Well, I tried Greg's suggestion of setting the limit weight to a very low number, but all that did is manage to scrounge up a few additional lux. This actually made things look worse, because my problems are with the really dense shades (which are already coming up with higher numbers than I'd expect).
I have a spreadsheet here with the results of all my tests so far. If anyone's interested in looking at it I'd be happy to share it. Perhaps there's a problem with my interpretation of this data. I dunno. But the tests indicate that whenever I try a 1% shade in any combination, I get very odd results from rtrace. And the whole reason I'm doing this test is because we were getting similar odd results on our real model; I was trying to simplify things so I could perhaps isolate something that was causing these oddities. So far, no joy. I shared my trans material definitions for all these tests earlier, but if anyone wants them to have a look I can send those as well.
My fear is that I'm simply trying to find too small a percentage of the available light, and relative comparisons between different shade materials will not be valid for this model. (?) I mean, this test model still has two pieces of diffuse glass in it, and the skylight is prety small, so already the building form/glazing is removing a ton of light from the space. Then when I start throwing shades in the skylight cavity, we're getting down to really tiny fractions of the exterior light. I dunno, I'd love to hear any thoughts.