fix(router): make acknowledgment messages role-aware per agent type

[aaight]

Summary

• Root cause: ACK_SYSTEM_PROMPT in src/router/ackMessageGenerator.ts only contained investigation/review-oriented example phrases ("checking", "looking into", "dig into"), biasing ALL agent acknowledgment messages toward review language regardless of which agent was triggered.
• Updated the system prompt with role-categorized example phrases (implementation → building/coding language, planning → design/mapping language, review → checking/examining language, splitting → breaking-down language, debug → tracing/investigating language).
• Added an explicit role-matching instruction to the user prompt so the LLM is reminded of the agent's specific role when generating the message.
• Added 4 new unit tests verifying the LLM receives correct role-specific prompt content for different agent types.

Card: https://trello.com/c/rUBjpU61

Test plan

• All 37 existing tests pass with no regressions
• 4 new tests verify system prompt contains role-category examples
• 4 new tests verify user prompt contains role reinforcement text per agent type (implementation, review, splitting)
• Lint and type checks pass

🤖 Generated with Claude Code

[nhopeatall]

🤖 On it — checking the ack messages for PR 616

[nhopeatall]

LGTM — Clean, well-scoped fix to a real problem (all ack messages sounding like review-agent messages). The system prompt now covers all agent role categories, and the user prompt reinforcement ensures the LLM gets a clear signal about which role it's generating for. Tests cover the important assertion points. No issues found.