selkies

test(display): Give the live tests after the Openbox one their DISPLAY back

live_openbox_is_read_and_restarted points DISPLAY at a private Xvfb it stops at the end, and left it there, so x11::tests' live_output_layout and live_resize_round_trip, run after it in the same process, opened a dead display and failed. It restores the DISPLAY the process had. All seven live tests pass in one run on a slot's Xvfb.

fix(pixelflux): Decline the GPU paths where libEGL is missing, not panic

smithay opens libEGL.so.1 the first time any of its EGL calls runs and panics where it cannot (ffi.rs: "Failed to load LibEGL"). On a host without libEGL that took down selkies gpu-probe (Ubuntu 24.04 without libegl1 and Alpine without mesa-egl, Tesla K80), the Wayland compositor's GPU bring-up, and the X11 DRI3 capture's NVENC import: the whole server, where every other library the server opens at run time only turns its path down.

EGL displays are made through egl_display, which opens libEGL.so.1 first (once, and keeps it open for smithay's own open) and returns an error where it is missing; gpu_render_init and the DRI3 capture decline on it as on any other EGL failure. crates/pixelflux/clippy.toml disallows smithay's EGL entry points elsewhere: EGLDisplay::new, EGLDevice::enumerate, make_sure_egl_is_loaded and get_proc_address (the one call left is made under a current context, so libEGL is open).

Every other library opened with dlopen already fails without a panic: pixelflux's and the server's own opens of CUDA, NVENC, NvFBC, VA-API, Vulkan, PipeWire, X11 and XCB, and the GPU probe's GL and EGL, return errors; nvml-wrapper's Nvml::init returns one; wayland-backend checks for libwayland-server before its Display::new and returns NoWaylandLib.

tests/tools/hide_libs.so makes dlopen fail for the libraries HIDE_LIBS names, and test_missing_libraries runs gpu-probe and an X11 and a Wayland WebSockets session with each of them hidden, with all of them at once and with none: the probe exits 0, the sessions send frames and the server stays up without a panic (without libwayland-server the Wayland server says it has no compositor). On an NVIDIA L4 with Xvfb, main fails exactly four of its 65 checks (gpu-probe with EGL hidden and with everything hidden, and the Wayland session's server panicking with EGL hidden) and this passes all 65. On a private NVIDIA Xorg (DRI3, NV-GLX), main's X11 session with libEGL hidden panics in two capture threads and sends nothing; with this the DRI3 capture declines ("no EGL display ... for the NVENC import"), NvFBC takes the capture and the frames flow, and with everything hidden it streams on x264. Without a render node, as on CI, the sessions run with nothing and with everything hidden.

test(packaging): Have the stand-in server report every crate's version

server.sh fails a build whose server does not name the release version for Selkies, pixelflux, and pcmflux alike (c0ea5fc9e), and the packaging simulation's stand-in Cargo left a server that named Selkies' alone, so every simulated package build stopped at that check. The stand-in reports the versions the workspace manifest and the pixelflux and pcmflux manifests beside it were stamped with, in the form selkies --version prints; the simulation stages all four packages again.

build(packages): Build the codecs once per architecture and C library

pixelflux's codecs are its pinned releases with its patches, the same archive whichever distribution links them. A codecs job builds all six (libx264, x265, libvpx, SVT-AV1, dav1d, and libde265) once for each of x86_64 and aarch64 and glibc and musl, in the wheels' manylinux and musllinux images, whose C library is the oldest any package targets (scripts/ci/build-codecs.sh), and every native package and the AppImage link those static archives through SELKIES_CODECS. A package build takes only the libraries that are its distribution's from its distribution, so the package scripts install no codec package and none of the tools only the codecs build with, and codecs.sh builds every codec rather than consulting the distribution. The wheel, built in those same images, builds its own with codecs.sh.

codecs.sh undefines FINITE_MATH_ONLY, which -ffast-math predefines, so an archive built against glibc before 2.31 calls the math functions by their own names and links against any newer glibc.

Verified on x86_64: the glibc and musl codec sets built by build-codecs.sh; the el9 rpm, bookworm deb, and Alpine apk built against them and each passed verify-package.sh in a pristine container; the AppImage built from the same prefix and passed verify-appimage.sh locally and on CentOS 7; tests/packaging passes. The aarch64 builds run first in CI.