Generative evaluation suite (run-length, switch prob, logistic coeffs)

Build the generative-run evaluation: roll out trained models in the task environment, compute summary statistics, compare data vs. models.

**Scope**
- Switch probabilities under various conditions
- Choice run-length distribution
- Logistic-regression coefficients

**Done when**
- Reusable eval suite producing data-vs-model summary-stat comparisons (Fig 2d)