- Built a small benchmark matrix to compare CPU and GPU pipelines on the same video.
- Added --row so each path can run by itself for debugging and stable testing.
- Switched the benchmark to a more realistic preview-style workload instead of full-size-only transforms.
- Added PNG dump support so rows can save sample output frames for visual verification.
- The custom Vulkan path now does real preview scaling and alpha overlay, and can stay fully on-GPU for Vulkan decode.
- Added runtime detection for VAAPI, CUDA, and Vulkan, so unsupported rows skip cleanly on different systems.
- Added CUDA benchmark rows for NVIDIA systems, while keeping VAAPI and Vulkan conditional at runtime.
Jonathan Thomas