When the primary (Claude) and secondary (Gemini) assertion models reach opposite verdicts, Passmark calls a third arbiter model to break the tie. This commit makes that behavior configurable via
configure({ assertions: { consensusPolicy } }).

• "consult-arbiter-on-disagreement" (default): unchanged; the arbiter decides. Existing users see no behavior change.
• "fail-on-disagreement": treats any disagreement as a failure immediately, with no arbiter call. The synthesized reasoning includes both models' verdicts and confidence scores so the failure is actionable. Useful when you'd rather surface ambiguity in the UI under test than risk a single model swinging the result.