docs(benchmark): 落地 DeepSeek 24-cell 诊断报告

[627150795]

Problem

Issue #22 已经记录了 24-cell DeepSeek 0731 对照与收敛消融，但结果、冻结条件、机制命中情况和统计限制尚未进入仓库内的正式报告，后续引用和复核不方便。

Value

把已有公开收据版本化，保留结果边界和失败/超时信息，避免把本轮小样本观察误读成官方成绩、稳定提速或总体排名。

Approach

• 新增 docs/research/OPENPI_BARE_PI_DEEPSEEK_24_CELL_DIAGNOSTIC.md。
• 记录实验契约、最终对照表、逐题结果和实际命中的收敛机制。
• 记录 ActiveEvidence、recovery/trajectory、delayed projection 三组消融及被拒绝的原因。
• 保留源码、runner、runtime guard hash、timeout 和 artifact 可审计边界。
• 本 PR 只整理 Issue [Benchmark] OpenPI 10/12 vs Bare Pi 6/12：DeepSeek 0731 下执行动态持平，tokens -37.7% #22 已公开的材料，不重新跑实验、不修改运行时代码。

Validation

• git diff --check 通过。
• 报告一致性检查通过：24-cell、模型、三个 hash、非官方成绩限制和原始 artifacts 边界均存在。
• 已逐项对照 Issue [Benchmark] OpenPI 10/12 vs Bare Pi 6/12：DeepSeek 0731 下执行动态持平，tokens -37.7% #22 的公开数字、限制和下一步，不新增实验结论。
• Biome 不处理 Markdown 文件；没有用格式化工具改写报告内容。

Impact

• 用户可见行为：新增一份研究报告。
• 模型可见工具与运行时：无变化。
• 配置、artifact 和兼容性：无变化。
• 报告只声明 Issue [Benchmark] OpenPI 10/12 vs Bare Pi 6/12：DeepSeek 0731 下执行动态持平，tokens -37.7% #22 的公开收据能够支持的窄结论。

Fixes #22

[tt-a1i]

This Benchmark report remains open under #277. Its merge gate is source-evidence review against Issue #22: frozen revisions, task/model/verifier identities, accounting, failure classification, limitations, and evidence references must all read back correctly. Existing CI is repository validation, not independent validation of the Benchmark conclusion.

[tt-a1i]

Blocking evidence finding on exact head a6112a0d58e83e6439cbc38c66de31425af2aed5:

The prose matches Issue #22, and the 24-cell local evidence directory still exists, but openpiContentSha256 alone is not a retrievable source revision and the PR intentionally omits the artifacts needed to inspect or recompute the 10/12 vs 6/12 tables. Under #278's Benchmark contract, keep this report unmerged until it is moved to docs/benchmarks/, given the governed metadata/status, linked to a retrievable credential-free evidence receipt/archive identity, and tied to a source commit or immutable source package. Do not bulk-publish raw Sessions, traces, credentials, or candidate workspaces.

Also replace the historical tt-a1i/openpi Issue links with the canonical openpi-dev/openpi identity (or explicitly label them as historical redirects).

No runtime finding; this is a publication/evidence-state gate.