ci: preserve timing evidence from interrupted Python shards

[loopx-agent]

Full Python shards need usable timing evidence even when tests fail or the runner interrupts them. Stream pytest-owned phase observations only when the CI plugin is explicitly loaded, and provide a bounded TypeScript log reader that keeps complete failed execution separate from successful qualification and interrupted lower bounds.

Records use a full-node SHA-256 identity and a bounded parameter-free display name. Parent call durations include subtests once; quiet/verbose serial and xdist progress preserve one record per line. The reader accepts GitHub timestamps, ANSI, CRLF and line-leading BOMs from concatenated log chunks, while rejecting malformed identity, duplicate/out-of-order phases and invalid durations. Logs never become a timing cache or a test-selection authority.

Validation: 32 complete observer cases passed on Python 3.11, including serial/two-worker v/q, subtests, interruption, oversized IDs and both plain/BOM negative evidence. The BOM regression reproduced four fixture failures before the fix. The updated reader consumed two untouched real hosted logs: one completed all 5,415 cases with pytest exit 1; one cancelled after 3,657 finished cases remains incomplete with no exit code. Prior default-off parity, full collection/split/coverage and workflow validations are retained; required gates, workers and deadlines are unchanged.

Premerge passed four direct and14 selected checks. Its CLI budget check initially failed because the host interpreter lacked pytest; the exact smoke passed in the source Python3.11 test environment. The initial failed aggregate is retained.

Full hosted qualification remains open: the observed run has failures, cancellation and queued shards. Frozen shared-lane and release-capacity acceptance is unchanged; the change-quality receipt preserves that hold. The previous review binds the old head; this updated head requires its own independent review. No installed or overall release-readiness claim is made.

The related refactor pass reused the existing observer/reader and test fixtures; one anchored transport normalization replaces no semantic owner. Frontend/Lark entry points are unaffected by this internal CI provider diagnostic.

[loopx-agent]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

[P2, non-blocking] Reuse the completed-case predicate for longest_cases[].finished. A log with start/collection, passed setup, teardown and end but no call currently returns execution_complete=false and finished_cases=0, while its sole longest_cases entry says finished=true. The aggregate correctly protects incompleteness, so this does not bypass CI or invalidate the real interruption result, but the per-case flag misleads a diagnostic consumer into treating partial work as finished. Use the same predicate as the finished counter and add a regression that checks both fields for the existing missing-call case.

动机

在完整 Python 分片超时或进程被终止后诊断耗时的维护者会遇到这个问题。例如分片仍在逐项执行，但 controller 被终止时 JUnit 和最终 duration summary 尚未生成，维护者只能看到测试名称，无法回读已完成 phase 的原生耗时。 本 head 在明确加载 observer 的 CI 命令中流式保存 pytest 原生 phase 耗时；实际 process-group 中断后，reader 仍能回读一项已完成工作和部分计时，同时明确标记 execution_complete=false。它只提供测量证据，不改变分片、worker、超时、coverage、选择缓存或资格门禁，也不授予 Goal、宿主或 provider 操作权限。真实 hosted Linux logging 成本及中断读回、完整 CI critical-path 改善仍未验证；单例 finished 对缺失 call 的日志与总计不一致，是下面保留的非阻塞修正意见。

改动思路

以改动前 docs/development/testing-and-quality.md、完整冻结基线 c980badbc4f9803099a22dc933691f06761d21a9 的 Quality Layers / 质量分层, Local Validation Environment / 本地验证环境, Pull-Request Baseline / PR 基线 为依据，并结合 owning PR/CI 契约和 S8/S12 路线。三项旧规格约束在当前范围内保留；修改后的文字不作为自己的独立验收 oracle。

JUnit 与最终 summary 都在结束后才写；复用 pytest controller 已有 TestReport 流并提供一个有界只读 reader，能保留中断前的测量。另建 timing cache、改变 splitter 或扩大 deadline 不能解决证据丢失，而且会引入不必要的决策 owner。

本次交付内部 CI 测量与下载日志的只读回读：原资格 owner 不变；partial worker-seconds 是下界，不是 runner-minutes、CPU 或性能改善证明。

具体改动

完整审查 head c5cf3fb034893e710e461d6ed88540541a721fa0。七个文件 +405/-0：一条 workflow load、75 行 pytest adapter、107 行 TS reader、182 行真实进程/损坏输入测试、两份 developer docs 和两处测试环境适配。没有新 job、cache 或 Core capability。

