chore(training): add air_approach_weight and restart stage-5 intercepts

Stage 5 blocked all three attempts on productive_air_touch_fraction
stuck exactly at 0.0 across a continuous 180M-step lineage, while
goal_rate/upright_fraction/forward_motion_fraction kept improving on
the same budget. forward_velocity_to_ball_weight (the term that solved
Stage 4's ground pursuit) is hard-gated below GROUND_HANDLING_HEIGHT
and does nothing in the air, so Stage 5's air_intercept_chance had no
matching aerial incentive to learn from. air_approach_weight adds the
airborne mirror (nose-first 3D closing speed, no uprightness
multiplier) and folds into HANDLING_REWARD_FLAGS so Stage 6 inherits
it too. Deleted the three blocked attempts and reset state to resume
Stage 5 from the Stage-4 checkpoint with the new term.

Game/scripts/ship_ai_controller.gd

@@ -46,6 +46,20 @@ extends AIController3D
 # frozen checkpoints keep their original objective; generation 5 handling
 # turns it on while reducing the orientation-agnostic term.
 @export var forward_velocity_to_ball_weight := 0.0
+# Aerial mirror of forward_velocity_to_ball_weight: nose-first closing speed
+# on the ball, active above GROUND_HANDLING_HEIGHT instead of below it (the
+# two are mutually exclusive by altitude, never both active on the same
+# tick). Generation 5's intercepts stage added air_intercept_chance without
+# an airborne equivalent of the term that actually solved ground handling;
+# above 3m the only remaining approach incentive was the generic, orientation
+# -agnostic velocity_to_ball_weight (0.02-0.04), which three consecutive
+# 60M-step attempts (180M cumulative, resuming each time) showed produces
+# zero learnable gradient toward touching an aerial ball at all —
+# productive_air_touch_fraction stayed exactly 0.0 the whole time while every
+# other metric kept improving on the same budget. Uses the full 3D nose
+# vector rather than the planar-only one, since a real aerial requires
+# pitching away from level.
+@export var air_approach_weight := 0.0
 @export var ball_velocity_to_goal_weight := 0.004
 # Per-tick penalty scaled by distance to the ball (full value at the arena's
 # far diagonal, 0 on top of the ball). Run04 lesson: with idling worth a flat
@@ -357,6 +371,21 @@ func _physics_process(delta):
 			reward += forward_velocity_to_ball_weight * forward_speed * facing_ball \
 				* approach_uprightness * handling_ground_factor
 
+	# Aerial shaping: nose-first 3D closing speed on the ball (see
+	# air_approach_weight). Mirrors the ground block above but with the full
+	# nose vector instead of the planar one, and no uprightness multiplier —
+	# a genuine aerial approach requires pitching away from level, so paying
+	# only while upright would oppose the exact behaviour this rewards.
+	if air_approach_weight > 0.0 and ship.global_position.y >= GROUND_HANDLING_HEIGHT \
+			and to_ball.length_squared() > 0.0001:
+		var nose_forward := -ship.global_transform.basis.z
+		if nose_forward.length_squared() > 0.0001:
+			nose_forward = nose_forward.normalized()
+			var to_ball_dir := to_ball.normalized()
+			var air_facing_ball: float = maxf(nose_forward.dot(to_ball_dir), 0.0)
+			var air_closing_speed: float = maxf(ship.linear_velocity.dot(to_ball_dir), 0.0) / ship.max_speed
+			reward += air_approach_weight * air_closing_speed * air_facing_ball
+
 	# Dense penalty: distance to the ball, so idling far away bleeds reward
 	# instead of scoring a safe zero (see ball_distance_penalty).
 	if ball_distance_penalty > 0.0: