diff --git a/pretab/core/selectors.py b/pretab/core/selectors.py
index bf172e4..d828a61 100644
--- a/pretab/core/selectors.py
+++ b/pretab/core/selectors.py
@@ -133,16 +133,29 @@ def _trim_over_max(self, points: list[float], context: object, max_count: int) -
         raise NotImplementedError
 
     def _enforce_spacing(self, split_points: list[float], x: np.ndarray) -> list[float]:
-        """Drop locations closer than ``min_location_spacing`` of the range."""
+        """Drop locations closer than ``min_location_spacing`` of the range.
+
+        The distance test runs against *every* location kept so far rather than
+        only the most recent one, which makes the filter independent of the
+        order in which candidates arrive. That matters because subclasses
+        deliberately order their candidates differently -- location order for a
+        single tree, gain-descending order for a boosted ensemble -- and an
+        order-sensitive test silently dropped every candidate positioned below
+        the previously kept one, collapsing a gain-ranked set into a small
+        clustered subsequence.
+
+        For already-ascending input (the single-tree path) this is equivalent to
+        comparing against the last kept location, so that behaviour is unchanged.
+        """
         if len(split_points) <= 1:
             return split_points
 
         x_range = float(x.max() - x.min())
         min_distance = self.min_location_spacing * x_range
 
-        spaced = [split_points[0]]
+        spaced: list[float] = [split_points[0]]
         for point in split_points[1:]:
-            if point - spaced[-1] >= min_distance:
+            if all(abs(point - kept) >= min_distance for kept in spaced):
                 spaced.append(point)
 
         return spaced
diff --git a/tests/test_location_selectors.py b/tests/test_location_selectors.py
index 15fe2e3..1a3e96e 100644
--- a/tests/test_location_selectors.py
+++ b/tests/test_location_selectors.py
@@ -119,3 +119,76 @@ def test_lightgbm_matches_knot_adapter(data):
         X, y, task="regression", min_count=adapter.min_knots, max_count=adapter.max_knots
     )
     np.testing.assert_array_equal(from_adapter, from_selector)
+
+
+# --------------------------------------------------------------------------- #
+# Spacing must not depend on the order candidates arrive in.
+#
+# ``_ordered_candidates`` returns location order for a single tree but
+# gain-descending order for a boosted ensemble, so an order-sensitive spacing
+# filter discarded every candidate sitting below the previously kept one.
+# --------------------------------------------------------------------------- #
+def test_enforce_spacing_is_order_independent():
+    x = np.linspace(0, 10, 500).reshape(-1, 1)
+    selector = CARTLocationSelector(min_location_spacing=0.01)
+    ascending = [1.0, 2.0, 3.0, 8.0, 9.0]
+
+    from_ascending = selector._enforce_spacing(list(ascending), x)
+    from_shuffled = selector._enforce_spacing([9.0, 1.0, 8.0, 3.0, 2.0], x)
+
+    assert sorted(from_shuffled) == sorted(from_ascending) == ascending
+
+
+def test_enforce_spacing_still_drops_close_neighbours():
+    x = np.linspace(0, 10, 500).reshape(-1, 1)  # min_distance = 0.1
+    selector = CARTLocationSelector(min_location_spacing=0.01)
+
+    assert selector._enforce_spacing([1.0, 1.01, 5.0], x) == [1.0, 5.0]
+
+
+def test_enforce_spacing_matches_legacy_result_on_ascending_input():
+    # The single-tree path always supplies ascending candidates; comparing
+    # against every kept location must be equivalent there.
+    rng = np.random.RandomState(0)
+    x = rng.uniform(0, 10, size=(500, 1))
+    points = sorted(rng.uniform(0, 10, size=40).tolist())
+    selector = CARTLocationSelector(min_location_spacing=0.05)
+
+    min_distance = 0.05 * float(x.max() - x.min())
+    legacy = [points[0]]
+    for point in points[1:]:
+        if point - legacy[-1] >= min_distance:
+            legacy.append(point)
+
+    assert selector._enforce_spacing(points, x) == legacy
+
+
+def test_lightgbm_locations_cover_the_feature_range(data):
+    pytest.importorskip("lightgbm")
+    X, y = data
+    x_min, x_max = float(X.min()), float(X.max())
+
+    lgbm = LightGBMLocationSelector(n_estimators=30).select(
+        X, y, task="regression", min_count=3, max_count=8
+    )
+    cart = CARTLocationSelector().select(X, y, task="regression", min_count=3, max_count=8)
+
+    def span(locations):
+        return (locations.max() - locations.min()) / (x_max - x_min)
+
+    # Previously the gain-ordered candidates collapsed into a narrow cluster;
+    # coverage should now be comparable to the single-tree selector.
+    assert span(lgbm) > 0.5 * span(cart)
+
+
+def test_lightgbm_keeps_the_highest_gain_locations(data):
+    pytest.importorskip("lightgbm")
+    X, y = data
+    selector = LightGBMLocationSelector(n_estimators=30)
+
+    candidates, _ = selector._ordered_candidates(X, y, "regression")
+    spaced = selector._enforce_spacing(candidates, X)
+    # Gain-descending order survives the spacing filter, so the trim keeps the
+    # top-ranked entries rather than whichever happen to be numerically lowest.
+    assert spaced == [c for c in candidates if c in set(spaced)]
+    assert selector._trim_over_max(spaced, None, 5) == spaced[:5]