Even GPT-5.2 Can't Count to Five: The Case for Zero-Error Horizons in Trustworthy LLMs