Arrange Keys: drag-to-reorder the key rail and visionOS cluster

[jhen0409]

Summary

• Arrange Keys (⋯ menu → Arrange Keys…): the keys wiggle and become drag sources; drop a key onto the key it should sit beside. It is the window tabs' drag and drop — UIDragInteraction on each TerminalTallyKeyControl, UIDropInteraction targets whose delegate is one shared KeyBarDropCoordinator (nearest other key by centre lights like a tab cell; landing parks the displaced keys for UIKit's drop animator). The rail (iPad/iPhone) and the visionOS cluster context adopt it through KeyBarDropSurface.
• The order is a permutation of slots (KeyBarOrder, KeyBarSlot — keyboard/mic share one slot, TMUX/HRDR another), stored device-local (KeyBarOrderStore); gaps and widths never move. Rails reconcile the order in layout (never part of the render signature), so a drop, RESET or a VoiceOver move re-creates no key.
• visionOS: a drop on the other slab's key crosses the UMD (both slabs plus the UMD/bar slabs are drop hosts); the ARRANGE KEYS bar (lamp · RESET · DONE) is its own ornament slab below the console row, one per tab, observing the store for RESET itself. GUIDE moves from the console row into the ⋯ menu (guideIsDirectChip, one decision for rows and menu).
• Tabs and keys share the process-local drag item recipe (TerminalChromeDragPayload / TerminalChromeDragItem); the pane's file-drop gate refuses every chrome drag with one check; the tab strip's drop rule now runs through the same pure RowDropGeometry.

Verification

• Vision Pro device: the pinch-and-drag across the UMD and back verified by hand.
• build.sh build + targeted unit suites green on visionOS and iPad (key control, rail/cluster incl. fake UIDropSession drives of the coordinator, order model, UMD, window, pane, tab strip); build.sh lint 0 violations; check-metadata.rb OK.
• New build.sh uitest [vos|ipad] → MultiplexUITests/ArrangeKeysUITests: real-touch press-and-drag of ESC onto TAB passes on the final code (iPad Pro 11-inch (M5), iOS 26.4 simulator: escape control tab → control tab escape, RESET restores, DONE). The visionOS simulator cannot synthesize a lift for a drag interaction — the test skips the drag there.
• Headless Vision Pro sim proofs (debug.arrangekeys, debug.arrangekeysmove): bar below the console row, a key carried across the UMD; console row shows GUIDE gone and ⋯ present.

Notes

• Not in CI: the XCUITest lane needs the dev-sshd harness.