Implement Tier 2 LLM-based intent routing

[mingshl]

Originally created by @mingshl in mingshl/opensearch-agent-server#19

Description

Implement LLM-based intent detection in the router to classify user messages when page context alone is insufficient.

Current state

Currently the router only does:

• Tier 1: Exact page context match
• Tier 3: Fallback agent

Desired behavior

• Tier 2: When page context doesn't match any agent, analyze the user's message with a lightweight LLM call to determine which agent is best suited
• Example: user asks about search relevance from the home page → route to ART agent

Considerations

• Should be fast (small model, cached decisions)
• Fallback gracefully if LLM call fails
• Configurable: can be disabled in favor of deterministic routing only

References

• Router: src/orchestrator/router.py
• Registry: src/orchestrator/registry.py