Handle matrix sequence conversion runtime errors

src/pyrecest/smoothers/abstract_smoother.py

@@ -58,7 +58,7 @@ def _normalize_matrix_sequence(  # pylint: disable=too-many-return-statements,to
 
         try:
             values_arr = asarray(values)
-        except (TypeError, ValueError):
+        except (TypeError, ValueError, RuntimeError):
             values_arr = None
         if values_arr is not None:
             if ndim(values_arr) == 0:

tests/smoothers/test_abstract_smoother_matrix_sequence_runtime_error.py

@@ -0,0 +1,30 @@
+import numpy as np
+
+import pyrecest.smoothers.abstract_smoother as abstract_smoother_module
+from pyrecest.smoothers.abstract_smoother import AbstractSmoother
+
+
+def test_matrix_sequence_falls_back_after_backend_runtime_error(monkeypatch):
+    matrices = [np.eye(2), 2.0 * np.eye(2)]
+    backend_asarray = abstract_smoother_module.asarray
+
+    def asarray_with_outer_conversion_failure(value):
+        if value is matrices:
+            raise RuntimeError("backend cannot stack the matrix sequence")
+        return backend_asarray(value)
+
+    monkeypatch.setattr(
+        abstract_smoother_module,
+        "asarray",
+        asarray_with_outer_conversion_failure,
+    )
+
+    normalized = AbstractSmoother._normalize_matrix_sequence(
+        matrices,
+        length=2,
+        name="transition_matrices",
+        matrix_dim=2,
+    )
+
+    np.testing.assert_allclose(normalized[0], matrices[0])
+    np.testing.assert_allclose(normalized[1], matrices[1])