So, rendering is probably simple enough. If we keep our current glass*refraction (so we can get coloured translucency), would simply adding the reflection (*some constant) work? Might do. Standing in the dark room means the reflection contribution will be small. Standing in the light room means the reflection contribution will be high. Yeah, I think that works from an intensity point of view - not sure about colour. Multiply the reflection by the colour of the glass, probably, just like the refraction.
Now, the other side of the coin I always like to toss - AI. Can we get the AI to take that reflection/refraction property into account?
Say you're in the dark room and the AI is in the light room. The AI shouldn't be able to see you. I'm wondering if we can use a sort of lightmapping to work that out. Create a lightmap for each side of the pane of glass, something that allows you to take a position on the glass and an angle and work out how much light is hitting that side of the glass at that angle (per-texel SH?). When you do your line-of-sight test through the glass, you perform two lookups, one for each side of the glass, and basically subtract the value of the near side from the value of the far side. The higher the result, the more you can see through the glass.