[loopx-agent]

A failed native Goal could retain a score in final_result.json, which the existing SForge scanner and page interpret as a completed attempt. The adapter now atomically preserves the full score and native/runtime outcome in partial_result.json for nonterminal native outcomes, keeping the completion marker for the existing handled terminal path. This repairs the reproduced post-merge finding in #7061.

The existing CLI regression now covers failed, usage-limited, cancelled, missing, complete, blocked, paused and budget-exhausted outcomes through ordinary SForge discovery and its actual HTTP/Jinja page. Documentation agrees on new-attempt source pinning, unchanged frozen histories and rollback. The writer remains with the existing adapter; no new classifier, scorer or provider abstraction is added.

Validation at this branch commit:

• Related native worker, adapter, feedback and report suites: 489 passed, 11 skipped. The first invocation lacked the existing tests/ import path; the corrected invocation passed without source changes.
• Same synthetic scoring input, pinned SForge source 193cecda05661380575b16c9fcc88a8c16ece9f3: all four abnormal cases changed from 1/1 done to 0/1 done / aborted, preserving their full scores; healthy control stayed 1/1 done. Sixteen complete non-native result payload pairs stayed equal.
• Regressions run against base 722bcc582043bdc2ab60746e4b0dc7a57130b716: four abnormal cases fail and four handled-stop controls pass; all eight pass with the repair.
• Ruff, diff hygiene and RFC bilingual staged gate pass. Public-boundary scans of all four files are clean; the broader global loopx check still reports unrelated legacy Goal authority problems.
• Repository standard premerge: four direct checks, Python compile, ten catalog canaries, eight risk-profile smokes and the public-boundary check pass. Its overall manual_review_required is solely benchmark_sensitive; the owner-authorized exact-head self-review is published, with native merge readiness still required. CI is not queried or waited for under this managed Goal policy.

The explicit Docker cases remain unqualified on this host: the existing provider's stream close raises OSError: ENOTCONN, independently reproduced without LoopX. No Linux Docker pass, installed rollout, live inference, benchmark scoring/submission or historical reclassification is claimed. This patch changes only result-file destination after collection, exercised with the real pinned consumer. Existing Docker assertions are preserved and adapted to that destination.

Bounded refactoring review: the existing writer, score serializer and atomic replace already own this boundary; adding a new abstraction would not improve it. Owner-authorized self-review and native exact-head merge readiness remain required before the authorized merge.

[loopx-agent]

Reviewer: actor_kind=model_agent; model=gpt-6.1-sol; provider=OpenAI; declaration_source=runtime_reported; reasoning_effort=max

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

精确提交 e41becf868425bd24a905a10e5202bf78a6665f5，不可变基线 722bcc582043bdc2ab60746e4b0dc7a57130b716。本记录为作者自评，不冒充独立维护者或 GitHub 正式批准。

动机

运行 native Goal 并通过既有 SForge 页面查看结果的维护者，需要同时保留失败结论与已取得的分数。

同一失败分数输入在基线显示绿色 done、1/1 done；修复后显示 aborted、0/1 done，完整评分仍保存在部分结果文件中。

真实消费者已不再把失败运行计为完成，健康完成对照及普通非 native 结果保持原行为。

本执行仍无 Linux Docker 全链路通过和已安装采用证明；既有外部 stream.close 故障及无关 canonical fixture 失败保留。

不更改评分、任务语义、native continuation 或外部执行权限，不改写冻结历史，也不把合成验证称为安装、Linux Docker 或长程模型收益资格。

依据是 #7061 的已复现审计、修复前 README 的失败不产生完成文件约定，以及固定 SForge 193cecda05661380575b16c9fcc88a8c16ece9f3 的真实消费者。未把本 PR 新增说明或作者测试结论当成独立规范。该修复减少错误完成数对后续判断的干扰；没有测量长程模型净收益。

改动思路

保留现有结果写入函数、RunResult 序列化及原子 replace，只按原有 terminal 状态选择产物名；没有第二套状态机或消费者兼容层。

本次交付界限是既有结果生产者与固定消费者之间的兼容修复；原状态分类、评分来源、归档和容器执行继续由原所有者负责。

不修改外部 SForge 是有依据的范围选择：它早已依赖最终文件判断完成。额外增加一个它不读取的标签不能修复问题；删除整个失败分数又会丢掉原需求。由既有投影所有者选择部分结果文件，保留完整序列化与原子写入，就是更小且可回滚的修复。相关 #6902 提交恢复仍归其原路径，本次不复制它的机制。

