The Lifecycle of LLM-as-a-Judge for Large-Scale Recommendation Explanations