Squashed from work: - video: hardware first-frame decode straight from RAM — no temp files - video: the encoder transform is reported once, not once per encoder - video: a single still does not need a whole AVAssetWriter
