test(nphd): add regression tests for NPHD metric persistence

Add comprehensive regression tests to catch future metric persistence bugs:

- test_load_preserves_nphd_metric(): Verifies that load() preserves the custom
  NPHD metric instead of falling back to standard Hamming distance
- test_view_preserves_nphd_metric(): Same test for the view() method

Both tests create vectors with known differences, verify correct NPHD distances
before save, then validate identical distances after load/view. This ensures
the custom metric is properly restored and not overwritten by usearch's default.

Covers the critical persistence bug where searches on loaded indexes were
returning incorrect results.

Author: titusz

tests/test_nphd_load.py

@@ -89,3 +89,40 @@ def test_loaded_index_supports_operations(tmp_path):
 
     matches = loaded.search(vector1, count=1)
     assert matches.keys[0] == 1
+
+
+def test_load_preserves_nphd_metric(tmp_path):
+    """Regression test: Load preserves custom NPHD metric instead of using standard Hamming."""
+    # Create index with two similar vectors (differ by 1 bit)
+    index = NphdIndex(max_dim=256)
+    vector1 = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.uint8)
+    vector2 = np.array([1, 2, 3, 4, 5, 6, 7, 9], dtype=np.uint8)  # Last byte differs by 1 bit
+    index.add([100, 200], [vector1, vector2])
+
+    # Search before save - should use NPHD metric
+    results_before = index.search(vector1, count=2)
+    distances_before = results_before.distances
+
+    # The NPHD distance should be small (around 0.015625 for 1 bit difference in 64-bit vector)
+    # If using standard Hamming, it would be 1.0
+    assert distances_before[0] == 0.0  # Exact match
+    assert distances_before[1] < 0.1  # Small NPHD distance, not 1.0
+
+    # Save and load
+    file_path = tmp_path / "index.usearch"
+    index.save(str(file_path))
+
+    loaded = NphdIndex(max_dim=256)
+    loaded.load(str(file_path))
+
+    # Search after load - MUST use same NPHD metric
+    results_after = loaded.search(vector1, count=2)
+    distances_after = results_after.distances
+
+    # Critical: Distances must match (proves NPHD metric is preserved)
+    np.testing.assert_array_almost_equal(
+        distances_before,
+        distances_after,
+        decimal=6,
+        err_msg="NPHD metric was not preserved after load() - distances changed",
+    )

tests/test_nphd_view.py

@@ -85,3 +85,40 @@ def test_viewed_index_supports_search(tmp_path):
 
     matches = viewed.search(vector, count=1)
     assert matches.keys[0] == 1
+
+
+def test_view_preserves_nphd_metric(tmp_path):
+    """Regression test: View preserves custom NPHD metric instead of using standard Hamming."""
+    # Create index with two similar vectors (differ by 1 bit)
+    index = NphdIndex(max_dim=256)
+    vector1 = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.uint8)
+    vector2 = np.array([1, 2, 3, 4, 5, 6, 7, 9], dtype=np.uint8)  # Last byte differs by 1 bit
+    index.add([100, 200], [vector1, vector2])
+
+    # Search before save - should use NPHD metric
+    results_before = index.search(vector1, count=2)
+    distances_before = results_before.distances
+
+    # The NPHD distance should be small (around 0.015625 for 1 bit difference in 64-bit vector)
+    # If using standard Hamming, it would be 1.0
+    assert distances_before[0] == 0.0  # Exact match
+    assert distances_before[1] < 0.1  # Small NPHD distance, not 1.0
+
+    # Save and view
+    file_path = tmp_path / "index.usearch"
+    index.save(str(file_path))
+
+    viewed = NphdIndex(max_dim=256)
+    viewed.view(str(file_path))
+
+    # Search after view - MUST use same NPHD metric
+    results_after = viewed.search(vector1, count=2)
+    distances_after = results_after.distances
+
+    # Critical: Distances must match (proves NPHD metric is preserved)
+    np.testing.assert_array_almost_equal(
+        distances_before,
+        distances_after,
+        decimal=6,
+        err_msg="NPHD metric was not preserved after view() - distances changed",
+    )