feat(*): Replace the test terrain arena with an enclosed standard-size space-platform arena (shared ArenaBoundary floor/walls/ceiling scene, starfield sky, ball CCD) and derive TrainingMode placement bounds from it, dropping the out-of-bounds reward guard

Game/scripts/arena_boundary.gd

@@ -0,0 +1,13 @@
+class_name ArenaBoundary
+extends StaticBody3D
+
+# The standard arena play volume (inner faces of the enclosure). Every arena
+# instances objects/arena_boundary.tscn so all arenas share one size; code
+# that needs field dimensions derives them from these constants rather than
+# restating numbers.
+const INNER_HALF_X := 12.0
+const INNER_HALF_Z := 18.0
+const INNER_HEIGHT := 12.0
+# Goal-centre distance from arena centre; the end walls sit 1 m behind, so a
+# ball pinned against them still overlaps the goal sensor.
+const GOAL_LINE_Z := 17.0

Game/scripts/arena_boundary.gd.uid

@@ -0,0 +1 @@
+uid://f32otkpo3lvr

Game/scripts/ship_observations.gd

@@ -11,8 +11,9 @@ extends RefCounted
 # The same rotation must be inverted when interpreting actions (see canon —
 # it is its own inverse).
 
-# Normalization scales. Arena bounds: goals at z ≈ ±15.56, ship spawns at
-# z = ±12; positions are soft-normalized to roughly [-1, 1].
+# Normalization scales. Standard arena volume (see ArenaBoundary): x ±12,
+# z ±18, height 12, goals at z ±17; positions are soft-normalized to roughly
+# [-1, 1]. Do not retune without retraining every model in Game/bots/.
 const POSITION_SCALE := Vector3(20.0, 10.0, 20.0)
 const BALL_SPEED_SCALE := 30.0
 const GOAL_DISTANCE_SCALE := 40.0

Game/scripts/training_mode.gd

@@ -24,24 +24,26 @@ extends GameMode
 @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
 
-# Placement bounds, inset from the arena (goals at z ≈ ±15.56).
-const FIELD_HALF_X := 10.0
-const FIELD_HALF_Z := 13.0
+# Placement bounds for randomized episode starts, derived from the standard
+# enclosure (ArenaBoundary). The inset keeps a randomly oriented ship (1x1x4
+# box, worst-case half-extent ~2.05) from spawning intersecting the walls,
+# ceiling, or goal sensors.
+const SPAWN_INSET := 2.5
+const FIELD_HALF_X := ArenaBoundary.INNER_HALF_X - SPAWN_INSET
+const FIELD_HALF_Z := ArenaBoundary.GOAL_LINE_Z - SPAWN_INSET
 const FIELD_MIN_Y := 1.5
-const FIELD_MAX_Y := 8.0
+const FIELD_MAX_Y := ArenaBoundary.INNER_HEIGHT - SPAWN_INSET
 const MAX_RANDOM_BALL_SPEED := 12.0
 const MAX_RANDOM_SHIP_SPEED := 8.0
 
 # Sim runs at 60 physics ticks per sim-second regardless of speedup.
 const TICKS_PER_SIM_SECOND := 60.0
 
-# The arena has no walls/ceiling yet (see TODO.md), so an untrained policy can
-# simply fly away. Leaving this volume ends the episode with a small penalty.
-const BOUNDS_HALF_X := 25.0
-const BOUNDS_HALF_Z := 25.0
-const BOUNDS_MIN_Y := -5.0
-const BOUNDS_MAX_Y := 25.0
-@export var out_of_bounds_penalty := 1.0
+# The arena is physically enclosed, so nothing should ever get this far out.
+# If a body escapes anyway (physics regression, boundary edit), it is warned
+# about and respawned with no reward change and no episode end — a multi-hour
+# training run must survive it, and the escape must not shape rewards.
+const ESCAPE_MARGIN := 15.0
 
 var _agents: Array[ShipAIController] = []
 
@@ -102,13 +104,10 @@ func _attack_goal_position(team: int) -> Vector3:
 
 
 func _physics_process(_delta):
+	_respawn_escaped_bodies()
 	if _eval:
 		_episode_ticks += 1
-		for ship in ships:
-			if is_instance_valid(ship) and _out_of_bounds(ship.global_position):
-				_place_body(ship, _ship_spawn_transforms[ship], Vector3.ZERO, Vector3.ZERO)
-		var ball_lost := is_instance_valid(ball) and _out_of_bounds(ball.global_position)
-		if _episode_ticks > int(episode_length_seconds * TICKS_PER_SIM_SECOND) or ball_lost:
+		if _episode_ticks > int(episode_length_seconds * TICKS_PER_SIM_SECOND):
 			_eval_draws += 1
 			_end_eval_episode()
 		return
@@ -128,23 +127,23 @@ func _physics_process(_delta):
 		_reset_episode()
 		return
 
-	# A ship leaving the play volume is penalized and respawned at its kickoff
-	# spawn (the episode continues); a lost ball ends the episode for both.
-	for agent in _agents:
-		if _out_of_bounds(agent.ship.global_position):
-			agent.reward -= out_of_bounds_penalty
-			_place_body(agent.ship, _ship_spawn_transforms[agent.ship], Vector3.ZERO, Vector3.ZERO)
-	if is_instance_valid(ball) and _out_of_bounds(ball.global_position) and not _agents.is_empty():
-		for agent in _agents:
-			agent.done = true
-		_reset_episode()
+
+# Escape failsafe: see ESCAPE_MARGIN.
+func _respawn_escaped_bodies() -> void:
+	for ship in ships:
+		if is_instance_valid(ship) and _escaped(ship.global_position):
+			push_warning("TrainingMode: ship escaped the enclosed arena — check boundary colliders")
+			_place_body(ship, _ship_spawn_transforms[ship], Vector3.ZERO, Vector3.ZERO)
+	if is_instance_valid(ball) and _escaped(ball.global_position):
+		push_warning("TrainingMode: ball escaped the enclosed arena — check boundary colliders")
+		_place_body(ball, arena.get_ball_spawn(), Vector3.ZERO, Vector3.ZERO)
 
 
-func _out_of_bounds(position: Vector3) -> bool:
-	return absf(position.x) > BOUNDS_HALF_X \
-		or absf(position.z) > BOUNDS_HALF_Z \
-		or position.y < BOUNDS_MIN_Y \
-		or position.y > BOUNDS_MAX_Y
+func _escaped(position: Vector3) -> bool:
+	return absf(position.x) > ArenaBoundary.INNER_HALF_X + ESCAPE_MARGIN \
+		or absf(position.z) > ArenaBoundary.INNER_HALF_Z + ESCAPE_MARGIN \
+		or position.y < -ESCAPE_MARGIN \
+		or position.y > ArenaBoundary.INNER_HEIGHT + ESCAPE_MARGIN
 
 
 func _on_goal_scored(conceding_team: int) -> void: