feat(training): add generation 5 curriculum

Game/scripts/ai_ship_controller.gd

@@ -34,6 +34,11 @@ var _policy: PolicyNetwork
 var _action := ShipAction.new()
 var _ticks_until_decision := 0
 
+# League mode revisits a small model pool every episode. Policies are
+# immutable after load, so share parsed networks by path instead of reading
+# and flattening the same 100KB+ JSON on every reset for every frozen ship.
+static var _policy_cache: Dictionary = {}
+
 var _ship: Ship
 var _teammates: Array[Ship] = []
 var _opponents: Array[Ship] = []
@@ -44,7 +49,24 @@ var _scene_refs_ready := false
 
 func _ready():
 	if not model_path.is_empty():
+		load_policy(model_path)
+
+
+# League training swaps a frozen opponent's policy between episodes without
+# despawning the ship. Reset the held action/decision cadence with the model
+# so no command from the previous opponent leaks into the next episode.
+func load_policy(path: String) -> void:
+	model_path = path
+	if model_path.is_empty():
+		_policy = null
+	elif _policy_cache.has(model_path):
+		_policy = _policy_cache[model_path]
+	else:
 		_policy = PolicyNetwork.load_from_file(model_path)
+		if _policy != null:
+			_policy_cache[model_path] = _policy
+	_action = ShipAction.new()
+	_ticks_until_decision = 0
 
 
 func get_action() -> ShipAction: