[tayllatheodoro]

Closes #458.

scripts/debug_track.py stopped working when #424 removed resolve_z_slice, and it had other problems too:

• Import: resolve_z_slice no longer exists, so the script failed with an ImportError at startup.
• Config keys: TrackingSettings forbids unknown keys, so the old z_range and mode would have been rejected.
• Function key: biahub.cli.track.mem_nuc_contour isn't in CUSTOM_FUNCTIONS. The same stale key was also in settings/example_track_settings.yml.

Changes

• Import from biahub.tracking (refactor(track): move the tracking library code into biahub/tracking #454) and call run_ultrack with keyword arguments (detection=, edges=, scale=).
• Replace the old keys with segmentation_method: foreground_contour and output_mode.
• Projection: track no longer selects Z-planes, so the script projects the FOV in memory with the biahub project API (no CLI, no output store).
  • find_focus() wraps _find_focus_per_t, and project_fov() calls resolve_z_indices / project_zyx per channel.
  • The defaults use the same window as nextflow/configs/a549/project.yml: a mean over 60 planes starting 20 below the Phase3D focus, found per timepoint.
  • Focus store: a virtual-stain store usually has no Phase3D, so the new focus_store_path lets the focus channel come from another plate. Without it, the script gives a clear error.
  • Like biahub project: every input_images entry with a path is projected, and the output is float32 when any channel uses mean/sum.
  • A store that's already projected (Z=1) is tracked as is.
• Tracks layer: in 3D the napari tracks layer now includes the z column and uses a (1, *scale) scale, so tracks line up with the (T, Z, Y, X) layers.
• Match the blank-frames CSV using the B/1/000000 FOV form, the same way track does.
• Fix the mem_nuc_contour key in the script and in example_track_settings.yml, plus some leftover typos. napari is now only imported when the viewer opens.
• Update the note about the script in the track-with-biahub skill (references/evaluation.md).

Testing

On synthetic plates (T=4, Z=70; nuclei_prediction, membrane_prediction, Phase3D):

• Projection:
  • With the virtual-stain channels in one store and Phase3D in another, the projection gives the same values as biahub project (project(..., cluster="debug")) on the combined plate. Without a focus store, the missing Phase3D raises a clear error.
  • The whole-stack mean fallback (no focus) equals mean over Z.
• Tracking: the script's pipeline produces 4 tracks over 4 timepoints from the 4 planted cells (2D).
• Viewer: the napari layers were built headless for 2D and 3D. The tracks layer has the image's dimensionality, with scale (1, *scale). napari.run() itself wasn't run.

Note: the script uses _find_focus_per_t and _FLOAT_METHODS, which are private to biahub/project.py, so it will need updating if they change.