feat(ai): Pest-evals-style assertions on the agent test double

[tmgbedu]

What

Adds Pest-evals-inspired assertions to the agent test double (AgentFake / AgentRecordFake), on top of the create_agent state-dict contract.

Deterministic

• assert_json() — response content parses as valid JSON (toBeJson)
• assert_follow_trajectory(expected) — the tools called across the whole session match the expected ordered sequence (toFollowTrajectory)

LLM judge (reuse the existing JudgeAgent)

• assert_satisfy(expectation) — response satisfies a natural-language expectation (toSatisfy)
• assert_relevant() — response is on-topic for the last prompt (toBeRelevant)
• assert_safe() — response is free of harmful content (toBeSafe)
• assert_prompt_judged(expectation) — grade the prompt itself

Notes

• The judged assertions' provider/model now default to the agent under test (overridable per call), so the no-arg forms work without repeating model config.
• assert_response_judged now grades the whole response — answer text plus tool calls when present — instead of only the final text. Text-only responses are graded as before, so existing behaviour/tests are unchanged.
• Judge verdicts are cached in a sidecar <cassette>.judge.json file, keeping recorded conversations free of grading noise.

Tests

• New tests/ai/test_eval_assertions.py (11 tests) covering each assertion, the model/provider defaulting, whole-response grading, and sidecar caching.
• Full AI suite green: 340 passed (tests/ai/ + tests/support/), ruff clean.

Base

Targets ai-package-fix (not main) because it builds on the refactored ai.state helpers / state-dict contract that only exist on this branch.

[codecov]

Codecov Report

❌ Patch coverage is 98.50746% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...astapi_startkit/src/fastapi_startkit/ai/testing.py	98.50%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[tmgbedu]

CI status

• Run tests (fastapi_startkit): ✅ green — the framework package this PR changes, including all tests/ai (new eval assertions + regression tests).
• config-app ✅, database-app ✅, Ruff ✅, Pyright ✅, codecov/patch ✅.
• Run tests (agents): ❌ — all 11 failures are ConnectionError: PostgreSQL server at "127.0.0.1:3306" rejected SSL upgrade, a pre-existing CI database-environment misconfiguration (asyncpg pointed at :3306). The same job fails identically on the base ai-package-fix branch (run 30494801363), so it is not introduced here.

The judged example tests that my first push briefly broke (test_router_agent::test_the_router_with_initial_messages, test_job_search_integration summarizer) are fixed by the backward-compat commit — pre-existing inline/plain-text judge verdicts replay again without a live call. Verified locally: those tests pass; the only remaining example failures are the DB-env ones above.