具体改动

完整四文件 +66/-20：一个生产函数 +5/-3，其余为原有回归与两处说明；无生成物、私有运行数据、状态框架或新 CLI。

关键代码讲解

1. _write_native_final_result 保留原 eligibility、完整 RunResult.to_dict()、原生标签和 best-only 字段。第101行仅按既有 terminal 选择 final_result.json，其余已经具备写入条件的 native 结果进入 partial_result.json，第102至104行仍以临时文件原子替换。运行前失败和零运行时的旧边界不扩展。
2. main 的结果交接 仍先根据实际退出/原生 outcome 判定状态，再将同一结果交给写入函数；runtime-receipt.json 和0/130/1退出码不变。正常新 attempt 使用原有排他建目录，修复不清理或改名旧成绩。
3. 原 CLI 回归扩展到 failed、usage_limited、cancelled、缺失 outcome、complete、blocked、paused、budget_exhausted；实际执行正常目录发现与原生 HTTP/Jinja 页面，断言异常为 aborted/零完成，正常停止保留 final。另核完整失败 score 字段与原子临时文件消失；既有 Docker 断言保留并适配真实产物名。
4. README 与 RUNTIME 统一说明部分结果、非完成标记、资格边界、新 pin、新 attempt、冻结历史与回滚；没有压缩掉 continuation、预算或权限义务。

正向路径是：真实终态 -> 原写入 -> 原最终文件 -> 正常 scanner -> 既有页面计完成。异常路径是：原分类器输出失败/取消 -> 同一完整评分写部分文件 -> scanner 无 final -> 页面 aborted/0 done；同一 runtime receipt 和非零退出仍可观察。部分文件的存在不授予评分验收、真实实验或模型调用权限。

逐项规范映射：benchmark/edgebench/README.md 的 completion-marker-compatibility 由异常不产生 final 与实际零完成读回满足；preserve-score-and-outcome 由完整评分字段、部分文件及原生标签读回满足；historical-and-nonnative-isolation 由新 attempt 排他目录、冻结证据保留与16组普通路径对照满足。

对主干的风险

验证与失败归因： 当前头六组相关测试 489通过、11跳过；首条命令遗漏既有 tests/ 导入路径导致5个导入失败，补齐测试调用路径后无需产品修改即可通过，保留原失败记录。Ruff、语义词汇 advisory、双语 RFC 与 diff 检查通过。仓库 standard premerge 的4项直接检查、编译、10项 catalog canary、8项风险 smoke及公开边界检查通过；整体仍返回 manual_review_required，唯一原因是 benchmark_sensitive 的显式维护者审查，不能说成工具全绿。当前具名修复已有维护者自修复/自合并授权，本精确头作者评审记录履行该审查，随后仍须原生 readiness。

反例灵敏度： 在不可变基线上运行这八个新 CLI 用例，四个异常分支失败，四个正常停止对照通过；当前八个都通过。同一外部消费者、同一合成评分、同一真实 HTTP 页面再次验证，四类异常从 done1变0，healthy保持1，完整分数不丢失；16组普通非 native 完整载荷逐项相同。未归一化状态、分数或完成计数。消费者测试用合成 scorer/backend，真实 writer、scanner、server 与模板均实际运行；它不等于真实模型或 Docker 全链路。

环境边界： 既有显式 Docker 三例此前实际执行，在外部 SDK stream.close 报 ENOTCONN；不加载 LoopX、只执行 printf 的原 SForge 后端也复现。相关 provider、backend、worker 因果源码在此前观测、基线及当前头一致；本次按原环境开关跳过，不声称 Linux Docker 已通过。此前独立 base/head 相同 canonical authority fixture 红灯亦保留，本次没有借此修改无关 authority。全局临时 loopx check 仍报告别的旧 Goal authority 问题；本 diff 的隔离 premerge 公开扫描通过。按管理 Goal 策略未查询、轮询或等待远端 CI。

语义对齐与隔离： 文件名是 benchmark 内部派生投影，复用既有 typed outcome，不从报错正文猜类别，不新增通用领域义务。异常默认产物的变化已披露；既有 terminal、非 native、提示词、continuation、权限、评分和冻结证据边界保留。正常 blocked/paused/budget 是合法停止，保持旧 final 行为，不借此声称任务完成或分数验收。风险是旧页面仍不展示完整 partial 内容；操作者可读取全量文件与现有 runtime summary，页面至少不再误报完成。新来源 pin 仅用于新 attempt，回滚同理。

有界重构判断：现有 writer、serializer、atomic replace 已形成足够清晰的 seam，继续抽象会增加维护成本，本次不新增。无需 UI 配置、新存储、外部 provider 修改或历史迁移。

我的整体评价

APPROVE（作者自评结论）。该提案完整修复已审计的结果交接回退，保留失败分数、实际失败读回和原成功/普通路径，范围与问题相称。当前没有未解决的代码阻断项；保留 Docker、旧 fixture、安装与真实研究资格边界。允许合并与检查通过是分开的门：仅在头仍为上述提交且原生 merge-readiness 为 ready 后，使用本次已授予的自合并权限。

English verdict: APPROVE at e41becf868425bd24a905a10e5202bf78a6665f5 (author-owned COMMENTED self-review, not formal GitHub approval). The existing writer preserves full abnormal native scores in partial_result.json while restoring the final marker's consumer-compatible meaning. Matched real scanner/HTTP rendering changes four abnormal runs from done1 to aborted/done0; healthy and 16 ordinary payload pairs remain unchanged. 489 related tests pass with 11 optional skips; native local checks pass with the explicit benchmark manual-review hold recorded. Existing Docker environment and historical fixture limitations remain disclosed; no installed, Linux Docker or live benchmark qualification is claimed. Merge still requires unchanged-head native readiness and owner authorization.