Could someone help explain why the sample BRDF and sample light importance samplings fail at the ones they do fail (small light on diffuse surface for BRDF sampling and large light on glossy surface for light sampling)?
When a ray hits the diffuse surface, we sample the BRDF to get a random direction for the ray to bounce. With a small light, the probability of that bounce hitting the light next is pretty small. So most samples on the surface end up missing the light and thus it fails.
When a ray hits the glossy surface, we sample a random direction for the ray to bounce that's guaranteed to hit the light next. But since this is a glossy BRDF, only the mirror of the incoming ray's direction will be passed along by the BRDF, and it'll zero out all other directions. With a big light, there's a larger number of valid directions to sample from, so the chances of picking the mirror direction is small. So most samples on the glossy surface end up getting zeroed out by the BRDF and thus it fails.