Weak-GPU hardware (Raspberry Pi 4 / v3dv, Haswell / hasvk) exposes a
real Vulkan adapter, so the old binary wgpu-or-software detection handed
it the slowest possible path: femtovg-wgpu over an immature Vulkan
driver. Mesa's GL path is the fast one on that class of hardware.
- Three-tier chain: wgpu -> femtovg GL -> software. Auto-detect
classifies adapters (weak-driver markers incl. hasvk, integrated-on-arm,
tiny texture limits); adapters only reachable through wgpu's GL backend
never select the wgpu tier (the renderer masks Backends::GL out, so
detection must too). renderer-femtovg (GL) feature now compiled on
Linux.
- wgpu tier requests PowerPreference::LowPower (WGPU_POWER_PREF still
wins).
- Per-window swapchain alpha in the vendored femtovg renderer: only the
miniplayer gets a transparent (blended) swapchain; the main window
keeps an Opaque one, sparing a full-window compositor blend per frame.
- Settings > Appearance gains a Renderer row (Linux only): Auto
(default) / GPU / GPU compatibility / Software, persisted in
ui_prefs.json, applied on restart. QBZ_RENDERER env still overrides
(now also accepts gl|gles|femtovg).
- Auto-revert sentinel: a non-auto choice is armed before backend init
and disarmed after the window survives 5s; if the previous start died
before that, the setting reverts to Auto with an explanatory toast, so
a bad renderer pick can never lock the user out of Settings.
- Developer tools > Diagnostics now reports the active renderer, why it
was chosen, and the enumerated GPU adapters (panel + markdown report).
cc