I've always wondered... I mean technically speaking, for a planar reflection, how much more gpu intensive is it, really, if sort of like a green screen, the object (eg, puddle) isn't rendered from the player perspective on initial pass - skipped, that is, and instead the 2nd camera views the scene from the reflected perspective (opposite z of the reflection plane), only rendering that clipped region, having its result flipped for placement in the original scene. You're not rendering the entire scene twice, you're not tracing reflection angles, only determining what objects are visible within the rendered viewport from camera 2, and only rendering that view. Environmental alterations could be applied before or after merging with camera 1 (whether there are significant plane mesh variances which alter the visible renderable objects for camera 2).

Heck even if the puddle takes up most of the screen, the render time isn't doubled (cam 1 renders the 'frame', cam 2 renders the puddle), only the visible object pass would effectively be executed twice. (my terminology I know is completely way off - the last time I programmed 3D graphics was high school, and it was a FarCry *ducks* from FPS engines :) SO I could be completely off base, or this could be one of simplest concepts already employed, heh... just thinking out loud about the logic of optimized 3d reflection rendering.