fix: prevent floating-point truncation in spatial location indexing

src/deepwave/location_interpolation.py

@@ -68,8 +68,15 @@ def _get_hicks_for_one_location_dim(
             - location
             + int(location)
         )
-        locations = (location + x).long()
-
+        locations = (
+            torch.arange(
+                -halfwidth + 1,
+                halfwidth + 1,
+                device=beta.device,
+            )
+            + int(location)
+        )
+
         if key in hicks_weight_cache:
             weights = hicks_weight_cache[key].clone()
         else:

tests/test_location_interpolation.py

@@ -389,3 +389,30 @@ def test_get_hicks_for_one_location_dim_invalid_n_grid_points() -> None:
             [-0.5, 19.5],
             -1,
         )
+
+
+@pytest.mark.parametrize(
+    "location",
+    [3.22, 7.35, 15.2, 31.4, 63.16, 127.23],
+)
+def test_hicks_locations_not_truncated_by_float32(location, halfwidth=2):
+    """Grid indices must not be truncated by float32 round-off.
+
+    The old formula ``(location + x).long()`` can yield 1.999... for the
+    leftmost cell, which ``.long()`` truncates (e.g. 1 instead of 2 at
+    location 3.01). Indices must be integer arithmetic around ``int(location)``.
+    """
+    betas = [0.0, 1.84, 3.04, 4.14, 5.26, 6.40, 7.51, 8.56, 9.56, 10.64]
+    beta = torch.tensor(betas[halfwidth - 1], dtype=torch.float32)
+    locs, _ = _get_hicks_for_one_location_dim(
+        {},
+        location,
+        halfwidth,
+        beta,
+        [False, False],
+        [-0.5, 199.5],
+        200,
+        True,
+    )
+    expected = torch.arange(-halfwidth + 1, halfwidth + 1) + int(location)
+    assert torch.equal(locs.cpu(), expected)