# 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).