• pytest_configure（pytest_timing.py:20）只在显式 -p 下注册 controller observer，worker 不重复写报告。
• TimingObserver.pytest_runtest_logreport（pytest_timing.py:61）消费 pytest 原生 duration/outcome；完整 node id 用 SHA-256，显示前 256 字符，并立即 flush。
• summarizeTimingLog（pytest_timing.ts:16）验证 start/collection/phase/end 顺序、数量、hash、phase/outcome 和非负有限时长；CLI 限制 32 MiB/200,000 observations，拒绝损坏输入。
• shard workflow 只新增显式 plugin load；删去这一行后，其余完整 workflow bytes 与冻结 B18 完全相同，包括四分片、每片两 worker、60 分钟、fork permissions、JUnit/coverage 和失败门禁。

573 项相关 Python 测试、11 项 CI 合同、strict TS、Ruff 均独立通过。另用同一私有合成 fixture 实际运行 B-default/H-default/H-enabled 三臂：11 个用例的身份、pass/fail/skip/setup-error 结果和 exit=1 完全一致，默认关态均无 timing rows；on 态 phase 总和与同一次 pytest 的原生 duration cache 精确一致。遍历完成而 exit=1 仍代表失败，没有替代 qualification。

真实 B/H process-group SIGKILL 都发生在 JUnit 生成前。B 无 timing evidence，当前 reader 明确拒绝把它当零成本；H 保留 selected=2、finished=1、部分原生 phase 时长、没有 end/exit/JUnit，并报告 execution_complete=false。原生测试另验证超长共同前缀参数仍是不同 SHA 身份、显示有界，以及缺失、损坏、超量、重复、错序输入。

对主干的风险

上述 P2 是局部诊断标志不一致：passed setup+teardown、无 call 的 incomplete log 返回 finished_cases=0、execution_complete=false，但 longest_cases[0].finished=true。aggregate 没有误报完整或绕过 CI，因此不阻塞已验证的中断测量价值；单例应复用同一 completed predicate，补双字段回归，避免下游分析误用。

CI 命令默认新增明确 observer load，普通 local pytest 仍需要 per-command -p，实际三臂已验证其选择/结果不变。phase 总和是 worker-seconds；partial 是下界，不是 CPU、runner-minutes、完整 elapsed 或性能提升。新增 stdout/序列化的 hosted 成本尚未测量；未继承作者 20,000 合成样本为本次云端证据。

当前 premerge 4 direct、15 selected，其中 14 selected 通过、1 项 advisory_inherited_failure。ratchet 本身仍失败；冻结 B18 独立复现相同未变 goal_frontier 模块预算项和相同 source blob，native 保留继承 advisory，未修改阈值。其余完整 semantic、CLI budget、profile 与 public-boundary 检查通过。advisory 只扫描 core loopx，因此报告 0 changed source 不证明本地 pytest 词汇；Phase/Outcome 复用 pytest owner，事件是本地 CI adapter 合同。

CI 按当前 wait_for_ci=false 未查询或等待；真实 hosted logging/中断、完整 CI critical-path 以及 Windows kill 未验证。现有所有原始失败/qualification holds 保留。这个内部 CI 测量入口没有新的前端/Lark/权限或安装态交付主张。

我的整体评价

APPROVE，保留 P2 非阻塞单例标志意见。维护者已有一个有界真实读取路径，可在进程中断后获取完成工作的测量并保留未完成状态，而原 test/coverage/cache/selection owners 不变；这是现有 CI 诊断的可用增量，不声称性能优化或总体 CI 已完成。

近邻重构不需要新框架：后续复用一个 completed-case predicate 修正两个字段即可。特殊 pytest provider adapter 保留 Python，读回由 TS 拥有；这种归属比重复测试/排序 authority 更小且可回退。关态、真实失败、原生时长和实际中断均独立验证，hosted/安装/merge 授权保持独立。

English verdict: APPROVE — c5cf3fb034893e710e461d6ed88540541a721fa0 is independently qualified for its bounded source-rule/CI-observation outcome. The P2 above is non-blocking; model/installed adoption or hosted-cost/whole-CI qualification remain unverified. This COMMENTED author-account fallback grants no merge authority.

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

[loopx-agent]

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

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

No blocking finding.

