That's exactly how we exercise it. We have special tables with different NULLs distribution over columns, and we use these tables to exercise the 3VL logic and compare against the upstream. Also, for contentious cases we consult ChatGpt/Claude to compare the original and the rewritten shapes. We also employ LLMs to also generate test cases for the rewrite layer.

This is one place where LLMs have truly helped out with testing.