Which implementation answers PRIMARY. A peer to the fallback cascade, not a replacement for it: whichever mode is primary, a failure still walks the same cascade (stale session → session history → cloud_fallback → the static guess) unmodified.

On means the router may use this machine's GPU. Turn it off when you stop Ollama to play a game: every local call is then skipped rather than attempted and timed out — classifier, health probe, local verification, the local-vision fallback, and local dispatch rows, which drop out of the candidate set instead of being picked and then failing.

Turning it off requires classifier.cloud_fallback. Skipping the local classifier does not make classification remote — without a cloud classifier it stops classifying, and every request falls through to a static guess recorded as general_chat, which is indistinguishable from a real classification afterwards.

The cloud classifier replaces the local model, not the free steps ahead of it: this session's cached classification and this session's history in route_decisions are both real classifications of this same session, so paying to re-derive an answer already held would be spending money for nothing.

A cloud classifier used only when the local classifier is unreachable. Independent of the mode choice above — not a primary-mode option.