core: fix security and concurrency issues found in a backend audit (#2805)

* core: fix security and concurrency issues found in backend audit

Security:
- privesc: pipe the sudo password via stdin (sudo -S) instead of
  embedding it in the command string, so it no longer appears in argv
  (readable by any local user via /proc/<pid>/cmdline or ps)
- greeter: tokenize a session .desktop Exec= line into argv and execve
  directly instead of running it through /bin/sh -c, closing a command-
  injection path via user-writable ~/.local/share/wayland-sessions
- plugins: reject path-separator/.. in plugin id/name before joining
  into a filesystem path, closing an arbitrary-directory-delete in the
  uninstall/update fallback
- keybinds/hyprland: always quote unrecognized bind actions/keys when
  writing generated Lua; only re-emit genuine round-tripped custom Lua
  verbatim (tracked via an explicit flag), closing a Lua-injection path
- desktop/mimeapps: reject newline/bracket in mime/desktop-id fields so
  they can't inject fake sections into the shared mimeapps.list

Robustness / concurrency:
- server: recover panics in the request-dispatch path so one bad
  handler can't crash the daemon and drop every client
- go-wayland: recover panics in the shared dispatch choke point so a
  malformed compositor event can't crash CLI tools / the daemon
- server: per-connection D-Bus client ID instead of a shared constant,
  fixing cross-client signal delivery and subscription teardown
- network: guard the NetworkManager device maps with a mutex (a
  concurrent map read/write here is an unrecoverable fatal error)
- cups: close the event channel on Stop() so Unsubscribe() of the last
  subscriber no longer deadlocks; allocate the fresh channel in Start()
- freedesktop: reuse the shared session conn for the settings watcher
  and tear it down in Close(), fixing a per-Manager conn+goroutine leak
- clipboard: mutex-guard lazy dbusConn creation
- geolocation: use WithMatchMember for the GeoClue2 LocationUpdated
  signal (was WithMatchSender with an interface.member string, so the
  match never fired and live location updates never arrived)
- screenshot: set failed=true on buffer/pool creation errors so the
  dispatch loop doesn't wait forever for a ready/failed that never comes

* apply code review comments

---------

Co-authored-by: bbedward <bbedward@gmail.com>

core/internal/plugins/manager.go

@@ -64,7 +64,20 @@ func (m *Manager) findInstalledPath(pluginID string) (string, error) {
 	return m.findInDir(systemDir, pluginID)
 }
 
+// isSafePluginPathComponent rejects ids that aren't a single path component,
+// so filepath.Join can't resolve (and later RemoveAll) outside the plugins dir
+func isSafePluginPathComponent(s string) bool {
+	if s == "" || s == "." || s == ".." {
+		return false
+	}
+	return !strings.ContainsAny(s, "/\\")
+}
+
 func (m *Manager) findInDir(dir, pluginID string) (string, error) {
+	if !isSafePluginPathComponent(pluginID) {
+		return "", fmt.Errorf("invalid plugin id: %q", pluginID)
+	}
+
 	// First, check if folder with exact ID name exists
 	exactPath := filepath.Join(dir, pluginID)
 	if exists, _ := afero.DirExists(m.fs, exactPath); exists {
@@ -507,6 +520,10 @@ func (m *Manager) findInstalledPathByIDOrName(idOrName string) (string, error) {
 }
 
 func (m *Manager) findInDirByIDOrName(dir, idOrName string) (string, error) {
+	if !isSafePluginPathComponent(idOrName) {
+		return "", fmt.Errorf("invalid plugin id/name: %q", idOrName)
+	}
+
 	// Check exact folder name match first
 	exactPath := filepath.Join(dir, idOrName)
 	if exists, _ := afero.DirExists(m.fs, exactPath); exists {