Function prepare_default_partitioned_runtime 

pub fn prepare_default_partitioned_runtime<A, B, M, G, W, P>(
    input: PartitionedSessionFactoryInput<A, G, W>,
    source_architecture: Option<A>,
    target_parallel_layout: LocalModelLayout,
    source_parallel_layout: Option<LocalModelLayout>,
    selected: &SelectedReplicatedTextRealization,
    topology: &PromptCacheTopology,
    scope: PartitionedUnitScope,
    addressable_parameters: &[String],
    mechanisms: &mut M,
    context: &<<B as NeuralBackend>::Tensor as Tensor>::Context,
) -> Result<PreparedPartitionedRuntimeComponents<A, G, W, M::State, P>, PartitionedRuntimeConstructionError>
where
    B: SubmissionBackend<Executor = <<B as NeuralBackend>::Tensor as Tensor>::Context>,
    M: ReplicatedTextSessionMechanisms<A, B, ResidentPolicy = P, BoundedPolicy = P>,
    A: LayeredArchitecture<B, M::State>,
    A::Error: Display,
    M::Error: Display,
    P: LayerwisePolicy<B, A::Unit, Error = M::PolicyError> + Clone,

Runs the reusable cold-path lifecycle for one exact partition.

Architecture authority and tasks arrive together from prepare_partitioned_session_runtime. This driver chooses the exact unit set, constructs target and optional transform-source units, prepares and partitions materialization, realizes local state, and selects the resident or bounded policy. The caller receives only the completed typed components needed to create native communication and executor mechanisms.