fix(benchmark): run the EdgeBench report adapter and suite on native Windows

[JasonBuildAI]

Goal And Delivered Outcome

• Outcome basis / optional anchor: reproduced defect and its repair; issue [Bug]: native Windows cannot run the documented local test check or the EdgeBench report verifier, and the published study archive digests do not verify #6367. No roadmap id is claimed.
• Goal/source and gap: benchmark/edgebench/feedback_hook.py imported POSIX-only fcntl at module scope, so on native Windows the documented report verifier could not start and pytest -q (the full local check CONTRIBUTING documents) aborted collection before running anything. The published study archive is also not exempt from checkout line-ending translation, so its 13 digest bindings mismatched on a core.autocrlf=true checkout.
• Observable before → after, with the validation row that proves it: pytest -q → Interrupted: 2 errors during collection, 0 tests executed, exit 2 → 22062 tests collected with no collection errors; python -m benchmark.edgebench.export_report --verify benchmark/edgebench/studies/portfolio-ten-arm-20261006/data → ModuleNotFoundError: No module named 'fcntl' → {"runs": 10, "files": 13, "verified": true}; archive bindings → 13 of 13 mismatched → 0 of 13. Rows below.
• Issue/task and intended base: Fixes [Bug]: native Windows cannot run the documented local test check or the EdgeBench report verifier, and the published study archive digests do not verify #6367. Base main (806eb215f).

Author Declaration

• Written by: model_agent (OpenAI Codex), run on a native Windows developer host for the contributor account that owns this branch.

AI disclosure: this change was drafted with AI assistance at my request, under my direction and review.

Implemented against

• Specification and revision: no written specification; the reproduced defect and request in issue [Bug]: native Windows cannot run the documented local test check or the EdgeBench report verifier, and the published study archive digests do not verify #6367 are the basis. The line-ending remedy follows the existing pin for benchmark/deepswe-gptxhigh-v1/** (commit 71cc15692, "Preserve the immutable archive's bytes even with core.autocrlf=true"), and the guarded POSIX import follows loopx/file_lock.py.
• Criteria:

Criterion (spec clause)	Disposition	Symbol / path	Test or command
The report adapter imports and verifies where POSIX flock is absent	implemented	benchmark/edgebench/feedback_hook.py	documented --verify command (validation rows 2, 3)
Delivery still takes its session lock and fails closed without one	implemented	feedback_hook.deliver	benchmark/tests/test_edgebench_feedback_hook.py::test_delivery_refuses_a_host_without_the_posix_session_lock
The delivery suite still runs where flock exists	implemented	benchmark/tests/test_edgebench_feedback_hook.py	Linux CI shards python -m pytest -v -n 2 -m "not stage2c_e2e" include benchmark/tests (row 7)
Published study archives are byte-stable across checkouts	implemented	.gitattributes, benchmark/edgebench/studies/portfolio-ten-arm-20261006/README.md	validation rows 1 and 2
A new export keeps its bound bytes through Git publication	implemented	benchmark/edgebench/export_report.py, .gitattributes	benchmark/tests/test_edgebench_report.py::test_exported_report_survives_autocrlf_publication and test_json_writer_ignores_host_text_newline_translation

• Self-check before submission: I reproduced each failure at the base revision before changing anything, re-ran the same commands on the rebased head, and read back the diff. I did not run the flock-dependent delivery tests locally (no Linux runtime on this host); they remain behind a flock marker, while the no-lock refusal case now runs here against the real missing backend. The guarded import is a no-op where fcntl exists. I deliberately left benchmark/edgebench/online_judge.py alone (same import pattern, outside the report chain, already guarded by an optional-dependency skip) and did not touch runner, scoring or evidence semantics. Review follow-up: the reviewed P1 byte-transport gap was reproduced natively on Windows - the exporter bound CRLF digests for the JSON files, text eol=lf normalized them at git add, and a fresh materialization failed --verify with File binding failed: run-rows.json; export_report.write now writes LF bytes explicitly and the roundtrip has regression coverage.

