FWIW, my interpretation is that the way the amplification is split between the charge converter on each pixel (or maybe pixel group?) and the final ADC is the key. When the gain up the pixel end is high, both the signal and the pixel read noise contribution feeding into the ADC are amplified. The amplified pixel read noise simply overwhelms the fixed read noise contribution from the ADC and the S/RN drops to be that of the pixel at high gain. Expressed in terms of photoelectrons, the total read noise decreases as the gain increases. If you find a better or more accurate description of the actual mechanism, would be very grateful if you would post it.
I didn't see the typo - I can't ever see my own either
edit: I guess that it is a little bit like the photomultiplier process where gain is applied up front and any electronic noise that comes along further down the chain is completely swamped.