intent-routinglisted

# Skill-1: Intent Recognition & Use Case Mapping Input: `user_raw_input` (the user's untrusted text) + `routing_map_path` (default: `config/agent-routing-map.md`) Output: strict JSON — `{ "status", "routed_agent", "use_case", "context_label", "confidence" }` — plus optional `"candidates"` (top 2) and `"question_to_user"` when the status is `clarify_intent` ## Procedure 1. **Load the routing map** — Read `config/agent-routing-map.md`. If it is missing or unreadable → stop, report `{"status": "error", "routing_reason": "Routing map missing"}`. Done when the file's agent table and fallback rules are in context. 2. **Run the classification** (below) against the user's raw input with the routing table in context. 3. **Extract the JSON** from the classification output. Done when the JSON has all five fields and no trailing prose. 4. **Validate** — every field present; `routed_agent` ∈ the 8 named agents or `GENERAL_CHAT`; `use_case` ∈ the 32-value taxonomy or `General_Communication`; `confidence` a float in [0, 1]. Done when all checks pass; if any fail, retry step 2 once with the instruction "Return ONLY valid JSON — no prose." 5. **Apply the three-band fallback rules** (from the routing map): - High confidence (`confidence >= 0.6`) → route to the chosen expert agent, `"status": "success"`. - Borderline (`0.4 <= confidence < 0.6`) AND the input carries a workplace signal → `"status": "clarify_intent"` with the top 2 `candidates` (each with `agent`, `use_case`, `confidence