Each recording is drawn as the tapestry's speaker lane: the moderator track, the participant track, each clip in the turn's shipped scene colour (the keyframe pass, as the pipeline computes it today). On top, at the chosen opacity and clipped to the clips, run the scene stripes: one column per sample, the mean colour of what was on screen at that moment. The source is either the keyframes the pipeline already decodes, or every frame at 4 fps, which is one sample per pixel at the tapestry's maximum zoom. The floor hides a stripe whose colour is within that ΔE of its clip's colour, so a steady scene stays flat and only change shows. The raw row above each lane is the same data at full strength, the movie-barcode view, for reference. Notches above the lane mark a scene change: the mean colour of a window before the moment against the window after it, above the edge threshold, measured on the untoned colours as the pipeline does.
Worst cases to look for: a uniform web session (should look like the lane does today); a scrolled or animated share (shimmer: try the median window); a gallery call where people move (camera drift: the floor should hide it); the 4K screen capture (its keyframes are already a second apart, so the keyframe source should be enough there).