ferrum_interfaces/vnext/operation/

dispatch.rs

use sha2::{Digest, Sha256};
use std::{collections::BTreeSet, sync::Arc};

use super::super::{
    classify_device_error, AllocationKind, AllocationLifetime, BackingInitializationEncodeError,
    BufferUsage, CompletionHandle, CompletionReaper, CompletionReservation,
    DefinitelyNotSubmittedRetryAuthority, DeviceBatchingForm, DeviceCommandBatch,
    DeviceCommandLogicalWork, DeviceComputePathRequirement, DeviceReusableExecutionCapture,
    DeviceReusableExecutionInvocation, DeviceReusableExecutionProgram,
    DeviceReusableExecutionProgramId, DeviceReusableExecutionTopologyFingerprint, DeviceRuntime,
    DeviceTimingMode, ExecutablePlanView, ExecutionIdentityEnvelope, ExecutionIdentityParts,
    ExecutionLane, InvocationResourceLease, LaneSubmitOutcome, NodeId, NodeInvocationId,
    OperationId, ParticipantNodeKey, PreparedStepSubmissionWave, ProgramBindingNodeBinding,
    ProviderId, ResourceId, SpanId, StepParticipantFrameAssignment, SubmissionWavePurpose,
    TrustedActiveSequenceBinding, VNextError, EXECUTION_IDENTITY_VERSION,
};
use super::backing_upload::encode_submission_wave_backing_upload;
use super::determinism::{
    SubmissionWaveDeterminismHandle, SubmissionWaveDeterminismReadbackPlan,
    SubmissionWaveDeterminismRestore,
};
use super::dispatch_contract::{
    BoundDeviceSubmissionAttribution, DisabledSubmissionWaveDispatchTimingSink,
    DispatchRetryAuthority, OperationDispatchError, ProfiledSubmissionHandle,
    SubmissionExecutionPolicy, SubmissionScratchInitialization, SubmissionWaveDispatchError,
    SubmissionWaveDispatchStage, SubmissionWaveDispatchStageTimer,
    SubmissionWaveDispatchTimingSink, SubmissionWaveInputUpload,
};
use super::foundation::invalid_operation;
use super::invocation::OperationInvocationResources;
use super::workspace_encoding::{
    encode_provider_workspace_initialization, encode_submission_wave_workspace_initializations,
};
use super::{
    translate_step_participant_upload_range, BatchOperationIdentity, BatchOperationNodeIdentity,
    BatchOperationParticipantIdentity, BatchedOperationInvocation, BoundOperationProvider,
    OperationInvocation, ProviderReplayEquivalence, ResolvedValueRole, ReusableExecutionTopology,
    ReusableExecutionTopologyRequest, TensorAccess,
};

/// The only public path from a resolved plan to an operation kernel.
fn validate_program_binding_patch(
    resolved: &dyn ExecutablePlanView,
    node_identity: &BatchOperationNodeIdentity,
    program_binding: Option<&ProgramBindingNodeBinding>,
    encoded_program_binding_count: usize,
    program_binding_resources: &mut BTreeSet<ResourceId>,
) -> Result<(), VNextError> {
    if program_binding.is_some() != (encoded_program_binding_count == 1) {
        return Err(invalid_operation(
            "compiled program binding slot and provider patch cardinality differ",
        ));
    }
    let Some(program_binding) = program_binding else {
        return Ok(());
    };
    let (plan_node_index, binding_resource) = resolved
        .execution_plan()
        .payload()
        .nodes()
        .iter()
        .enumerate()
        .find(|(_, node)| node.id() == node_identity.node_id())
        .and_then(|(node_index, node)| {
            node.binding_resource()
                .map(|resource| (node_index, resource))
        })
        .ok_or_else(|| {
            invalid_operation("program binding requires a provider-owned binding workspace")
        })?;
    if program_binding.node_index() != plan_node_index
        || program_binding.slot().node_id() != node_identity.node_id()
        || program_binding.slot().resource_id() != binding_resource
    {
        return Err(invalid_operation(
            "provider program binding differs from its compiled node slot",
        ));
    }
    if !program_binding_resources.insert(binding_resource.clone()) {
        return Err(invalid_operation(
            "program bindings from different nodes alias one binding workspace",
        ));
    }
    Ok(())
}

