Function prepare_partitioned_session_runtime 

pub fn prepare_partitioned_session_runtime<A, B, R, S, G, W, E, F>(
    architecture: A,
    selected: SelectedReplicatedTextRealization,
    partition: ArchitecturePartition<G, W>,
    communication: CommunicationManifest,
    expected_tasks: Option<&[ReplicatedTextMaterializationTask]>,
    topology: PromptCacheTopology,
    output_selection: ReplicatedTextOutputSelection,
    context: &<B::Tensor as Tensor>::Context,
    factory: F,
) -> Result<PreparedPartitionedSessionRuntime<R, S>, PartitionedSessionPreparationError<E>>
where
    B: NeuralBackend,
    S: RuntimeState<B>,
    A: LayeredArchitecture<B, S>,
    A::Error: Display,
    F: FnOnce(PartitionedSessionFactoryInput<A, G, W>, &SelectedReplicatedTextRealization, &<B::Tensor as Tensor>::Context) -> Result<(R, S), E>,

Consumes exact architecture authority into a rank-local runtime and state.

The selected realization, architecture, partition, communication manifest, and optional architecture-precomputed task proof are consumed in one operation. Payload work is re-derived from the consumed architecture and partition before the factory runs. A task proof, when supplied, must match that derivation exactly. The factory therefore cannot be paired with a different architecture after admission.