Fix flaky doublet test (#3436)

tests/external/test_hashsolo.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import numpy as np
+import pandas as pd
 from anndata import AnnData
 
 import scanpy.external as sce
@@ -27,9 +28,11 @@ def test_cell_demultiplexing():
     sce.pp.hashsolo(test_data, test_data.obs.columns)
 
     doublets = ["Doublet"] * 10
-    classes = list(
-        np.repeat(np.arange(10), 98).reshape(98, 10, order="F").ravel().astype(str)
-    )
+    classes = np.repeat(np.arange(10), 98).reshape(98, 10, order="F").ravel().tolist()
     negatives = ["Negative"] * 10
-    classification = doublets + classes + negatives
-    assert test_data.obs["Classification"].astype(str).tolist() == classification
+    expected = pd.array(doublets + classes + negatives, dtype="string")
+    classification = test_data.obs["Classification"].array.astype("string")
+    # This is a bit flaky, so allow some mismatches:
+    if (expected != classification).sum() > 3:
+        # Compare lists for better error message
+        assert classification.tolist() == expected.tolist()