The overall reason for my skepticism is that
1) Using a 14-bit ADC instead of 12-bit in the 1DMkII would have added very little to the cost of that camera,
2) There is plenty of room for a 14-bit ADC: after all Sony is putting thousands of 12-bit ADC's on its new EXMOR sensor.
3) Despite rather widespread 'internet arrogance' ("we know better how to design cameras than even the most successful camera making companies"), the engineers at Canon were probably quite aware of the idea of using a 14-bit ADC instead of 12-bit in the 1DMkII, and if that option really could have increased DR by up to three stops, and significantly reduced shadow noise at low ISO speeds, they probably would have done it!
According to Roger's analysis, total read noise has two components: ADC noise introduced by the analog to digital converter and sensor read noise, which is introduced by the sensor.
As far as I can tell, Clark ignores the possibility that noise from the pre-amplifier contributes to an overall maximum attainable electrical S/N ratio. Pre-amp output signal with S/N ratio of about 4000 would itself limit noise in AD output to a minimum of about one level.
Clarks units of electrons are potentially misleading: he actually has data in terms of A/D converter output levels. So what the really has is something like
ISO 100: 1.28 "levels" of noise in A/D output, which he combines with gain of 13.02e per level in 12-bit output, to get 1.28x13.02e = 16.61e as his stated noise value.
ISO 3200: 9.59 "levels" of noise in A/D output, combined with gain of 0.41e per level to get his 3.93e of noise.
At other ISO levels, the actual noise measurements in A/D output ranges from 1.28 to 9.59 levels.
So Clark's data are consistent with a noise floor in the input to the ADU of not much less than 1.28 levels with 12-bit, or S/N not much better than 3190:1, independent of the ADU's performance.
In fact, such a limit seems the most likely reason why Canon did not bother to use a more precise ADC.