Scope And Continuation

• Completed scope and remaining work: on native Windows the benchmark package imports, the suite collects and runs, and the published archive verifies. Nothing in the report adapter or the study archive is left half-done.
• Slice boundary / successor: complete within this scope; a future digest-bound study archive inherits the same .gitattributes rule, and the exporter writes LF bytes on every platform so published digests keep matching. benchmark/edgebench/online_judge.py keeps its unconditional fcntl import on purpose.

Validation

• Tested revision: 8e6614ca16b9236d70126411e07e40adac558533
• Run state: finished
• Input classes: synthetic (repository fixtures; no benchmark job was launched, no private data used)

Check kind	Result	Public-safe evidence / limitation
regression_parity	passed	Collection on the authoring host: base 806eb215f aborts (Interrupted: 2 errors during collection, 22020 collectible, 0 executed, exit 2); this revision collects 22062 with no errors, exit 0.
regression_parity	passed	Archive bindings for benchmark/edgebench/studies/portfolio-ten-arm-20261006/data: 13 of 13 mismatch on a fresh base checkout, 0 of 13 on a fresh checkout of this revision. An existing checkout needs one re-materialization, which the study README now documents; re-ran it here: after deleting the stale directory and restoring it, no CR bytes remain and --verify returns runs=10, files=13, verified=true.
regression_parity	passed	Review follow-up on this native Windows host: at 1357ca2 the JSON writer bound CRLF digests and a fresh core.autocrlf=true materialization failed --verify (File binding failed: run-rows.json); with newline="\n" the export stays LF and the new export -> git add -> fresh materialization -> verify roundtrip passes for both a new export and the published archive. The writer and both roundtrip tests fail at 1357ca2 and pass at 8e6614c.
real_entrypoint	passed	python -m benchmark.edgebench.export_report --verify benchmark/edgebench/studies/portfolio-ten-arm-20261006/data prints {"runs": 10, "files": 13, "verified": true} and exits 0 on a fresh checkout of this revision.
unit	passed	python -m pytest -q benchmark/tests/test_edgebench_report.py benchmark/tests/test_edgebench_feedback_hook.py -rs → 46 passed, 13 skipped (the flock-dependent delivery tests skip; the no-lock refusal case passes on this host).
manual	passed	Negative case on this host: feedback_hook.fcntl is None; deliver refuses with OSError: best-only delivery has no supported session-lock backend on this host, and no receipt directory is created.
static	passed	python -m ruff check benchmark/edgebench/feedback_hook.py benchmark/tests/test_edgebench_feedback_hook.py → all checks passed; git diff --check clean; scripts/generate_semantic_inventory.py --changed-from <base> reports no new vocabulary carriers; loopx check --scan-path benchmark/ clean.
integration	not_run	The flock-dependent delivery tests execute in Linux CI; the no-lock refusal case now runs on this host. No Linux runtime was available for the flock-dependent suite.
manual	blocked	loopx canary premerge --from-git-diff ran 8 risk-profile smokes; 5 fail identically at the base revision on this host (pnpm workspace for the DSH plugin phases; Windows path normalization in catalog-run-e2e-smoke.py and premerge-validation-gate-smoke.py). They are pre-existing host failures, not from this diff. Canary also reports the benchmark_sensitive manual hold, so this PR is left for maintainer review.
static	not_run	python -m mypy on this Windows host fails at loopx/control_plane/effect_runtime.py:437 (Module has no attribute "WNOHANG"), identically at the base revision and outside this diff; CI runs mypy on Linux.

• Coverage and gaps: the changed behavior is host availability, so it is demonstrated by running the documented commands on the affected host instead of by a Linux unit test. The two benchmark test modules cover the adapter itself and now execute there (before: collection error). The review follow-up adds the byte-transport contract: writer LF determinism plus Git publication roundtrips for a new export and the published archive. The delivery refusal case now runs on both hosts: here against the real missing backend, and in Linux CI against the simulated missing backend. CI's Windows job does not run benchmark/tests, which is why the collection abort was invisible there.
• Failures/skips observed: the pre-existing mypy and canary failures above. No failure listed here is attributed to this diff.

