agent: add anti-duplication guardrails to delegation policy

[shayne-snap]

Add three guidelines to the delegation policy in the system prompt to prevent unnecessary subagent spawning:

• Weigh context overlap before spawning: a child agent starts cold and cannot see anything the parent has already read or concluded. Only spawn when the overlap is genuinely low.
• Never delegate understanding: do not spawn a subagent to redo analysis or verification the parent has already performed.
• Do not chain spawn_subagent calls for the same unresolved question: if a child's findings are insufficient, synthesize and continue directly.