Ok, I made the change, only the spot light uses the light_shaft material now, other lights use stock falloff_exp2. https://we.tl/t-lGvhXwFaYI
As for the performance, the window is 338 tris, shadow tris in that scene take up around 500. It's easier to use a physical model right now, as you don't have to line up things. Fake projection it will be a PITA anyway, since you'll have to make a window noshadows, carve some space behind to line up the light with projection etc. IMO the order of testing should be: 1) a physical model, 2) a plane with alphatest transparency, 3) a light with projection texture.
Makes sense. Not an issue for small rooms.