[P2, non-blocking] Keep observation rows line-delimited with quiet xdist progress. On this exact head, a real 3-item / 9-parent-report run with python -m pytest -q -n 2 -p scripts.ci.pytest_timing completes with exit 1, but emitted call/end rows are prefixed by progress text (.LOOPX_PYTEST_TIMING and [100%]LOOPX_PYTEST_TIMING). The reader accepts only a prefix at column zero and returns finished_cases=0 / pytest_exit_code=null with only partial phase cost. The configured CI lane uses -v and is independently qualified, and this never bypasses completion or CI gates, so this is non-blocking. Ensure every observation occupies its own line, or document the required verbose mode; add a real quiet/verbose xdist regression.

动机

检查中断 CI 分片耗时的维护者会遇到这个问题。分片还没写完 JUnit 就被取消时，原流程只能看到零散进度；旧观察器又把嵌套子测试当普通 phase，导致日志被当成矛盾记录拒绝。 当前观察器保留收集项的原生耗时，子测试成本只计入父 call；强制中断后仍能读回选中两项、完成一项的部分记录，缺 call 的条目不再标记完成。 该观察结果不替代 JUnit、coverage 或 CI 必过门禁，不进入 timing cache，不改变分片选择，也不代表 runner 成本下降。 当前 hosted 传输与完整 CI critical path 尚未独立验证；quiet 模式的记录前缀丢失已复现并作为非阻塞意见公开；首次三个子进程超时记录保留，原因未确认，未靠延长超时或修改代码取得通过。

改动思路

保留原生 pytest owner 和单个 TypeScript reader；Python 只适配 pytest 插件钩子，过滤其明确的 SubtestReport 类型，复用一个完成谓词即可修复本轮问题，无需新 cache 或第二个 test gate。 完整七文件源码提案是 CI 可观察性增量；独立证明真实分片入口和失败读回，完整 hosted 成本及全量 CI 修复仍由既有 CI owner 验证。

入口为 python-tests.yml pytest -v -n2 -p scripts.ci.pytest_timing → controller TestReport observation → summarizeTimingLog CLI → diagnostic reader。规范状态与副作用由 pytest owns selection, TestReport timing/outcome and exit; local CI observation adapter owns bounded readback; existing workflow owns qualification 拥有；Local observation start→collection→ordered unique setup/call/teardown→optional end; missing/null evidence never becomes successful qualification. One caseFinished predicate feeds aggregate and per-case completion. 架构判断保留现有责任边界，One local observer and bounded reader; no cache migration, state authority or provider framework. 近邻重构判断：Applied bounded simplification in this PR: one caseFinished predicate and typed native SubtestReport exclusion; no new scheduling layer or cache.

具体改动

审查完整 head b3dfe660c0bc8fd49649ca6b176312b7c468e40b，实际冻结 merge base 2e3ab4c4a5665378b9e422f9483a820066fca40b，7files+481/-0,189production adapter lines,239focused test lines, explicit workflow hook and bilingual disclosure.

关键代码讲解

• TimingObserver.pytest_runtest_logreport（scripts/ci/pytest_timing.py:65）：Only parent reports emit cost; full identity hash and bounded payload-free display preserve exact cases.
• summarizeTimingLog（scripts/ci/pytest_timing.ts:11）：Validate record order, preserve unknown/incomplete evidence and share the finished predicate.
• CI shard pytest invocation（.github/workflows/python-tests.yml:555）：Explicit plugin activation on existing complete four-way partitions; gates/cache unchanged.

修改前规格 docs/development/testing-and-quality.md：Quality Layers / 质量分层 → implemented：Independent real entrypoint, native oracle and negative cases；Pull-Request Baseline / PR 基线 → implemented：Current exact whole diff and risk-based required checks；Local Validation Environment / 本地验证环境 → implemented：Selected source interpreter/dependencies and actualminimumpytest。当前修改的说明是披露材料，未用它替代冻结规格。

Own5realruns: B/H default and H-enabled same11test identities/outcomes/exit1; H enabled sum equals same-run native duration cache. Own verbose nested old/head fixture:9parent reports/6subtests, old rejects duplicate phases, H3finished/exit1 and0.1999987069worker-seconds exactly equal its same-run native parent total; nested0.1162699170seconds is not added again. Payload-free displays and distinct hashes verified.

