Right, let's start out with a basic reflective surface. No turbulence, very obvious bilinear filtering of a low res reflection buffer. A slight bleed.
So what I think you're doing right now is adding the normal map to the screen coordinates to add some noise that roughly follows the surface. Tangent space has no relevance to screen space, so it rarely distorts in the right direction, but nobody seems to notice.
If we work out how much of the buffer is filled with useful data in script and pass it in as a parameter, we can scale the buffer about the screen centre, causing a slight zoom effect in the reflection and adding a padding border around the edge.
If the noise is thrown back on with this padding and the noise is scaled back enough to not escape the padding, it's a pretty convincing result. The scale issue isn't pretty however.
Given we know where the buffer ends from our parameter passed in by script, we might as well just clamp it there instead. Also adding a slight offset to get away from the bleed.