An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint