KWin's libeis backend logs a single, contextless line when it fails to add
an input device to an EIS client's session:

kwin_libeis: Libeis: Failed to add device, disconnecting client

No device type, no error code, and no further detail precedes or follows it
in the journal. The effect is that the entire EIS client connection is torn
down, not just the one device that failed to add. Every input-injection
call the client makes afterward fails at the socket level (broken pipe /
EPIPE on the client side) for as long as that client keeps trying to use
the now-dead connection.

ENVIRONMENT (where this was observed)
- kwin-wayland 4:6.3.6-1 (Debian 13/trixie)
- libeis1 1.3.901-1
- xdg-desktop-portal-kde 6.3.5-1
- Client: lamco-rdp-server, an RDP server using libei (via the
  RemoteDesktop portal) for input injection

OBSERVED
Fired twice within a 12-hour period on the same machine, both times within
under a minute of the client establishing or re-establishing its EIS
session. In the more visible case, 35 seconds after an RDP client
reconnected, every mouse and keyboard injection attempt failed with a
broken pipe (500+ occurrences logged client-side) until the RDP client
itself gave up and disconnected roughly 7 minutes later. To the end user
this presented simply as "keyboard and mouse stopped working."

UPSTREAM STATUS
Filed upstream as https://bugs.kde.org/show_bug.cgi?id=525307. It was
auto-closed RESOLVED/UNMAINTAINED by KDE's bug-janitor bot within half an
hour, on version-support grounds (6.3.6 predates KDE's current support
window), not on technical merits -- the bot suggested reporting to Debian
instead, which is what this is.

FRESH EVIDENCE THIS MAY BE TRIXIE-SPECIFIC
I reproduced the same disconnect/reconnect sequence against a KWin 6.7.4 /
libeis1 1.5.0-3+b1 host (Debian sid) today. Three consecutive
disconnect-then-reconnect cycles against the same lamco-rdp-server client
produced zero "Failed to add device" occurrences in the KWin journal --
clean reconnects every time, where the trixie host fails close to
reliably. So this looks like it may already be fixed somewhere between
libeis 1.3.901 and 1.5.0, or between KWin 6.3.6 and 6.7.4, rather than
being a live upstream bug -- but I don't have the means to bisect which
package or which exact change fixed it.

REQUEST
Given the above, this may be resolvable either by backporting whatever
fixed it to trixie's libeis/kwin, or simply noting it as fixed upstream if
that's confirmed. Happy to provide the full journal excerpt from either
machine, or test further, if that helps narrow it down.