Task 4 evidence - agent-aware resolution and persistence

What was tested

• resolveAgent covers fallback resolution, direct model precedence, ordered model alternatives, disabled and unknown agents, unavailable registries, explicit-model persona resolution, tool-rule filtering, and resolved_model.source: "agent".
• A fresh RecordStore instance re-reads an agent-sourced record from disk through parseTaskRecord.
• Unknown-target rendering carries both active agent and category rosters.
• Curated names are rejected by team-member validation.
• Registry parsing was extracted to agent-model-registry.ts, leaving resolve-agent.ts at 176 lines.

What was observed

• The focused resolve-agent.test.ts run completed with 7 pass and 0 fail; the full package gate covered the record-store, task execution, and member-validator cases.
• Mutation proof: temporarily removing case "agent" from the record parser made the fresh-store round-trip fail; restoring it returned the suite to green.

Why it is enough

The tests cross the model, persona, tool-policy, persistence, error-reporting, and team-eligibility boundaries that consume agent resolution.

What was omitted

No category lookup occurs inside resolveAgent; planner compatibility remains task 7.