wayland/lock: recover graphics initialization failures

[acrogenesis]

Graphics initialization failure on a session-lock surface aborts Quickshell, leaving the session locked without an authentication UI.

Handle sceneGraphError and recreate the failed native window: Qt cannot retry RHI initialization on that window. Preserve the compositor lock, QML input state, geometry, visibility, color and focus. Retries back off from 1 to 30 seconds, reset after a rendered frame, transfer on reload, and stop after unlock or output removal.

Tested with Clang 22 and Qt 6.11.2:

• All-features Debug build and eight focused Qt tests pass across offscreen and opt-in Wayland runs.
• Dedicated Hyprland VM lock fixture: EGL_BAD_ALLOC injection, reload during pending retries, recovery, keyboard input and unlock pass with basic and threaded render loops.
• Full offscreen suite: 9/10 executables pass; existing popupwindow::moveWithParent failure remains.
• Formatting passes. Local lint reports only existing findings outside changed lines; custom tidyfox is unavailable.

Physical NVIDIA memory exhaustion, older Qt and multiple monitors remain untested. Persistent allocation failure still requires resources to become available.

[outfoxxed]

Have you observed failure paths besides OOM on launch? I'd rather just trigger a wayland unlock if we fail to allocate all lock surfaces initially instead of doing an exponential backoff on vram.

If you're specifically observing a case where a scene graph error happens on unsuspend or something then that requires more consideration.

[acrogenesis]

This happened when idle locking an already running desktop, rather than on resume:

The sequence of events was:

1. vLLM occupied ~30 GB of VRAM, leaving roughly 388 MB free.
2. The desktop reached its idle timeout and requested a lock.
3. Qt hit EGL_BAD_ALLOC during graphics initialization and aborted Quickshell.
4. Quickshell crashed seven times, exhausting the launcher's restart attempts. Hyprland remained alive.
5. I recovered the shell using software OpenGL.

The backoff was my proposed recovery approach, rather than something required by an observed resume failure. Your suggestion to unlock if initial lock-surface allocation fails fits the confirmed case.