chore(training): retune stage-4 handling penalties and restart from Stage-3 foundation

Stage 4's upright/forward-motion telemetry plateaued flat across all three
blocked attempts because ground_tilt_penalty (0.003) was too weak to matter
and nothing penalized sideways/reverse motion at all. Raise
ground_tilt_penalty to 0.05 and add a new non_forward_penalty term
(ship_ai_controller.gd) that directly costs non-forward planar velocity near
the floor, independent of the ball. Delete the three blocked attempts'
checkpoints/logs/exports and reset generation5_state.json so the next run
starts fresh from the Stage-3 foundation checkpoint instead of continuing
from the drifted retry2 weights.

training/train.py

@@ -358,6 +358,11 @@ def parse_args():
         "--ground-tilt-penalty", type=float, default=None,
         help="Low-altitude-only tilt cost that fades to zero by the handling-height threshold",
     )
+    curriculum.add_argument(
+        "--non-forward-penalty", type=float, default=None,
+        help="Overrides ShipAIController.non_forward_penalty (low-altitude dense cost on sideways/reverse "
+        "planar velocity, independent of the ball)",
+    )
     curriculum.add_argument(
         "--speed-reward-weight", type=float, default=None,
         help="Overrides the orientation-agnostic own-speed reward (generation 5 handling sets it to zero)",
@@ -391,6 +396,7 @@ def _curriculum_kwargs(args) -> dict:
         "team_size": args.team_size,
         "ai_tilt_penalty": args.tilt_penalty,
         "ai_ground_tilt_penalty": args.ground_tilt_penalty,
+        "ai_non_forward_penalty": args.non_forward_penalty,
         "ai_velocity_to_ball_weight": args.velocity_to_ball_weight,
         "ai_forward_velocity_to_ball_weight": args.forward_velocity_to_ball_weight,
         "ai_ball_distance_penalty": args.ball_distance_penalty,