Add goal-based regression cases evaluated across repeated runs

[Doringber]

Stacked on #2, which adds scorers and pass rates. Merge that one first.

Why

Today a regression case describes one exact path: required tools, order, counts. For AI behavior the useful contract is different. It says what the user must achieve, what must never happen, and how often the agent must get it right. This lets the same YAML files express "expected behavior", not only "expected result".

What

A regression case gains optional fields. Existing cases load and evaluate exactly as before.

Field	Meaning
goal	What the user must accomplish
persona	Who is asking
runs	Executions per case in case.run(); default 1, max 100
behavior.min_pass_rate	Fraction of runs that must pass; default 1.0
behavior.required_terms / forbidden_terms / required_fields / case_sensitive	Built-in deterministic checks
behavior.scores	scorer_name: min_score for consumer-provided scorers

Evaluation works at three levels:

• case.evaluate(result) stays sync and deterministic. It now also applies the built-in behavior checks and returns scores per check.
• await case.evaluate_runs(results, scorers=[...]) evaluates each run, applies named scores, and compares the pass rate to min_pass_rate. It returns a RegressionRunEvaluation with the goal, persona, pass rate, and each run's failures and scores.
• await case.run(executor, scorers=[...]) runs evaluate_runs over run_repeatedly with one idempotency key, regression:<id> by default. It rejects missing or duplicate scorer names before calling the agent, so a config mistake does not spend agent calls.

RegressionRunEvaluation.assert_passed() failures report the pass rate and the failures of each failing run.

Also included:

• examples/behavior_regression_cases.yaml, which a test loads to keep it valid.
• A README section and public API exports: RegressionBehavior and RegressionRunEvaluation.

Testing

• tests/test_regression_behavior.py adds 21 tests. They cover schema validation and loader error context, built-in checks, pass-rate boundaries, named scores, missing or duplicate scorers, idempotency key propagation, and default behavior for cases without the new fields.
• The existing tests/test_regression_cases.py passes unchanged.
• Coverage for regression.py is 93%.
• Full suite: 225 passed, 19 skipped, 3 failed. The 3 failures are the same ones that fail on main: the version check (0.1.1 vs 0.1.2) and one malformed-payload case in the cursor client.
• ruff check and ruff format are clean. mypy --strict shows only the error in reporting/json_report.py:117 that already exists on main.