This came via a UDN.
Very small triangles can generate very small lods, which make query time really slow if your self collision thickness is significantly larger. You will do X^3 checks per query where X is the ratio of thickness/lodSize. (lodSize ~ triangle size). Nearly degenerate triangles can create cases where you're doing billions of queries per vertex.
We should cap the smallest lod based on the self collision thickness (known at hierarchy build time). This would prevent a single tiny triangle from creating this situation.
There's no existing public thread on this issue, so head over to Questions & Answers just mention UE-183992 in the post.