Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- suite/runner stay as the library; cases run as parametrized pytest tests
(per-case reporting, -k filtering, exit-code CI gating for free)
- drop typer/rich CLI and the hand-rolled scorecard
- model chosen via --model / MODEL env; base-url/api-key default to ollama

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

pydantic-ai over any OpenAI-compatible endpoint (ollama/vLLM/LM Studio).
scores behaviour — tool selection, arg routing, abstention, parallel,
numeric fidelity, dependent chains — not prose. ships a gemma3 Modelfile
that exposes tools through ollama's otherwise tool-less stock template.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>