fix(computer-use): win32 input actions fail truthfully; targeted mouse-down presses

[EvanProgramming]

Summary

Fixes #5908 (relates to #5896). The Windows computer-use backend reported {action_sent: true} even when the underlying PowerShell subprocess never delivered the input.

Three root causes, all in crates/tui/plugins/computer-use/src/backends/win32.mjs:

1. User32 type never loaded in the action process. create() only ran Add-Type in a bootstrap process, but every ps() spawns a fresh powershell.exe, so [User32]::mouse_event/SendInput/GetCursorPos were undefined there and the call failed silently.
2. withUser32 ignored the exit code. It just returned ps(...), so a nonzero exit still produced {action_sent: true}.
3. left_mouse_down precedence bug. target ? A : "" + B put the mouse_event(LEFTDOWN) in the no-target branch, so a targeted call only sent SetCursorPos.

Changes

• USER32_DEF prepends the User32 P/Invoke definition to every User32-backed action so each is self-contained in its own process.
• withUser32 throws an ExecError on a nonzero exit — no more false success.
• left_mouse_down rewritten so a targeted call both moves the cursor and presses.
• cursor_position now routes through withUser32 (loads the type, fails truthfully instead of lying about a position).
• create({ exec }) accepts an injectable runner so the backend is testable on any host without powershell.exe (mirrors the harmonyos backend's pattern).

Tests

New crates/tui/plugins/computer-use/tests/win32.test.mjs (runs on any host via a mocked runner):

• User32 actions define the type in their own process.
• A nonzero subprocess exit rejects the action (no false action_sent: true).
• left_mouse_down({ target }) sends both SetCursorPos and LEFTDOWN.
• cursor_position loads the type in-process and never reports a fake position.

Verified: node --test tests/win32.test.mjs → 4/4 pass; node --test tests/backends.test.mjs → 10/10 pass (no regression).

[github-actions]

Thanks @EvanProgramming for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.

[Hmbown]

Closing with credit — this bug (#5896) was also fixed independently and has just merged as #5903 (self-contained Add-Type prelude per action, psOk/psJson throwing on nonzero exit/spawn failure, targeted mouse-down always pressing after the optional move). Your take arrived while ours was in CI, so rather than let it sit: thank you — and specifically, your instance-local injectable runner is a cleaner test seam than the fake-powershell.exe-on-PATH approach we merged, and we may adopt that pattern for future backend tests. If you'd like to rebase your DI seam on top of main as a follow-up refactor PR, it would be welcome. Also happy to review your other PR (#5911) — that one's untouched territory.