It's not surprising to me that it struggles with a language where there aren't billions of lines of code available to use as training data.

I wonder if that's all it is, or if the lack of context you mention is a more fundamental issue.