Interesting read, even if I didn't fully understand everything.
From the examples 8x SGSSAA looks closest to the reference image. Can't see what the 4×4 OGSSAA adds with my eyes.
Does any game on consoles use SGSSAA even?
What I - guess - didn't understand is where downsampling falls. It's definitely not analytic-based. With downsampling a pixel doesn't get multiple samples, right? It's just many more pixels that get "shrinked". Wait, is that sample-based? ... urgh, I don't get it.