It seems like there are two issues contributing to the artifacts:
Non-manifold complexes
Interpolation between "unlit" normals and "lit" normals
When you see random isolated black spots that seem like they should be lit, it's because the vertex is actually non-manifold, and its normal was estimated to point more in the "unlit" direction than the "lit" direction.
Then the interpolation will make that darkness bleed out where it might not makes sense. In particular, on sharp edges. If the normal is estimated as "unlit" that will bleed around the edge.
I think the sharp edge issue is actually a bigger problem. Non-manifold complexes are always undesirable. But sharp edges are certainly something we would like to look correct.
It seems like there are two issues contributing to the artifacts:
When you see random isolated black spots that seem like they should be lit, it's because the vertex is actually non-manifold, and its normal was estimated to point more in the "unlit" direction than the "lit" direction.
Then the interpolation will make that darkness bleed out where it might not makes sense. In particular, on sharp edges. If the normal is estimated as "unlit" that will bleed around the edge.
I think the sharp edge issue is actually a bigger problem. Non-manifold complexes are always undesirable. But sharp edges are certainly something we would like to look correct.