Class DiscoveryLabelMatchOracle

discovery-label-match: scores an agent-discovery run against its label, for either arm.

• An agent label is correct when the chosen agent is the labelled agent: the Choice's pick for the decision, whatever its confidence, or the first row Find Candidate Agents would return for the semantic-search baseline.
• A none label is correct when discovery would not inject a suggestion (the run's WouldInject): production's own verdict, DecisionDiscoveryFromResult, not confident once the answers are calibrated for the answering model, or that model having no calibration, or the discovery finishing after production's timeout (decision); or no candidate reaching the action's similarity floor (baseline).
• A run with no usable answer fails, naming why, with correct: null.

details carry { arm, label, kind, expectedAgentId, chosenAgentId, confidence, anyApplies, wouldInject, topScore, correct }, so the scorecard can aggregate without re-running anything. It takes no configuration.