Add Jev semantic assertions on agent replies

[Doringber]

Summary

Agent replies change wording even when the meaning stays the same. This adds AgentExecutionResult.assert_means, which sends the response text to pytest-jev and records each claim's probability on assertion_outcomes so the HTML report shows it.

How to use it

pip install "agent-test-kit[jev]"
export OPENROUTER_API_KEY=...   # or TYPESAFE_API_KEY

async def test_refund_reply(agent_scenario, jev):
    result = await agent_client.execute({"ticket": "I was charged twice"})
    agent_scenario.attach_execution_result(result)
    result.assert_success()
    result.assert_means(
        jev,
        holds=[
            "apologizes to the customer",
            "says the duplicate payment was refunded",
        ],
        lacks=["asks for a password or a full card number"],
    )

• String responses are judged as-is. Dict responses use the first non-empty field among text, summary, message, content, reply, and suggested_prompt (result.response_text).
• Tests that request the jev fixture are marked jev. pytest -m "not jev" stays offline. Without an API key, pytest-jev skips those tests; --jev-require fails CI instead.
• Copy-paste example: examples/jev_semantic_example.py.

Tests

tests/test_semantic_assertions.py uses a stand-in judge, so the suite does not call Jev or need an API key.