I'm sorry, I have more important things to do on a Sunday so just a quick fire response..

The extra cache on K6-2+ means it can setup triangles multiple times faster than the K6-2. It also means it can spin through the orchestrating software layers (e.g. Direct3D) faster. That brute force is going to boost FPS on the K6-2+ regardless of PCI/AGP bus.

I had a 3Dfx Banshee SGRAM, so 3Dfx definitely put SGRAM on AGP cards too. It's not a PCI/AGP differentiator, but it shows that test cards need to be carefully hand picked before running PCI/AGP comparisons.

In the later PCI-via-southbridge vs 8xAGP-3.0 comparison, not only is the PCI slot severely handicapped by being architecturally demoted to the chipset periphery, but the actual cards compared a 64-bit VRAM PCI card vs 128-bit VRAM AGP card. As a PCI/AGP comparison it's totally ludicrous!