fix(labels): defer float-dtype rejection to spatialdata's model validation

src/spatialdata_plot/pl/render.py

@@ -2189,12 +2189,9 @@ def _render_labels(
             is_label=True,
         )
 
-    if np.issubdtype(label.dtype, np.floating):
-        raise ValueError(
-            f"Label element '{element}' has dtype {label.dtype}. Label arrays must use an "
-            f"integer dtype (e.g. int32 or uint16). Cast before plotting, e.g.:\n"
-            f"    sdata['{element}'] = sdata['{element}'].astype('int32')"
-        )
+    # Label dtype is validated upstream: spatialdata rejects non-integer label rasters at the model
+    # boundary (parse / SpatialData construction / __setitem__), so a validly built element always
+    # reaches here with an integer dtype. No local guard needed (see #606, resolved upstream).
 
     # rasterize spatial image if necessary to speed up performance
     if rasterize:

tests/pl/test_render_labels.py

@@ -710,23 +710,6 @@ def test_render_labels_lognorm_with_zeros_does_not_crash(sdata_blobs: SpatialDat
     plt.close(fig)
 
 
-@pytest.mark.parametrize("dtype", [np.float16, np.float32, np.float64])
-def test_render_labels_rejects_float_dtype(dtype):
-    # Regression test for #606: float-dtype labels must raise a clear
-    # ValueError naming the element and dtype, not a cryptic skimage TypeError.
-    arr = np.zeros((20, 20), dtype=dtype)
-    arr[3:8, 3:8] = 1
-    arr[12:17, 12:17] = 2
-    sdata = SpatialData(labels={"lbl": Labels2DModel.parse(arr, dims=["y", "x"])})
-
-    fig, ax = plt.subplots()
-    try:
-        with pytest.raises(ValueError, match=r"Label element 'lbl'.*integer dtype"):
-            sdata.pl.render_labels("lbl").pl.show(ax=ax)
-    finally:
-        plt.close(fig)
-
-
 def test_render_labels_rejects_background_instance_id_in_table():
     # Regression test for #607: table row with instance_id=0 (background)
     # used to crash with obnscure error.

tests/pl/test_utils.py

@@ -628,12 +628,20 @@ def test_unmatched_instance_ids_warn_and_write_nan(self, sdata_blobs: SpatialDat
             measure_obs(sdata_blobs, "blobs_labels")
         assert np.isnan(table.obsm["spatial"]).all()
 
-    def test_float_dtype_labels_supported(self, sdata_blobs: SpatialData) -> None:
-        # #3: a float-typed (but integer-valued) labels raster must not crash np.bincount.
-        arr = np.asarray(sdata_blobs["blobs_labels"].data).astype(np.float32)
-        sd = _labels_sdata(arr)
-        measure_obs(sd, "lab", table_name="t")
-        assert np.isfinite(sd["t"].obsm["spatial"]).all()
+    def test_float_dtype_labels_handled_by_centroid_stats(self, sdata_blobs: SpatialData) -> None:
+        # #3: an integer-valued but float-typed raster must be cast to int, not crash np.bincount.
+        # spatialdata now rejects float labels at the model boundary, so this can no longer reach
+        # `measure_obs` through a SpatialData; the cast lives in `_stream_label_centroid_stats`, so
+        # exercise it there: a float raster must yield exactly what its integer counterpart yields.
+        from spatialdata_plot.pl.utils import _stream_label_centroid_stats
+
+        arr_int = np.asarray(sdata_blobs["blobs_labels"].data).astype(np.int64)
+        lbl_i, x_i, y_i, area_i = _stream_label_centroid_stats(arr_int)
+        lbl_f, x_f, y_f, area_f = _stream_label_centroid_stats(arr_int.astype(np.float32))
+        np.testing.assert_array_equal(lbl_f, lbl_i)
+        np.testing.assert_allclose(x_f, x_i)
+        np.testing.assert_allclose(y_f, y_i)
+        np.testing.assert_array_equal(area_f, area_i)
 
     def test_existing_nonnumeric_column_raises_before_any_write(self, sdata_blobs: SpatialData) -> None:
         # #4: a non-numeric collision raises BEFORE obsm is mutated (no half-written table).