Clarify r3 deduction-type output contract / 明确 r3 扣分类型输出契约

[yyuan308]

Summary / 摘要

• The r3 runtime contract now lists every allowed deduction_type token and explicitly forbids invented aliases.
• r3 运行时契约现在列出全部允许的 deduction_type token，并明确禁止模型自造别名。
• This addresses observed output-vocabulary validation failures from DeepSeek V4Flash without changing any rubric, score, human review decision, or gold reference.
• 该改动解决 DeepSeek V4Flash 已观察到的输出词汇校验失败；不改动 rubric、分数、人工复核决定或 gold 参考。

Validation / 验证

• python -m unittest discover -s tests/benchmark/core — 478 passed.
• A real zero-data, non-full-credit DeepSeek V4Flash probe passed the complete deduction_trace_v1 contract after the prompt clarification.
• 使用合成且自动删除的零数据“非满分”探针，DeepSeek V4Flash 已通过完整 deduction_trace_v1 契约。

Privacy and limits / 隐私与局限

• This PR contains only public runtime-prompt code and a synthetic regression test. No Data/, submissions, identities, transcripts, gold scores, raw outputs, or metrics are included.
• 本 PR 仅包含公共运行提示代码和合成回归测试；不包含 Data/、作答、身份、转录、gold、原始输出或指标。
• It makes no grading-accuracy claim. Interrupted private DeepSeek attempts remain non-reportable audit artifacts.
• 本 PR 不作评分准确率主张；已中断的私有 DeepSeek 尝试仍仅为不可报告的审计材料。

Next step / 下一步

• After merge, restart DeepSeek V4Flash from a clean private output directory against the already rebuilt r3 development G1 packet; heldout remains untouched. M1-Codex continues under its separately validated route.
• 合并后，使用已重建的 r3 开发集 G1 packet 从全新私有输出目录重新启动 DeepSeek V4Flash；heldout 保持不访问。M1-Codex 按已验证路线继续运行。