feat(evals): compare agent tool-use across models #8471

Problem

The live eval layer runs one model. There is no way to compare tool-use
reliability, latency, or cost across models, so model-selection and routing
changes (including sim-auto) are unmeasured.

Proposal

Run the same scenarios across a list of models and report a scenario × model
matrix.

• EVAL_MODELS=provider:model,... selects the models; a bare model id defaults
  to DeepSeek.
• A small provider registry (DeepSeek, OpenAI, Groq, OpenRouter) reads each key
  from <PROVIDER>_API_KEY, so adding a provider is one entry.
• The report carries per-model pass rate, iterations, latency, and tokens, plus
  a per-scenario pass-rate matrix.
• Opt-in live suite, never in CI. EVAL_MIN_PASS_RATE fails a model below a
  floor.

Acceptance criteria

• EVAL_MODELS runs two or more models
• Comparison JSON/Markdown with per-model metrics and a scenario matrix
• A model is selectable by env var with no code change
• README documents the spec format and supported providers