Then why is he wrong so often?
I don't think he's wrong as what he says matches other sources on the matter.
The entire point of raising ISO is to decrease read noise.
You don't decrease the read noise by raising ISO at all, that's the point. You can raise ISO after ADC by multiplying the digital output, or before ADC by analog amplification. Both ways, the reading has already happened so you can't reduce the read noise by raising ISO.
Moreover, after raising of ISO, the resulting value has both signal and noise amplified. That's why high-ISO images are literally noisier.
You just contradicted yourself.
Well capacity depends on the pixel size, but also on the quantum efficiency, so the size isn't the only factor. Newer sensors may have better QE, that's why comparing sensors of different generations from different manufacturers isn't relevant.
Shouldn't be, since read noise drops with well capacity, keeping DR constant when you change pixel size. That's why pixels that are 1/20th the size of other pixels can still have the same DR. If pixel size was a key driving factor, reducing pixel size by a factor of 20 would change DR, no matter what. It doesn't.
DR isn't kept constant at all because the read noise is actually low and almost constant, worst case linear, and well capacity grows quadratically as the pixel size grows.
Also apart from the read noise there's a very significant digital conversion noise. As your well capacity drops, at some point you simply don't have enough electrons to feed the full range of ADC. If it's 14 bits, you need 16384 different values from electron counting, and very small pixels on small sensors just don't have enough capacity to feed the full range even for 12-bit ADCs.
