diff --git a/README.md b/README.md
index 9eca6a9..bce5d23 100644
--- a/README.md
+++ b/README.md
@@ -14,7 +14,7 @@ python -m pytest
 python scripts/run_all.py
 ```
 
-Python 3.11+. SQLite is still the stdlib path for the main lab. DuckDB now runs in CI on that compact fixture.
+Python 3.11+. SQLite is the stdlib path and the tests that do not need DuckDB run after `pip install -e .`. The DuckDB comparisons need the extra; they are skipped if it is not installed. CI installs `.[duckdb]`.
 
 ## Query design
 
diff --git a/tests/test_cross_engine_parity.py b/tests/test_cross_engine_parity.py
index 8f71bd3..eb74ab2 100644
--- a/tests/test_cross_engine_parity.py
+++ b/tests/test_cross_engine_parity.py
@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import pandas as pd
+import pytest
 
 from sqlfeat.cross_engine import (
     duckdb_point_in_time_features,
@@ -12,6 +13,17 @@
     sqlite_rank_results,
 )
 
+try:
+    import duckdb as _duckdb  # noqa: F401
+except ImportError:
+    _HAS_DUCKDB = False
+else:
+    _HAS_DUCKDB = True
+
+_duckdb_only = pytest.mark.skipif(
+    not _HAS_DUCKDB, reason="install sqlfeat[duckdb] for the second-engine checks"
+)
+
 
 def _normalise(df: pd.DataFrame) -> pd.DataFrame:
     out = df.copy().sort_values("prediction_id").reset_index(drop=True)
@@ -22,17 +34,21 @@ def _normalise(df: pd.DataFrame) -> pd.DataFrame:
     return out
 
 
-def test_sqlite_duckdb_and_pandas_agree_on_point_in_time_features() -> None:
+def test_sqlite_and_pandas_agree_on_point_in_time_features() -> None:
     sqlite = _normalise(sqlite_point_in_time_features())
-    duckdb = _normalise(duckdb_point_in_time_features())
     expected = _normalise(independent_expected_features())
-
     pd.testing.assert_frame_equal(sqlite, expected, check_dtype=False, atol=1e-12, rtol=0)
-    pd.testing.assert_frame_equal(duckdb, expected, check_dtype=False, atol=1e-12, rtol=0)
+
+
+@_duckdb_only
+def test_duckdb_agrees_with_sqlite_on_point_in_time_features() -> None:
+    sqlite = _normalise(sqlite_point_in_time_features())
+    duckdb = _normalise(duckdb_point_in_time_features())
+    pd.testing.assert_frame_equal(duckdb, sqlite, check_dtype=False, atol=1e-12, rtol=0)
 
 
 def test_future_sentinel_is_excluded_without_using_model_performance() -> None:
-    result = _normalise(duckdb_point_in_time_features())
+    result = _normalise(sqlite_point_in_time_features())
     row = result.loc[result["prediction_id"] == "p1"].iloc[0]
     assert row["txn_count"] == 2
     assert row["spend"] == 30.0
@@ -48,13 +64,17 @@ def test_no_history_preserves_zero_aggregates_and_null_recency() -> None:
     assert pd.isna(row["recency_days"])
 
 
-def test_window_rank_semantics_match_across_engines() -> None:
+def test_sqlite_rank_semantics() -> None:
     sqlite = sqlite_rank_results().sort_values("id").reset_index(drop=True)
-    duckdb = duckdb_rank_results().sort_values("id").reset_index(drop=True)
-    pd.testing.assert_frame_equal(sqlite, duckdb, check_dtype=False)
-
     # Tied score 9 gets rank 1 in both rows; the next ordinary rank is 3,
     # while dense rank advances to 2. ROW_NUMBER is made deterministic by id.
     assert list(sqlite["rank_value"]) == [1, 1, 3, 4]
     assert list(sqlite["dense_rank_value"]) == [1, 1, 2, 3]
     assert list(sqlite["row_number_value"]) == [1, 2, 3, 4]
+
+
+@_duckdb_only
+def test_window_rank_semantics_match_across_engines() -> None:
+    sqlite = sqlite_rank_results().sort_values("id").reset_index(drop=True)
+    duckdb = duckdb_rank_results().sort_values("id").reset_index(drop=True)
+    pd.testing.assert_frame_equal(sqlite, duckdb, check_dtype=False)