[BigDataDZ]

Goal And Delivered Outcome

On native Windows, the local Codex observer and visible conversation readers open their rollout with ordinary Python read sharing, preventing the host from renaming or natively replacing that file during observation or history/content readback. The existing rotation regression also used os.replace, which cannot overwrite an open Windows destination even with delete sharing. A read-only observer should allow native host rotation and recover the replacement's state.

The provider now opens Windows rollout files with GENERIC_READ, OPEN_EXISTING and read/write/delete sharing. The returned binary stream owns a noninheritable handle and cannot write or create the source. The same transport now covers visible history paging and content reads, repairing the companion gap in this PR. Existing activity identity checks discard a mid-read replacement; the next observation recovers the new completed event. Visible reads retain the opened snapshot; subsequent paging resets for the replacement identity and old content references are rejected. POSIX keeps its existing Path.open("rb") behavior. The Windows fixture uses real ReplaceFileW, with and without observation caching; POSIX retains os.replace.

• Outcome basis: reproduced native Windows failure under the existing bounded local Codex activity contract; intended base main, implementation baseline 19ccb632845ceae49ca6463ef26e868e2dcc1a45. No issue closeout is claimed.
• Owner and placement: provider-specific OS IO in loopx/hosts; shared typed activity/state decisions remain unchanged. This stable Python provider adapter does not add a parallel semantic owner, capability or configuration input.

Author Declaration

• Written by: model_agent (OpenAI Codex), with AI assistance under the author's direction.
• Specification and revision: Bounded local Codex activity reads.

Criterion	Disposition	Evidence
Read-only provider does not interfere with native rotation	implemented	Actual rename/native replacement; reader write attempts fail, source remains intact, handle closes and is noninheritable.
Replacement invalidates old facts and recovers the current source	implemented	Existing race regression, optional cache off/on: unknown during replacement, then exact replacement completion timestamp.
Bounded reads, source identity, session/header and cache boundaries remain	implemented	Full cursor and host-activity suites, including real byte-budget instrumentation and synthetic SQLite thread lookup; existing visible-history paging, content, stale-reference and privacy cases pass.

Scope And Continuation

Complete within this scoped repair. A small provider IO seam keeps Win32 handle mechanics separate from the existing cursor, visible-history projection and typed decisions. The bounded future-facing pass reuses that seam at both previously omitted history call sites without a new module or semantic owner. The owning reference and existing native Windows CI now cover the reader/cursor suites. No frontend, Lark, CLI option, scheduler, model or execution authority changes; no new screen or control is needed. Maintainer review/merge and installed live-host qualification remain separate.

Validation

• Tested integration revision: 48c24a21d2093cb7d969a28250dca68aa0dafb0c. Fetched-main parent: 59464f9741fe177381743eadaaffebbb54e5472e. The automatic integration preserved the original reader implementation; the only new authored delta is durable native constructor-failure coverage.
• Run state: finished. Input classes: synthetic.

Check kind	Result	Public-safe evidence / limitation
real_entrypoint / real_backend	passed	Native Windows, CPython 3.12.10, local file IO and SQLite 3.49.1 in disposable synthetic homes. tests/test_codex_rollout_file.py, tests/test_codex_rollout_cursor.py, tests/test_host_thread_activity.py, tests/presentation/test_codex_attached_conversation.py: 159 passed, 32 skipped after current-main integration. Production observe_codex_threads, read_messages and read_content use real file/mmap/SQLite backends. Desktop Unix-peer HTTP cases skip on Windows; packaged frontend and installed live-host validation were not run.
regression_parity	passed	Original reader behavior fails both native replacement/cache cases with sharing violation 32; corrected reader passes. Both newly added visible message/content reader cases were first run red with native ReplaceFileW sharing violation 32, then pass with the shared reader, including replacement reset/readback and stale-content rejection. Independent native probes confirm ordinary reader blocks rename and ReplaceFileW, while shared reader permits both. os.replace of an open Windows destination remains unsupported by that OS operation.
static / unit	passed	Ruff and whitespace pass for the companion delta. Earlier workflow YAML and committed bilingual RFC check; advisory coinage scan has no supported new vocabulary carriers; full semantic vocabulary smoke passes with its existing coverage limitations.
manual / causal replay	passed	Previous independent author-side native execution on db4ed16a: 147 passed, 25 skipped. Same-head intervention restoring only the three production openers to ordinary Path.open("rb"): 142 passed, 25 skipped, 5 failed, all native sharing violations. This is an old-opener intervention, not a separately checked-out baseline suite. Native handle probes confirm read-only/noncreating access, noninheritance and closure.
hosted native Windows (prior head)	passed	Exact-head Windows job, Windows Server 2025 / Python 3.11.9: named rollout qualification step executes reader/cursor modules, 46 passed. Other CI failures remain; no all-green claim.
constructor failure / mutation control	passed	Real native handles are allocated before injected CRT adoption or stream-construction failures. Both failures preserve the original exception, close the handle and retain source bytes. Temporarily removing both cleanup calls process-locally makes both cases fail (2 failed); the durable reader module is 6 passed with the implementation intact. No production cleanup changes were needed.
historical broader integration	failed	Broader module-budget/workflow run: 533 passed, 22 failed initially. Adding existing Git Bash to the test session PATH resolves 21 missing-shell cases (21 passed on rerun); the remaining test_six_shards_execute_each_test_once_and_merge_portable_coverage expects a diagnostic stack marker absent on this Windows host. The identical assertion fails when replayed against the unchanged baseline workflow. No timeout, assertion or CI gate was weakened.

