Running an LLM as a Pipeline Step Without 2am Surprises

Treat an LLM call like a flaky external dependency: structured output, evals-as-gates, retries, and cost control.