I guess the op was implying that is something fixable fairly easily?

(Which is true - it's easy to prompt your LLM with the language grammar, have it generate code and then RL on that)

Easy in the sense of "it is only having enough GPUs to RL a coding capable LLM" anyway.