Aha!
Now we’re talking.
You’re right - now that I’m running it the second time, it looks much more like you described.
Without seeing the proper result, I was under the assumption that you’re doing screen-space reflections, but they’re actually raytraced. Not bad.
Is this the paper you implemented? http://graphics.cs.umass.edu/pubs/pg08.pdf
Can you add a teapot or a curved surface? This would show off the accuracy of the reflections much better than a simple plane.
About mip-maps: if I got the paper right, you’d get rid of the noise across the surfaces of reflected objects, since the mipmaps are for the shading in its texture space, but you’d still have noise along objects’ sihouettes, where you have no prefiltered texture to sample from.
They also mention this on page 8 and I guess solve it by sampling the sihouettes more excessively.
Good stuff though!
Unfortunately such techniques are still way to memory- and computation-heavy for a game, but maybe on the next next gen, we have enough memory and horse-power to drop rasterizers for good.