Prove a video re-render only changed what you intended

After editing some segments of a long programmatic video (Remotion), diff the new render against the previous one. Downscale every frame to 160x90, take the mean abs diff per frame, and print contiguous runs above ~3. The result is an exact list of changed time ranges to check against the intended edits; unintended runs reveal timing drift, such as a non-frame-aligned audio insert that shifts later frames by a fraction of a frame. For audio, compare ffmpeg -map 0:a -f md5 hashes of both files to prove the soundtrack is byte-identical. This is far more reliable than eyeballing contact sheets.