I don't know the reason of the difference from the "ideal" result; I know nothing of the hardware, but I think the "read noise" has several sources and I guess not all are constant.
Seems so, here's one explanation whyhttp://astrosurf.com/buil/5d/test.htm
And some more data herehttp://astrosurf.com/buil/50d/test.htm
A few points worth noting imho
Buil's protocols are the correct ones, from the point of view of maximizing SNR, but hardly practical for standard photography (who wants to substract master biases made of 20+ frames from a shoot of a building at night).
Maximizing SNR involves, at the most basic level, minimizing noise whatever the origin is, and getting as many photons as one can for a given exposure. In other words, this is exposing as far "right" as one can without overexposing and staying in the sensor linear response zone if photometric measures are to be made, probably of little practical interest for photography... except when testing to the limit for the sake of it.
The behaviour of cameras changes somewhat as the sensor heats (dramatically as far as the thermal noise component is concerned). It may seem again like a non-issue for photography, but if one does extensive testing consisting of almost continuously shooting, live view usage, etc, the sensor will be warmer after a while if there's no cooldown period between shots and basic camera use.
Camera manufacturers use lots of tricks to minimize noise, this is why Nikon cameras, with their "artificially" low noise are not suitable, or at least the best choice, for astrophotography. The Nikon strategy could very well be the best for everyday photography, but it also means that the ideal "ETTR" strategy will vary with the camera model just as the ideal ISO, as far as maximizing SNR is concerned, varies between cameras of the same brand.
And beyond deep pixel peeping, the factor that matters most, in my observations and for my purpose, is heat. The difference in image quality between a warm sunny afternoon and a colder morning is striking. Keeping the camera in cold storage, if practical, beats ETTR by a wide margin. But of course, one can try to do both.