feat(window): show the taskbar button and Dock icon outside stealth mode

[alpha5611331]

Closes #83

The taskbar button and the Dock icon now follow stealth mode instead of being suppressed all the time, and the Ctrl+Shift+F8 restore shortcut that existed to work around their absence is gone.

	Windows taskbar button	macOS Dock icon / Cmd+Tab	Desktop shortcut
Normal mode	shown	shown	never created
Stealth mode	hidden	hidden	never created

Commits

1. feat(hotkeys) - remove Ctrl+Shift+F8. restoreWindow() itself is unchanged; relaunching the app still routes to it through the single instance lock.
2. feat(window) - applySurfaceVisibility() replaces hideFromTaskbar() and drives both surfaces from the stealth flag.
3. test(stealth) - a manual Windows probe for the part unit tests cannot reach.

Each commit stands on its own; pnpm test:main passes at every one.

Two traps worth knowing about

LSUIElement had to go. It pins the packaged mac app to accessory from launch, so there would be no Dock icon to give back. The activation policy is switched at runtime instead, alongside dock.show()/dock.hide() - the policy is what actually moves the app, dock.show() alone does not lift an accessory app back out.

macOS drops a Dock call made within one second of the previous one, silently (docs). My first pass had this bug: toggling stealth twice quickly - mashing the hotkey - left the Dock icon on screen for the rest of the session, no error, nothing logged, in the one mode where that matters most. Now no-op calls are skipped so window events cannot spend the budget a real toggle needs, and a call landing inside the window schedules a re-assert. The activation policy carries no such limit and is applied immediately, so the icon still goes at once even in the swallowed case.

Testing

Live Windows taskbar, 9/9. Unit tests can only prove we pass the right argument to setSkipTaskbar; whether the shell acts on it is what actually broke before, and it fails without an exception. test/manual/taskbar-probe.mjs drives the real compiled service in a real Electron process and reads the taskbar back through UI Automation:

ok  the window starts on the taskbar
ok  cycle 1/2/3: stealth takes the button away      (absent)
ok  cycle 1/2/3: leaving stealth gives it back      (present)
ok  rapid toggles settling in stealth               (absent)
ok  rapid toggles settling in normal mode           (present)

Two false readings on the way there, in case anyone repeats this: screenshotting the taskbar is useless if another Electron app is running (same icon), and counting taskbar buttons is wrong because Windows groups all Electron apps under one button (shared AppUserModelID). Counting the windows that button reports is the reliable signal.

Automated (pnpm test:main, 118 checks):

• stealth-toggle - 10 full cycles with a window event between each, asserting the button tracks the mode every time. One pass in each direction proves little here; the original bug only appeared after switching.
• stealth-dock (new) - 5 rapid cycles, the swallowed-call re-assert, no-op suppression, and that the re-assert does not re-arm itself. Loads a second copy of the service via loadMainAs('darwin', ...), since these branches are dead code on the Linux runner CI uses.
• stealth-surface - updated to pin the runtime policy switch, the absence of LSUIElement, and that no F8 shortcut is registered.

pnpm lint, both tsc configs and vite build pass.

Not covered

The macOS Dock is verified at the call level only - the stub records what we ask for, not what macOS does with it, and the rate-limit repair rests on Electron's documented behaviour. Someone with a Mac should confirm the icon comes and goes across repeated toggles, including two fast ones inside a second. window-all-closed still quits on every platform, unchanged and still justified: stealth mode can be active when the last window closes.