I stopped testing the GF4 at this point as it was clear that it would not be able to catch the AGP card, being limited by the bandwidth of the PCI bus.
The only test where the GF4 was significantly better was the High Polygon Count tests, especially with 8 lights. Presumably this is due to the better T&L on the GF4 core.
The first test I ran on the default GF256 I got 2588 - thought this was a little low, so went to update drivers.. "Error: Your drivers are not compatible with Windows XP and the default drivers are being used" even though they were Det52s! So just using the nVidia drivers boosted me nearly 1000 and the overclock gave me nearly another 1000. Not bad for a graphics card going on 6 generations old :P
The moral of the story is that PCI cards never have enough bandwidth to be able to take on even an inferior AGP card.
Interesting comparison, but bear in mind that the GF256 DDR isn't exactly "inferior" to the MX420. It has twice the memory bandwidth and single-textured pixel fillrate per clock.
The reason the MX420 wins the high poly test is probably because it's a geometry-limited situation. AFAIK, its T&L engine isn't any better than that of the GF256, so it's probably just clockspeed that's making the difference.