Clarify r3 deduction-trace retry feedback / 明确 r3 扣分依据重试反馈

[yyuan308]

Summary / 摘要

• On a structural retry, r3 grading now repeats the exact deduction-trace invariant: full credit uses deduction_trace=null; every non-full leaf needs a non-empty trace whose deducted points sum exactly to max_score - score.
• r3 评分在结构重试时会重申精确扣分依据约束：满分使用 deduction_trace=null；每个非满分叶子项必须有非空 trace，且扣分总额严格等于 max_score - score。
• The feedback contains contract rules only. It never includes a prior response, answer evidence, student identifier, or private validator details.
• 该反馈只包含契约规则，不包含先前模型输出、作答证据、学生标识或私有校验细节。

Validation / 验证

• python -m unittest discover -s tests/benchmark/core — 479 passed.
• Adds regression coverage for the retry invariant and uses the same feedback in text and headless runners.
• 增加重试约束回归测试，并让文本与 headless 运行器使用同一反馈。

Privacy and limits / 隐私与局限

• Public code and synthetic tests only; no Data/, submissions, identities, transcripts, gold scores, raw outputs, or metrics.
• 仅含公共代码和合成测试；不包含 Data/、作答、身份、转录、gold、原始输出或指标。
• This is output-contract reliability work, not a rubric revision or an accuracy claim.
• 这是输出契约可靠性改进，不是 rubric 修订，也不构成准确率主张。

Next step / 下一步

• After merge, run DeepSeek V4Flash from a fresh private output directory against the frozen r3 G1 packet. M1-Codex continues separately; heldout remains untouched.
• 合并后，从全新私有输出目录使用冻结的 r3 G1 packet 运行 DeepSeek V4Flash。M1-Codex 独立继续，heldout 保持不访问。