fix(ai): map team-relative rotation actions

Game/scripts/ship_action_codec.gd

@@ -8,6 +8,13 @@ extends RefCounted
 # "do not fork this logic" role for observations; the train/inference seam
 # broke once before over exactly this kind of divergence (commit 8c15c46).
 #
+# Ship thrust is body-local, so its axes must not be mirrored for team 1.
+# Ship rotation, however, is applied directly as world-space torque in
+# Ship.apply_rotation_forces(). Team 1 observes a canonical frame rotated
+# 180 degrees about world Y, so its canonical pitch/roll outputs must be
+# rotated back to world space before they reach the ship. apply_team_frame()
+# is the shared training/inference seam for that conversion.
+#
 # Curriculum generation 4 replaces the old continuous Gaussian action space
 # (Box(7), see the "continuous" path below) with a per-axis MultiDiscrete
 # space: PPO's Gaussian std reliably collapsed to ~0.13-0.15 within the first
@@ -100,6 +107,17 @@ static func from_logits(logits: Array, noise: float) -> ShipAction:
 	return result
 
 
+# Map a policy's canonical-frame rotation intent back into the physical
+# team's world frame. A 180-degree Y rotation negates X and Z and leaves Y
+# unchanged. Translation remains untouched because Ship applies it through
+# the ship's local basis rather than as a world-space vector.
+static func apply_team_frame(action: ShipAction, team: int) -> ShipAction:
+	if team == 1:
+		action.rotation.x = -action.rotation.x
+		action.rotation.z = -action.rotation.z
+	return action
+
+
 # Legacy continuous decode — moved verbatim from ai_ship_controller.gd so
 # every model exported before generation 4 (no "action_space" block in its
 # JSON, e.g. Game/bots/promoted/easy.json) keeps behaving byte-identically.