qwen 2.5 72b q4_k_m generates valid rust 73% of the time vs q5_k_m at 84%

tested on mbpp rust subset (487 examples), temp 0.0, context size 8k. q4_k_m gets 73.1% valid compilable rust, q5_k_m gets 84.3%. the 11 point gap is way bigger than i expected for one quant step up. llama.cpp b4729, same prompts, same seed. anyone else seeing this big a gap on code generation tasks or is my setup cursed