feat(training): add airborne_penalty and a stage-6 "unmask" curriculum run

Stage 5 (aggression) passed (41-47 vs grounded curric-s2-defend, within
the lenient gate but not yet a clear win). Rather than keep the locomotion
mask on indefinitely, stage 6 reopens full 3D controls on top of the
aggression retune and pairs it with a new dense airborne_penalty (scaled
by height above the floor) so the policy learns to prefer staying grounded
through incentives instead of a hard mask — same regime shift that
regressed stage 3, but this time with a mitigation and ~12x the training
time (~240M timesteps / ~24h vs ~20M / ~2h) to actually re-converge
instead of stalling mid-shift.

airborne_penalty follows the existing SHIP_AI_OVERRIDES pattern: default
0 (off) on ship_ai_controller.gd, exposed via train.py's new
--airborne-penalty flag, added to training_mode.gd's allow-list. Also adds
a per-stage timesteps override in curriculum.py (STAGES[n]["timesteps"])
since this is the first stage to need a different budget than the rest.

Game/scripts/ship_ai_controller.gd

@@ -70,6 +70,14 @@ extends AIController3D
 # makes running the clock out strictly worse than scoring as soon as a
 # chance appears, instead of a free way to keep collecting dense reward.
 @export var time_penalty := 0.001
+# Per-tick penalty scaled by height above the floor (0 on the floor, full
+# value at the arena's ceiling) — distinct from the locomotion mask, which
+# only discards *thrust*-driven vertical/pitch-roll input; a masked ship can
+# still be launched airborne by collisions (ball impacts, ship-vs-ship
+# knockback, the wall/ceiling surface-pull field), and nothing previously
+# penalized time spent up there. Default 0 (off) so ordinary runs are
+# unaffected; the floor-lock curriculum stage turns it on.
+@export var airborne_penalty := 0.0
 
 # Locomotion curriculum: when false, the corresponding action axes are
 # discarded in set_action before reaching the ship, so the ship stays
@@ -196,6 +204,14 @@ func _physics_process(delta):
 		var uprightness: float = ship.global_transform.basis.y.dot(Vector3.UP)
 		reward -= tilt_penalty * (1.0 - uprightness) * 0.5
 
+	# Dense penalty: height above the floor (see airborne_penalty). The
+	# floor sits at world y = 0 (see training_mode.gd's FIELD_MIN_Y/
+	# _escaped bounds); normalized so the worst case is pinned at the
+	# ceiling.
+	if airborne_penalty > 0.0:
+		var height := maxf(ship.global_position.y, 0.0)
+		reward -= airborne_penalty * height / ArenaBoundary.INNER_HEIGHT
+
 
 func _wall_or_ceiling_contact() -> bool:
 	var state := PhysicsServer3D.body_get_direct_state(ship.get_rid())

Game/scripts/training_mode.gd

@@ -188,7 +188,7 @@ const SHIP_AI_OVERRIDES := [
 	"ball_touch_reward", "ball_touch_cooldown_ticks", "ball_touch_direction_floor",
 	"velocity_to_ball_weight", "ball_velocity_to_goal_weight", "ball_distance_penalty",
 	"wall_contact_penalty", "tilt_penalty", "speed_reward_weight", "time_penalty",
-	"allow_vertical", "allow_pitch_roll",
+	"airborne_penalty", "allow_vertical", "allow_pitch_roll",
 ]
 
 
@@ -238,6 +238,7 @@ func _ai_default(name: String) -> Variant:
 		"tilt_penalty": return 0.002
 		"speed_reward_weight": return 0.004
 		"time_penalty": return 0.001
+		"airborne_penalty": return 0.0
 		"allow_vertical", "allow_pitch_roll": return true
 		_: return null