[bug] Keep cache keys stable across processes on Python 3.14

ccflow/examples/tpch/data_generators.py

@@ -103,10 +103,10 @@ def get_table(self, table: TPCHTable) -> pl.DataFrame:
         return pl.from_arrow(tbl_arrow)
 
     def get_answer(self, query_id: int) -> pa.Table:
-        row = self._conn.query(f"SELECT answer FROM tpch_answers() WHERE scale_factor={self.scale_factor} AND query_nr={query_id}").fetchone()
-        if not row:
+        rows = self._conn.query(f"SELECT answer FROM tpch_answers() WHERE scale_factor={self.scale_factor} AND query_nr={query_id}").fetchall()
+        if not rows:
             raise ValueError(f"No TPC-H answer found for scale_factor={self.scale_factor}, query_nr={query_id}")
-        return pc.read_csv(io.BytesIO(row[0].encode("utf-8")), parse_options=pc.ParseOptions(delimiter="|"))
+        return pc.read_csv(io.BytesIO(rows[0][0].encode("utf-8")), parse_options=pc.ParseOptions(delimiter="|"))
 
 
 class TPCHTableProvider(CallableModel):

ccflow/tests/utils/test_tokenize.py

@@ -265,6 +265,25 @@ def f(self):
 
         assert compute_behavior_token(A) == compute_behavior_token(B)
 
+    def test_skips_annotate_methods(self):
+        """PEP 649 __annotate__/__annotate_func__ are skipped (injected here to run pre-3.14)."""
+
+        def _annotate(fmt):
+            return {"x": int}
+
+        class A:
+            __annotate__ = _annotate
+            __annotate_func__ = _annotate
+
+            def f(self):
+                return 1
+
+        class B:
+            def f(self):
+                return 1
+
+        assert compute_behavior_token(A) == compute_behavior_token(B)
+
 
 def _helper_add(x):
     return x + 1

ccflow/utils/tokenize.py

@@ -433,6 +433,11 @@ def compute_cache_token(*, data_values: Iterable[Any] = (), behavior_classes: It
         "__get_pydantic_core_schema__",
         "__get_pydantic_json_schema__",
         "model_post_init",
+        # PEP 649: the auto-generated __annotate__ closure is PYTHONHASHSEED-sensitive,
+        # which would make the behavior token differ between processes. Annotations are
+        # already captured through other paths, so skipping these is lossless.
+        "__annotate__",
+        "__annotate_func__",
     }
 )