Class AgentDecisionOracle

agent-decision-match — did the agent decide the right thing?

Reads the decision off whichever channel the model used (envelope or native tool calls) and compares it against an encoding-independent expectation. That normalization is what lets a baseline cell and a native cell be compared on an identical corpus case: the yardstick does not move between them, only the model's behavior does.

The result's details carry the components separately — decision kind, action accuracy, param fidelity, per-matcher outcomes — so §6.1's metrics aggregate out of persisted ResultDetails without re-running anything.

Configuration is the DecisionExpectation itself.