Embedding-based agent router

[shreydevkar]

The current router is a classifier prompt against a small Groq model. Works fine for ~4 agents; won't scale past ~10. Replace with an embedding nearest-neighbor match across agent router_description vectors.

Acceptance criteria:

• One-time embedding generation on agent registration.
• Per-request embedding of the user prompt + cosine match against agent vectors.
• Fallback to the current classifier prompt if the embedding provider isn't configured.