fix: keep DMA-BUF renderer on Wayland + NVIDIA, disable explicit sync

[kipavy]

Refs #224. Complements #313, which is still needed.

What

When linux_gfx.rs detects NVIDIA on a native Wayland session, it now sets __NV_DISABLE_EXPLICIT_SYNC=1 and leaves WebKit's DMA-BUF renderer on. Before, it set WEBKIT_DISABLE_DMABUF_RENDERER=1.

Everything else is unchanged:

• X11 + NVIDIA still disables DMA-BUF.
• AppImages still disable DMA-BUF. Their bundled GTK hook exports GDK_BACKEND=x11, so they always run on X11.
• Non-NVIDIA systems are not affected.
• Env vars set by the user still win, and VOLTIUS_NO_LINUX_GFX_WORKAROUNDS=1 still turns everything off.
• If the user explicitly sets __NV_DISABLE_EXPLICIT_SYNC to 0 or to an empty value, DMA-BUF is disabled as before.

Why

On Wayland + NVIDIA, the DMA-BUF renderer fails because the compositor raises a protocol error and disconnects the app:
wp_linux_drm_syncobj_surface_v1 error 4, "explicit sync is used, but no acquire point is set".
NVIDIA's egl-wayland turns on explicit sync for the GTK toplevel surface. GTK then commits a wl_shm buffer to that same surface without an acquire point. Upstream bug: https://bugs.webkit.org/show_bug.cgi?id=324551

The fallback renderer (DMA-BUF off) shows WebGL canvases one frame late (https://bugs.webkit.org/show_bug.cgi?id=324549). That caused the ~0.6 s terminal echo lag in #224. With explicit sync disabled, the DMA-BUF renderer works and WebGL frames appear on time.

Testing

Tested on an RTX 5070 with driver 610.57.04, KDE Plasma 6.7.4 and webkit2gtk-4.1 2.52.6.

Config	Result
Wayland, DMA-BUF off (current dev)	~0.5 s echo lag; with cursor blink off, each key shows the previous character
Wayland, DMA-BUF on, no env var	App exits with the protocol error above
Wayland, DMA-BUF on + __NV_DISABLE_EXPLICIT_SYNC=1	Renders correctly, instant echo even without #313, no protocol error in WAYLAND_DEBUG
X11, DMA-BUF on + the env var	Blank page ("Failed to create GBM buffer"), so DMA-BUF stays off on X11
X11, DMA-BUF off	Echo 559–640 ms on dev, 22–34 ms with #313 (measured by injecting keys and polling screen pixels)

• Wayland latency was checked by eye; there's no pixel-polling probe for Wayland.
• Resizing, maximizing and switching between terminal tabs all looked fine.
• CPU use was lower with the env var than with DMA-BUF off:
  • Idle, app + web process: ~1% vs ~8–9%.
  • seq 1 1000000 drains in 2.3 s vs 2.8 s.
  • I don't know why.
• Using the new binary with no env overrides: the web process gets __NV_DISABLE_EXPLICIT_SYNC=1 and DMA-BUF stays on. GDK_BACKEND=x11 and a user-set WEBKIT_DISABLE_DMABUF_RENDERER=1 still turn DMA-BUF off.
• New unit tests for the decision logic and for GDK_BACKEND parsing.
• cargo fmt, cargo clippy --all-targets, cargo build, cargo test for linux_gfx (8/8) and pnpm run build all pass.

#313 is still needed for X11 + NVIDIA, for all AppImage users, and for anyone who sets WEBKIT_DISABLE_DMABUF_RENDERER=1 themselves.

🤖 Generated with Claude Code