The 9B model spent its whole 4000-token budget on reasoning and came back
model_incomplete. Call /api/chat with format=json, think=false and num_predict,
and read prompt_eval_count/eval_count/done_reason. Structured extraction no
longer needs chain-of-thought.
- DecisionOllamaGatewayClient calls the self-hosted Ollama on
server-decision-01 (tailnet HTTP only for loopback/RFC1918/100.64-127),
keeps the same local admission and reports providerKind=local.
- createInf0Gateway prefers it over INF0/LiteLLM when configured.
- Evidence is computed deterministically from values when a small model omits
offsets; a value must occur exactly once in the source or the result goes to
review. This makes a 9B local model usable without weakening verification.
Tests: 4 new gateway cases + 2 evidence-derivation cases; full typecheck green.