Unless I misunderstand the bug report (if I did please feel free to submit again), the issue is coming from a limitation of our system.
The silhouette is visible because the refracted point in screen space ends up on an object that is in front of the refractive object (the capsule or the big cube), so in those situations we sample the color pyramid at the entry position of the ray (i.e. as if the ray has not been refracted).
We could fallback on cubemap on those situation, but you'll still end up seeing the silhouette and in most cases the data within the area that is covered by an object in front of the camera might end up being even less correct-looking.
If I misunderstood the bug, feel free to re-open it.