对主干的风险

Owned process-group SIGKILL before JUnit leaves H2selected/1finished/noend/partial while B has no observations; malformed, oversized and contradictory logs reject. Old missing-call counterexample reports per-case finishedtrue despite aggregate0; current same input returns both0/false. CI shard logging intentionally enabled; local phases mirror pytest and do not create a shared vocabulary, scheduler obligation or broader authority. New bilingual docs disclose nested cost, parameter display and qualification limits. Actual local source pytest9.1.1 and independent minimum9.0.3; no hosted CI query/poll/wait, no paid model or benchmark job. Initial3timeouts retained and same575/28full reruns pass unchanged.

本轮原生 risk-based premerge 的直接检查通过，selected 中保留一项 goal_frontier maintainability metric 失败，被当前工具标为 inherited advisory。相同冻结基线先前实际运行同命令失败，致因模块与 smoke 的源码 blob 仍一致；没有改预算、删断言或将失败写成通过。其他选定检查通过。首次575例运行出现3个子进程超时/572通过；同源码不改30秒deadline的575完整复跑及28单独复跑都通过，最低pytest9.0.3另28通过。首次失败保留，未确认历史根因不是本次独立通过的替代物。旧c5cf→新head的完成flag、subtest和display修复与当前base2e3→head七文件整体验收分别判断。

CI按 wait_for_ci=false 未查询、轮询或等待。Useful worker-seconds must not be advertised as complete compute or runner cost; hosted and quiet-mode limits remain, and retries do not erase initial failures. 该观察结果不替代 JUnit、coverage 或 CI 必过门禁，不进入 timing cache，不改变分片选择，也不代表 runner 成本下降。

我的整体评价

APPROVE：当前完整 head 的有限用户结果和责任归属均有独立证据，未发现当前阻塞。long_horizon improved：Independent current real reports and refusal cases preserve usable failure evidence without another scheduler/cache/qualification authority. user_experience improved：A downloaded log can explain completed versus missing work; parent display stays compact while hashes keep exact identities, and diagnostic limitations remain explicit. 当前 hosted 传输与完整 CI critical path 尚未独立验证；quiet 模式的记录前缀丢失已复现并作为非阻塞意见公开；首次三个子进程超时记录保留，原因未确认，未靠延长超时或修改代码取得通过。 完整七文件源码提案是 CI 可观察性增量；独立证明真实分片入口和失败读回，完整 hosted 成本及全量 CI 修复仍由既有 CI owner 验证。 Local adapter Phase/Outcome/event sets reuse pytest semantics; no new shared persisted vocabulary. quiet-mode 意见不影响当前 -v CI 的独立结论。 这是 COMMENTED 作者账户回退，不授予 merge 或部署权限。

English verdict: APPROVE — exact head b3dfe660c0bc8fd49649ca6b176312b7c468e40b. Configured verbose native timing, same-run parent cost, disabled-path parity and actual interruption qualify this observation slice; hosted/full-cost claims remain unverified; the quiet-output limitation is recorded above.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; runtime_reported; reasoning_effort=xhigh.

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

APPROVE — 当前 9795f25ba40bbde2d2a632d0df23951b65dcfc15 未发现阻塞性问题；此前 quiet 输出 finding 已由独立实际进程验证修复。旧 b3dfe660c0bc8fd49649ca6b176312b7c468e40b 的结论不继承。

动机

查看失败或中断 Python CI 分片的维护者，以及显式加载本地观察器的开发者。 完整 Python 分片在 JUnit 写出前中断会丢失耗时证据；旧 observer 的 quiet 进度还会给记录加前缀，使严格 reader 拒绝已观察到的工作。 新提交在实际 quiet/verbose、串行/双 worker 输出中提供独立完整行，phase 耗时与 pytest 原生记录一致，中断记录仍明确不完整。 原生测试仍决定成功与失败，这些记录只帮助定位已观察成本。Hosted 日志开销、Windows 物理中断和全套 CI critical path 尚未验收；旧矛盾/进度前缀坏日志不能变成合格证据。 No test exclusion, worker/cost/deadline change, new gate, cache, authority, capability UI, retrospective log repair or merge.

