Add check for badly formatted vcz.

Just looks for the dimensions of the call_genotype array, but that's most likely to be wrong (e.g. 2d by mistake)

tests/test_variantdata.py

@@ -670,6 +670,14 @@ def test_sgkit_ancestor(small_sd_fixture, tmp_path):
 
 
 class TestVariantDataErrors:
+    def test_bad_zarr_spec(self):
+        ds = zarr.group()
+        ds["call_genotype"] = zarr.array(np.zeros(10, dtype=np.int8))
+        with pytest.raises(
+            ValueError, match="Expecting a VCF Zarr object with 3D call_genotype array"
+        ):
+            tsinfer.VariantData(ds, np.zeros(10, dtype="<U1"))
+
     def test_missing_phase(self, tmp_path):
         path = tmp_path / "data.zarr"
         ds = sgkit.simulate_genotype_call_dataset(n_variant=3, n_sample=3)

tsinfer/formats.py

@@ -2311,6 +2311,11 @@ def __init__(
                 # Assumed to be a VCF Zarr hierarchy
                 self.path = None
                 self.data = path_or_zarr
+            else:
+                raise ValueError(
+                    "Expecting a VCF Zarr object with 3D call_genotype array: "
+                    "see https://github.com/sgkit-dev/vcf-zarr-spec/"
+                )
         except AttributeError:
             self.path = path_or_zarr
             self.data = zarr.open(path_or_zarr, mode="r")