fix(benchmark): isolate native profile runtime temporary paths

[huangruiteng]

The formal native Codex profile isolated HOME and CODEX_HOME but left the installer, CLI and app-server on the shared process temporary directory. Equal-source profiles consequently reused one Effect runtime, so stopping one fixture could affect another and its directory could still be written during parallel full-public cleanup.

Bind TMPDIR, TMP and TEMP to the existing profile-owned HOME in both established environment builders. Add a real two-profile install/CLI test that stops one runtime and reads the other as still running; retain the existing explicit runtime teardown and provider/privacy assertions. No scoring, task semantics, model turn or benchmark job changes.

Placement: built-in benchmark-toolkit provider, existing native Codex transport/profile owner. Python continues to adapt the installed host environment; the typed runtime authority and global defaults remain unchanged.

Validation passed on the exact head before authorized merge: immutable base/head temp-scope readback, two real isolated profiles, the formal installed-profile/app-server smoke, focused release qualification tests, static/boundary checks and native premerge.

15 focused tests and actual two-profile runtime/shutdown pass. Final clean-head formal install and Codex app-server skills/Goal readback pass with no model turn. Ruff, configured mypy19, TypeScript typecheck, docs governance, boundary and exact quality pass; all19 native premerge checks pass. The conservative benchmark-path manual review is covered by the published exact-head review and maintainer authorization; original classification remains recorded. No frontend consumer or setting changes: this is derived environment in the existing explicit host/provider adapter. Final merged-source full release qualification remains required.

[huangruiteng]

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

Reviewed exact head 5226@e8ee52003a7554bd2f5ab2c26d4f67c87335f747.

动机

1.2.2 的并行全量公开验证仍遇到正式安装 profile 清理失败。此前 #5224 增加停止 callback，但 HOME/CODEX_HOME 隔离并不能证明 runtime 独占：现有内容寻址 runtime locator 使用系统临时目录，相同源码 profile 因而能共享 writer；其中一个退出后另一个仍可重新写入已清理目录。原始失败记录保留，维护者已授权本次发行的必要修复和合并。

改动思路

由既有 benchmark-toolkit 原生 Codex profile 安装/环境适配器拥有这项边界。两个现有环境 builder 都将 TMPDIR/TMP/TEMP 指向已存在的私有 HOME，安装和后续 CLI/app-server 采用相同 scope。复用原 TypeScript runtime locator、doctor 停止结果和标准临时目录解析；没有新增状态键、全局环境设置、进程 supervisor 或第二个决策 owner。Python 留在已有专用 host/安装 transport，TypeScript 权威逻辑不变。

具体改动

三个文件、57 行增加和 1 行删除：生产适配器 8 行；一项真实双安装 profile 测试 41 行；现有 provider 文档更新 8 行、替换 1 行。实际相同源码的两个正式安装 profile 分别启动 runtime，停止 first 返回 stopped 且 locator 位于 first HOME，second 仍 running；最后只停止各自 owner。15 项相关测试通过，生产代码在已测试功能提交与最终文档提交之间完全相同；最终干净 head 的正式安装/app-server 技能及 Goal 读回再次通过，无 model turn。对同一个实际已安装 fixture，immutable base 的导出环境使 Python tempdir 位于共享系统目录，最终 head 则位于私有 HOME。Ruff、mypy（19 源文件）、TypeScript typecheck、文档治理、公开边界、精确质量以及全部 19 项 native premerge 通过。原始 benchmark-sensitive 人工门禁不被改写；本评审核对没有评分、task、runner 提交、benchmark 作业或生产 Goal 操作，结合维护者授权满足此人工要求。

对主干的风险

变更仅对显式创建的资格验证 profile 生效；通用 runtime、正常 CLI/App 环境、profile schema、技能内容和权限 owner 均保持原字节。主要风险是安装与后续调用 scope 不一致，双 builder 同步和真实安装测试覆盖了这个路径；停止 pending/失败仍会拒绝，不能伪装清理成功。HOME 由已有安装路径预先创建；临时目录不携带上游凭证。实际模型工作仍需原 OS/provider isolation，环境隔离本身不授予凭证、网络沙箱或执行权限。没有 frontend 配置项或应用 consumer：这是已有显式 provider 适配器派生的环境边界，实际安装 CLI 和 app-server 已验证。此局部通过不代替合入后的全量 CI、真实模型、公开下载及签名验收。

我的整体评价

当前精确 head 可以批准。相关未来改进检查选择直接复用已有 HOME 和两个 builder，没有新增字段或抽象；回归测试保护真实安装后的并行 shutdown ownership。#5224 的串行清理结论需补充这项并行边界修正，最终发行仍在当前任务中继续完成。

English verdict: APPROVE

[huangruiteng]

Qualification readback update for the existing native-profile lifecycle owner:

The private HOME and runtime-locator scope verified by this PR still holds, but a later Linux full-public attempt on ee9dad81b14c6d15d32b95851b5d38fcc485e732 failed during the disposable installed-profile smoke's TemporaryDirectory cleanup (Directory not empty: profile). Original attempt.

The complete five-shard rerun of the same immutable source passes all 514 public smokes with the original budgets, and eight additional concurrent local runs of the real installed-profile smoke pass. Same-source complete rerun. No cleanup error was ignored, assertion removed, timeout raised, or production Goal changed.

This successful rerun qualifies the current source; it does not prove elimination of the intermittent Linux fixture-quiescence race. The remaining maintenance boundary stays with this existing profile lifecycle owner: identify the exact late writer/process under Linux, then prove it is quiescent before removing the owned fixture. The final bilingual release validation preserves the first failure and this limit.