SIGSEGV on vulkan fullscreen swapchain recreation (wayland-native, NVIDIA 595.91.07, mpv 0.41.0) — deterministic #18577

Summary

On wayland-native with --gpu-api=vulkan, ANY fullscreen swapchain recreation to panel-native size segfaults deterministically — every time, both vo=gpu and vo=gpu-next, driver 595.91.07 (also 595.84). The last log line before the crash is always (Re)creating swapchain of size 2560x1440.

This is a sharper repro of the behavior previously seen as a cross-monitor maximize crash: the trigger is any fullscreen swapchain recreation on native wayland; cross-monitor move+maximize was merely one route into it. Xwayland (X11 context) is fully immune (20 maximize + 5 fullscreen cycles clean).

Minimal repro

# any h264 clip; toggle fullscreen a few times on a wayland-native display
mpv --no-config --vo=gpu-next --gpu-api=vulkan --msg-level=gpu=v clip.mp4
# press f (or maximize) once or twice → SIGSEGV

Our measured runs: fullscreen-only torture → SIGSEGV on cycle 1, 3/3. maximize-only → 20/20 clean. Xwayland → all clean.

Environment

• mpv 0.41.0, libplacebo 7.360.0
• NVIDIA RTX 3070, driver 595.91.07 (and 595.84 — same crash)
• Ubuntu, GNOME/mutter wayland session, native (not XWayland) window

Full verbose logs + the torture scripts: happy to attach (they're in our repo's lane evidence dir). This looks like the NVIDIA vulkan driver's wayland swapchain recreation path; an mpv-side guard (refuse/recreate differently) may also be possible.