Week 4 Day 7: Evaluate Observable Outcomes

[skyzh]

Why

Close Week 4 with one learner-visible way to judge a completed agent run from declared observable outcomes, without grading hidden reasoning or one exact transcript.

Change

• add a compact reference evaluator plus solution-free TODO starter for final, file, result, and named receipt facts
• return stable structured checks and a deterministic aggregate report while keeping evaluation read-only
• compare expected UTF-8 file content from exact bytes, preserving CRLF/LF distinctions and rejecting malformed UTF-8
• add one scripted edit-and-validate scenario, causal evidence regressions, cumulative starter sync, and the Day 7 learner chapter/navigation

Review note

Final exact-head review is complete on the strict-UTF-8 successor.

Base b107310ab977b0a434e54f6b694b2f91b0f1652e; head 2a8942a6e29547e96730803e5dd5f0d01312e38e; tree b60e25dbd2401d21c54ef0a22d96799a64c168fe. The successor from rejected c47407f19cd13580602fc64ae7dcce776b5d91bf changes only tests_refsol/test_week_4_day_7.py (28 insertions); every production/API/starter/prose/navigation/Day 1–6/stack byte is unchanged. Total main-relative delta: 11 paths, 957 insertions, 260 deletions.

Final evidence: Day 7 plus sync 51/51; cumulative Week 4 plus sync 108/108; full reference 443 passed with 8 intentional skips; Ruff/format/diff clean; learner copy byte-identical and expected-red 22/22 before cleanup; manual mdBook build passed. Independent malformed-byte cases fail if strict decoding is mutated to errors="ignore" or errors="replace"; the prior nine causal controls remain green. Hosted macOS run 31600359755, job 94125932294, succeeded on the exact head.

AI-Assisted: GPT-5.6 Sol + Forge
Reviewed-by: GPT-5.6 Terra + Oracle (consistency)
Reviewed-by: GPT-5.6 Sol + Sage (correctness)
Reviewed-by: GPT-5.6 Sol + Tuner (performance)
Reviewed-by: GPT-5.6 Terra + Scholar (learner)