frames/ — frame cache for steps 2 and 3 (companion repo GMLRVigil/BenchCheck-Pool-Frames, public)

Layout after unpacking: frames/pool_frames/<video_id>/frame_<k>.jpg (JPEG, long side 448 px) plus frames.json (grid positions, keys, mid_k = the single frame used by step 2, w, h). One directory per normalized video referenced by a step-1 item. 16 tar shards frames_<0-f>.tar, sharded by the first hex character of video_id; frames_manifest.json lists sha256 and counts. coverage.json gives the snapshot coverage: videos that were not normalized on the origin cluster at snapshot time have video_id "k:..." in the items table and no frames here; steps 2/3 skip them (they are finished on the origin cluster later).