SIENTIAPDE-2201: fix: keep the timestamp column in train/test data CSVs

generate_report wrote train_data.csv and test_data.csv with index=False, which
dropped the UTC time index. laborious drift reads test_data.csv from MLflow as
its reference and needs a timestamp column; without it, every model trained here
drifted against the first 30% of its own target window. Both CSVs now write the
index as timestamp. If a column already has that name, the CSV is written as
before and a warning is logged, so training never fails over drift.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/utils/repository/test_data_manager_repository.py

@@ -514,6 +514,58 @@ def test_generate_report_adds_target_alias_for_reports(tmp_path):
         assert test_csv['target'].equals(test_csv['t'])
 
 
+_UTC_INDEX = pd.DatetimeIndex(['2024-01-01 02:00', '2024-01-01 01:00'], tz='UTC')
+
+
+def _dated_result(**train_columns: list[float]) -> TrainModelResult:
+    """Frames indexed by UTC time, like prepare_training_data + predict produce them."""
+    columns = train_columns or {'v1': [1.0, 2.0], 't': [1.0, 2.0]}
+    return TrainModelResult(
+        params=TrainModelParams.from_dict(_minimal_dict_for_prepare()),
+        train_data=pd.DataFrame(columns, index=_UTC_INDEX),
+        val_data=pd.DataFrame(columns, index=_UTC_INDEX),
+        y_train_pred=pd.DataFrame({'t': [1.0, 2.0]}, index=_UTC_INDEX),
+        y_pred=pd.DataFrame({'t': [1.0, 2.0]}, index=_UTC_INDEX),
+        run_name='testrun',
+    )
+
+
+def _generate_report(repo, tmr: TrainModelResult, tmp_path) -> TrainModelResult:
+    with (
+        patch.object(repo, '_get_reports_directory', return_value=str(tmp_path)),
+        patch('model_manager.utils.repository.data_manager_repository.Reports'),
+    ):
+        return repo.generate_report(tmr, {})
+
+
+def test_generate_report_csvs_keep_the_time_index_as_timestamp(tmp_path):
+    """laborious drift reads test_data.csv as its reference and needs ``timestamp``."""
+    repo = dmr.DataManagerRepository(MagicMock())
+
+    out = _generate_report(repo, _dated_result(), tmp_path)
+
+    for csv_path in (out.train_data_path, out.test_data_path):
+        csv = pd.read_csv(csv_path)
+        assert list(csv.columns) == ['timestamp', 'prediction', 'v1', 't', 'target']
+        assert pd.to_datetime(csv['timestamp']).tolist() == list(_UTC_INDEX)
+
+
+def test_generate_report_csvs_skip_timestamp_when_a_column_already_has_the_name(tmp_path):
+    """A variable literally named ``timestamp`` wins; training must not fail over drift."""
+    repo = dmr.DataManagerRepository(MagicMock())
+    tmr = _dated_result(timestamp=[5.0, 6.0], v1=[1.0, 2.0], t=[1.0, 2.0])
+
+    with patch.object(repo, 'warning') as warning:
+        out = _generate_report(repo, tmr, tmp_path)
+
+    for csv_path in (out.train_data_path, out.test_data_path):
+        csv = pd.read_csv(csv_path)
+        assert list(csv.columns) == ['prediction', 'timestamp', 'v1', 't', 'target']
+        assert csv['timestamp'].tolist() == [5.0, 6.0]
+    assert warning.call_count == 2
+    assert "Column 'timestamp' already exists" in warning.call_args.args[0]
+
+
 def test_generate_report_skips_equation_file_when_not_linear(tmp_path):
     repo = dmr.DataManagerRepository(MagicMock())
     p = TrainModelParams.from_dict(_minimal_dict_for_prepare())