native: restart a WSLg clipboard read when focus leaves its helper - #1849

WSLg serves its compositor at `/mnt/wslg/runtime-dir/wayland-0`. When
`XDG_RUNTIME_DIR` points elsewhere (a systemd user runtime directory,
for example), libwayland cannot find the socket, the clipboard falls
back to the WSLg X11 bridge, and an image paste gives a 0-byte text.
Reporters worked around it with
`WAYLAND_DISPLAY=/mnt/wslg/runtime-dir/wayland-0`.

On WSL only, a failed connect now tries
`/mnt/wslg/runtime-dir/$WAYLAND_DISPLAY` before it reports
`unsupported`. Off WSL nothing changes. A connection made this way is
not the inherited `WAYLAND_SOCKET`, so the idle drain may close it.
The source grows by 13 lines for this branch, which the user asked
for. `allow_core_data_device` becomes `is_wsl`, because it now gates
both WSL paths.

Issue: 383

The setup table did not notice two plausible breaks. Trying the WSLg
socket before the default one on WSL passed, and so did forgetting a
failed connect so that every later drive connects again.

The table now counts connect calls and drives each connection three
more times. A failed connect must stay at one attempt (two on WSL)
until the host closes the connection. A new WSL row with a working
default socket must not touch the WSLg path.

Issue: 383

With `XDG_RUNTIME_DIR` unset and a relative `WAYLAND_DISPLAY` (the
anomalyco/opencode#57724 environment), libwayland refuses the connect
and logs "XDG_RUNTIME_DIR is invalid" to stderr, which is the
terminal. Since unsupported results are retried per operation, that
line appeared on every paste instead of once per process.

The default connect now runs only when libwayland can resolve it: an
inherited `WAYLAND_SOCKET`, an absolute `XDG_RUNTIME_DIR`, or an
absolute `WAYLAND_DISPLAY`. Otherwise it goes straight to the WSLg
socket on WSL, or reports unsupported. Five reads in that environment
print 0 lines, down from 5 (1 on the base). The source grows by five
lines net for the check and its comment.

Issue: 383

The WSLg fallback and the core data device both depend on the host
copying `libraries.is_wsl` into the connection, but no test went
through that assignment. Dropping it, or setting it to true, passed
the native suite while a real socket probe showed the wrong result
on and off WSL.

A host test now swaps in fake libwayland symbols, drives a connection
from the service for both flag values, and checks the flag the
connection received.

Issue: 383

A Wayland core data-device read needs a focused helper surface. When
keyboard focus left that surface while the read waited, for example
because another window took focus on WSLg, the read ended as `failed`
with "Wayland clipboard protocol failed".

The read now destroys the helper and maps a new one, which takes focus
again and receives the current selection. It does so at most twice per
acquisition, the bound the stale-image retry uses, so a compositor that
keeps taking focus still ends the read as `failed`. On WSLg with a
second clipboard client taking focus, failed reads went from 3-11 of
48 to 0 of 48.

The source grows by 8 lines: the restart, its counter and its comment.
The host test drives a core read through replayed compositor events
for the first time, so the tests grow by its table.

Issue: 385

Losing the bound seat's keyboard fails the connection in the same
branch, and both readers of `core_focus_lost` return on a failed
connection first, so setting the flag there had no effect. Now only a
`leave` for a live helper sets it.

Issue: 385

The restart lives in the connection, so a connection test that replays
focus losses shows it as well as a host read did: the restarted read
takes the fresh offer, and the third loss fails it. That replaces the
host test, and the listener functions are private again.

A host row now checks that a data-control read takes no helper and
stays on Wayland, which no test caught before.

Issue: 385