refactor: Benchmark should compare models, not prompt strategies

[sangicook]

Context

The live benchmark (run_live_benchmark.py) currently runs a 2-arm experiment comparing Typed vs Raw prompts on the same model (Gemini Flash). We already know Typed is strictly better — Raw prompts invent YAML paths and fail consistently.

The benchmark should instead compare models on the same (Typed) prompt strategy.

Current State

configs = [
    BenchmarkConfig(prompt_builder=build_typed_action_prompt, model="gemini-flash", label="Typed Gemini Flash"),
    BenchmarkConfig(prompt_builder=build_consultation_prompt, model="gemini-flash", label="Raw Gemini Flash (control)"),
]

Proposed Change

configs = [
    BenchmarkConfig(prompt_builder=build_typed_action_prompt, model="gemini-flash", label="Gemini Flash"),
    BenchmarkConfig(prompt_builder=build_typed_action_prompt, model="gpt-4.1-mini", label="GPT-4.1 Mini"),
    # etc.
]

Key decisions

• Drop the Raw arm entirely — build_consultation_prompt and collect_agent_proposals (raw pipeline) can be deprecated
• All arms use Typed actions — build_typed_action_prompt + collect_typed_proposals only
• Expand MODEL_REGISTRY with candidate models to benchmark
• Report should rank models by resolution rate, CQS lift, cost, and cop-out rate
• Cost-per-kept-change becomes the key efficiency metric

Notes

• This is a refactor of the evaluation harness only — the core auto-eval loop (System 1) is unaffected
• Related: feat: Switch AI benchmark to google.genai Batch API for 50% cost reduction #1 (Gemini Batch API migration) would make multi-model benchmarking cheaper