nx_llm_route.nx

nx_llm_route.nx -- MIXED-LLM ROUTING (companionchat CC6, 2026-08-30): a DATA ROW can send a chat turn to a stronger backend FIRST; the local encoder-LLM stays the fallback; and the reply records which one answered (model_route), so a route is never silent and an absent backend is never a mystery. Operator ask: "the companion intelligent, use your mixed llm". CONF llm_routes.conf, rows `tag|trigger|a.b.c.d|port`: the FIRST row whose trigger appears (case-insensitive) in the user turn selects the backend; trigger `*` matches every turn (a default route). Comments `#` and blank lines ignored. The conf is read PER TURN (data, no rebuild) from the daemon cwd, then from the nishihost knowledge tree. No conf, no match, or a malformed row = NO ROUTE, never a guess. The backend must speak the same OpenAI-shape /v1/chat/completions the local worker speaks; nothing here re-implements dispatch -- the orchestrator prepends the routed candidate to its own list. Contract symbol: go_llm_route. Gate: nx_llm_route_gate (fixture confs under /tmp, neg-controls for no-match and no-conf). license_tier: ORIGINAL