AM-1 (dense refs): Every ActorMeshRef has exactly one actor ref per
rank in its Region. Managed refs materialize those refs from their cast
domain, while data refs store them in a ValueMesh.
AM-2 (slice materialization):RankedSliceable::sliced has no
caller context, so it carries only a raw cast-domain descriptor. The first
cast through that ref materializes the descriptor with the caller context
before sending the cast message, sequencing setup and delivery on the same
sender stream.
AM-3 (explicit direct monitoring): Each call to
ActorMeshRef::monitor creates an independent direct actor monitor. The
managed variant separately retains its legacy controller supervision.
The id of the actor mesh an actor was spawned into. Data meshes label
their members’ supervision events with it, so the owner can name the
failed mesh.
Liveness watchdog for the supervision stream. If no
supervision message (healthy or unhealthy) is observed within
this duration, the controller is assumed to be unreachable and
the mesh is treated as unhealthy. This timeout is about
detecting silence, not slow messages.
This value must be > poll frequency + get actor state timeout + get proc state timeout
or else it is possible to declare the controller dead before it could
feasibly have received a healthy reply.