judge: embed seed test file in the prompt; reframe tests as joint review

[samkeen]

Summary

Closes #16. Lands two paired changes that fix the symptom (false rejection on file-existence ACs) and the deeper problem (judge can't verify the assertions actually pin down the criteria).

• Embed the seed test in the judge prompt. _build_judge_user_message globs tests/test_t<NNN>_*.py (reusing validators._task_test_glob, so the judge sees exactly what pytest ran) and embeds it under ## Seed acceptance test (already on disk, not in this diff). Capped at JUDGE_TEST_FILE_MAX_CHARS = 8000.
• Reframe the judge prompt. Tests and implementation are now joint review subjects. Two new rejection categories — Weak test and Tests-pass-but-wrong — plus a per-AC↔assertion correspondence rule and an explicit "files in HEAD count as exists" rule that resolves the Judge can't review seed acceptance test files; rejects on file-existence criteria #16 symptom directly.

The new _build_judge_user_message is pure on (task, worktree, diff) so it's testable without mocking the LLM client.

What's intentionally not in this PR

The larger responsibility-matrix question (Gherkin output from prep-feature, worker writes its own tests, etc.) stays open. The plan is to run two or three demo sessions with the new judge framing, look at what it rejects on, and decide based on data whether the seeder's tests or the worker's gaming is the bottleneck. Phase 2 walks through that door with evidence, not a guess.

Test plan

• pytest tests/ — 220 passed (7 new in test_judge_seed_test_embed.py)
• ruff check tilth/ tests/ — clean
• Run the demo session and observe whether the file-existence rejection goes away on T-001
• Watch a session where the seed test is weak (e.g. only checks return code on an AC that names stderr too) and confirm the judge flags it

🤖 Generated with Claude Code