[Router] llm router classifier + routing.router desugaring (L0a)

[ramkrishna2910]

Goal

llm router classifier + routing.router desugaring — the L0(a) "LLM-as-router" on-ramp, the starting point for most users. Kept inside the one-engine invariant (everything compiles to a Decision).

Scope

• New classifier type: "llm": runs a (small, local) chat LLM with an author-supplied NL prompt, constrained to choose among the candidate names; parses the structured choice and returns it as the label (score 1.0) + the model's short rationale.
• routing.router sugar: at load time the parser desugars { "router": { "type":"llm", "model", "prompt" } } into one llm classifier (labels = routing.candidates) + auto-generated identity rules (label X → route_to X). Engine stays single-path.
• Runs the chat LLM through the existing Router chat path (extends the M·svc wiring, [Router] ClassifierServices to Router wiring (M-svc) #2384). The LLM model is a bundled component (declared as routing.router.model; role = classifier-model).
• Trace records the chosen candidate + rationale — so even L0(a) is auditable.

Out of scope

The L0(b) offline NL→policy compiler (separate, deferred). Determinism guarantees — L0(a) is explicitly the least-deterministic/slowest tier; docs steer steady-state users to L0(b)/L1–L3.

Acceptance

• A routing.router (llm) policy with N candidates routes a request to a candidate the LLM selects; the decision trace shows the chosen label + rationale.
• Tested with a fake LLM-invoke returning a fixed choice (unit) + one live end-to-end path.
• Invalid choice (LLM returns a non-candidate) falls back to default_model.

Example — L0(a) collection.router JSON (LLM-as-router)

The lean local form — components reference already-registered models by name (like the built-in Ultra Collection); no models[] needed:

{
  "model_name": "user.Router-Auto",
  "recipe": "collection.router",
  "components": ["Qwen3-1.7B-GGUF", "Qwen3-8B-GGUF", "Qwen3.5-35B-A3B-GGUF"],
  "routing": {
    "candidates": ["Qwen3-8B-GGUF", "Qwen3.5-35B-A3B-GGUF"],
    "default_model": "Qwen3-8B-GGUF",
    "router": {
      "type": "llm",
      "model": "Qwen3-1.7B-GGUF",
      "prompt": "You route requests. Reply with ONLY a model name. Use Qwen3-8B-GGUF for everyday questions; use Qwen3.5-35B-A3B-GGUF for hard reasoning, coding, or long context."
    }
  }
}

The router LLM (Qwen3-1.7B-GGUF) is a classifier-model, not a candidate. No rules — routing.router desugars to one llm classifier + identity rules. For Hugging Face redistribution, embed each component's full def in a models[] array (as LMX-Omni-52B-Halo.json does); locally it's omitted.

---

Depends on: #2379 (classifier registry), #2383 (parser/desugaring), #2384 (Router chat wiring)