改动思路

Python 留在 pytest provider，TypeScript 留在有界诊断 reader；复用原 terminal writer 的行宽和原父项报告类型，不放宽 reader 或另建 timing cache。 本 PR 交付已有 CI 诊断路径的可用观察与严格回读，保留完整分片、worker 数、限时、覆盖率与失败门禁。 先比较最终 JUnit、保留流式观察和宽松日志 salvage：只有流式父项报告能保留已完成工作，而宽松读回会把坏记录当证据。当前选择保持严格 reader；quiet 补行使用 pytest writer 已有事实，不另造路径状态。父项和子测试复用原 SubtestReport discriminator，不能把嵌套 call 重新计费。诊断记录和 execution_complete 都不是 scheduler、test gate 或验收权威。

具体改动

完整 PR 七文件 +495/-0；本轮不只检查最后四行 emitter。判定依据 docs/development/ci-impact-selection.md，不可变版本 509c38c502d3d0b4eafb01550efcdc4b5b5a1d17，并核对同版本 docs/development/testing-and-quality.md 的完整分片与测量边界。Four complete Python shards：implemented，完整 collection/four shards/two workers 保留；Override and evidence：implemented，失败、取消、缺失/矛盾输出仍不能资格化，原 JUnit/coverage/mandatory aggregate 保持。main509 integration 的相关调用者与 workflow 都重新检查，strict reader blob 与 b3 完全一致。

关键代码讲解

• pytest_timing.py:34 TimingObserver.emit 在 ensure_newline 后读取现有 writer 实际行宽，补 quiet 进度的换行，再写有界 JSON 并 flush；不修改原 report。
• pytest_timing.py:68 pytest_runtest_logreport 跳过 native SubtestReport 的重复嵌套成本，父项 node-id 的 SHA256 保留身份，显示去除参数并截长。原 detailed test result 仍由 pytest 输出。
• pytest_timing.ts:15 summarizeTimingLog 校验 start/collection/phase/end 顺序、重复和合法值；汇总完整 predicate 复用于单项和总体，exit1 的完成遍历不等于成功。

对主干的风险

独立实际进程在精确 b3 provider 上复现 quiet 串行/双 worker 均失败、verbose 均通过；当前 H 的四组合可读，耗时等于同次 pytest native cache，而非从 observer 生成 oracle。基线与当前默认关闭臂的11个原生 case 身份、结果和 exit1 一致，无观察输出。实际 POSIX process-group kill 在 quiet/verbose 两模式都留下 selected2/finished1、非零已观察耗时、null final exit 和无 JUnit；不能伪称测试成功或全套成本。

完整 observer/workflow 当前重跑 579 passed /48.22s；独立完整 observer 32 passed /10.34s，覆盖父子测试、setup error、skip、失败、超长参数和损坏输入。最初同 head 全跑为575pass/4个30s xdist超时，仍保留且未声称根因已修复；随后原命令、原断言、原期限的完整重跑、四臂 off/old/head 同原生6-case输入及14实际模式运行给出了当前 changed invariant 的独立证据，未只挑选成功用例。初次独立 TS 调用缺失 node ambient types，补为明确 --types node 的 strict NodeNext 检查后通过，源码和 decoder 未变。

语义与 CI 对齐

本地 native premerge 4 direct，15 selected 中14必需项通过、1可选 maintainability advisory失败；相同 immutable base ratchet 也保留同一 frontier module_metric_budget，未修改限额或删验证。开发 vocabulary advisory 后 full semantic census、public boundary、Ruff/strictTS通过。复用本地 pytest phase/outcome vocabulary，无新共享状态或权限；workflow 只增加显式 -p，collection、worker、60分钟、coverage 和 fork 权限不变。没有查询或等待 hosted CI。

我的整体评价

long_horizon=preserved，user_experience=improved：日志可直接严格读回，部分工作不再伪装成零成本或完整验收，后续测试选择、timing cache 与成本门禁保持。Python 留在 pytest provider，TypeScript 留在有界诊断 reader；复用原 terminal writer 的行宽和原父项报告类型，不放宽 reader 或另建 timing cache。 本 PR 交付已有 CI 诊断路径的可用观察与严格回读，保留完整分片、worker 数、限时、覆盖率与失败门禁。 future-facing pass 复用单一 native 判别与 completion predicate，未加入通用框架。Hosted 日志开销、Windows 物理中断和全套 CI critical path 尚未验收；旧矛盾/进度前缀坏日志不能变成合格证据。 最初超时的历史原因、物理 Windows 和 hosted成本仍未知，但当前同期限完整负载及独立反例覆盖不支持把它们归为该 framing 修复的已知回归。此源码资格不关闭整个 CI/capacity/CQR 未验收边界，也不授权合并。

English verdict: APPROVE. At 9795f25, independently reproduced old quiet failures are repaired through native writer framing; the strict reader, default-off behavior, native outcomes and incomplete evidence boundaries remain intact. Current579/32-case qualification passes; four earlier timeouts remain unroot-caused history, and hosted/Windows/cost acceptance and merge authority remain separate.