fn validate_determinism_replay_coverage(
    program: &DeviceReusableExecutionProgram,
    node_count: usize,
    eager_boundary_node_indices: &[u32],
    batch_identity: &BatchOperationIdentity,
) -> Result<(), VNextError> {
    if node_count == 0
        || eager_boundary_node_indices
            .windows(2)
            .any(|pair| pair[0] >= pair[1])
    {
        return Err(invalid_operation(
            "determinism replay topology has no nodes or non-canonical eager boundaries",
        ));
    }
    if program.node_count() as usize != node_count
        || program.eager_boundary_node_indices() != eager_boundary_node_indices
    {
        return Err(invalid_operation(
            "determinism reusable program coverage differs from the live provider topology",
        ));
    }
    if !program.is_determinism_ready() {
        let gaps = program
            .gaps()
            .iter()
            .take(8)
            .map(|gap| {
                let node_index = gap.node_index() as usize;
                format!(
                    "node_index={} node_id={} provider_id={} operation_id={} reason={}",
                    gap.node_index(),
                    batch_identity
                        .node_id_at(node_index)
                        .map_or("<missing>", NodeId::as_str),
                    batch_identity
                        .provider_id_at(node_index)
                        .map_or("<missing>", ProviderId::as_str),
                    batch_identity
                        .operation_id_at(node_index)
                        .map_or("<missing>", OperationId::as_str),
                    gap.reason().as_str(),
                )
            })
            .collect::<Vec<_>>()
            .join("; ");
        return Err(invalid_operation(format!(
            "determinism reusable program has {} non-resident replay-eligible gap(s): {gaps}",
            program.gaps().len()
        )));
    }
    let mut coverage = vec![0_u8; node_count];
    for node_index in eager_boundary_node_indices {
        let node_index = usize::try_from(*node_index)
            .map_err(|_| invalid_operation("determinism eager boundary exceeds usize"))?;
        let marker = coverage.get_mut(node_index).ok_or_else(|| {
            invalid_operation("determinism eager boundary is outside the submission wave")
        })?;
        *marker = 1;
    }
    for segment in program.segments() {
        let start = usize::try_from(segment.start_node_index())
            .map_err(|_| invalid_operation("determinism replay segment start exceeds usize"))?;
        let end = usize::try_from(segment.end_node_index())
            .map_err(|_| invalid_operation("determinism replay segment end exceeds usize"))?;
        let range = coverage.get_mut(start..end).ok_or_else(|| {
            invalid_operation("determinism replay segment is outside the submission wave")
        })?;
        if range.iter().any(|marker| *marker != 0) {
            return Err(invalid_operation(
                "determinism replay segment overlaps another segment or a declared eager boundary",
            ));
        }
        range.fill(2);
    }
    if coverage.iter().any(|marker| *marker == 0) {
        return Err(invalid_operation(
            "determinism reusable program does not cover every replay-eligible node",
        ));
    }
    if !coverage.iter().any(|marker| *marker == 2) {
        return Err(invalid_operation(
            "determinism replay topology contains no resident replay node",
        ));
    }
    Ok(())
}

struct ReusableExecutionWaveAuthority {
    program_id: DeviceReusableExecutionProgramId,
    eager_boundary_node_indices: Vec<u32>,
}

pub struct OperationDispatch;

impl OperationDispatch {
    fn trusted_submission_node_identity(
        resolved: &dyn ExecutablePlanView,
        active: &TrustedActiveSequenceBinding,
        frame: StepParticipantFrameAssignment,
        node_index: usize,
    ) -> Result<ExecutionIdentityEnvelope, VNextError> {
        active.ensure_open_for_emission()?;
        let plan = resolved.execution_plan();
        let node_count = u64::try_from(plan.payload().nodes().len())
            .map_err(|_| invalid_operation("immutable plan node count exceeds u64"))?;
        let node_index = u64::try_from(node_index)
            .map_err(|_| invalid_operation("submission wave node index exceeds u64"))?;
        let node = plan
            .payload()
            .nodes()
            .get(node_index as usize)
            .ok_or_else(|| invalid_operation("submission wave node index is out of bounds"))?;
        let completed_frames = frame.frame_id().get() - 1;
        let node_invocation = completed_frames
            .checked_mul(node_count)
            .and_then(|value| value.checked_add(node_index))
            .and_then(|value| value.checked_add(1))
            .ok_or_else(|| invalid_operation("node invocation id space is exhausted"))?;
        let node_invocation_id = NodeInvocationId::try_from(node_invocation)?;

