feat(*): retune scoring incentives and add finishing reps

Game/scripts/training_mode.gd

@@ -18,11 +18,22 @@ extends GameMode
 # "EVAL_RESULT {...}" line is printed before quitting.
 
 @export var episode_length_seconds := 30.0
-@export var goal_reward := 10.0
+# Run01-vs-run02 eval (see training/eval_history.json) came back 87.5% draws:
+# with a 30s episode, the dense per-tick terms on ShipAIController can sum to
+# several times this value before it was raised, so scoring and forfeiting
+# the rest of the episode's farmable reward was worse than never finishing.
+# Raised well above that ceiling so a real scoring chance always beats
+# continuing to farm dense reward for however long is left in the episode.
+@export var goal_reward := 40.0
 
 # Episode-start state mix; remaining probability = fully random state.
 @export_range(0.0, 1.0) var kickoff_state_chance := 0.2
-@export_range(0.0, 1.0) var ball_near_goal_chance := 0.2
+# Raised from 0.2: fixing the reward incentive to score (see goal_reward,
+# ball_touch_reward, time_penalty on ShipAIController) only helps if the
+# policy also gets enough reps at actually finishing. At 0.2 that scenario
+# was 1 in 5 episode starts; most training time was spent in generic
+# midfield play where a finish never comes up.
+@export_range(0.0, 1.0) var ball_near_goal_chance := 0.35
 
 # Placement bounds for randomized episode starts, derived from the standard
 # enclosure (ArenaBoundary). The inset keeps a randomly oriented ship (1x1x4