Align Continue labels with price plan outcomes

training/trader_training/state_continue_experiment.py

@@ -380,8 +380,9 @@ def _state_rows_for_age(frame: pd.DataFrame, stop_bps: float, target_bps: float,
     target_price = np.where(long_mask, entry_price * (1.0 + target_bps / 10000.0), entry_price * (1.0 - target_bps / 10000.0))
     distance_to_stop = np.where(long_mask, (current_price / stop_price - 1.0) * 10000.0, (stop_price / current_price - 1.0) * 10000.0)
     distance_to_target = np.where(long_mask, (target_price / current_price - 1.0) * 10000.0, (current_price / target_price - 1.0) * 10000.0)
-    expected_edge = frame["future_return_bps"].astype(float) - cost_bps
-    continue_target = ((expected_edge >= min_continue_edge_bps) & (frame["mae_bps"].astype(float) < stop_bps)).astype("int8")
+    # Continue must score the first price-plan outcome from the current state, not the raw horizon close.
+    expected_edge = frame["gross_edge_bps"].astype(float) - cost_bps
+    continue_target = ((expected_edge >= min_continue_edge_bps) & (frame["stop_hit"].astype(int) == 0)).astype("int8")
 
     out = frame[
         [
@@ -601,6 +602,8 @@ def _source_manifest(
             "uses_same_round_model_prediction_as_feature": False,
             "entry_predicted_edge_bps": "baseline frozen ENTRY ONNX output selected by side",
             "entry_direction_prob": "baseline frozen DIRECTION ONNX output selected by side",
+            "expected_continue_edge_bps": "price-plan gross edge minus cost; target/stop/timeout outcome is respected",
+            "continue_target": "expected_continue_edge_bps >= threshold and stop is not the first path barrier",
             "path_features": "position path shape and side-adjusted market pressure at current state time",
             "add_count": "synthetic first-position diagnostic, fixed to 0",
             "minutes_since_last_add": "synthetic first-position diagnostic, fixed to 9999",