        // RequestAccepted and PlanBuilt consume the first two journal rows.
        // Every immutable-plan frame then emits FrameStarted, three rows per
        // node, and FrameCompleted. This operation identity is therefore the
        // exact future OperationSubmitted row and remains stable across a
        // definitely-not-submitted retry of the same frame.
        let events_per_frame = node_count
            .checked_mul(3)
            .and_then(|value| value.checked_add(2))
            .ok_or_else(|| invalid_operation("execution event sequence space is exhausted"))?;
        let sequence = completed_frames
            .checked_mul(events_per_frame)
            .and_then(|value| value.checked_add(node_index.checked_mul(3)?))
            .and_then(|value| value.checked_add(5))
            .ok_or_else(|| invalid_operation("execution event sequence space is exhausted"))?;

        let span_root = format!("vnext/request/{}", active.fingerprint());
        let node_span = SpanId::new(format!(
            "{span_root}/frame/{}/node/{node_invocation}",
            frame.frame_id()
        ))?;
        let operation_span = SpanId::new(format!("{node_span}/operation"))?;
        let provisioning = active.static_provisioning_identity();
        ExecutionIdentityEnvelope::new(ExecutionIdentityParts {
            version: EXECUTION_IDENTITY_VERSION,
            run_id: active.run_id().clone(),
            request_id: active.request_id().clone(),
            sequence,
            plan_id: Some(plan.payload().plan_id().clone()),
            plan_hash: Some(plan.plan_hash().clone()),
            frame_id: Some(frame.frame_id()),
            node_invocation_id: Some(node_invocation_id),
            node_id: Some(node.id().clone()),
            operation_id: Some(node.operation_id().clone()),
            provider_id: Some(node.selection().selected_provider().clone()),
            device_id: Some(plan.payload().device_id().clone()),
            resource_pool_id: active.static_pool_id(),
            resource_pool_identity_fingerprint: active
                .static_pool_identity_fingerprint_ref()
                .map(str::to_owned),
            provisioning_run_id: provisioning.map(|identity| identity.run_id().clone()),
            provisioning_request_id: provisioning.map(|identity| identity.request_id().clone()),
            transaction_id: provisioning.map(|identity| identity.transaction_id().clone()),
            active_sequence_slot: Some(active.sequence_authority().sparse_id()),
            admission_generation: Some(active.sequence_authority().generation()),
            activation_epoch: Some(active.activation_epoch()),
            runtime_implementation_fingerprint: Some(
                active.runtime_implementation_fingerprint().to_owned(),
            ),
            active_sequence_fingerprint: Some(active.fingerprint().to_owned()),
            completed_sequence_fingerprint: None,
            aborted_sequence_fingerprint: None,
            resource_id: None,
            resource_generation: None,
            resource_batch_fingerprint: None,
            span_id: operation_span,
            parent_span_id: Some(node_span),
            async_links: Vec::new(),
        })
    }

