fix(desktop): stop bundling glib in the Linux libmpv slice

[chrisdebian]

What

scripts/mpv-linux/stage.sh bundles libglib-2.0.so.0 (2.72, from the Ubuntu 22.04 stage container) into the Linux libmpv slice because it wasn't in SYSTEM_LIBS. On a newer glib host (e.g. Ubuntu 24.04, glib 2.80), the bundled glib claims the soname before AWT's XDesktopPeer.init() can dlopen the system libgio-2.0.so.0, which then fails with an undefined symbol and disables java.awt.Desktop for the rest of the process — breaking all 23 external-link call sites via LocalUriHandler.

This is the root cause already documented in CLAUDE.md's 2026-07-31 entry ("Bundled glib disabled java.awt.Desktop on Linux"), which also names the cure directly: "The actual cure is to stop bundling glib — add it to SYSTEM_LIBS, which needs the Linux tarball rebuilt, republished and re-pinned in mpvNativesChecksums." This PR is that first half.

Adding libglib-2.0.so.0 to SYSTEM_LIBS excludes it from the bundle, so the loader falls through to the host's own glib — same mechanism as every other libc-family exclusion already in that list.

What this does NOT do

• Doesn't touch the Desktop.isDesktopSupported() workaround in DesktopApp.kt — still needed for anyone on the currently-published tarball, and removing it here would break them.
• Doesn't rebuild/republish the actual tarball or re-pin mpvNativesChecksums — per CLAUDE.md, that only happens through mpvBundleAll, which I don't have access to.

Verification

I don't have the mpvBundleAll pipeline available, so this hasn't been tested against a real staged build. bash -n and shellcheck both clean on the modified script — no new warnings introduced. Happy to help however's useful once it's staged on your end (e.g. confirming the fix against a built artifact, if you can share one).

AI disclosure

This PR was drafted with Claude Code's help — it traced the fix against your own CLAUDE.md root-cause entry, and I reviewed and confirmed the change before opening this.

[maxrave-dev]

Thanks — this is exactly the cure already documented in CLAUDE.md's 2026-07-31 entry, and it checks out against the real staged bundle. Verified before approving:

• Only libharfbuzz.so.0 links glib in the whole closure (pulled in via libass; Ubuntu's harfbuzz is built with glib support). No libgobject/libgio/libgmodule ever gets bundled, so excluding libglib-2.0.so.0 alone fully cuts the soname clash.
• harfbuzz imports only 11 ancient glib symbols (g_bytes_* ≥ 2.32, g_unichar_* / g_unicode_script_* ≈ 2.30) — so resolving against the host's glib is safe in both directions: newer hosts (2.80) are a superset, and any distro glib since ~2012 satisfies it. No "fixed on 24.04, broken on older hosts" trade hiding here.
• is_system() matches tokens exactly and ldd prints exactly libglib-2.0.so.0, so the one-token change is mechanically sound; the in-container resolution check and smoke test still pass, since the stage container has its own system glib.

Two notes for the record:

1. libpcre.so.3 will remain in the bundle as a ~470 KB orphan — it only ever entered the closure as glib 2.72's own dependency (verified: nothing else NEEDs it), but collect() walks ldd's transitive output, so it still gets copied. Deliberately NOT asking for it in SYSTEM_LIBS: pcre1 no longer exists on modern hosts (Ubuntu 24.04 ships only pcre2), so excluding it would hide a genuinely missing dependency in-container and detonate on users — the exact trap this PR escapes. Dead weight is the safer failure mode.
2. As you noted, this only takes effect at the next tarball rebuild (rebuild → republish to simpmusic-files → re-pin mpvNativesChecksums). We'll fold that into the next mpv bump rather than doing a dedicated rebuild, since the isDesktopSupported() probe plus the xdg-open fallback already mask the symptom for current users. Correct call leaving that workaround untouched.

Appreciate the honest scoping and the AI disclosure — this is precisely the kind of AI-assisted-but-human-reviewed contribution CONTRIBUTING.md welcomes.