migrate layer schema from float seconds to integer frame numbers

fix frame timing: use round() to match JS Math.round for frame range calc

fix audio: drop PyAV audio mux, use ffmpeg stream copy from source

reduce ZeroGPU duration to 10s

revert to 7.2s build (pre-multiplied cache, pinned memory, threaded audio)

optimize: fp16, torch.compile blend, double-buffered decode, batch=64, torch.no_grad

optimize: pre-multiplied static composite cache, parallel layer render, pinned memory, threaded audio

optimize: static UV cache, threaded encoder, batch=32, animated warm-start

full GPU pipeline: batched Newton solve + grid_sample on CUDA, precomputed pixel grid

match JS bilinear quad warp (lerpQuad) instead of perspective transform

replace kornia with cv2.warpPerspective — exact same logic as original render_editor_video.py

fix text wrap: use font.getbbox() to match original render_editor_video.py exactly

fix homography: use get_perspective_transform(src,dst) + align_corners=True to match OpenCV exactly

Fix warp_perspective homography direction: dst→src for backward warp

Fix homography direction: src=layer corners, dst=frame corners

Fix layer tensors not on DEVICE inside GPU fn, fix src_pts off-by-one

Fix frame normalization: divide by 255 after torchcodec decode

libx264 ultrafast encode, remove PyNvVideoCodec

Add PyNvVideoCodec GPU encoder with libx264 fallback

Fix audio muxing: use codec name instead of unsupported template= kwarg

Replace decord with torchcodec for native CUDA frame decode

Batched homography per layer + h264_nvenc encoder with libx264 fallback

Add application file

add frame-compositor space

initial commit