    #[allow(clippy::too_many_arguments)]
    fn bind_node_identity<'binding, R, I>(
        resolved: &dyn ExecutablePlanView,
        participant_identities: Vec<ExecutionIdentityEnvelope>,
        active_bindings: I,
        resources: OperationInvocationResources<'_, R>,
        lane: &Arc<ExecutionLane<R>>,
        node_index: u32,
        participant_start: u32,
    ) -> Result<BatchOperationNodeIdentity, VNextError>
    where
        R: DeviceRuntime,
        I: ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        let plan = resolved.execution_plan();
        let node_id = resources.node_id()?;
        let node = plan
            .payload()
            .nodes()
            .iter()
            .find(|node| node.id() == node_id)
            .ok_or_else(|| invalid_operation(format!("plan has no node `{node_id}`")))?;
        // Wave construction proves every node shares this authority by Arc identity.
        let validate_common_authority = match resources {
            OperationInvocationResources::Invocation(_) => true,
            OperationInvocationResources::Wave {
                node_index: resource_node_index,
                ..
            } => resource_node_index == 0,
        };
        let participant_count = resources.participant_count()?;
        let frames = resources.participant_frames()?;
        let plan_identity_matches = validate_common_authority
            .then(|| {
                resources.plan_identity_matches(
                    plan.payload().plan_id(),
                    plan.plan_hash(),
                    plan.payload().device_id(),
                )
            })
            .transpose()?;
        if participant_identities.is_empty()
            || participant_identities.len() != participant_count
            || participant_identities.len() != frames.len()
            || participant_identities.len() != active_bindings.len()
            || resources.prepared_participant_count()? != participant_count
            || validate_common_authority && {
                !plan_identity_matches.expect("common wave authority requested plan evidence")
                    || !Arc::ptr_eq(resources.runtime(), lane.runtime_arc())
                    || resources.step_resources().execution_lane().id() != lane.id()
                    || lane.descriptor() != resolved.device()
                    || lane.descriptor() != resolved.capabilities().device()
                    || lane.descriptor().runtime_implementation_fingerprint
                        != plan.payload().device_runtime_implementation_fingerprint()
            }
        {
            return Err(invalid_operation(
                "batch node identity inputs differ from submission resources, plan, or lane",
            ));
        }
        let mut participant_projections = Vec::with_capacity(participant_identities.len());
        for (local_index, (identity, active)) in participant_identities
            .into_iter()
            .zip(active_bindings)
            .enumerate()
        {
            let participant = resources.participant(local_index)?;
            let frame = *frames
                .get(local_index)
                .ok_or_else(|| invalid_operation("operation participant frame is missing"))?;
            let session = validate_common_authority
                .then(|| resources.participant_session_identity(local_index))
                .transpose()?;
            let key =
                ParticipantNodeKey::new(frame.participant(), frame.frame_id(), node.id().clone());
            let parts = identity.parts();
            if key.sequence_authority() != participant.sequence_authority()
                || key.request_authority() != participant.request_authority()
                || key.frame_id() != frame.frame_id()
                || validate_common_authority && {
                    let session = session
                        .as_ref()
                        .expect("common wave authority requested session evidence");
                    active.sequence_authority() != participant.sequence_authority()
                        || active.coordinator_id() != resources.coordinator_id()?
                        || active.run_id() != participant.run_id()
                        || active.request_id() != participant.request_id()
                        || !active.matches_sequence_session(session.0, session.1)
                        || active.plan().plan_id() != plan.payload().plan_id()
                        || active.plan().plan_hash() != plan.plan_hash()
                        || active.plan().device_id() != plan.payload().device_id()
                        || active.runtime_implementation_fingerprint()
                            != plan.payload().device_runtime_implementation_fingerprint()
                }
                || parts.run_id != *active.run_id()
                || parts.request_id != *active.request_id()
                || parts.plan_id.as_ref() != Some(plan.payload().plan_id())
                || parts.plan_hash.as_ref() != Some(plan.plan_hash())
                || parts.frame_id != Some(frame.frame_id())
                || parts.node_invocation_id.is_none()
                || parts.node_id.as_ref() != Some(node.id())
                || parts.operation_id.as_ref() != Some(node.operation_id())
                || parts.provider_id.as_ref() != Some(node.selection().selected_provider())
                || parts.device_id.as_ref() != Some(plan.payload().device_id())
                || parts.active_sequence_slot != Some(active.sequence_authority().sparse_id())
                || parts.admission_generation != Some(active.sequence_authority().generation())
                || parts.activation_epoch != Some(active.activation_epoch())
                || parts.runtime_implementation_fingerprint.as_deref()
                    != Some(active.runtime_implementation_fingerprint())
                || parts.active_sequence_fingerprint.as_deref() != Some(active.fingerprint())
                || parts.completed_sequence_fingerprint.is_some()
                || parts.aborted_sequence_fingerprint.is_some()
                || parts.resource_id.is_some()
                || parts.resource_generation.is_some()
                || parts.resource_batch_fingerprint.is_some()
            {
                return Err(invalid_operation(format!(
                    "batch node {node_index} participant {local_index} differs from its resource, frame, session, or plan identity"
                )));
            }
            let local_index = u32::try_from(local_index)
                .map_err(|_| invalid_operation("batch participant index exceeds u32"))?;
            participant_projections.push(BatchOperationParticipantIdentity::new(
                participant_start.checked_add(local_index).ok_or_else(|| {
                    invalid_operation("physical batch participant index overflows u32")
                })?,
                key,
                identity,
            ));
        }
        BatchOperationNodeIdentity::from_validated(
            node_index,
            node.id().clone(),
            node.operation_id().clone(),
            node.selection().selected_provider().clone(),
            node.provider_implementation_fingerprint().to_owned(),
            node.provider_execution_semantics(),
            resources.work_shape()?.fingerprint().to_owned(),
            participant_projections,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn bind_batch_identity<R>(
        resolved: &dyn ExecutablePlanView,
        participant_identities: Vec<ExecutionIdentityEnvelope>,
        active_bindings: &[TrustedActiveSequenceBinding],
        invocation_resources: &InvocationResourceLease<R>,
        lane: &Arc<ExecutionLane<R>>,
    ) -> Result<BatchOperationIdentity, VNextError>
    where
        R: DeviceRuntime,
    {
        let plan = resolved.execution_plan();
        let resources = OperationInvocationResources::Invocation(invocation_resources);
        let node_identity = Self::bind_node_identity(
            resolved,
            participant_identities,
            active_bindings.iter(),
            resources,
            lane,
            0,
            0,
        )?;
        BatchOperationIdentity::from_validated(
            resources.batch_step_id(),
            resources.batch_invocation_id(),
            plan.payload().plan_id().clone(),
            plan.plan_hash().clone(),
            plan.payload().device_id().clone(),
            plan.payload()
                .device_runtime_implementation_fingerprint()
                .to_owned(),
            lane.id(),
            resources.backing_fingerprint().to_owned(),
            vec![node_identity],
        )
    }