Coverage and gaps: native reader ownership, source replacement/append, bounded recovery and actual provider readback are covered. Local POSIX execution, live UNC shares and an actual installed Codex writer were not exercised. Constructor-failure injection covers the two handle-ownership transfer boundaries, not arbitrary allocator failures. Native synthetic fixtures establish file/observer mechanics, not live-host adoption. The remaining workflow diagnostic failure is preserved; the prior-head scoped hosted Windows qualification passes, current-head hosted CI and maintainer re-review are pending. Historical broader CI failures are preserved. No live Codex store, real Goal, model/provider Turn, global upgrade, publication or self-merge was performed.

Frontend / Visual Evidence

UI impact: none. Before/after/first viewport/craft review: N/A; this changes an existing provider's file access, with no presentation change.

Type / Area / Direction

Bug fix, documentation and focused validation; host/runtime integration and CI. Core control-plane reliability through the existing provider boundary. No shared-authority fixture or registered vocabulary change.

[loopx-agent]

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

Exact head: db4ed16a88e0c254a63ae5b9bbd47a7b146b4cd6

动机

在 Windows 上读取 Codex 活动状态和可见历史的用户。 宿主在观察器或历史读取器打开日志期间轮换文件时，读取器应放行原生替换，随后丢弃旧状态、恢复新日志；可见历史仍应保持已打开快照，下一页重置。 本轮真实 SQLite、文件、mmap 及公开读者的 POSIX 对照保持相同，替换后重置和旧内容引用拒绝通过；Windows 原生分支尚无本轮独立执行结果，不能据此认证目标平台问题已修复。 本轮不安装到真实 Codex、不启动宿主写者，也不将作者的 Windows 声明或 POSIX 通过称为独立 Windows 通过。

改动思路

把平台句柄规则放在现有 hosts provider IO 边界是合理的；活动游标和历史投影共用同一个只读打开器，源身份、分页和状态判断仍由各原 owner 持有。Windows 默认申请读取访问并允许读写删除共享，POSIX 原路径不变。 当前边界是三个真实日志读入口的文件打开、替换和恢复，不应新建状态协议、配置开关或把文件共享提升为执行权限。

No change leaves the reported Windows sharing defect. A shared OS provider seam at all three real readers is smaller than rewriting cursor/history ownership. The remaining action is native execution evidence, not speculative new machinery. 原权威输入和失败恢复边界保留：只读观察器不会成为宿主写者；打印安装指导不会成为自动安装授权。这里用文字说明前后行为即可，不引入未核实图示或新的用户确认步骤。

具体改动

接受的修改前规格为 docs/reference/protocols/host-integration-surface-v0.md，不可变版本 19ccb632845ceae49ca6463ef26e868e2dcc1a45。exact-session-read-only：implemented，pr7222-original-tests.log; source inspection; actual POSIX source byte hashes unchanged；bounded-completed-event-recovery：implemented，172 original tests and 5 complete base/head consumer pairs；replacement-cache-source-validation：deferred，POSIX source replacement/cache paths pass, but the new Windows transport and mapped visible history replacement were not independently executed here.。新增文档和作者自定义清单不是独立验收来源；平台实现仍需符合原始只读、来源恢复或既有目录与回滚契约。

