Predictive Auditing of Hidden Tokens in LLM APIs via Reasoning Length Estimation