So this means I won't see that much speed up from combining my identical particle quads into one big mesh entity?
Far be it from me to argue with the guy who created the language, but I think I'm going to argue with the guy who created the language. I wrote 2 demos a long time ago and posted them here for people to test. One created 500 identical sprites ( sharing the same, texture, colour, shininess, etc ) and one that created 500 identical quads, all attached to the same surface. If I recall, the speed increase ranged between 200% and 400% on average. With larger numbers of sprites, it was much more even.
So, assuming what Mark says is true, the code which discards redundant renderstate changes is.. fallible.