Investigating whether explicit, lightweight grammar priors can improve syntactic correctness in large language models without modifying model weights