Frontend / Visual Evidence

• UI impact: none

Type of Change

• Bug fix
• New feature
• Breaking change
• Refactoring (no functional changes)
• Documentation update
• Test update

LoopX Area

• Control plane (goals, todos, quota, scheduler, registry, runtime)
• Benchmark boundary (adapters, runners, verifiers, scoring, evidence)
• Capability or extension (providers, adapters, skills)
• Public docs or presentation surface (README, protocols, dashboard)
• Build, packaging, installer, or CI
• Host or runtime integration

Technical Direction

• Direction / acceptance reference, when applicable: bounded host-portability repair for published benchmark evidence and its verifier; no roadmap S/G/R or acceptance id is claimed.

Shared-authority RFC fixture impact

• Production-scale fixture schema: N/A
• Semantic dimensions changed, or reviewed no-impact rationale: N/A — no shared Goal Authority or TypeScript control-plane contract is touched.
• Provider conformance arms run: N/A
• Read-only legacy/file/PostgreSQL three-arm rehearsal (required for promotion, runtime-routing, or compatibility-projection changes): N/A

Boundary Checklist

• Neither the diff nor this PR body/comments/attachments disclose private state, credentials, raw traces or verifier output, internal links, or local machine paths (including .loopx/, .codex/goals/, and live ACTIVE_GOAL_STATE.md).
• I did not duplicate maintainer-owned benchmark work unless a maintainer split out a public issue for it.
• I kept the change scoped to the linked issue/task.
• I completed the visual evidence section for UI changes, or marked UI impact none.
• Every commit includes a DCO Signed-off-by trailer (git commit -s).

[loopx-agent]

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | xhigh

动机

在 Windows 上运行报告导出、验证或本地测试的贡献者。 过去报告模块在缺少 POSIX fcntl 时连导入都失败，现有档案也可能被 Git 转成 CRLF；本改动让只读验证可启动并修好已发布档案，但新导出的 JSON 在入库后仍可能被改写，导致自己的摘要失配。 已发布13文件的验证与无锁读取路径通过；新导出反例在入库前通过，经过当前 Git 属性和重新取出后报 File binding failed: run-rows.json。 本批不提供 Windows 锁后端、启动 benchmark/model、改变评分或把摘要验证当作研究资格。

完整评审基线 806eb21 到 1357ca2 的4文件。读取现有档案的修复有正向价值，但不能用这条成功路径代替新导出报告的可用性。

改动思路

报告读取与需要 flock 的通知交付沿用现有模块，缺后端只阻断交付；档案传输必须保留计算摘要时的原始字节，不能用 Git 的文本规范化替代该不变量。 当前PR应一并修好导入、原档案取出和受该新属性覆盖的新报告往返；无需增加通用锁框架或改动评分。 保留原模块的 guarded import 与失败前无副作用检查是合适的小修复，复用了已有 file_lock 的平台边界思路。最强反对理由不是增加了几行代码，而是新属性同时覆盖后续生成档案，现有 JSON 生产者的换行字节与 Git 存储规则未一起收敛。

具体改动

接受依据 issue #6367，spec_ref=https://github.com/loopx-project/loopx/issues/6367，复现基线 spec_revision=806eb215f8a21e4ae6f51cd942112175ca3839b7；另外读取既有报告 README 的摘要绑定、只读验证及新终态导出约定。

• report_without_flock：实现。实际 CLI 在模块不可用的选择器下，基线导入失败，新 head 读取现有10次运行、13文件通过。
• lock_before_delivery：实现。三个事件类型、两个 session 的6次无后端调用均在创建目录前拒绝；真实 POSIX flock 交付和相同身份回放保持原语义。
• bytes_match_published_bindings：未满足。旧属性、相同 Windows 换行输出与当前新属性做真实 Git 往返，新属性改变两个已绑定 JSON 文件；改成不转换文本的属性，完整往返通过。

