fix(desktop): make Linux AppImage GStreamer work on non-Debian distros

[dcadenas]

Problem

The Linux AppImage opens a blank / fully transparent window on non-Debian
distros (Arch, Fedora, …): the transparent Tauri shell shows the desktop
through it because the WebKitWebProcess crashes (SIGABRT) during media-pipeline
init, logging:

GStreamer element appsink not found
GStreamer element appsrc not found
GStreamer element autoaudiosink not found

The .deb is unaffected; only the AppImage is broken, and only off
Debian/Ubuntu.

Root cause

Two things combine in desktop/scripts/fix-appimage.sh (the CI post-build
step):

1. linuxdeploy's compiled AppRun.wrapped force-sets
   GST_PLUGIN_SYSTEM_PATH_1_0=$APPDIR/usr/lib/gstreamer-1.0. A set
   GST_PLUGIN_SYSTEM_PATH_1_0 replaces GStreamer's compiled-in default
   search path — it does not add to it.
2. fix-appimage.sh populated $APPDIR/usr/lib/gstreamer-1.0 as an absolute
   symlink to the Debian multiarch path /usr/lib/x86_64-linux-gnu/gstreamer-1.0,
   which only exists on Debian/Ubuntu — so on Arch/Fedora the symlink dangles.

Result on non-Debian: GStreamer searches only a nonexistent directory → zero
plugins → appsink/appsrc/autoaudiosink missing → WebKit aborts → blank
window. The old comment claimed a dangling symlink "safely falls back to
default discovery"; that is false — a set GST_PLUGIN_SYSTEM_PATH_1_0 disables
the default.

(fix-appimage.sh already strips the bundled GStreamer core libs so the app
uses the host's GStreamer; the bug is purely that the plugin path was pinned
to a Debian-only location.)

Fix

Complete the script's existing "rely on the host's GStreamer" design. Instead
of pinning the plugin path to a hardcoded distro layout, install a small
launcher shim in front of the app binary that removes the bundle-pointing
GST_PLUGIN_* overrides after AppRun.wrapped sets them, so the system
GStreamer resolves plugins via its own default path — correct on Debian, Arch,
and Fedora alike.

Why a shim and not just unsetting the var in AppRun: AppRun.wrapped is a
compiled binary that rewrites GST_PLUGIN_SYSTEM_PATH_1_0 after every
apprun-hook, discarding any value set before it (verified empirically — a
runtime value passed in does not survive). The only place to neutralize the
override is between AppRun.wrapped and the real binary, i.e. on the exec
target usr/bin/buzz-desktop (moved aside to buzz-desktop.bin).

The shim only unsets values that point into the AppImage, so a user's own
GST_PLUGIN_* env is preserved. Fail-loud guards were added (verify
AppRun.wrapped still injects the override; verify the app binary exists;
refuse double-install) so a future tauri/linuxdeploy change is caught in CI,
not by users. Also removed a now-unused multiarch-detection block and renamed
the CI step.

Validation

Proven on Arch Linux (x86_64):

• gst-inspect-1.0 with the override pointed at the dangling Debian dir →
  appsink/appsrc/autoaudiosink NOT FOUND; with it unset → all FOUND.
• The shim, exercised through the real launch chain
  (AppRun → AppRun.wrapped → shim → app) against a fresh extract of the
  shipped AppImage: the forced override is stripped and the app process sees
  all three elements.
• bash -n passes on the script and the generated shim.

Not yet tested (needs CI):

• The full fix-appimage.sh extract → repack → sign path locally (no
  appimagetool on hand; the only local AppImage is already fix-processed,
  which the libwayland guard correctly refuses to re-process).
• Actual GUI window paint from a freshly-built AppImage on a non-Debian distro.

Files

• desktop/scripts/fix-appimage.sh — shim replaces the dangling symlink;
  guards; corrected comments; removed unused multiarch block.
• .github/workflows/release.yml — step name updated.

[wpfleger96]

hey Daniel — the root-cause analysis here is right and well-evidenced (a set GST_PLUGIN_SYSTEM_PATH_1_0 replaces GStreamer's default search path, so the old Debian-multiarch symlink was a Debian-only accident), and the shim placement after AppRun.wrapped is the only spot that works. The Arch validation through the real AppRun → AppRun.wrapped → shim → app chain is exactly the right disproving test. The fail-loud guards are great too.

Requesting changes for two small hardenings (comments 1 and 2 inline — one flag and one character respectively). Comments 3 and 4 are nits, take or leave. None of this changes the approach.

[wpfleger96]

the bare exec makes argv[0] …/buzz-desktop.bin. GTK derives prgname → WM_CLASS from argv[0], so taskbar icon matching / window grouping against the .desktop file can break, and pidof/killall buzz-desktop stop working. One flag fixes it:

exec -a "buzz-desktop" "$here/buzz-desktop.bin" "$@"

Separately, the kernel comm comes from the exec'd file's basename (not argv[0], so exec -a doesn't cover this leg): buzz-desktop.bin truncates to buzz-desktop.bi, which no longer matches DESKTOP_BINARY_NAMES in desktop_is_alive_for_instance (runtime.rs:761 / :955). I checked and this isn't a new regression — release Linux builds already fail that function's cmdline leg (the instance id only appears in /proc/pid/cmdline for tauri dev --config launches), so release desktops are already invisible to reap_dead_instance_agents on main. A companion line adding buzz-desktop.bi to DESKTOP_BINARY_NAMES would close the comm leg for whenever the cmdline leg gets fixed — fine as a fast-follow if you'd rather keep this PR tight.

[dcadenas]

Addressed in 8f3003f. The shim now uses exec -a "buzz-desktop" to preserve argv[0], and I also added the Linux-truncated buzz-desktop.bi name to DESKTOP_BINARY_NAMES with a focused regression test.

[wpfleger96]

*"$appdir"* is a loose substring match — if someone runs the extracted AppDir from e.g. /opt/buzz, a user-set GST_PLUGIN_PATH=/opt/buzz-plugins would falsely match and get unset. *"$appdir/"* still matches the injected $APPDIR/usr/lib/gstreamer-1.0 and kills the false positive. One character.

[dcadenas]

Addressed in 8f3003f. The match now requires $appdir/, so the injected in-AppImage paths are removed without falsely matching sibling paths such as /opt/buzz-plugins.

[wpfleger96]

nit (non-blocking): the shim adds a hard runtime dep on bash + readlink -f (coreutils). Both are universally present on the distros we target so I wouldn't hold anything on this, but a POSIX-sh shim would be marginally more portable if you feel like it.

[dcadenas]

Keeping Bash intentionally here: the blocking process-name fix relies on Bash's exec -a, which POSIX sh does not provide portably. bash and readlink -f remain available on the Linux distributions we target.

[wpfleger96]

nit (non-blocking): this guard also fires when AppRun.wrapped is missing entirely (grep's exit 2 under if ! — which is good, that's the right behavior), but then the message says the file "no longer references" the var when the file doesn't exist. Behavior's right, wording's slightly off — maybe cover both cases in the message.

[dcadenas]

Addressed in 8f3003f. The guard now says AppRun.wrapped may be missing or may no longer reference the variable, covering both grep failure cases accurately.