diff --git a/Game/scripts/arena_boundary.gd b/Game/scripts/arena_boundary.gd
index 32de28ce..de7446d4 100644
--- a/Game/scripts/arena_boundary.gd
+++ b/Game/scripts/arena_boundary.gd
@@ -98,6 +98,9 @@ const DECK_SHADER_PATH := "res://shaders/arena_deck.gdshader"
 # _process() drives. Both stay null in headless runs, which never render.
 var _shell: MeshInstance3D
 var _field_material: ShaderMaterial
+# Cached active camera for _process(), mirroring ship_camera.gd's _get_ball()
+# pattern so the viewport lookup isn't repeated every frame.
+var _camera: Camera3D
 
 
 func _ready() -> void:
@@ -153,12 +156,26 @@ func _process(_delta: float) -> void:
 	# individually. Collision is untouched.
 	if _field_material == null:
 		return
-	var camera := get_viewport().get_camera_3d()
+	var camera := _get_camera()
 	if camera == null:
 		return  # headless (RL/CI) has no camera
 	_field_material.set_shader_parameter("camera_local_pos", to_local(camera.global_position))
 
 
+# Caches the viewport's active camera; a plain is_instance_valid revalidation
+# is enough because exactly one ship_camera_rig (Camera3D) is spawned per
+# game-mode run (GameMode.spawn_camera_rig, called once each from
+# free_play.gd/match_mode.gd/spectate_mode.gd) and never re-spawned mid-match
+# — there's no active-camera-switch scenario today. ArenaBoundary itself is
+# destroyed/recreated per scene change, so the cache naturally resets with it;
+# there's no stale-cache-across-scenes concern. Revisit this if split-screen
+# or multiplayer camera-switching lands (see TODO.md's multiplayer section).
+func _get_camera() -> Camera3D:
+	if not is_instance_valid(_camera):
+		_camera = get_viewport().get_camera_3d()
+	return _camera
+
+
 # --- shared layout -----------------------------------------------------------
 # Consumed by both the collider rings and the visual shell, so the two can
 # never end up describing different geometry.
diff --git a/TODO.md b/TODO.md
index 36ce1c23..1ba511f8 100644
--- a/TODO.md
+++ b/TODO.md
@@ -28,7 +28,6 @@ Bugs found in an adversarial review. None are gameplay- or physics-affecting, so
 - [ ] Shared per-team materials instead of `Ship._apply_team_color()` allocating a fresh `StandardMaterial3D` and assigning it as `material_override` (and running at least twice per ship — once from `_ready`, once from the `team` setter). Removes the allocation and lets same-team ships batch.
 - [ ] Merge each goal's visuals into one `ArrayMesh` with a hull surface and a net surface — currently 11 `MeshInstance3D`s and 4 materials per goal, ~22 draw calls for a static prop. `arena_boundary.gd:333` already demonstrates the `SurfaceTool` technique in this codebase.
 - [ ] Merge the four non-tinted ship meshes (Hull/Canopy/EngineGlowL/R) into one — 6 draw calls per ship down to 3. Irrelevant at 1v1; 36 calls before VFX at 3v3.
-- [ ] Cache the camera in `ArenaBoundary._process` instead of a `get_viewport().get_camera_3d()` tree lookup every frame (`ship_camera.gd` already caches the ball this way).
 - [ ] Name collision layers in `project.godot` and assign them — nothing configures `collision_layer`/`collision_mask` today, so every body tests against every other.
 - [ ] Measure `nebula_dust.gdshader`'s per-fragment depth-texture sample across 500 large soft billboards before adding more particle work.
 - [ ] Bake `ArenaBoundary`'s ~160 runtime-generated `CollisionShape3D` nodes into the scene. Costs a load hitch on every arena entry and repeats in every parallel headless training env. **Blocked on the trained-bot decision below** — `arena_boundary.gd:235` notes this geometry is what the shipped policies were fitted against, so the bake must be verified byte-identical.