GPU correctness fixes and optimizations

draw/backdrop.odin

@@ -684,7 +684,13 @@ upload_backdrop_primitives :: proc(device: ^sdl.GPUDevice, pass: ^sdl.GPUCopyPas
 		sdl.GPUBufferUsageFlags{.GRAPHICS_STORAGE_READ},
 	)
 
-	prim_array := sdl.MapGPUTransferBuffer(device, GLOB.backdrop.primitive_buffer.transfer, false)
+	// cycle=true: this is a persistent per-frame streaming transfer buffer. The previous
+	// frame's UploadToGPUBuffer is almost certainly still in flight when we map here
+	// (allowedFramesInFlight defaults to 2 on Metal). Without cycling, the CPU memcpy below
+	// races the GPU's blit read on the same MTLBuffer.contents. Cycling rebinds the
+	// container's active internal buffer to an unbound one (or allocates a new one) — O(1)
+	// in steady state, no fence wait. See SDL_gpu_metal.m's METAL_INTERNAL_PrepareBufferForWrite.
+	prim_array := sdl.MapGPUTransferBuffer(device, GLOB.backdrop.primitive_buffer.transfer, true)
 	if prim_array == nil {
 		log.panicf("Failed to map backdrop primitive transfer buffer: %s", sdl.GetError())
 	}