I'm working off the 2d soft shadows article here on GDNet which uses the alpha channel to store the light values before it renders the actual geometry on top of that. I can't quite seem to get the system to both blend the geometry correctly when using multiple layers of geometry as well as blending the multiple render passes. The only solution I can think of is to render all the geometry to a texture once, fully bright without any lighting. Then render each light using the target texture as the geometry source. I should be able to get the effect I want but it will result in a bit of a performance hit because I'm going to have to change the render target twice per scene - hopefully it'll still be fast enough. If its done well it might even be faster than the other way.
So I'll get that working tomorrow.