Stop trusting a 99% uptime claim: the LLM API relay failures we hit, and how we test stability