Verify the libglvnd RPATH fix for the EGL dlopen actually builds on NixOS (#508 follow-up)

[EtienneLescot]

Context

#508 adds an EGL-based DRM modifier enumeration whose libEGL.so.1 is dlopen'd at runtime (electron/native/pipewire-capture/csrc/dmabuf_modifiers.c:44). Nothing links libEGL, so nothing puts it on the binary's RPATH, and the dlopen fails on any host without an ld.so.cache entry for it — NixOS being the case we package for.

The PR handles this: nix/pipewire-helper.nix adds libglvnd to the patchelf'd rpath alongside the libpipewire entry that already existed for the same reason. The fix is present. It has never been built.

Why it can't be verified in CI today

nix-build.yml runs on workflow_dispatch, push to main, and a Monday 06:00 UTC cron — not on pull_request. So a packaging change merges unverified and is first exercised either by the weekly cron or by a NixOS user.

The failure mode is quiet, which is what makes it worth tracking rather than assuming:

• the dlopen fails, only LINEAR/INVALID modifiers get advertised
• mutter falls back to shm, and [Bug]: Whole-screen recordings freeze/judder on GNOME/Wayland (AMD) — capture forced onto shm, mutter under-delivers vs dmabuf #507's fullscreen freeze persists on NixOS
• the probe still reports dmabuf-import available: true, and nothing appears in the log unless OPENSCREEN_PIPEWIRE_DEBUG is set

So the symptom on NixOS is "the fix didn't work", with the diagnostics actively pointing the other way.

What would close this

1. Build nix/pipewire-helper.nix on a NixOS host (or in a Nix container) and confirm the helper's RPATH resolves libEGL.so.1 — patchelf --print-rpath on the output binary, then an actual recording.
2. Confirm the enumeration returns more than LINEAR/INVALID there, i.e. that the dmabuf path actually engages.

Note that fixupPhase runs patchelf --shrink-rpath, and build.rs passes -Wl,--disable-new-dtags on purpose — the existing comments in pipewire-helper.nix explain the interaction. A change here that looks correct can still be shrunk away, which is precisely why reading the expression is not a substitute for building it.

Related

• Adding pull_request to nix-build.yml would close the general gap, not just this instance. Worth considering separately — it is the reason this issue has to exist at all.
• Follow-up to fix(capture): zero-copy dmabuf→VAAPI capture — fix whole-screen recording freeze on GNOME/AMD (#507) #508 (thread on dmabuf_modifiers.c:44).