关键代码讲解

• open_rollout：OS IO and handle ownership；Windows reads gain sharing DELETE; POSIX direct Path.open unchanged。输入由原调用者提供，错误/重试仍归原 owner；未把内部助手变成新执行能力。

• _rollout_activity：Bounded exact-source activity observation；same pre/post source guards and cursor, now shared opener。输入由原调用者提供，错误/重试仍归原 owner；未把内部助手变成新执行能力。

• read_messages：Visible opened-snapshot mmap and cursor reset；same source stamps and projection through shared opener。输入由原调用者提供，错误/重试仍归原 owner；未把内部助手变成新执行能力。

• read_content：Read authorized content_ref against file source；same stale-ref rejection and display limit through shared opener。输入由原调用者提供，错误/重试仍归原 owner；未把内部助手变成新执行能力。

完整审查7个路径 +206/-12，包括全部生产分支、相邻未修改消费者、持久测试和文档/工作流影响。没有新增共享状态、枚举、CLI参数或第二策略owner；Python保留为现有平台IO或安装渲染适配层，不扩展无关迁移。

对主干的风险

暂缓批准：新增 Win32 分支的关键后置条件尚未独立执行。 这不是已经证明的代码缺陷，也不要求为本机平台限制修改源码。当前机器只能进入 POSIX 分支；还原旧 Path.open 后同样172通过，5组完整观察也同样通过，恰好说明这些结果不能区分新的 Windows 修复是否有效。最小补充是在同一个不可变头上提供可独立重放的原生 Windows base/head 结果：普通读取控制应出现原共享冲突，头部应允许真实 ReplaceFileW；缓存关/开分别由unknown恢复到替换日志的精确完成时间；可见消息/mmap仍返回已打开快照，下一页重置，旧内容引用被拒绝；读取不能创建或写入源，句柄关闭且不继承。可直接重跑四个现有模块及新增具体用例，不需要新协议、迁移或放宽断言。

独立本机四模块本头 172 passed；相同套件接回旧打开器为 172 passed。5组真实SQLite/文件/mmap公开观察完整相等，无语义归一化，覆盖open、idle、abort、不完整尾记录、错误header；另实际替换后分页重置、旧内容拒绝且读者未改写源。local quick premerge 6直接+4选中检查通过。这些支持POSIX兼容和架构放置，不能认证Win32目标平台。

作者声明的原生Windows 147 passed/25 skipped、red/green sharing violation32和42项补充验证均已读，来源明确保留为作者声明；其更广533通过22失败、21项shell PATH恢复及1项基线诊断断言失败也未重写为本轮已验证。这些不是新代码缺陷的证明。本轮未查询或等待远程CI，当前wait_for_ci=false；没有用远程CI缺失作为阻塞。

API静态核对依据：Microsoft CreateFileW 将共享删除和本句柄访问权分开；Microsoft ReplaceFileW 描述被替换文件的访问/共享条件。这支持代码所用机制的方向，但并非实际Windows调用、CRT句柄转移或mmap恢复的执行证明。

我的整体评价

REQUEST_CHANGES（目标平台验证缺口），未发现已复现的新代码缺陷；problem_context 为 not_yet_proven，long_horizon 与 user_experience 的 Windows 实际改善同样未证明。架构判断 retain，设计边界和代码量合理，当前 hold 来自最强主张的真实后端证据缺口。 把平台句柄规则放在现有 hosts provider IO 边界是合理的；活动游标和历史投影共用同一个只读打开器，源身份、分页和状态判断仍由各原 owner 持有。Windows 默认申请读取访问并允许读写删除共享，POSIX 原路径不变。 当前边界是三个真实日志读入口的文件打开、替换和恢复，不应新建状态协议、配置开关或把文件共享提升为执行权限。

检查了现有活动、历史和内容读取路径；无页面、布局、组件或视觉素材改动，视觉评审不适用。 本轮未独立执行 Windows CreateFileW/open_osfhandle/ReplaceFileW、Windows mmap 轮换、UNC/长路径和失败清理，也未执行 live Codex 写者或已安装 App。 通过、失败与未测试范围分别保留；合并、安装采用和其他PR验收仍须单独满足其现有边界，本轮未合并或安装到个人宿主。

English verdict: REQUEST_CHANGES for an unresolved target-platform validation gap, not a demonstrated code defect. POSIX parity and local repository checks pass, but they never execute the changed Win32 path. Independently replay the unchanged-head native replacement, mapped-history and recovery cases; author-reported Windows runs are clearly attributed and are not this reviewer’s execution.

