Add behavior scorers and pass-rate assertions

[Doringber]

Why

AI agents can give different correct answers to the same input. Today the kit can only check exact expectations: tool order, exact response equality, and exact trace equality. This PR adds the first step from "expected result" to "expected behavior". A run passes when its score reaches a threshold, and a repeated run passes when enough of its runs pass.

The package stays lightweight. There are no new dependencies and no built-in LLM provider. Judgment-based scoring, such as an LLM judge or embeddings, lives in the consumer repo behind a Scorer protocol. This follows the same pattern as SideEffectVerifier.

What

• Scorer protocol and Score. score(result) can be sync or async and returns a value in [0.0, 1.0] with an optional reason and details. Out-of-range values, NaN, booleans, and wrong return types are rejected.
• Built-in deterministic scorers:
  • RequiredTermsScorer scores the fraction of terms found in the response.
  • ForbiddenTermsScorer scores 1.0 or 0.0.
  • RequiredFieldsScorer scores the fraction of fields present and non-null, with dotted paths.
• await result.assert_score(scorer, min_score=...) is recorded in assertion_outcomes. The score, threshold, and reason appear in the existing JSON and HTML assertion tables with no reporter changes.
• RepeatedExecutionResult.assert_pass_rate(check, min_rate=...) checks a pass rate across repeated runs. A check that raises AssertionError counts as a failed run, so existing result.assert_* calls can be reused as checks.
• await RepeatedExecutionResult.assert_score_rate(scorer, min_score=..., min_rate=...) checks that a scorer reaches its threshold on enough runs.
• README section, public API list, and a repeated-run example in examples/consumer_test_example.py.

Existing APIs are unchanged, including assert_responses_stable and assert_traces_stable.

Testing

• tests/test_scoring.py has 44 parametrized unit tests covering the scorers, threshold boundaries, validation, sync and async scorers, pass-rate failure messages, and the path of score outcomes into the JSON report.
• Coverage is 99% for scoring.py, 94% for repetition.py, and 95% for models/execution.py.
• Full suite: 204 passed, 19 skipped, 3 failed. The 3 failures also fail on main before this change and are unrelated:
  • test_version.py expects 0.1.1, but VERSION is 0.1.2.
  • One case in test_cursor_agent_client.py for malformed payloads.
• ruff check and ruff format are clean on all touched files. mypy --strict shows no new errors; the one existing error in reporting/json_report.py:117 is untouched.