Observability / LLMOps: per-call token + cost + latency metering for every LLM invocation, persisted for aggregation, with optional Micrometer metrics. Plugs into the model module via ChatModelDecorator.