As I understand it, there are three broad types of noise; photon noise; dark noise and read noise. It's the additional read noise (the process of quantifying the electronic signal) which might be unavoidable if one substitutes fewer bigger pixels for a lot of smaller ones.
you make an important point: perhaps you read the Roper site more thoroughly that I did! I only considered the main sources of noise that affect the electron counts recorded in the wells of the sensor, photon noise
and dark current noise
, but not read noise
from the subsequent analogue amplification before A/D conversion.
Perhaps it is time to add one mathematical fact; for constant exposure time, aperture ratio, and sensor size, photon noise and dark current noise scale with the square root of photosite area, and so does discretization error in the A/D converter. Any noise source that scales with this square root behaviour is cancelled by binning or downsampling, because the S/N of combined data decreases in proportion to the square root of the number of values combined.
But it is possible that as photosite size increases, the read noise increases more slowly than this square root pattern, and if so, increasing photosite size can give a visible benefit. In this situation, increasing site size enough would lead to read noise becoming distinctly smaller than the other sources, so noise would then be back to roughly "square root" scaling, at which point further site size increases would have no further benefit for image quality.
So considerations of read noise (and that "well fraction" issue) will indicate a natural site size, below which noise problems get noticably worse; above which there is no significant benefit. Clearly this size limit decreases with time, as chip feature size continues to shrink and amplifiers improve, and that trend probably underlies the steady shrinkage of site size across the whole spectrum of digital sensors, with 9microns the current upper limit even in US$30,000+ digital backs.
I speculate that Sony might have pushed a bit beyond this limit with the 2.7micron pixel pitch of its 8MP 2/3" format sensor, on the slender evicence that, according to DPReview, the 5MP downsampled output option of the Sony 828 has higher noise than the native 5MP output of the Sony 717.