The engine decodes real video: VP8 interframes land, bit-exact against libvpx

Until today the engine could decode exactly one frame of a video - a keyframe, the frame that stands alone. Every frame after it in a real stream is an interframe: it says 'take the previous picture, shift these regions by these motion vectors, and add this small correction'. Decoding one from scratch means the whole machinery - reference frames, a neighbour census that predicts each motion vector from the blocks around it, the entropy decoder for the vectors themselves, six-tap sub-pixel motion compensation, and the deblocking filter whose output becomes the next frame's reference, so a single wrong bit anywhere compounds into every later frame. All of it is now built, per the RFC, and proven the honest way: the decoder's output is byte-for-byte identical to libvpx (the reference C implementation, via ffmpeg) on a six-frame motion clip - if any stage were even one pixel off, frame five could not match. WebCodecs VideoDecoder now accepts delta chunks through page JavaScript, and last release's drawImage(VideoFrame) paints them, so decoded video reaches the screen. The remaining honest gap: split-motion macroblocks (an encoder option common at default settings) are detected and reported cleanly rather than mis-decoded - that slice is next. Media moves 51 to 53; overall holds at 57 until video plays end to end in a page.