Normalize zero-point deduction traces / 规范化零分扣分依据条目

[yyuan308]

Summary / 摘要

• Normalize one non-semantic provider transport variant: a deduction-trace entry with points_deducted: 0 is removed before validation.
• 规范化一种不改变语义的模型输出形式：在校验前移除 points_deducted: 0 的扣分依据条目。
• This never adds or changes points. A non-full leaf still fails unless its remaining positive trace entries sum exactly to max_score - score.
• 此改动绝不补分或改分。非满分叶子项仍必须由剩余正分 trace 精确满足 max_score - score，否则校验失败。

Validation / 验证

• python -m unittest discover -s tests/benchmark/core — 479 passed.
• Regression tests cover full-credit empty traces, zero-point noise removal, and continued rejection of non-full leaves without a valid trace.
• 回归测试覆盖满分空 trace、零分噪声清理，以及对无有效 trace 的非满分叶子项继续拒绝。

Privacy and limits / 隐私与局限

• Public runner and synthetic tests only; no Data/, submissions, identities, transcripts, gold scores, raw outputs, or metrics.
• 仅包含公共运行器和合成测试；不包含 Data/、作答、身份、转录、gold、原始输出或指标。
• This is a narrowly scoped output-contract compatibility fix, not a scoring-rule, rubric, or accuracy change.
• 这是严格限定的输出契约兼容性修复，不改变评分规则、rubric 或准确率结论。

Next step / 下一步

• After merge, restart DeepSeek V4Flash from a fresh private output directory under the frozen r3 plan. M1-Codex continues separately; heldout remains untouched.
• 合并后，在冻结的 r3 计划下从全新私有输出目录重启 DeepSeek V4Flash。M1-Codex 独立继续，heldout 保持不访问。