…retry)

When the monitor's USB hub power-cycles (screen off→on), the kernel removes
and re-enumerates the keyboard + mouse. yserver picked up the keyboard but the
mouse stayed dead for many seconds — until an unrelated keypress — both under
libseat (Cinnamon) and direct (lightdm) mode.

Root cause (HW-traced on bee, confirmed by udevadm-vs-yserver timestamps): the
re-enumerated device's libseat/logind open is DEFERRED by a lagging udev uaccess
ACL, so libinput surfaces no DEVICE_ADDED on that wake. There is no later fd
readiness edge once the ACL settles, so the idle core loop (and the direct-mode
input thread, blocked on EpollTimeout::NONE) never re-dispatches libinput to
retry the open — it waits for unrelated input. The recently-landed idle-sleep
made this visible; it is a latent libinput-hotplug bug, not a regression of the
compose path.

Fix — a bounded retry window armed on any libinput device add/remove, so a
deferred open is retried without needing unrelated input, then it converges and
the loop returns to true idle:

- Libseat (core loop): new `KmsBackendV2.libinput_hotplug_retry_until`;
  `on_libinput_ready` arms it (2.5s) on DeviceAdded/Removed; new
  `Backend::poll_deferred_input` (default no-op; called once per core-loop
  iteration in run.rs) re-dispatches while armed and clears on expiry;
  `next_wakeup` adds a ~250ms retry cadence while armed (ungated by scanout, so
  input recovers regardless of screen state). Only device churn extends the
  window → idle-sleep preserved.
- Direct (input_thread.rs): the loop blocked on EpollTimeout::NONE forever; now
  epoll_wait takes a ~250ms timeout while a retry window is armed (armed on a
  batch containing DeviceAdded/Removed), so the existing per-iteration dispatch
  retries the deferred open. (Level-triggered epoll alone was insufficient — the
  udev event is consumed without surfacing the add.)

Unit tests cover the scheduling seam (retry cadence in next_wakeup, window
cleared on expiry). HW-verified on bee: monitor power-cycle under both Cinnamon
(libseat) and lightdm (direct) — mouse returns on its own within the retry
window, no keypress, and compose returns to 0 fps afterward (idle preserved).

Also adds repro/diagnostic Justfile recipes: yserver-cinnamon-hotplug-probe,
yserver-input-hotplug-probe, yserver-fvwm3-idle, yserver-e16-idle-trace,
xorg-e16-idle-trace.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…nic!

The idle-compose fix added a `#[ignore]`d placeholder whose body was
`panic!("enable this when Repaint::Clipped returns again")`. Plain `cargo test`
skips it, but CI runs `--include-ignored`, which force-runs it → guaranteed
failure.

Replace it with a real, non-ignored test that asserts `pick_repaint_region`
returns `Repaint::Full` unconditionally today (it does) and carries the
cursor-specific message. It passes now and fails — pointing at the requirement —
the day `Repaint::Clipped` is re-enabled, which is exactly the tripwire the
idle-compose spec wanted, without breaking `--include-ignored`.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>