Evaluate a small local decision model on routing dataset v2 #57

Compare the installed gemma3:1b local model with the current heuristic and Nomic embedding result on the frozen routing dataset v2.

Scope:

• Experiment-only Ollama adapter; no product runtime integration or model download.
• Preserve exact meta commands, deterministic edits, lane selection, and all product safeguards.
• Ask the model only for otherwise-unknown scope classification.
• Record semantic and near-miss accuracy, abstention, cold/warm latency, and high-confidence wrong direct actions.
• Use family-isolated splits and do not tune on holdout.

Decision rule: keep the result as evidence. Do not adopt unless it materially beats Nomic while preserving 100% vague near-miss safety and acceptable latency/model size.