[ieivanov]

Summary

Two additions for a step that segments the plate it reads and writes its labels into that plate's label images (Biohub/biahub#413 moves cellpose segmentation out of biahub track into biahub segment, which does exactly that and needs to resume after a preemption like every other step):

• process_single_position(..., output_label=<name>) writes the transform's result into the label image labels/<name> of the output position instead of its image. Everything else is unchanged: the same shard-aligned write units, torn-shard repair and --resume markers (kept under the label's name beside the image's), the same extra_metadata on the position. A label image is TZYX, so the transform returns one channel per timepoint (dropped on the way in) and a single input channel group feeds it; output_channel_indices is ignored. The label image must exist; a missing one, several channel groups, or a multi-channel result are refused with a message.
• create_empty_labels(store_path, position_keys, label_names, ...) is the label-image half of create_empty_plate for a plate that already has its positions: it scaffolds labels/<name> without restating the plate's channels, shape or scale (each label takes its own position's image geometry, as create_empty_plate(label_names=...) already does for existing positions) and records the step's provenance on each label group through label_metadata. The label-creation and layout-validation code is shared with create_empty_plate.

Test plan

• tests/ngff/test_ngff_utils.py: test_process_single_position_writes_a_label_image_of_the_input (serial and 2 workers; values, image untouched, resume skips, markers under labels/<name>), test_process_single_position_label_output_refusals, test_create_empty_labels_in_existing_positions (0.4 and 0.5; explicit layout, transforms, existing label untouched, metadata on the group, second call a no-op), test_create_empty_labels_refusals.
• biahub segment --init / biahub track --init on this branch create their label images with create_empty_labels, and biahub segment fills them with --resume (biahub#413's tests, and an end-to-end run on a real plate).
• Note: while running the full file, hypothesis found a pre-existing failure in test_apply_transform_to_tczyx_and_save that reproduces on main (c9832a7) with the same example (channel chunks of 2, shards_ratio=(2, 1, 1, 2, 2), a slice over all 3 channels → RuntimeError: incompatible offset): apply_transform_to_tczyx_and_save does not refuse a channel-sharded store the way process_single_position does. Not touched here.

🤖 Generated with Claude Code