    fn bind_submission_wave_identity_from_envelopes<'binding, R, I>(
        resolved: &dyn ExecutablePlanView,
        participant_identities: Vec<Vec<ExecutionIdentityEnvelope>>,
        active_bindings: I,
        wave: &PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
    ) -> Result<BatchOperationIdentity, VNextError>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        let plan = resolved.execution_plan();
        if wave.nodes().is_empty()
            || wave.execution_lane_id() != lane.id()
            || participant_identities.len() != wave.nodes().len()
            || active_bindings.len() == 0
            || wave.claimed_backing().node_count() != wave.nodes().len()
            || wave.claimed_backing().plan_node_count() != plan.payload().nodes().len()
            || wave
                .nodes()
                .windows(2)
                .any(|pair| pair[0].plan_node_index() >= pair[1].plan_node_index())
            || wave.nodes().iter().any(|prepared| {
                plan.payload()
                    .nodes()
                    .get(prepared.plan_node_index())
                    .is_none_or(|planned| prepared.node_id() != planned.id())
            })
        {
            return Err(invalid_operation(
                "submission wave identity must cover one exact canonical immutable-plan node scope",
            ));
        }
        let mut participant_start = 0_u32;
        let mut nodes = Vec::with_capacity(wave.nodes().len());
        for (node_index, (identities, _node)) in participant_identities
            .into_iter()
            .zip(wave.nodes())
            .enumerate()
        {
            let node_index = u32::try_from(node_index)
                .map_err(|_| invalid_operation("submission wave node index exceeds u32"))?;
            let participant_count = u32::try_from(identities.len())
                .map_err(|_| invalid_operation("submission wave participant count exceeds u32"))?;
            let node_identity = Self::bind_node_identity(
                resolved,
                identities,
                active_bindings.clone(),
                OperationInvocationResources::Wave {
                    wave,
                    node_index: node_index as usize,
                },
                lane,
                node_index,
                participant_start,
            )?;
            participant_start = participant_start
                .checked_add(participant_count)
                .ok_or_else(|| {
                    invalid_operation("submission wave participant index space overflows u32")
                })?;
            nodes.push(node_identity);
        }
        BatchOperationIdentity::from_validated(
            wave.batch_step_id(),
            wave.batch_invocation_id(),
            plan.payload().plan_id().clone(),
            plan.plan_hash().clone(),
            plan.payload().device_id().clone(),
            plan.payload()
                .device_runtime_implementation_fingerprint()
                .to_owned(),
            lane.id(),
            wave.fingerprint().to_owned(),
            nodes,
        )
    }

