[triggeredcode]

Closes #57

What changed

• adds an experiment-only Ollama decision-model routing candidate
• preserves deterministic and explicit-state routes
• records calibrated confidence and model evidence in reports
• documents the Gemma 3 1B calibration result

Result

Gemma 3 1B was eliminated on calibration: 12/16 routes, 50.0% semantic scope accuracy, and 0.0% vague near-miss safety. Wrong answers were reported at 95% confidence, so the holdout was intentionally not opened. Product runtime remains unchanged.

Validation

• pnpm format:check
• pnpm lint
• pnpm typecheck
• pnpm test (1531 passed, 22 skipped)