Take a lens, and make resolution tests with finer and finer pixel pitches until, say, the MTF10 in cy/ph stops changing. Then say that the number of pixels on the sensor just before the MTF10 stopped changing is the pixel resolving power of the lens.
It occurs to me that, since this method results in asymptotically approaching a certain cy/ph, you could argue that it never actually converges. That would be pedantic. But it certainly would be true to say that the answer you get depends entirely on what you decide is an inconsequential improvement.