Firstly, Michael's photographic tests show, better than any theoretical comparison of Airy disc diameters and pixel dimensions, that at apertures of f/11 and smaller, diffraction is reducing the resolution of this combination of lens, sensor and demozaicing algorithms (Bayer interpolation) noticeably below what is possible at f/5.6. Maybe also at f/8, but f/5.6 vs f/8 is a closer call for my eyes.
Secondly, theoretical calculations based on the Nyquist frequency of a given pixel pitch (74lp/mm for the 6.8 micron pitch in this case) ignore the reduction of MTF and resolution caused by anti-aliasing filters and demozaicing algorithms. A rough but pessimistic guideline is three pixel widths per line pair, in stead of the two pixel widths per line pair of the crude Nyquist frequency.
Also, photo-sites are square, not round, so the Nyquist frequency is different for lines at different angles: the standard number only applies to horizontal and vertical lines. I believe that it is lower for any other orientation of the lines: these "6.8 micron" photo-sites are 9.5 microns long on the diagonal.
Thirdly, there is not an abrupt line between "diffraction limited" and "not diffraction limited", just as there is not between "lens out-resolves sensor" and "sensor out-resolves lens". Instead, there is a transition zone where the different factors limiting resolution (sensor, diffraction, geometric aberrations of the lens) have roughly equal MTF at a given spatial frequency, and in this zone, the combined MTF is less than the separate MTF of diffraction alone, or sensor alone, etc.
One thing that Michael's evidence seems to show is that at f/11 and smaller (and maybe at f/8) diffraction is a very visible factor in the overall system resolution, probably the dominant limitation on sharpness, and that for optimum sharpness (ignoring OOF effects) one is better off with larger apertures like f/5.6 and f/8.