fix(gamecli): wrap angle in reward to prevent unbounded spin penalty

GameCli/Reward.cs

@@ -3,8 +3,10 @@ namespace GameCli;
 /// <summary>
 /// Reward = -distance_to_target
 ///        - λ_v · speed
-///        - λ_θ · |angle|
+///        - λ_θ · |wrap(angle)|
 ///        - λ_fuel · engine_on
+/// The angle is wrapped to [-π, π] so a ship that has completed N revolutions
+/// gets the same upright-penalty as a physically identical pose that hasn't.
 /// Coefficients are public so training scripts can log/inspect them.
 /// </summary>
 public static class Reward
@@ -19,11 +21,12 @@ public static class Reward
         float dy = target.Y - s.Y;
         float distance = MathF.Sqrt(dx * dx + dy * dy);
         float speed    = MathF.Sqrt(s.VX * s.VX + s.VY * s.VY);
+        float wrapped  = MathF.Atan2(MathF.Sin(s.Angle), MathF.Cos(s.Angle));
         float engineOn = action == Physics.ActionNoop ? 0f : 1f;
 
         return -distance
              - LambdaVelocity * speed
-             - LambdaAngle    * MathF.Abs(s.Angle)
+             - LambdaAngle    * MathF.Abs(wrapped)
              - LambdaFuel     * engineOn;
     }
 }