fix(desktop): keep windows process PATH for managed agents

[Chessing234]

Summary

• on Windows, login_shell_path() returns None on purpose so children inherit the real PATH
• build_augmented_path still returned a managed-only PATH (~/.local/bin + exe parent), and runtime overwrote the child PATH with it
• system Node then disappeared ('"node"' is not recognized)
• when shell PATH is missing, append the process PATH; add regression tests

Closes #2327

Test plan

• unit tests for process-PATH fallback / shell-path preference
• on Windows with system Node only: launch a Claude Code managed agent and confirm harness starts

Made with Cursor

[Chessing234]

when login-shell path is missing on windows we now append the process path so system node stays visible.

[wpfleger96]

🤖 hey @Chessing234 — thank you for this contribution. You diagnosed this correctly: Command::env("PATH", …) replaces rather than extends, so on Windows the child was losing every system binary and npm shims died with 'node' is not recognized.

The fix landed in #2563 (cca16635d) while your PR was open, so I'm going to close this as superseded. Worth calling out that main ended up going a bit further than your version: you gated the inherited-PATH fallback on #[cfg(windows)], and main applies it on Unix too, because a login shell that exits non-zero or prints nothing hits the same failure mode there and leaves the child without curl/sh/tar. The logic now lives in should_use_inherited() and compose_path_entries() in desktop/src-tauri/src/managed_agents/runtime/path.rs, with a truth table in the compose_tests module.

I'd welcome more Windows work from you if you're up for it — nvm-windows discovery is still genuinely broken (find_nvm_default_bin only understands POSIX ~/.nvm) and that one's wide open.

[Chessing234]

not really looking forward to more windows work, but happy to help