[songoow]

Problem

A fresh default EdgeBench cohort currently fails before provider import with EdgeBench compatibility patch preimage differs. The default follows a moving integration branch while its full-index checkpoint patch still names older upstream blobs. Reapplying by context or dropping the guard would make the qualified source ambiguous.

Change

• Pin the default source to reviewed integration commit e278cda3f4d2da3cf06aabf2df4daa99b9749501.
• Refresh the bundle as nonfinite-checkpoint-v2 against that exact source. Its added/deleted provider code, tests and documentation are byte-identical to the previous patch; only full-index identities and hunk locations change.
• Treat the default revision and patch as a reviewed pair. A remote branch advance no longer changes a new default cohort. Explicit source options still select unmodified provider code.
• Add a synthetic real-Git regression for upstream branch movement, refusal of a new pin paired with an old patch, successful paired refresh, and unchanged historical cohort reuse.

The strict preimage, digest, complete-diff and deterministic-commit checks are unchanged. Existing cohorts validate their own frozen patch and receipt, independently of the current bundle.

Validation

Baseline: 46f1175fbbb67bf2d0d358f62efc8ade0652bbbc.

• Reproduced the original failure through the real default bootstrap against current upstream: exit 2; no source directory was published.
• Prepared two independent fresh default cohorts through the actual bootstrap. Both produced effective provider commit d0863478c09d7f929c0b2ab905e030aafb6c8c14 and identical receipts.
• Bundled patch SHA-256: 02d21096a2e55f625d3c8beee2335728caf08aa6c75a5a27b1ab412bc46323bf.
• Source selection suite: 47 passed, including offline reuse, explicit-source isolation, tamper rejection and concurrent cohort preparation.
• Patched upstream native checkpoint suite: 82 passed.
• LoopX checkpoint judge and actual HTTP capture/export/restore regressions against the freshly prepared source: 14 passed, including nonfinite reports, retained history/opportunities and fresh receiver authority.
• Ruff, public-boundary scan and staged git diff --check: passed.

The native HTTP checks use synthetic grading and an external backend stub. No model, real task or Docker grading was launched. The environment emitted a Starlette TestClient deprecation warning. Full required CI remains a separate qualification.

Scope and risk

This is an internal benchmark dependency selection and compatibility repair, with no new product-facing UI or CLI option. The observable default change is intentional: new default cohorts stay on this reviewed source until the pair is updated. Existing cohorts, scoring, selection, budget/cooldown rules and permissions retain their contracts.

This qualifies the source/transport pairing; it does not requalify historical checkpoints or establish long-horizon research benefit. Future upstream updates, including an upstream nonfinite implementation, require a fresh pairing review and the same native transport checks.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T01:59:31.272955Z	5de466a	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[songoow]

Integrated main at 2f2c4efadf0c3ac21c51f7fc09f4985c8447642e with a signed merge that preserves the existing PR history. Current head: 7616fb1480cb65108abfbffc488d0090f92237bc; tested tree: 997a8fd0fc60128dbf44223ae9b69eed0efe4db5.

The post-merge source qualification passed 47 tests. The v2 compatibility patch and provider pin remain the proposed fix; this merge also preserves the current upstream README changes.

The merge was conflict-free. The changed production/test blobs from the prior PR head remain identical; where needed, the documentation merge was inspected separately. RFC bilingual and diff hygiene checks passed. The API-published tree matches the locally tested tree.

Earlier browser CI failures remain recorded as failures; these focused checks do not turn them into passes or establish a new CI result. Required checks and independent review must qualify this new head. No merge was performed.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max

动机 / Motivation

需要从默认入口准备新 EdgeBench 实验环境的维护者和研究操作者。旧默认跟随远端分支，而兼容补丁仍要求旧文件前像；远端改到被补丁触及的文件后，新环境准备直接失败。新版本把源码提交和对应补丁固定成一对，新环境可重复准备，已有环境继续读取各自的冻结收据。已验证默认准备成功、两个新环境生成相同有效提交，错误的源码与补丁组合仍在发布前被拒绝。本 PR 不证明模型效果或净效率收益，不升级既有实验，不改变评测机会、冷却或预算，也不授权启动实验、安装到正在使用的 App 或合并。

改动思路 / Approach

故障来自默认依赖对不一致；更新现有 provider 配置和补丁前像即可恢复真实入口，无需新增决策 owner、宽松匹配或自动回退。当前 PR 只交付新默认环境的固定源码与兼容补丁对、原有冻结环境的可验证复用及其文档和回归检查。权威输入仍是现有 source-default.json 和已经冻结的 source receipt（记录实际源码身份的收据）；prepare_source 先识别既有收据，再处理新环境的默认或显式来源，最后在原有锁内校验并原子发布。没有增加额外的配置副本、自动换源或宽松补丁回退。

不修改源码/补丁就会保留已复现的失败；只改补丁会再次受远端分支推进影响，只改当前 pin 则仍过不了旧补丁前像。固定旧源码也是可行回退，但这版通过复核当前集成源码及其配套补丁，恢复当前原生入口。代价是今后的默认升级需要维护者再审查并资格化一对输入；自动跟随分支不再替代这一步。provider IO 和失败恢复仍属于原 source owner，校验失败是必须停止的完整性边界。

