Keep root attributes when appending mask parameters

`to_zarr(mode="a")` replaces the store's root attributes with those of
the appended dataset. `ds_props` has none, so every store lost
`map_image_id_to_filename` and the rest of ethology's and the script's
metadata. Copy the stored attributes onto `ds_props` before appending,
in a new `append_mask_props` shared by `main` and the round-trip test,
which now checks that the attributes survive.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: sfmig

scripts/generate_masks_from_bboxes.py

@@ -350,6 +350,23 @@ def compute_mask_props(labels_zarr, ds_bboxes):
     )
 
 
+def append_mask_props(path_to_zarr, ds_props):
+    """Append the data variables of ``ds_props`` to the store.
+
+    Only the data variables are appended: the coordinates are already in
+    the store, and xarray checks that rewritten variables keep their stored
+    dtype. `space` is <U1 here, but reads back from the store as a
+    variable-length string dtype (StringDType in recent xarray/zarr), so
+    appending it would raise a dtype mismatch.
+
+    ``to_zarr(mode="a")`` replaces the store's root attributes with those of
+    the appended dataset, so the stored attributes are copied onto it first.
+    """
+    ds_props = ds_props.drop_vars(list(ds_props.coords))
+    ds_props.attrs = dict(zarr.open_group(path_to_zarr).attrs)
+    ds_props.to_zarr(path_to_zarr, mode="a")
+
+
 def predict_masks_across_images(
     model,
     image_array,
@@ -536,12 +553,7 @@ def main(args):
 
     # Add mask and ellipse parameters per bbox, measured on the stored masks
     ds_props = compute_mask_props(labels_zarr, ds_bboxes)
-    # append only the data variables: the coordinates are already in the
-    # store, and xarray checks that rewritten variables keep their stored
-    # dtype. `space` is <U1 here, but reads back from the store as a
-    # variable-length string dtype (StringDType in recent xarray/zarr),
-    # so appending it would raise a dtype mismatch
-    ds_props.drop_vars(list(ds_props.coords)).to_zarr(output_zarr, mode="a")
+    append_mask_props(output_zarr, ds_props)
 
     print(f"Saved bboxes, masks and mask parameters to {output_zarr}")
 

tests/test_unit/test_generate_masks_from_bboxes.py

@@ -8,6 +8,7 @@
 
 from scripts.generate_masks_from_bboxes import (
     ImageArrayLazy,
+    append_mask_props,
     compute_mask_props,
     create_dataset_store,
     flatten_masks_smallest_wins,
@@ -88,15 +89,20 @@ def test_compute_mask_props(labels, ds_bboxes):
 def test_store_round_trip(tmp_path, ds_bboxes, labels):
     """Run the three passes of main, without SAM2."""
     store = tmp_path / "bboxes_and_masks.zarr"
-    labels_zarr = create_dataset_store(store, ds_bboxes, (100, 120), {})
+    labels_zarr = create_dataset_store(
+        store, ds_bboxes, (100, 120), {"occlusion_policy": "smallest_wins"}
+    )
     labels_zarr[:] = labels  # what predict_masks_across_images does
     ds_props = compute_mask_props(labels_zarr, ds_bboxes)
-    ds_props.drop_vars(list(ds_props.coords)).to_zarr(store, mode="a")
+    append_mask_props(store, ds_props)
 
     ds = xr.open_zarr(store)
     mask = ds.labels.sel(image_id=7) == ds.label_of.sel(id=2)
     assert int(mask.sum()) == ds.mask_area.sel(image_id=7, id=2)
     np.testing.assert_array_equal(ds.labels.values, labels)
+    # root attributes survive the append
+    assert ds.attrs["map_image_id_to_filename"] == {"5": "a.png", "7": "b.png"}
+    assert ds.attrs["occlusion_policy"] == "smallest_wins"
 
 
 def test_create_dataset_store(tmp_path, ds_bboxes):