[Windows][P0] 当前插入链路没有真正的 current-caret 保证，只是盲发 Ctrl+V

[Cooper-X-Oak]

Summary

当前 Windows 输出路径没有保存原目标窗口/控件，也没有恢复焦点或验证 caret 所属控件；本质上只是“写剪贴板 + 在当前焦点处发 Ctrl+V”。同时 Windows 端 insertStatus 永远降级成 copiedFallback，系统自己也无法证明“已经成功插入”。

Repro

1. 让某个目标输入框先获得焦点。
2. 录音过程中切换窗口，或让 OpenLess 主窗 / capsule / 其他窗口抢前台。
3. 结束录音，观察文本是否回到原输入位置。

Expected

• 插入前应有“原目标窗口/控件仍为活跃输入目标”的保证。
• 至少应能区分“已插入”与“仅复制待手动粘贴”。

Actual

• Windows 路径没有保存目标窗口，没有恢复前台窗口，也没有验证当前 caret。
• 非 macOS 路径即使 simulate_paste() 成功，也固定返回 CopiedFallback。

Evidence

Verified:

• openless -all/app/src-tauri/src/coordinator.rs:543-624
• openless -all/app/src-tauri/src/insertion.rs:64-90
• openless -all/app/scripts/windows-real-asr-insertion-smoke.ps1

补充：现有 smoke 仍依赖把目标窗口重新置前后再做回读，且脚本本身也接受 insertStatus == copiedFallback。

Impact

这是 Windows 主链的核心可用性缺口：文本可能粘到错误窗口、完全没粘上，或者只能留在剪贴板里等用户手动二次 Ctrl+V。

Owner Note

Owner intent: @Cooper-X-Oak plans to fix this personally. Please do not pick this up or reassign it unless asked.

[Cooper-X-Oak]

Claiming this for personal implementation.

Please do not work on this in parallel or open a competing fix/PR unless we coordinate first.

[Cooper-X-Oak]

Progress update from current upstream-aligned branch (codex/windows-upstream-inplace):

Implemented on the Windows path:

• Added a distinct pasteSent insertion status so we no longer misreport a successful paste attempt as copiedFallback.
• Capture the foreground window at session start and attempt to restore it immediately before insertion.
• On Windows, try to show the capsule with a no-activate path (SW_SHOWNOACTIVATE / SWP_NOACTIVATE) before falling back to regular show().
• Combined with the clipboard restore work, this reduces OpenLess-originated focus/clipboard interference around insertion.

Current status of #96:

• This does not fully prove current-caret insertion across arbitrary Windows hosts yet.
• It does reduce self-inflicted focus theft and removes one major source of misleading fallback semantics.

Runtime note:

• A previously observed symptom where the capsule appeared to open and not close could not be reproduced anymore on the current build in manual retest.
• I do not want to overclaim a full fix for all hosts, but the current version is behaving better than the earlier repro case.

Local validation completed:

• npm run build ✅
• cargo +stable-x86_64-pc-windows-gnu check --manifest-path .\src-tauri\Cargo.toml ✅
• cargo +stable-x86_64-pc-windows-gnu check --manifest-path .\src-tauri\Cargo.toml --tests ✅
• rustfmt --check --edition 2021 ✅
• git diff --check ✅

Still not fully green end-to-end:

• GNU cargo test --no-run is still blocked by a separate linker issue: ld.exe: export ordinal too large
• So this is not yet a claim of full CI/runtime closure.

I am continuing to own this issue personally. Please do not open a parallel fix unless we coordinate first.

[Cooper-X-Oak]

Status note after re-check:

Commit 920a31f (fix(windows): 收敛插入焦点与剪贴板副作用) is already contained in main, and it improved the Windows insertion path by:

• reducing OpenLess-originated focus theft around insertion
• distinguishing pasteSent from copiedFallback
• combining that with clipboard-restore behavior

However, this issue is staying open on purpose.

Reason:

• the current evidence still does not fully prove true current-caret insertion across arbitrary Windows hosts
• the earlier progress note on this issue already made that boundary explicit

So the partial mitigation is merged, but the stronger mechanism-level closure requested by #96 is not yet proven complete.