fix(tests): make compile-cache symlink checks independent of umask

[loopx-agent]

The existing compile-cache symlink test requests a 0755 target using mkdirSync, then assumes that mode survived umask. Under a valid 077 umask the target starts at 0700, so the test reports a permission-repair failure even though the unchanged runtime never changed it.

Explicitly set and read back the intended mode before loading the real preload. Reuse the existing case for public/private directory targets and a dangling symlink, retaining disabled-cache, source execution, empty-target, unchanged-permission and link-identity checks. Runtime, provider defaults, cache policy and authority behavior remain unchanged. No new fixture owner, API or abstraction is added.

Validation:

• Immutable-base original case: passes under 022, fails under 077. Direct real-Node mkdir comparison leaves existing symlink target permissions unchanged. This corrects the earlier permission-mutation attribution; no speculative runtime fix is retained.
• Final existing Node suite: 9 passed under each umask. Full TS suite: 4093 passed, 31 skipped, 0 failed. TS typecheck passed.
• Existing source cache/Effect integration tests: 81 passed. Fresh wheel outside the checkout: 10 real launcher/authentication/restart/cache/shutdown cases passed, plus the 9 TS cases under each mask; installed preload hash equals source.
• First wheel build rejected a stale generated Chat bundle; the documented Chat build repaired it, then the fresh wheel was built and tested. No generated assets enter this PR.
• Native premerge: 3 direct checks passed, 0 catalog checks selected for this test-only scope; the actual Node, full TS and installed runtime checks above provide the affected coverage. Exact-scope change-quality receipt recorded. Diff/privacy scans passed.

Windows symlink cases retain their existing privilege-based skip; no Windows host qualification is claimed. Frontend/Lark/CLI settings and entrypoints are unchanged. This repairs a validation prerequisite for recoverable canonical migration and last-caller retirement; it does not qualify SQLite defaults, complete writer retirement, or certify sustained storage operation.

Future-facing pass: consolidate the stronger no-touch cases in the existing test rather than add a parallel smoke or modify an unbroken runtime. Reverting this test-only commit requires no state migration. DCO signed; dedicated branch from a00509bd7269cfc3d669a8341c454b90b307cb9a.

中文：原测试把 mkdir 的请求 mode 当作实际权限，077 umask 下产生假失败。现在先显式建立并读回权限，再验证真实 preload 对普通、私有和悬空链接均不跟随、不改权限、不创建目标。产品源码和默认行为保持不变；原失败与新的对照均保留，完整 TS 和隔离安装态已验证。

[loopx-agent]

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

English verdict: APPROVE - exact head 2206a234b79593459f81b4c433bea249d89320f9; the original failure is an umask-sensitive fixture error, corrected without runtime changes.

Approval conclusion (author-owned PR; GitHub blocks formal self-approval). No blocking finding.

动机

维护者在权限较严格的宿主上运行编译缓存验收时，会遇到这项假失败。 修复前，目标目录在测试开始时已被 umask 限制为 0700，测试却把它当作被运行时从 0755 改坏；修复后，先建立并读回目标权限，再验证真实 preload 没有修改它。 同一现有测试在 022、077 两种 umask 下都通过，且继续检查链接、目标权限、目录内容和原源码执行。 这次不改变缓存运行时、provider 默认、权限准入或用户操作，也不把它当作完整 SQLite 迁移验收。 完整默认切换、旧 writer 最后调用方退役及持续恢复资格仍由现有 R5/T4 工作验收。

Maintainers running cache qualification on a restrictive host receive a false permission-repair failure. Establishing the actual precondition makes the no-touch check meaningful; it does not weaken the expected runtime result or qualify the parent migration.

改动思路

先反证产品归因，再修测试前提。不可变基线的原测试在022下通过、077下失败；实际Node mkdir对照没有改变目标权限。现在chmod和lstat都在真实子进程执行前发生，之后独立核对源执行、disabled状态、链接身份、目录内容及准确权限。继续复用fixture/run和原preload，没有新增缓存、判决owner、配置或手工同步路径；最小修复比无故修改运行时更合适。

具体改动

关键代码讲解

fixture.run继续用真实Node进程与--import加载preload。既有symlink case现在以0755、0700和不存在目标参数化：先建立并读回权限，run之后检查目标未被修改；悬空链接要求目标仍为ENOENT，不能用空目录代替。只修改一个既有测试文件，+22/-11；产品0/0，没有新smoke、兼容层或生成资产。

规范依据：docs/reference/local-delegation.md，固定修订 a00509bd7269cfc3d669a8341c454b90b307cb9a。cache-no-touch：默认cache为symlink/非私有/不可用时执行原源码且不修改权限，两种umask及安装模块均覆盖。source-cache-recovery：源指纹、鉴权、关闭/coverage和原退出恢复保持权威，由未修改的源码与隔离安装态实际launcher验证。完整R5/T4默认与恢复验收继续开放。

对主干的风险

最强反例是测试在run后恢复权限而掩盖产品副作用；当前建立权限发生在run前，run后断言仍独立，因此没有这种掩盖。私有target和悬空target也不能错误启用cache或被创建。原错误归因已在相关评审更正，原失败保留。当前同一Node套件在022/077各9项通过；完整TS4093通过、0失败、31跳过；类型检查通过，源码cache/Effect81项通过；全新wheel在源码外实际启动/鉴权/重启/退出10项通过，安装preload与源码hash相同、两种mask各9项通过。第一轮wheel因旧Chat资产失败，文档规定的build后重新构建通过，没有跳过包装验证。premerge3项直接检查通过、该测试范围catalog选择0项，实际覆盖来自上述测试，不能把0项当作运行时验收。

Windows链接case保持原权限限制跳过，未声明Windows真实host资格；完整套件31项跳过不冒充通过。没有API/默认/前端/Lark/CLI或actor授权改动；无新协议词汇、分类器、义务或隐藏开关。公开边界扫描及DCO通过；CI按当前契约未查询。

我的整体评价

这是justified_increment：修复一个真实验证缺口，避免后续错误归因和无效运行时修改。long_horizon为improved，质量证据在严格权限环境仍可可靠使用；user_experience为preserved，没有新增操作，原实际launcher/recovery行为保持。相邻简化已应用：在既有case里集中三种有意义的负例，没有平行smoke或兼容结构。下一步继续既有R5/T4最后调用族与安装态恢复工作；本评审不关闭Goal或赋予合并权限。

The unchanged runtime is independently exercised, rather than mocked or rewritten to fit the test. The bounded validation repair is proportionate and reversible; Windows and the retained skips remain explicit gaps.