feat(a1z): a1z learning workflow

CONTEXT.md

@@ -98,3 +98,25 @@ _Avoid_: Interpolated episode, repaired episode
 **Training Frame**:
 One fixed-rate set of aligned observations, an applied action, task text, and complementary information.
 _Avoid_: Camera frame, raw sample
+
+## Policy Rollout
+
+**Policy rollout**:
+One active execution of one configured checkpoint against live observations, from an explicit start until stop, timeout, failure, or control preemption.
+_Avoid_: Policy task, inference request
+
+**Rollout supervisor**:
+The lifecycle supervisor that translates semantic operator actions into policy and control-task activation requests.
+_Avoid_: Policy runtime, control task
+
+**Policy arm command**:
+An absolute arm-joint target routed to the coordinator's activation-gated policy trajectory task.
+_Avoid_: Applied command, measured state
+
+**Policy gripper command**:
+A normalized gripper opening routed to the coordinator's activation-gated policy gripper task.
+_Avoid_: Native gripper position, arm command
+
+**Control preemption**:
+A control tick in which a higher-priority task wins joints claimed by another task. Activation-gated tasks deactivate themselves after preemption.
+_Avoid_: Rejected command, hidden output