This. I think problem is that the LLMs really struggle with 3d scene understanding, so what you would need to do is generate code that generates code.

But also I suspect there just isn't that much openscad code in the training data, and the semantics are different enough to python or any of the other languages that are well-represented that it struggles.