Fix r3 strict deduction-trace schema compatibility / 修复 r3 严格扣分依据 Schema 兼容性

[yyuan308]

Summary / 摘要

• Make the r3 deduction_trace_v1 output schema compatible with Codex strict structured output: nullable trace and attention fields are required structurally, while remaining semantically optional when not applicable.
• 使 r3 deduction_trace_v1 输出 schema 兼容 Codex 严格结构化输出：扣分依据和注意事项字段在结构上必填，但在不适用时可为 null，其评分语义不变。
• Normalize only the provider spelling deduction_trace: [] to null for full-credit leaves. Non-full leaves with an empty trace still fail validation.
• 仅将满分叶子项中的 deduction_trace: [] 规范化为 null；非满分叶子项的空 trace 仍会严格校验失败。

Validation / 验证

• python -m unittest discover -s tests/benchmark/core — 477 passed.
• Real zero-data r3 contract probes passed for both Codex and official stable DeepSeek V4Flash. The probes used only synthetic content and were automatically deleted.
• 真实零数据 r3 契约探针已分别通过 Codex 和官方稳定版 DeepSeek V4Flash；探针仅使用合成内容并自动删除。

Privacy and reproducibility / 隐私与可复现性

• This PR contains only public runner, schema, and synthetic-test changes. No Data/, submissions, identities, transcripts, gold scores, raw model output, or metrics are included.
• 本 PR 仅包含公共运行器、schema 和合成测试改动；不包含 Data/、作答、身份信息、转录、gold、原始模型输出或指标。
• Existing interrupted r3 attempts are private audit artifacts only and do not count as experimental results.
• 已中断的 r3 尝试仅保留为私有审计材料，不计入实验结果。

Limitations and next step / 局限与下一步

• This fixes output-contract compatibility only; it does not change the frozen rubric, human review decisions, or make any accuracy claim.
• 本修复仅解决输出契约兼容性；不改动冻结 rubric、人工复核决定，也不作任何准确率主张。
• After review and merge, rebuild the private r3 development packets with new hashes, then run M1-Codex, G1-Codex, and G1-DeepSeek V4Flash with aggregate-only progress. Heldout remains untouched.
• 审核合并后，将用新 hash 重建私有 r3 开发集 packet，再运行 M1-Codex、G1-Codex 与 G1-DeepSeek V4Flash，仅报告聚合进度；heldout 保持不访问。