tifffile LZW TIFFs need imagecodecs; decimating high-rate real data gives known-truth tests

Two notes from validating an image-analysis tool on public datasets.

1. tifffile.imread raises ValueError("<COMPRESSION.LZW: 5> requires the imagecodecs package") on LZW-compressed TIFFs. Mixed datasets often have older uncompressed frames and newer LZW ones, so a run can silently skip a whole subset (here, every newer class). Install imagecodecs up front, and record per-file errors instead of dropping them, then check which classes the failures fall in.

2. To test how a video method degrades at low frame rate with real (not synthetic) truth: take a high-rate public recording, establish the answer at full rate (ideally cross-checked by an independent sensor sampled faster), then keep every k-th frame and rerun with frozen settings. Score silent errors (wrong with no warning) separately from flagged ones. This exposes aliasing failures that synthetic data built by the same author can hide.

Also: a multiprocessing.Pool whose workers exit unexpectedly can leave the parent blocked forever with zombie children and no traceback. Independent shard processes, each writing its own output file with try/except per item, are more robust for long batch evaluations.