Decode the dependencies before selecting the clip

A demuxer separates encoded packets by stream. A decoder consumes those packets under its API contract and can produce zero, one, or several frames in response; it can also buffer frames before output. FFmpeg 7.1 documents send/receive behavior and end-of-stream draining. PyAV v14.2.0's decode path iterates over packet decoding results and includes flushing packets at the end. An implementation that assumes exactly one image for each packet is therefore structurally wrong even before considering sampling policy.

Keep compressed-byte I/O, CPU decode, color conversion, resizing, and model transfer as separate measured stages. A slow extraction can be caused by remote seeks or repeated reference decoding rather than by the model. Hardware and software decoders can have different support and numerical image outputs, so record the selected path and relevant color/range/orientation handling. A successful frame count does not prove that the pixel interpretation or timing matches the intended dataset.

PyAV's pinned seek documentation describes offsets in stream timebase units when a stream is supplied, otherwise in the container's defined unit, followed by decoding toward the desired position. The seek result is not a promise that the next image has exactly the requested PTS. Start from an appropriate random-access entry point supported by the format and codec, rebuild the required state, and select decoded frames by the stated presentation-time rule.

GOP structure matters because some pictures reference others outside the requested presentation interval. A validated closed independent segment is different from an arbitrary I-picture or keyframe flag in an open dependency structure. Do not claim that every such flag guarantees independent extraction across all codecs. Test the actual source class, seek behavior, and boundary artifacts. When the target is near the end, drain delayed decoder output according to its API rather than assuming the final packet has already yielded every relevant frame.

Use a small source with known reordered pictures, a long interval between independent entry points, and a clip beginning between them. Compare random-access extraction with a complete sequential decode using the same frame-selection rule and pixel oracle. Record actual selected PTS values and whether delayed frames were drained. Repeated extraction should also expose leaked decoder state after a seek, unsupported streams, and corrupt-input handling instead of returning silently shortened clips.

The cache lab isolates repeated working-set access as one reason locality can matter. Reset, set both C=4,s=4, predict twelve hits in sixteen initial requests, then Play or Step. Reset and set both C=3,s=4 for zero initial hits. The stream caps at 4,096 requests and edits retain time. It contains no codec, GOP, seek, decode dependencies, or pixel conversion; its automatic LRU eviction does not imply real reference pictures can be discarded without correctness or recomputation costs.