I managed to get some more info out of Terrasoft on this. It looks like some of our assumptions were off-base..
First off, most importantly, the relative performance figures (7x the framerate, 80x the fragment rate) are comparing performance of their technique implemented with Rapidmind versus without Rapidmind. The idea of the project was to implement a system that takes OpenGL fragment shaders and turns the interpreter for them into a compiler that compiles them into SPE code, and offloads that code to the SPEs. They implemented it with and without Rapidmind, and the relative performance figures shown are comparing those implementations - not MESA's speed before and after. The figures by no means reflect the performance you would get in the general case going to this new version of MESA, not by a long shot. It wouldn't accelerate something like Q3 on the system since it's just addressing fragment shaders (which Q3 doesn't use)..and for apps using fragment shaders it would just make them go from being really really really slow to really really slow

Their system also doesn't yet allow all operations in the fragment shaders, there are still some holes to be filled. In absolute terms it's all still
quite slow apparently, they still have a lot of work to do to make this part run well, and comprehensively.
The guys working on this at the hack-a-thon haven't had a chance to get back into it since then either, so it's hard to tell if it's an on-going concern on their part or not. But it is something they think would require a lot more work if it were to be a generally accelerated version of MESA.