fix(windows): skip unsupported parent-directory fsync

[Duang777]

Summary

• run parent-directory fsync only on POSIX in four atomic writers
• preserve file fsync and os.replace() on every platform
• add one parameterized regression for Lark state, decision context, extension projection, and heartbeat automation
• run the regression in the native Windows Python job

Root cause

Windows rejects opening a directory through os.open(). These writers called it after os.replace(), so the file was committed before the command reported PermissionError.

Closes #4997.

Verification

• red: tests/test_windows_atomic_writes.py failed all 4 cases before the production fix
• green: uv run --extra test python -m pytest -q tests/test_windows_atomic_writes.py (4 passed)
• related suites: 143 passed
• Ruff: changed Python files passed
• diff hygiene and Python compile: passed
• risk-profile and public-boundary premerge checks: passed

The full local premerge did not pass because the fresh worktree lacked the prebuilt Chat bundle and Node parser dependencies. install-local-smoke.py timed out during dependency preparation, codex-cli-packaged-install-smoke.py could not find the Chat bundle, and semantic-vocabulary-drift-smoke.py could not start its TypeScript parser. The changed Python tests and checks passed.

[Duang777]

Exact head 401f1b2b6854bf4af840a28176d3b550672567f7 is ready for CI.

Local evidence:

• regression red before fix: 4/4 Windows directory-open cases failed after the target file was written
• regression green after fix: 4 passed
• related Lark, decision-context, presentation, and heartbeat suites: 143 passed
• Ruff, diff checks, Python compile, risk-profile smokes, and public-boundary scan passed

All GitHub workflows are queued with zero executed steps. Please approve the fork workflows so the native Windows job can validate the new regression.

[huangruiteng]

动机

公开问题 #4997 记录了 Windows 上 Lark 私有 JSON 写入的真实故障：目标文件已被 os.replace 更新，随后打开父目录做 fsync 却抛 PermissionError，使一次成功替换被报告为失败。此 PR 不改变数据格式或权限模型，目标是让这类本地原子写入在不支持目录 fd 的平台正确返回。

改动思路

保留“临时文件写入 → 文件 fsync → 原子替换”的现有流程，只让替换后的父目录 fsync 在 POSIX 执行。四个写入器仍各自由原来的调用边界管理路径、载荷和失败重试；这与仓库内其他已有的 POSIX guard 一致，没必要为一个平台效果引入新的全局写入器。

具体改动

关键代码讲解

• loopx/extensions/lark/private_json.py:20 的 write_private_json_atomic：Lark 私有 JSON 的文件 fsync 与替换保持无条件执行，Windows 不再进入父目录 os.open。
• loopx/capabilities/decision_context/private_state.py:163 的 _write_private_json_atomic：游标持久化应用同一边界，避免写入后报错造成重试歧义。
• loopx/control_plane/heartbeat/automation_upgrade.py:61 的 _atomic，以及 loopx/extensions/presentation.py:728 的 _atomic_write_projection：覆盖同形的自动化提示与扩展投影写入。
• 参数化测试用真实临时文件、替换和读回，只模拟 Windows 对父目录打开的拒绝；Windows 工作流也纳入该测试。

对主干的风险

我用同一四入口用例在精确 base 90f21a5 跑出 4 个预期的 PermissionError，在精确 head 401f1b2 跑出 4 个通过；额外 POSIX spy 确认每个入口仍有一次目录打开、两次 fsync。相邻的 Lark、投影、决策游标、自动化测试共 95 个通过，git diff --check 干净，解释器与包源都来自本 worktree。未在本机原生 Windows/NTFS 上独立运行，因此平台特有的 ACL/文件系统行为仍是合并前的剩余验证；这不是我已验证 CI 的声明。没有发现阻断性代码问题。

我的整体评价

APPROVE，针对精确 head 401f1b2b6854bf4af840a28176d3b550672567f7。修复范围与已证实的失败相称，既不吞掉其他写入错误，也没有放弃 POSIX 上原有的目录持久化效果。未来相关的小型重构检查结论：四个写入器保留各自领域所有权更易局部回滚，本 PR 无需再抽象共享层。

English verdict: APPROVE — exact head 401f1b2; no blocking finding. Native Windows validation remains the principal residual risk.