    /// Binds one immutable-plan submission wave without accepting caller-made
    /// execution envelopes. Frame, node, provider, provisioning, span, and
    /// invocation identities are minted from core-owned plan/session evidence.
    pub fn bind_submission_wave_identity<'binding, R, I>(
        resolved: &dyn ExecutablePlanView,
        active_bindings: I,
        wave: &PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
    ) -> Result<BatchOperationIdentity, VNextError>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        if active_bindings.len() == 0 {
            return Err(invalid_operation(
                "submission wave requires a non-empty active participant set",
            ));
        }
        let participant_identities = wave
            .nodes()
            .iter()
            .enumerate()
            .map(|(node_index, node)| {
                if active_bindings.len() != node.participant_frames().len() {
                    return Err(invalid_operation(format!(
                        "submission wave node {node_index} active binding count differs from its participant frames"
                    )));
                }
                node.participant_frames()
                    .iter()
                    .copied()
                    .zip(active_bindings.clone())
                    .map(|(frame, active)| {
                        Self::trusted_submission_node_identity(
                            resolved,
                            active,
                            frame,
                            node.plan_node_index(),
                        )
                    })
                    .collect::<Result<Vec<_>, _>>()
            })
            .collect::<Result<Vec<_>, _>>()?;
        Self::bind_submission_wave_identity_from_envelopes(
            resolved,
            participant_identities,
            active_bindings,
            wave,
            lane,
        )
    }

    /// Derives the exact reusable-program identity for the current wave without
    /// materializing device buffers or entering dispatch.
    ///
    /// Provider topology remains opaque. Core binds each topology row to its
    /// immutable node/provider position before aggregating it, so two providers
    /// cannot accidentally alias the same program variant. Eager boundaries
    /// remain part of the program identity while resident segments on either
    /// side stay eligible for typed reuse.
    pub fn reusable_execution_program_id_for_wave<R>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        wave: &PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
    ) -> Result<Option<DeviceReusableExecutionProgramId>, VNextError>
    where
        R: DeviceRuntime,
    {
        Ok(
            Self::reusable_execution_wave_authority(providers, resolved, wave, lane)?
                .map(|authority| authority.program_id),
        )
    }

    fn reusable_execution_wave_authority<R>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        wave: &PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
    ) -> Result<Option<ReusableExecutionWaveAuthority>, VNextError>
    where
        R: DeviceRuntime,
    {
        let plan = resolved.execution_plan();
        let plan_nodes = plan.payload().nodes();
        if providers.is_empty()
            || providers.len() != wave.nodes().len()
            || wave.execution_lane_id() != lane.id()
            || lane.descriptor() != resolved.device()
        {
            return Err(invalid_operation(
                "reusable execution topology requires one exact provider per wave node and lane",
            ));
        }
        let Some(program_id) = wave.claimed_backing().reusable_execution_program_id(
            &lane.descriptor().runtime_implementation_fingerprint,
            lane.id(),
        )?
        else {
            return Ok(None);
        };