[loopx-agent]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

中断的 Python shard 需要可信的逐阶段观察，不能把缺少记录当作零成本或执行完成。本次完整评审 head c93b704c19097bdc38ffe23d170a31d918a296c6；依据改动前 docs/development/ci-impact-selection.md @ 509c38c502d3d0b4eafb01550efcdc4b5b5a1d17 的 Four complete Python shards 与 Override and evidence，作者新增文档不作为验收 oracle。

面向维护者及获授权的 CI 诊断操作人员。 原先中断日志缺少可靠 phase cost，旧 reader 丢弃 BOM 记录；当前显式观察保留 native phase/exit，reader 恢复 BOM framing 而仍拒绝矛盾记录。 当前独立实际入口证明 native exit1 与所有 phase/outcome 保留，同一 BOM 输入恢复，重复阶段仍拒绝。 不授予合并权限，不证明整次 hosted CI 成本、CPU 成本、性能收益或旧 timeout 已解决。 Hosted 自然中断、整次成本、物理 Windows 与旧 unrelated shard timeout 原因未验证。

改动思路

复用 pytest 的 collection、TestReport、父级 duration 和 exit status，增加显式 controller observer 与严格 reader。当前 BOM 修复只处理行首一个 BOM，再处理时间戳/ANSI，不放宽 phase、identity 或完整性判断。最小方案是保留原 pytest/gate owner；另建计时缓存或宽松 salvage parser 会引入重复权威。

只读 CI 观察和 framing 修复是本次有界交付。 复用 pytest 和严格 reader 即足够；宽松 salvage 或另建 timing cache 会重复权威并丢失完整性义务。

具体改动

TimingObserver.emit 保证 quiet progress 后独立输出并 flush；pytest_runtest_logreport 排除 SubtestReport 重复计费，完整 nodeid 用摘要绑定、显示不包含参数；summarizeTimingLog 复用同一 case-finished 判断处理缺 call/null timing/缺 end。全七文件与调用链已读。

我的 source 验证是 579 passed，另有严格 NodeNext typecheck、Ruff、完整 semantic/public-boundary 验证。独立真实 pass/fail/skip pytest：base/default 和 head/default 无 timing，enabled 保留全部 native identity/phase/outcome 和 exit1，计时和等于同进程原生报告。把自己的真实记录加 BOM/时间戳/ANSI后，旧 reader exit1，当前 reader 与 plain 的完整 JSON 一致；重复 setup 仍 exit1。serial/xdist 的 v/q framing 与缺 call/end 等负例通过。没有借用作者下载的 hosted 日志或计数。

对主干的风险

current PR boundary 是只读 CI 观察和 framing 修复。工作流除显式 plugin 行外的完整字节保持原样；测试 subprocess 的 PYTHONPATH 只定位此 checkout。四 shard、JUnit/coverage、timeout 和稳定 gate 未改，完整遍历失败仍 exit1。新增 usage 明确 worker-seconds 不是 CPU、完整 runner cost 或性能收益。关闭方式是移除显式 -p。

Premerge 4 direct、14 selected required checks 通过；原 frontier 2007/2000 advisory 仍失败，完整 base/head report 一致且对应 owner/metric 未改，不将它写成通过或调高预算。Hosted 自然中断、整次成本、物理 Windows 与旧 unrelated shard timeout 原因未验证；CI 未查询或等待。future-facing pass 保留 native owner 和共享完成谓词，无需额外框架。

我的整体评价

APPROVE 当前完整 head：旧 quiet-output/per-case-completion 问题和这次 BOM 问题均有当前独立反例及实际入口证据。长程价值是保留可用于下一次排查的真实部分证据，用户收益是分清“失败/未完/缺观察”；不宣称整个 CI 成本或故障已解决。无 actionable code blocker；批准不授予 merge 权限。

English verdict: APPROVE c93b704c19097bdc38ffe23d170a31d918a296c6. Native pytest truth, explicit activation and strict partial-evidence semantics are preserved; same-input BOM recovery is independently demonstrated. Hosted whole-run cost and unrelated failures remain unqualified.
Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=xhigh