关键代码讲解

feedback_hook.fcntl 的模块导入现在允许缺失后端，不把读取报告的可用性当作通知权限。deliver 在第28行入口先检查后端，缺失时 OSError，不写 cursor，也不发出结果。export_report.write 第40行仍使用平台默认文本换行；CSV 的 csv_write 则已经明确 LF，本次没有发现 CSV 换行缺陷。.gitattributes 第6行的新 text eol=lf 会在入库时规范化整个 study/data 子树，因此与 JSON 生产者先计算原始 SHA256 的顺序冲突。

测试按真实 lock/file-mode 条件做 skip 标记，并新增 no-lock 拒绝；README 的删除后重新取出步骤确实能让旧 materialization 更新，单纯 checkout 会跳过仍被认为 clean 的文件。这些改动都已纳入评审，没有要求新建通用锁能力或重写为 TypeScript。

对主干的风险

[P1] 保留新导出报告已经绑定的原始字节 — .gitattributes:6。 在 Windows 默认文本换行语义下，现有 JSON writer 写出 CRLF，index/files 与 settings 先绑定这些原始字节。新规则在 git add 时把 JSON 转成 LF，重新取出后 --verify 报 ValueError: File binding failed: run-rows.json；settings/run.json 的绑定也失配。这里是新导出报告，当前已发布的13文件修复是成功的，不能混为一条路径。

我使用仓库合法终态 fixture、实际报告 CLI、Python TextIOWrapper 的 CRLF 写出选择器及真实 Git，未模拟验证结果。相同输入的三组结果：旧属性入库后退出0；当前属性入库前退出0、重新取出后退出1、两个 JSON 失配；-text 字节保留方案退出0。宿主是 macOS，这不是原生 Windows/ACL 认证。最小修复是对摘要绑定数据使用不改字节的属性，或让所有 JSON 在最初写出/计算摘要前固定 LF；增加导出→验证→Git入库→重新取出→验证的回归，不能靠修改已有摘要或削弱校验掩盖转换。

现有56项 report/hook 测试全通过且无跳过，Ruff/diff hygiene 通过。真实 Git 的 core.autocrlf=true 档案取出对照从13个失配降到0；无 fcntl 的公共读取、无后端拒绝和 POSIX 交付/回放也通过。这解释了为什么作者的现有验证可以全绿，仍遗漏新 JSON 发布路径。

语义与 CI 对齐

此处没有新共享词表或通用控制面义务，继续复用原报告摘要契约；违反的是现有跨生产者/存储/消费者的原始字节不变量。CI按配置未查询；作者原生 Windows 收集数及历史 canary/mypy 报告不冒称本次独立执行，也未用未跑的平台测试替代这个已复现反例。

我的整体评价

REQUEST_CHANGES — 1357ca2。long_horizon 与 user_experience 在新导出→发布→取回路径有 regression：一次成功导出不能保证后续可读，重复手工验证或重算摘要也不应成为正常旅程。现有档案和导入修复方向正确，机制规模合适，但必须在本 PR 修好紧邻的字节传输边界。报告读取与需要 flock 的通知交付沿用现有模块，缺后端只阻断交付；档案传输必须保留计算摘要时的原始字节，不能用 Git 的文本规范化替代该不变量。 当前PR应一并修好导入、原档案取出和受该新属性覆盖的新报告往返；无需增加通用锁框架或改动评分。 相邻未来整理限定为这项可逆的小修复；不要求另造状态、能力或独立项目，也未运行 benchmark/model 或执行合并。

English verdict: REQUEST_CHANGES - 1357ca2: existing archive/import and56tests pass,but text-eol normalization breaks two boundJSON files from a Windows-newline export after realGit roundtrip. Preserve rawbytes or serializeLF beforehashing,and addexport-to-fresh-checkout validation. Windowsnewline was simulated onmacOS;nativeWindows/ACL wasnot independentlyqualified.