[BigDataDZ]

Independent author-side native Windows follow-through on unchanged head db4ed16a88e0c254a63ae5b9bbd47a7b146b4cd6 addresses the platform-evidence gap in the outstanding review. This is additional author evidence, not approval by the previous reviewer.

• Native Windows 10.0.18363 / CPython 3.12.10: reader, cursor, activity and attached-conversation suites 147 passed, 25 skipped; skips are Unix-peer cases.
• Causal control at the same immutable head: restoring only the three production opener references to ordinary Path.open("rb") gives 142 passed, 25 skipped, 5 failed. All five failures are native sharing violation 32: rename, visible messages/mmap, content and activity rotation with caching off/on. This is a same-head old-opener intervention, not a separately checked-out baseline suite.
• Exact-head hosted Windows qualification ran both reader/cursor modules on Windows Server 2025 / Python 3.11.9: 46 passed. This proves the relevant step executed; other jobs still fail.
• Actual handle probes confirm read-only/noncreating access, noninheritance and closure. Native replacement tests cover UNKNOWN during the race, then the exact replacement timestamp; real mmap/content reads retain the opened snapshot, reset on replacement and reject stale content references.

Reproduce the four-suite run from this checkout using its compatible activated test environment:

python -m pytest -q tests/test_codex_rollout_file.py tests/test_codex_rollout_cursor.py tests/test_host_thread_activity.py tests/presentation/test_codex_attached_conversation.py

The old-opener control is process-local: before pytest imports test modules, set open_rollout in loopx.hosts.codex_rollout_file, loopx.codex_app_thread_activity and loopx.capabilities.native_chat.codex_history to lambda path: path.open("rb"), then run the same modules. Platform detection is unchanged.

Live installed Codex writers, UNC servers and allocation-failure injection remain untested. No code/head change or merge is included; maintainer re-review remains required.

[BigDataDZ]

Integrated fetched main 59464f9741fe177381743eadaaffebbb54e5472e at signed head 48c24a21d2093cb7d969a28250dca68aa0dafb0c, without changing the reader implementation. Added durable native constructor-failure coverage to the existing Windows-qualified reader suite.

• Fresh four-module native run: 159 passed, 32 skipped (Unix-peer cases).
• Both ownership transfers are exercised with actual Windows handles: injected CRT adoption failure closes the unadopted handle; injected stream-construction failure closes the adopted descriptor and underlying handle. Both preserve the original exception and source bytes.
• Process-local mutation removing both cleanup calls causes 2 failures; normal reader suite is 6 passed. An independent author-side reviewer reran all six on native Windows and found no actionable blocker. This is not formal GitHub approval.
• Ruff, whitespace, semantic advisory and bilingual RFC checks pass. The complete PR delta remains seven owned paths against the integrated main. No production cleanup rule, permission or state contract was added.

The body now distinguishes fresh integrated evidence from the prior-head hosted Windows result. Current-head CI/re-review remain required. Live installed Codex writers and UNC servers are still untested; no live state, model turn, global upgrade or self-merge occurred.

[loopx-agent]

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

REQUEST_CHANGES — 当前 head 48c24a21d2093cb7d969a28250dca68aa0dafb0c 的阻塞项是 Windows 关键路径的独立验证缺口，未发现可独立复现的实现缺陷。 新增的构造失败清理用例有价值；本轮重新审查完整 head，不沿用旧 db4 结论。

动机

Windows 上查看 Codex 活动或历史的用户会在宿主轮换日志时遇到读取与替换互相干扰的风险。
旧路径直接打开日志；新路径在 Windows 请求只读访问及读、写、删除共享，让宿主可以继续重命名或替换，读取者仍返回已打开快照。
本轮已独立验证 POSIX 的真实 SQLite、mmap、活动投影和历史正文路径保持一致；Windows 关键分支及新清理用例尚未在本轮独立执行。
本 PR 不改变活动判定、扫描预算、配置或权限，也不要求重做整个 App、启动模型或完成其他路线图工作。

