feat(multitask)!: seed SpotOptim warm start from warm_start_lags list

src/spotforecast2/model_selection/spotoptim_search.py

@@ -877,7 +877,7 @@ def build_warm_start_x0(
             ``SpotOptimStrategy.prepare_forecaster``).
         forecaster: The pre-tuning forecaster; its ``estimator`` supplies the
             starting values for the numeric hyperparameter dimensions.
-        lags_seed: The lag configuration to seed (e.g. ``config.lags_consider``).
+        lags_seed: The lag configuration to seed (e.g. ``config.warm_start_lags``).
 
     Returns:
         A 1-D float array of length ``len(var_name)``, or ``None`` when the

src/spotforecast2/multitask/strategies.py

@@ -241,9 +241,10 @@ def prepare_forecaster(
             task: A `BaseTask` (or compatible) instance that supplies ``cv_ts``,
                 ``config``, ``logger``, ``save_tuning_results``, and
                 ``create_forecaster``.  The config must expose
-                ``n_trials_spotoptim``, ``n_initial_spotoptim``,
-                ``random_state``, ``warm_start_lags``, and optionally
-                ``lags_consider`` and the TensorBoard knobs described
+                ``n_trials_spotoptim``, ``n_initial_spotoptim``, and
+                ``random_state``; optionally ``warm_start_lags`` (the seed
+                lag set, sf2-safe >= 22.0.0; ``None``/empty = cold start),
+                ``max_time_spotoptim``, and the TensorBoard knobs described
                 above.
             target: Target column name; forwarded to ``task.create_forecaster``
                 and ``task.save_tuning_results``.
@@ -288,7 +289,7 @@ def prepare_forecaster(
                 n_trials_spotoptim=5,
                 n_initial_spotoptim=3,
                 random_state=0,
-                warm_start_lags=False,
+                warm_start_lags=None,
             )
             task = types.SimpleNamespace(
                 config=cfg,
@@ -318,15 +319,15 @@ def prepare_forecaster(
         search_space = self.search_space or _default_spotoptim_search_space()
         cv = task.cv_ts(y_train)
 
-        # Warm start: inject ``lags_consider`` as a candidate lag set and seed
-        # the optimizer's first evaluation with it.  Only dict search spaces
-        # with a ``"lags"`` list are eligible; anything else falls through to a
-        # normal cold-start run.
+        # Warm start: ``config.warm_start_lags`` (sf2-safe >= 22.0.0) is the
+        # seed lag set itself — it is injected as a search-space candidate and
+        # seeds the optimizer's first evaluation. ``None``/empty disables the
+        # warm start.  Only dict search spaces with a ``"lags"`` list are
+        # eligible; anything else falls through to a normal cold-start run.
         kwargs_spotoptim: Dict[str, Any] = {}
-        lags_seed = getattr(task.config, "lags_consider", None)
+        lags_seed = getattr(task.config, "warm_start_lags", None)
         if (
-            getattr(task.config, "warm_start_lags", False)
-            and lags_seed
+            lags_seed
             and isinstance(search_space, dict)
             and isinstance(search_space.get("lags"), list)
         ):

tests/test_multitask_strategies.py

@@ -75,7 +75,7 @@ def _make_fake_task(**config_extra):
         n_trials_spotoptim=2,
         n_initial_spotoptim=1,
         random_state=0,
-        warm_start_lags=False,
+        warm_start_lags=None,
         **config_extra,
     )
     return types.SimpleNamespace(

tests/test_spotoptim_warm_start.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2026 bartzbeielstein
 # SPDX-License-Identifier: AGPL-3.0-or-later
 
-"""Tests for warm-starting the SpotOptim search with ``lags_consider``.
+"""Tests for warm-starting the SpotOptim search with ``warm_start_lags``.
 
 Covers the ``build_warm_start_x0`` helper (full-dim seed point, clipping,
 graceful ``None`` cases) and an end-to-end check that the seeded lag
@@ -112,8 +112,7 @@ def test_warm_start_x0_round_trips_to_seed():
 
 
 class _MockConfig:
-    warm_start_lags = True
-    lags_consider = [1, 2, 24]
+    warm_start_lags = [1, 2, 24]
     random_state = 1
     n_trials_spotoptim = 4
     n_initial_spotoptim = 3
@@ -141,7 +140,7 @@ def create_forecaster(self, target):
 
 
 def test_strategy_injects_seed_and_forwards_x0(monkeypatch):
-    """The strategy injects lags_consider as a candidate and forwards x0."""
+    """The strategy injects warm_start_lags as a candidate and forwards x0."""
     from spotforecast2.multitask.strategies import SpotOptimStrategy
 
     captured = {}
@@ -186,7 +185,7 @@ def fake_search(**kwargs):
 
 
 def test_strategy_no_x0_when_flag_disabled(monkeypatch):
-    """With warm_start_lags False, no x0 is passed and lags is untouched."""
+    """With warm_start_lags None, no x0 is passed and lags is untouched."""
     from spotforecast2.multitask.strategies import SpotOptimStrategy
 
     captured = {}
@@ -207,7 +206,7 @@ def fake_search(**kwargs):
     )
 
     task = _MockTask()
-    task.config.warm_start_lags = False
+    task.config.warm_start_lags = None
     y = pd.Series(
         np.sin(np.arange(400) * 2 * np.pi / 24),
         index=pd.date_range("2022-01-01", periods=400, freq="h"),