feat(wayland): close the live-input gap with a sovereign virtual-input injector

[tannevaled]

Task 1 — close the Wayland live-input gap

The Wayland live screenshot proof already passed, but live input was left
pending-on-compositor: the headless sway seat exposed no keyboard, so an
injected key never reached the client. This closes it end to end, with the
behaviour captured (not claimed).

What changed

• Sovereign virtual-input bindings (internal/wayland/virtual.go): client
  side of zwp_virtual_keyboard_manager_v1/zwp_virtual_keyboard_v1 and
  zwlr_virtual_pointer_manager_v1/zwlr_virtual_pointer_v1. Attaching a
  virtual device to a seat makes it advertise the matching capability, so a
  plain client receives the injected events through the ordinary
  wl_keyboard/wl_pointer path. 100 %-covered (both endian paths).
• Dynamic input hot-plug: wl_seat gains an OnCapabilities callback; the
  window binds pointer/keyboard when a device appears after bring-up, not
  just at connect time — required for a device-less headless seat.
• Live test is its own injector (second connection, no wtype/ydotool):
  uploads a real libxkbcommon-compiled us keymap through the virtual
  keyboard, injects KEY_A + a left click, asserts the exact dispatched
  toolkit.Event.
• Deterministic floor kept: wldynamic_test.go proves the hot-plug path
  in-process; the original fake-compositor input test stays.
• CI: lane upgraded to live Wayland (input); installs libxkbcommon-tools
  (xkbcli), drops wtype.

Proof (captured, verified live under sway headless)

live input: EventChar 'a' dispatched from a real compositor key
live input: EventClick dispatched at (200,150) from a real compositor button
--- PASS: TestLiveWayland (2.01s)

internal/x11 and internal/wayland both at 100.0 % coverage under -race.

🤖 Generated with Claude Code