读取观察不能干扰宿主写入，这会直接影响后续活动判断及历史连续性。三个现有入口共用一个打开函数是有用、可回退的修复方向；本次没有因为 PR 小而否定它，也没有把整个 App 或其他路线图验收转嫁给作者。当前判断为 not_yet_proven：实现方向合理，本轮实际执行尚未覆盖它要修复的平台效应。

改动思路

先按既有规范判断：docs/reference/protocols/host-integration-surface-v0.md，接受的不可变版本 59464f9741fe177381743eadaaffebbb54e5472e，条目 Bounded local Codex activity reads。此标题没有数字验收编号；这里使用原文标题，要求从已选线程的宿主完整记录做有界只读观测，保留来源、header、cursor 校验和不可用时的 unknown，不向 Codex home 写入。当前实现保留了这些 owner；文档新增描述本身不作为证明。

统一 opener 是三个现有消费者之间最小的 IO 共享点；活动分类、分页和权限继续由原 owner 决定，没有新增并行状态或策略。
当前边界是 Windows 文件打开与所有权清理、POSIX 兼容性及现有读回路径；该边界可独立回退，但原生验证缺口仍在本 PR 内。

路径	旧行为	当前行为及验证边界
POSIX	Path.open("rb")	直接返回同一操作；本轮真实调用前后观察相同
Windows	普通二进制打开	CreateFileW 只读访问、共享读/写/删除，再转 CRT fd/stream；本轮未原生执行
转换失败	普通流负责关闭	fd 接管前关原生 handle；stream 接管前关 fd；两种新增原生用例本机跳过

这是从源码得出的控制流说明，表格不是执行证据。CreateFileW 的共享模式、ReplaceFileW 的访问要求与实现意图一致；_open_osfhandle 的所有权转移规则也支持两段清理。但 API 文档不能证明这份代码在实际内核、CRT、mmap 组合上通过。

具体改动

完整差异为七个路径、+259/-12：47 行 loopx/hosts/codex_rollout_file.py、活动及历史三个 opener 调用、两组 owner 测试、提供者契约说明，以及 Windows 工作流中的两个测试模块。没有配置、权限、schema、缓存或模型指令改动。检索 CreateFileW/open_osfhandle/FILE_SHARE_DELETE/open_rollout 及既有消费者后，未发现应复用的旧原生共享 opener；复用现有分类和分页 owner 比复制它们更合适。这个 OS IO helper 保持 Python 有实际放置理由，无需另起 TypeScript 决策 owner。

• open_rollout:10 接收已选路径。POSIX 立即返回；Windows 用 GENERIC_READ、共享位 0x7 和 OPEN_EXISTING，安全属性为空，再以 binary/non-inherit 标志接管。共享删除允许别的句柄操作，不给读取者写或删除权限。错误路径分别由当前资源 owner 清理，再抛出原异常。
• _rollout_activity 的调用 保留既有 exact-source、完整记录和字节预算。缺失或错 header 不得制造执行状态；部分末尾记录不替代已完成事件。
• read_messages 的打开点 和 read_content 的打开点 继续使用只读 mmap 及原来的来源校验。打开后的快照、下一页重置与旧 content ref 拒绝不能混为一个成功断言。

独立验证使用完整不可变 base/head、同一真实 synthetic SQLite/JSONL 输入：六种线程状态、cache 开关、实际 attach_host_thread_activity 投影、四组完整消息与独立正文读回，结果无语义归一化且完全相同。真实 mmap 打开期间替换文件后，旧快照保留，下一次读取采用新源，旧正文引用被拒绝；输入字节指纹未改变。这些证明 POSIX 兼容性，不能区分 Windows 修复与旧 opener。

当前四个 owner 模块为 189 passed / 2 skipped，进程退出 0；跳过的恰好是 descriptor/stream 两种原生所有权失败。标准本地 premerge 为 6 个直接检查 + 14 个选择检查通过，修改 Python/测试文件 Ruff 通过。本轮没有读取、轮询或等待 CI。首个私有探针误取 activity event 而非 assistant 消息，修正为独立 role/text fixture 断言后重跑；该准备错误保留，未算作产品失败或通过。

