Fix 0-row result when re-executing a cached prepared statement scanning a python object

src_cpp/include/pyarrow/pyarrow_scan.h

@@ -26,6 +26,11 @@ struct PyArrowTableScanSharedState final : public function::TableFuncSharedState
         : TableFuncSharedState{numRows}, chunks{std::move(chunks)}, currentChunk{0} {}
 
     ArrowArrayWrapper* getNextChunk();
+
+    // TableFunctionCall::copy() (used by the physical-plan cache on prepared
+    // statement re-execution) shares the same sharedState instance, so the
+    // chunk cursor must be rewound here or a re-executed scan yields 0 rows.
+    void resetState() override { currentChunk = 0; }
 };
 
 struct PyArrowTableScanFunctionData final : public function::TableFuncBindData {

test/test_scan_pandas_pyarrow.py

@@ -141,6 +141,18 @@ def test_pyarrow_primitive(conn_db_empty: ConnDB) -> None:
             pyarrow_test_helper(establish_connection, sf, thread)
 
 
+def test_pyarrow_scan_repeated_execution(conn_db_in_mem: ConnDB) -> None:
+    # Regression test: re-executing the same (implicitly cached) prepared
+    # statement that scans a python object returned 0 rows on the second run.
+    # The cached physical plan shares the scan's shared state, whose chunk
+    # cursor was never rewound between executions.
+    conn, _ = conn_db_in_mem
+    df = pd.DataFrame({"a": pd.Series([1, 2, 3, None], dtype="int32[pyarrow]")})
+    for _ in range(3):
+        result = conn.execute("LOAD FROM df RETURN count(*)")
+        assert result.get_next()[0] == 4
+
+
 def test_pyarrow_time(conn_db_readonly: ConnDB) -> None:
     conn, _ = conn_db_readonly
     col1 = pa.array([1000123, 2000123, 3000123], type=pa.duration("s"))