fix(agents): extract intermediate agent text + pin chain tests

[romamid]

Summary

• Intermediate agent replies in a multi-agent chain are collected via collect_full_response. The non-streaming path previously returned the raw chat.completion JSON envelope instead of the assistant text, causing the next agent to sometimes receive a serialized object as context.
• Extract /choices/0/message/content when present; fall back to the body when content is missing or the response is not JSON.
• Add mockito unit tests for execute_agent_chain and select_agents that simulate the sequential chain (route order, named assistant injection, last-agent-wins, single-agent short-circuit, fallbacks).

Working on extending Plano to support better multi-agent composition on the agent path (specifically subagent orchestration). Follow-up work will stay opt-in and will not change the Plano-Orchestrator prompt / {"route": [...]} contract.

Testing

• cargo test -p brightstaff --lib handlers::response
• cargo test -p brightstaff --lib handlers::agents
• future iteration: design a two-agent benchmarking listener to confirm intermediate context is working appropriate accordingly (communicated via plain text rather than json)

[romamid]

First iteration towards multi-agent composition. Design/optimization question: #1039