对主干的风险

[P2] 原生 Windows 效应尚未独立验收。 位置为 open_rollout 的 Windows 分支。全部 POSIX 用例都可能在原生共享标志、ABI、句柄清理或替换/mmap 交互有误时仍通过，因此现有本机结果不够消除该风险。这个结论不是“已证明 Windows 有 bug”，也不是要求等待远程 CI。

作者已经补充当前 48c 的原生 Windows 159 passed / 32 Unix-only skips、构造清理模块 6 passed 及去掉清理后 2 failed；另有旧 db4 的 147/25 和 old-opener 干预 142/25/5 个 sharing-violation 失败。这些材料逐条保留为作者报告，没有重新称作缺少作者说明，也没有把旧广域环境失败算到当前改动。它们尚不等于本轮独立验证；本机 Darwin 无 WinDLL/msvcrt，未观察到配置好的原生 Windows 执行路线。

最小补证：由具备原生 Windows 环境的独立验证者，针对这个完整 head 重放只读句柄、替换/mmap、三个实际消费者及两种构造失败，并保留源码、命令、环境和结果；对旧 opener 的相同干预应能使目标断言失败。 建议重放当前四模块命令 uv run --extra test python -m pytest -q tests/test_codex_rollout_file.py tests/test_codex_rollout_cursor.py tests/test_host_thread_activity.py tests/presentation/test_codex_attached_conversation.py，并明确 ReplaceFileW、可见 mmap 快照、content ref 与 cleanup 的负例；不需要重写业务代码或再造测试框架。实时宿主、UNC/超长路径若未执行，应继续标明，而不能由普通本地路径推断。回滚只涉及 opener 与三处调用，不需要迁移状态。没有新增通用状态分类、主体权限、default-off 声明或隐藏的调度义务。

我的整体评价

当前完整改动的体积和放置是合比例的：共享 IO、原有决策 owner、聚焦测试，结构上无需再拆。统一 opener 是三个现有消费者之间最小的 IO 共享点；活动分类、分页和权限继续由原 owner 决定，没有新增并行状态或策略。 当前边界是 Windows 文件打开与所有权清理、POSIX 兼容性及现有读回路径；该边界可独立回退，但原生验证缺口仍在本 PR 内。

本次没有修改渲染界面；已检查活动投影及历史消费者，UI 截图不能证明 Windows 内核共享与 CRT 资源行为。 Windows 原生分支的执行、实时宿主写入及 UNC 路径未由本轮独立验证；本机 POSIX 通过不能替代这些证据。 用户完成既有读取任务无需新增配置、复制或确认轮次；这保持了交互成本，但 Windows 读取连续性仍是未证实的目标收益。当前无独立复现的代码缺陷，也没有把测试数量当作长程效果或用户价值证明。待上述独立原生补证到达这个确切 head 后，可就缺口复审；本评审不授权合并、部署或启动模型任务。

English verdict: REQUEST_CHANGES at 48c24a21d2093cb7d969a28250dca68aa0dafb0c for an independently unqualified native Windows invariant, not a reproduced code defect. The shared IO seam and cleanup coverage fit the problem. Independent POSIX public-path parity, 189 passed/2 native skips and local premerge pass; author Windows reports are preserved separately. A source-bound independent native replacement/mmap/three-consumer/cleanup replay with a discriminating negative control is the minimum remaining validation. CI was not consulted; no merge authority is asserted.

[BigDataDZ]

Rechecked the outstanding review at unchanged head 48c24a21d2093cb7d969a28250dca68aa0dafb0c on native Windows / CPython 3.12.10. The remaining request is independent native validation, not a newly reproduced implementation defect; I am preserving that distinction.

Fresh author-side four-module run: 159 passed, 32 Unix-only skips. The same-head process-local old-opener intervention (excluding the two constructor tests that no longer apply to that opener) gives 152 passed, 32 skipped, 5 failed. All five failures are native sharing violations: rename, visible messages/mmap, content, and cached/uncached activity replacement. Separately removing just the two cleanup calls in memory gives 2 failed constructor ownership tests; the durable tests close the injected leaked resources. Tracked source bytes and head were not modified by either control.

This confirms the existing native evidence is reproducible here and distinguishes the repair from ordinary Path.open. It is still author-side evidence, not independent approval or dismissal of the current review. Independent exact-head Windows replay remains the review gate. Live installed Codex writers, UNC/network paths and allocation failures remain unqualified. No model, benchmark, active Goal or merge action was used.