🎉 (plaid-check) add a Global shape check for dataset consistency

Author: fabiencasenave

src/plaid/cli/plaidcheck.py

@@ -524,6 +524,10 @@ def check_dataset(
     target_splits = target_splits & dataset_splits
 
     checksum_report = {}
+    # Track shape of each Global feature across all checked splits/samples to
+    # detect inconsistencies (e.g. a Global stored as a scalar in one sample
+    # and as a vector in another).
+    global_shape_observations: dict[str, dict[tuple, list[str]]] = {}
     for split in sorted(target_splits):
         dataset = datasetdict[split]
         converter = converterdict[split]
@@ -595,6 +599,19 @@ def check_dataset(
                         issue,
                     )
 
+                # Record the observed shape of this Global so we can later
+                # detect dimension mismatches across all checked samples
+                # (across splits).
+                if value is not None:
+                    try:
+                        shape = tuple(np.asarray(value).shape)
+                    except Exception:
+                        shape = None
+                    if shape is not None:
+                        global_shape_observations.setdefault(
+                            global_name, {}
+                        ).setdefault(shape, []).append(f"{split}[{idx}]")
+
             for time in sample.get_all_time_values():
                 local_bases = sample.get_base_names(time=time)
                 for base in local_bases:
@@ -625,6 +642,25 @@ def check_dataset(
                                         issue,
                                     )
 
+    # Report Globals whose dimension/shape is not consistent across all
+    # checked samples (across splits).
+    for global_name, shape_to_locations in global_shape_observations.items():
+        if len(shape_to_locations) <= 1:
+            continue
+        details = "; ".join(
+            f"shape={shape} at {locations[:5]}"
+            + (f" (+{len(locations) - 5} more)" if len(locations) > 5 else "")
+            for shape, locations in sorted(
+                shape_to_locations.items(), key=lambda kv: str(kv[0])
+            )
+        )
+        report.add(
+            "error",
+            "GLOBAL_SHAPE_MISMATCH",
+            f"global/{global_name}",
+            f"Global '{global_name}' has inconsistent shapes across samples: {details}",
+        )
+
     # Compare checksums from every checked sample to flag identical sample data.
     checksum_values = list(checksum_report.values())
     if len(checksum_report) != len(np.unique(checksum_values)):

tests/cli/test_plaidcheck.py

@@ -633,6 +633,43 @@ def test_check_dataset_split_and_data_warnings_and_duplicates(
     assert any(msg.code == "DUPLICATED_DATA" for msg in report.messages)
 
 
+def test_check_dataset_reports_global_shape_mismatch(
+    tmp_path: Path, monkeypatch
+) -> None:
+    """Globals with inconsistent shapes across samples should be reported."""
+    dataset = _make_minimal_layout(tmp_path)
+
+    monkeypatch.setattr(
+        plaidcheck,
+        "load_infos_from_disk",
+        lambda path: _infos({"train": 2}),  # noqa: ARG005
+    )
+    monkeypatch.setattr(
+        plaidcheck,
+        "load_metadata_from_disk",
+        lambda path: ({"train": {}}, {"Var": {}}, {"train": {}}, None),  # noqa: ARG005
+    )
+    samples = [
+        _FakeSampleForCheck(global_value=np.array(1.0)),
+        _FakeSampleForCheck(global_value=np.array([1.0, 2.0])),
+    ]
+    monkeypatch.setattr(
+        plaidcheck,
+        "init_from_disk",
+        lambda path: (  # noqa: ARG005
+            {"train": _FakeDataset(2)},
+            {"train": _FakeConverter(samples)},
+        ),
+    )
+
+    report = check_dataset(dataset, splits=["train"])
+
+    mismatch_msgs = [m for m in report.messages if m.code == "GLOBAL_SHAPE_MISMATCH"]
+    assert len(mismatch_msgs) == 1
+    assert mismatch_msgs[0].severity == "error"
+    assert "G" in mismatch_msgs[0].location
+
+
 def test_check_dataset_sample_conversion_error(tmp_path: Path, monkeypatch) -> None:
     """Checker should emit conversion errors when converter fails on an index."""
     dataset = _make_minimal_layout(tmp_path)