具体改动 / Changes

审查 head：7616fb1480cb65108abfbffc488d0090f92237bc。规范依据为修改前的 benchmark/edgebench/README.md，固定 revision 2f2c4efadf0c3ac21c51f7fc09f4985c8447642e 的 Select and freeze the native provider。以下 criterion ID 是原文段落的审查映射，并非新增要求：source-freeze 已实现，首次进入原子冻结且并发启动共享同一来源；offline-reuse 已实现，旧 v1/v2 环境离线复用自己的源码、补丁和收据；strict-preimage 已实现，摘要、完整文件前像和完整 diff 校验在导入前执行，失败不发布半成品；explicit-source 已实现，显式 ref/revision/repository 仍选择未加默认补丁的原生源码。

关键代码讲解

1. source-default.json:3 把可移动分支改为 e278cda3f4d2da3cf06aabf2df4daa99b9749501，同时更新 v2 补丁位置、身份和摘要。两个输入必须配套，后续远端变化不会悄悄进入新的默认组。
2. prepare_source:190 实现没有变化。既有收据优先于当前默认，显式来源不依赖默认补丁；新环境只在全部校验完成后发布。worker、serve 和 online_judge 仍在 native provider 导入前调用这条链路。
3. _apply_patch:153 保留严格前像和可重复派生提交。v2 资产相对 v1 仅刷新 full-index blob 身份和 hunk 位置；独立比较的 421 条实际增删行完全一致，没有借换版本修改 checkpoint 逻辑。两次真实默认准备均得到上游 e278cda3f4d2da3cf06aabf2df4daa99b9749501、有效提交 d0863478c09d7f929c0b2ab905e030aafb6c8c14，补丁 SHA-256 为 02d21096a2e55f625d3c8beee2335728caf08aa6c75a5a27b1ab412bc46323bf，完整收据相同且 checkout 干净。
4. test_pinned_default_and_patch_advance_only_as_a_reviewed_pair:226 使用真实 Git 构造远端已改被补丁触及文件的情形：旧 pin 的新环境仍能准备；只推进 pin 会拒绝且不发布目录；更新配套补丁后恢复成功；此前旧环境收据字节保持不变。它保护长期依赖漂移，而不是冻结某次实验日志。

对主干的风险 / Risks

主要风险是把“准备成功”误当成 benchmark 已资格化，或在将来更新默认时误改历史实验。源码收据只证明代码身份；真实模型、Docker 评分和整场实验收益仍需各自验收。回滚应针对未来新环境恢复整对默认输入，或在新目录显式选择已审查的未修改源码；已有组保留完整冻结目录，不热换源。显式 ref/revision/repository 的三个分支均在默认补丁不可用时通过，旧 v1/v2 在当前资产缺失时仍可离线验证，错误收据、额外提交、脏源码和不配对前像继续拒绝。这是明确的默认修复，不作 default-off 隔离宣称。

独立执行的检查	结果与边界
source 管理全文件测试	PASS：47 passed，0 skipped；含并发、离线历史、显式来源、错误输入和配套升级
固定原生源码的 checkpoint 测试	PASS：82 passed，0 skipped；实际导入上述有效派生提交
LoopX checkpoint judge 与两个 controller HTTP 路径	PASS：14 passed，0 skipped；覆盖原点/新权限及非有限公开结果
不可变 base 的默认 source CLI 与 head 对照	PASS：base 复现 exit 2 / preimage differs / 无已发布来源；head 两个新目录均成功，完整收据相同
公共边界、ruff、compile、diff/RFC 检查	PASS：五个改动文件无私有材料命中；适用检查通过

没有查询或等待 GitHub CI，也没有启动科学实验。base 复现使用精确导出的 stdlib source CLI、原默认及 v1 资产，并由同一解释器执行；它不是完整 base 安装。原生与 HTTP 检查使用隔离依赖和合成运行资源，不替代 Linux Docker 评分。以前其他 head 的浏览器桥接失败没有被宣称修复；本次适用检查没有失败或跳过。

语义与 CI 对齐

此处复用已有 source/receipt owner，v2 是局部兼容资产身份，既有 edgebench_source_v1/v2 持久格式继续验证；未添加通用 Goal 状态、substring 分类或新权限。README 明示新默认不再跟随分支，严格校验继续是机器执行的拒绝条件。diff-scoped advisory 未发现支持识别的新词汇载体；该工具不能证明不存在语义变化，人工比较和上述反例是实际依据。CI 状态与本地审查结论、合并门分别保留。

我的整体评价 / Verdict

本 PR 达成了新默认环境可重复准备这一完整、有限的目标。long_horizon 为 improved，依据是新建不再依赖移动分支维持前像且旧环境可继续离线复用；user_experience 为 improved，普通命令恢复可用，没有增加所需信息重填或确认轮次。该判断不包含模型效益。已检查 benchmark/edgebench/source.py、source-default.json、README 及三个 CLI 调用入口；没有改变渲染页面或产品视觉布局。无本 PR 改动范围内的视觉缺口；已安装 App 体验与科学实验效果未在本次验证。

五文件 66+/22− 主要为文档、回归和资产元数据。相较新增 fallback/schema，直接更新现有配套输入是合适的收敛边界；旧持久收据兼容有真实消费者，保留合理。当前未发现应阻塞此 head 的可验证缺陷，建议 APPROVE；将来换源须重新资格化，合并与发布授权另行处理。

English verdict: APPROVE this exact head. The pinned source/patch pair repairs default preparation while preserving strict verification and frozen-source reuse; no model-benefit or deployment claim is made.