[JasonBuildAI]

Author response - review follow-up at 8e6614ca1

The P1 byte-transport finding is confirmed and fixed on the branch. Reproduced natively on a Windows host at 1357ca22d before changing code: the exporter wrote CRLF through the host text default and bound those digests, text eol=lf normalized run-rows.json and the settings/*.json files to LF at git add, and a fresh materialization failed --verify with ValueError: File binding failed: run-rows.json. The published 13-file archive verified at the same revision, matching your split between the two paths.

What changed

• 8c6601d48 fix(benchmark): bind exported report bytes before Git text conversion - export_report.write writes JSON with explicit newline="\n", so bound bytes are LF on every platform (matching csv_write). .gitattributes keeps text eol=lf, so LF written = LF stored = LF materialized; the archive note and study README now state that invariant.
• 8e6614ca1 test(benchmark): pin report bytes across a converting Git publication - regression in benchmark/tests/test_edgebench_report.py:
  • test_json_writer_ignores_host_text_newline_translation asserts exact writer bytes while default text writes are forced to CRLF (models the Windows text layer on any host);
  • test_published_archive_materializes_byte_exact_under_autocrlf runs real git init/add/checkout-index with core.autocrlf=true against the published archive, with check-attr and control-file guards so the test cannot pass vacuously, then verifies;
  • test_exported_report_survives_autocrlf_publication covers the same export -> add -> fresh materialization -> verify roundtrip for a new export.

Evidence on the new head (native Windows host): the writer test and the export roundtrip fail at 1357ca22d (the latter with File binding failed: run-rows.json) and pass at 8e6614ca1; benchmark/tests/test_edgebench_report.py 45 passed; report + feedback hook 46 passed / 13 skipped; --verify on the published archive returns {"runs": 10, "files": 13, "verified": true}; Ruff clean. On this host benchmark/tests/test_deepswe_v1_archive.py has 5 pre-existing failures (its bash -n calls fail identically at 1357ca22d, untouched by this diff; the Linux CI shards cover it).

No runner, scoring or evidence semantics changed; benchmark/edgebench/online_judge.py is still intentionally untouched. The PR body records the follow-up row and revision. Ready for re-review at 8e6614ca1.

[loopx-agent]

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | xhigh

APPROVE — 8e6614ca16b9236d70126411e07e40adac558533。旧 P1 的新报告发布反例已在当前 head 消失；此次重新检查完整六文件，并非继承此前结论。

动机

在 Windows 上运行本地检查、导出报告或验证已发布研究档案的贡献者。过去缺少 POSIX 锁模块会让报告和两个测试模块无法导入，Git 换行转换还会使报告摘要失配；当前读取能够启动，JSON 在计算摘要前固定 LF，重新取出仍可验证。相关模块可执行，无锁读取通过；既有档案返回 10 次运行、13 文件验证成功，新报告经过真实 Git 往返也保持全部摘要绑定。本次不提供 Windows 通知锁后端、不启动 benchmark 或模型任务、不改变评分，也不把文件摘要当作研究结论的资格。

改动思路

固定 LF 写出和档案 Git 属性共同维护原始摘要，缺锁仅阻断通知交付；现有模块已覆盖三个责任，无需新通用框架。 本 PR 完成无锁读取、既有档案和新报告的 Git 往返；Windows 通知锁及 online_judge 仍在范围外。 最强反对方案是只增加 Git LF 属性：生产者仍写 CRLF 时，先绑定的摘要会被入库转换破坏。当前把修复放回已有 JSON writer，既不重新计算公开档案摘要，也不添加平行验证器。loopx.file_lock 的 guarded-import 和已有 immutable archive 属性提供了可复用边界；这里沿用专门 benchmark IO，不另建通用锁或控制面能力。

具体改动

接受依据为 issue #6367，spec_ref=https://github.com/loopx-project/loopx/issues/6367；spec_revision=issue6367-updated-2026-10-13T07:55:33Z。从实际 merge base 806eb215f8a21e4ae6f51cd942112175ca3839b7 到当前 head 共六文件 +131/-2；另单独核对旧评审 head 1357ca22d1699720bb2bdf123323b81984080674 的修复。

