fix: 有界部分扣分 trace / bounded partial-credit traces

[yyuan308]

摘要 / Summary

• 执行型 rubric 的 criterion 分值改为“可扣分上限”：有作答证据时允许按 score step 的部分扣分，但不得超过 criterion 上限、不得重复 criterion、不得使用未声明 criterion。
• Treats execution-rubric criterion points as deduction caps: evidence-supported partial deductions are allowed on the score step, but may not exceed a cap, repeat a criterion, or use an undeclared criterion.
• 仍严格要求每题 deduction total 等于分差。仅当唯一一个已选 criterion 可在上限内修复加法笔误时，才做不改变 score 的私有归一化；有多个可能修复时仍拒绝并重试。
• Deduction totals still must match the score gap. Only a uniquely determined, cap-bounded arithmetic repair is normalized privately without changing the score; ambiguous cases remain rejected and retried.

验证 / Verification

• python -m unittest tests.benchmark.core.test_execution_contract_rubric tests.benchmark.core.test_deduction_trace_contract tests.benchmark.core.test_model_runner — 41 passed
• python scripts/run_ci_tests.py core — 486 passed
• 新增覆盖：部分扣分、超上限拒绝、唯一修复、歧义修复拒绝。

隐私、限制与下一步 / Privacy, limitations, and next step

• 只包含公共 runner 和合成测试；未暂存或提交 Data/、学生内容、成绩、转录、gold 或模型输出。
• Contains only public runner code and synthetic tests; no Data/, student content, scores, transcripts, gold, or model output is staged or committed.
• 不报告模型指标。合并后会从全新目录重启三条已对齐的开发集路线；此前所有部分输出仍不可报告。