To follow up on this, this happens to be slightly trickier than I originally thought as well. I believe the reason why this is happening is that the red light is getting around a 99% sampling probability in that middle region. However, the issue is that the red light only contributes to the red channel, so the much weaker blue light will still have a non-trivial impact on that middle region.
What this implies is that the heuristic should account for the contributions in the different channels, but the weighting is a little unclear. This issue is likely addressed in the paper by adaptive splitting, but I suspect that our version of adaptive splitting will encounter the same issue because we’re still only evaluating a single sample at a time.