Fix Windows tool navigation and managed browser reuse

[matt-greathouse]

Summary

• Improve Windows application resolution and native text input handling
• Update daemon and Lua prelude navigation behavior
• Add managed browser reuse coverage and refresh Windows build/documentation support

Testing

• Added or updated CLI, core, and managed-browser reuse tests
• Not run (not requested)

[matt-greathouse]

Reviewed the full diff (12 files, ~1000 added lines) across the Windows app resolver, desktop-session detection, native input hardening, managed-browser navigation Lua, and the Windows build/signing script. Ten findings inline, roughly in severity order.

The one to look at first is DaemonTextInput.cpp:160. The new ResolveKeycode pre-validation runs on all platforms, but the Linux implementation is a stub that returns -1 unconditionally — Linux input actually flows through NormalizeKey/XSendHotkey inside SendHotkey. Every press on Linux now fails with invalid_key. The existing tests are negative-only, so CI won't catch it.

Three other behavior regressions worth resolving before merge:

• RDP sessions lose native control (PlatformWindows.cpp:546) — turning on real console-session detection makes onConsole false under RDP, failing window_activate, app_activate_pid, and managed-browser launch with desktop_session_unavailable.
• focus_app injects a stray Shift on every call (LuaPrelude.cpp:2302) — force = true bypasses the daemon's alreadyReady short-circuit, so macOS posts a real key event even on an awake, unlocked desktop.
• navigate tests for "URL changed" instead of "URL matches" (LuaPrelude.cpp:2184) — false failure when the target canonicalizes to the current URL, false success on any unrelated redirect during polling.

The rest are narrower: exe names dropped from GetFrontmostApp (breaks wait { frontmost = "notepad.exe" }, and so the confirm half of focus_app("notepad.exe")), a lost pid on the ShellExecute fast path, continue-on-activation-failure falling through to a duplicate launch, an app-catalog cache that never caches an empty result on a 200 ms poll path, a possibly-missing propsys link (unverified — worth checking the Windows CI job), and a post-sign Status -eq "Valid" check that a self-signed dev cert can't satisfy.

The direction of the PR is good — the activation retry, the input_failed vs invalid_key split, and the AUMID-based catalog matching are all real improvements. Most findings are about edges those changes opened up rather than the approach itself.

---

Generated by Claude Code