Reasoning chains as first-class OTel spans, which is what makes autonomous agent production deployments auditable (rather than indefensible.)