As a huge proponent of constrained decoding for LLM reliability, I don't quite think it's the right approach for code. This is because in many programming languages, it is legal to use a function before its declaration. Since this is common in existing code, LLMs will also try to write code that way.