v0.0.3: full implementation of background-thread mode (start/stop/get_latest_frame), paced CFR iterator (frames(fps=N), exact 1/fps slot timing), and zero-copy GPU handle (grab_gpu returning a shared NT handle on a BGRA D3D11 texture). Plus a real benchmark (flip_demo + mover.py per zentape DDA_BENCHMARK_NOTES.md) that shows the gap honestly: on realistic moving content rustcam captures 93+ unique fps with 100% consecutive-changed while bettercam and dxcam stall at <1 fps.