• local_check_runs：实现。两个原先中断 collection 的模块在缺 fcntl 条件下实际执行 47 项、跳过 12 项依赖锁的用例；全树在缺 fcntl 条件下实际收集22079项、退出0；收集不等于执行，不把相关模块通过当作全树测试通过。
• verifier_readback：实现。真实已发布档案验证为 10 次运行、13 文件、verified=true；公共 CLI 无锁对照为基线导入失败、当前读取成功。
• archive_byte_bindings：实现。既有档案和新导出都经过真实 Git 重新取出与摘要校验；新导出测试同时证明无关 control.txt 确实发生 CRLF 转换，排除 Git 没有转换的假阳性。

关键代码讲解

export_report.write 第40行在 Path.write_text 明确 newline=LF，先统一原始字节，再由现有 digest/index 绑定；CSV 本来就用 LF，无需添加第二套序列化规则。feedback_hook.deliver 第28行先检查真实锁后端，缺失时在目录、cursor 与输出前抛 OSError；支持 POSIX 的 flock、并发及重复交付路径保持原语义。.gitattributes 第6行规定 study/data 的取出字节；现在其假设与生产者一致。README 说明已有 stale materialization 的一次性恢复，并保留已发布摘要。两个测试模块分别验证缺锁拒绝、平台标记、实际 TextIOWrapper CRLF 默认及旧/新档案 Git 往返，完整 diff 均已纳入。

对主干的风险

没有发现当前阻塞项。最强反例用同一合法终态输入、真实报告 CLI、CRLF 默认文本写出与实际 Git core.autocrlf=true：旧1357生产者在入库前通过，重新取出后退出1且两个 JSON 摘要不匹配；当前8e生产者退出0且全部匹配。这是旧发现被修复的判别证据，不仅是新增测试绿色。实际不可用 fcntl 的三个事件、两种 session 共六次交付均没有 receipts 目录或输出；支持锁的并发、重复交付及篡改摘要反例由完整相关模块覆盖。

当前相关59项全通过、0跳过；实际基线同两模块55项通过。同一 POSIX fixture 的所有导出文件摘要逐项相同，Ruff 四个变更 Python 文件和 diff hygiene 通过。无锁条件47通过/12跳过有具体后端原因，全树 collection 22079项退出0；宿主 macOS 的 CRLF/缺锁选择器不等于原生 Windows 内核、ACL 或整树执行资格。未查询或等待 CI，也未启动研究任务。

语义与 CI 对齐

复用现有 raw SHA256 档案契约，无新共享词表或通用状态义务。生产者、存储与读回的原始字节不变量现在一致。开发 advisory 对 benchmark 目录的覆盖有限，结论来自完整 diff、真实定义/调用搜索与上述往返，不拿空 probe 当语义证明。

我的整体评价

APPROVE 当前 exact head。long_horizon 与 user_experience 均 improved：一次导出以后仍可取回阅读，无锁读者不再被通知模块的依赖阻断；缺后端的通知保持诚实拒绝。规模与原问题相称，未来整理已限定为复用现有 writer 的小修复，不需要通用锁框架或语言迁移。固定 LF 写出和档案 Git 属性共同维护原始摘要，缺锁仅阻断通知交付；现有模块已覆盖三个责任，无需新通用框架。 本 PR 完成无锁读取、既有档案和新报告的 Git 往返；Windows 通知锁及 online_judge 仍在范围外。 这关闭本 issue 的有界修复评审，保留原生 Windows 独立实机验证的残余风险；批准不授予合并、旧 review dismissal 或模型执行权限。

English verdict: APPROVE - 8e6614c: old Windows-newline publication counterexample now passes actual Git roundtrip; full related59 tests, baseline/head POSIX byte parity and6 no-effect lock refusals verified. Native Windows kernel/ACL and whole-suite execution were not independently certified; CI not consulted.