The GPU compositor can present directly to a window, no readback
Floati has had a real GPU compositor for a while - it blends every layer of a page on the graphics card, byte-for-byte identical to the software path. But it was dark: after compositing on the GPU it copied the whole frame back across the bus to hand the shell a PNG, and that copy erased the entire speed advantage. This release builds the missing half - a present path that draws the composited frame straight to a texture the window owns, so the frame never leaves the GPU. A full-screen pass reads the composited pixels and writes them to the swapchain; because the target is a plain 8-bit-per-channel format, the round trip is a bit-exact copy, so the surface shows precisely what the readback path returned. That is proven, not hoped: an offscreen version of the exact present path is read back and compared to the software compositor across opaque, translucent, opacity-tie, edge-clipped, odd-width and 60 random layer stacks - all byte-identical. A SurfaceRenderer drives it from any window handle and degrades to software where there is no GPU. GPU moves 2 to 4; overall holds at 52 (the live view is wired next).