fix(vectors): clear error for an out-of-range layer in normalize_alpha / dose

src/steerbench/vectors.py

@@ -123,6 +123,10 @@ def normalize_alpha(
 
     Returns all three units. See :func:`dose` for the scalar full dose.
     """
+    if layer not in vector.directions:
+        raise ValueError(
+            f"layer {layer} not in steering vector; available layers: {sorted(vector.directions)}"
+        )
     norm = float(torch.linalg.vector_norm(vector.directions[layer]))
     by_vector = alpha * norm
     by_residual = by_vector / residual_norm if residual_norm is not None else None

tests/test_vectors.py

@@ -212,3 +212,21 @@ def test_gguf_superset_loads_in_repeng(tmp_path: Path) -> None:
     assert hint == "gpt2"
     names = {t.name for t in reader.tensors}
     assert names == {"direction.5", "direction.6"}
+
+
+def test_normalize_alpha_out_of_range_layer_names_available() -> None:
+    vec = SteeringVector(directions={3: torch.tensor([1.0, 0.0]), 7: torch.tensor([0.0, 2.0])})
+    with pytest.raises(ValueError, match=r"layer 999 not in steering vector.*\[3, 7\]"):
+        normalize_alpha(vec, layer=999, alpha=1.0)
+
+
+def test_dose_out_of_range_layer_names_available() -> None:
+    vec = SteeringVector(directions={3: torch.tensor([1.0, 0.0]), 7: torch.tensor([0.0, 2.0])})
+    with pytest.raises(ValueError, match=r"layer 999 not in steering vector.*\[3, 7\]"):
+        dose(vec, layer=999, alpha=1.0, residual_norm=10.0)
+
+
+def test_normalize_alpha_valid_layer_unchanged_by_guard() -> None:
+    vec = SteeringVector(directions={3: torch.tensor([3.0, 4.0])})  # norm 5
+    got = normalize_alpha(vec, layer=3, alpha=2.0)
+    assert got.by_vector_norm == pytest.approx(10.0)