you are underestimating one point. Measuring the noise in the RAW channels (something I also do
here) doesn't say everything about the noise after demosaicing.
Many algorithms (vector median was one of the first) do not use anymore the value of the pixel itself in the bayer grid (I mean for example that if we are rebuilding the (R,G,B ) values for a green pixel of the Bayer array whose value is v, we don't get G=v for that pixel after demosaicing).
As noise is a stochastic process, this can have for example an averaging effect on noise, and thus reduce the most important and visually unpleasant values.
Noise in the output depends of course on the noise in the RAW channels, but also on the demosaicing algorithm, even if no specific noise reduction is applied.
PS: Btw, I have no special interest in DxO, I even don't like some of its features such as, for example, their perspective correction algorithm or their specific noise reduction algorithm.