Clef-flash is a decision model, not a chat model. It serves Clef's own
System One / Jev API at /v1/systemone: a state plus a schema of typed
questions in, one probability per allowed option out, single prefill
pass, no decoding and no free-form text.
Router previously dispatched by body `model` only for
/v1/chat/completions, /v1/embeddings and /v1/rerank. A request to
/v1/systemone therefore fell through to the configured routes, found
none, and answered "no routes configured". Adding the path to the
dispatch predicate is the whole fix: routeByModel already sets
upstreamPath := path, so the upstream receives /v1/systemone unchanged.
There is deliberately no /rerank-style rewrite.
Config: `ornith:35b` is removed with the Ollama predictor, and its GPU
now serves clef-flash. `qwen2.5:3b-instruct` already pointed at qwen-cpu
and is unchanged. Callers of ornith:35b must move to `reasoning`, which
keeps its serving name while running Qwen3-30B-A3B.
Tests cover body dispatch, unknown_model parity with the other three
paths, and that the upstream path is not rewritten.