        const DOMAIN: &[u8] = b"ferrum.runtime-vnext.reusable-program-topology.v3\0";
        let mut digest = Sha256::new();
        let mut eager_boundary_node_indices = Vec::new();
        digest.update(DOMAIN);
        digest.update(
            u64::try_from(wave.nodes().len())
                .map_err(|_| invalid_operation("reusable topology node count exceeds u64"))?
                .to_le_bytes(),
        );
        for (wave_node_index, (provider, prepared_node)) in
            providers.iter().zip(wave.nodes()).enumerate()
        {
            let plan_node_index = prepared_node.plan_node_index();
            let node = plan_nodes.get(plan_node_index).ok_or_else(|| {
                invalid_operation("reusable execution wave node is absent from the immutable plan")
            })?;
            if !provider.matches_plan_node(plan.payload().plan_id(), plan.plan_hash(), node.id())
                || prepared_node.node_id() != node.id()
            {
                return Err(invalid_operation(
                    "reusable execution topology provider or node differs from the immutable plan",
                ));
            }
            let request = ReusableExecutionTopologyRequest::new(
                node.id(),
                node.operation_id(),
                node.attributes(),
                node.values(),
                node.scratch_resource(),
                node.binding_resource(),
                node.persistent_resource(),
                plan.payload().memory(),
                prepared_node.work_shape(),
                wave.claimed_backing(),
                wave.step_resources().backing_slices(),
            )?;
            let declared = node.provider_execution_semantics().replay_equivalence();
            let topology = provider.provider().reusable_execution_topology(request)?;
            if topology == ReusableExecutionTopology::EagerBoundary {
                eager_boundary_node_indices.push(u32::try_from(wave_node_index).map_err(|_| {
                    invalid_operation("reusable topology wave node index exceeds u32")
                })?);
            }
            let (topology_kind, topology_bytes) = match (
                declared,
                &topology,
            ) {
                (
                    ProviderReplayEquivalence::BitwiseEagerEquivalent,
                    ReusableExecutionTopology::Static,
                ) => (0_u8, &[][..]),
                (
                    ProviderReplayEquivalence::BitwiseEagerEquivalent,
                    ReusableExecutionTopology::Dynamic(topology),
                ) => (1_u8, &topology.as_bytes()[..]),
                (_, ReusableExecutionTopology::EagerBoundary) => (2_u8, &[][..]),
                (
                    ProviderReplayEquivalence::Ineligible,
                    ReusableExecutionTopology::Static | ReusableExecutionTopology::Dynamic(_),
                ) => {
                    return Err(invalid_operation(format!(
                        "provider `{}` returned reusable topology without a bitwise eager-equivalence contract",
                        provider.descriptor().provider_id()
                    )))
                }
            };
            let wave_node_index = u64::try_from(wave_node_index)
                .map_err(|_| invalid_operation("reusable topology wave node index exceeds u64"))?;
            let plan_node_index = u64::try_from(plan_node_index)
                .map_err(|_| invalid_operation("reusable topology plan node index exceeds u64"))?;
            let node_id = node.id().as_str().as_bytes();
            let provider_id = provider.descriptor().provider_id().as_str().as_bytes();
            let node_id_len = u64::try_from(node_id.len())
                .map_err(|_| invalid_operation("reusable topology node id exceeds u64"))?;
            let provider_id_len = u64::try_from(provider_id.len())
                .map_err(|_| invalid_operation("reusable topology provider id exceeds u64"))?;
            let topology_len = u64::try_from(topology_bytes.len())
                .map_err(|_| invalid_operation("reusable topology payload exceeds u64"))?;
            digest.update(wave_node_index.to_le_bytes());
            digest.update(plan_node_index.to_le_bytes());
            digest.update(node_id_len.to_le_bytes());
            digest.update(node_id);
            digest.update(provider_id_len.to_le_bytes());
            digest.update(provider_id);
            digest.update([topology_kind]);
            digest.update(topology_len.to_le_bytes());
            digest.update(topology_bytes);
        }
        Ok(Some(ReusableExecutionWaveAuthority {
            program_id: program_id.with_topology_fingerprint(
                DeviceReusableExecutionTopologyFingerprint::from_sha256(digest.finalize().into()),
            ),
            eager_boundary_node_indices,
        }))
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit<R>(
        provider: &BoundOperationProvider<'_, R>,
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: &[TrustedActiveSequenceBinding],
        mut invocation_resources: InvocationResourceLease<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, OperationDispatchError<R>>
    where
        R: DeviceRuntime,
    {
        let node_identity = batch_identity.single_node().ok_or_else(|| {
            OperationDispatchError::Contract(invalid_operation(
                "single-operation dispatch requires a one-node batch identity",
            ))
        })?;
        provider
            .validate_binding(resolved, node_identity.node_id())
            .map_err(OperationDispatchError::Contract)?;
        if active_bindings.is_empty()
            || active_bindings.len() != batch_identity.participants().len()
            || lane.id() != batch_identity.lane_id()
            || lane.descriptor().id != *batch_identity.device_id()
            || lane.descriptor().runtime_implementation_fingerprint
                != batch_identity.runtime_implementation_fingerprint()
        {
            return Err(OperationDispatchError::Contract(invalid_operation(
                "operation execution lane or participant set differs from batch identity",
            )));
        }
        invocation_resources
            .begin_dispatch()
            .map_err(OperationDispatchError::Contract)?;
        let mut completion = CompletionReaper::reserve(
            reaper,
            invocation_resources,
            Arc::clone(lane),
            batch_identity.clone(),
        )
        .map_err(OperationDispatchError::Contract)?;
        let runtime = lane.runtime();
        if !lane.current_descriptor_matches_snapshot() {
            return Err(OperationDispatchError::Contract(invalid_operation(
                "operation encode runtime differs from its execution lane snapshot",
            )));
        }
        let mut commands = DeviceCommandBatch::with_capacity(3);
        completion
            .encode_backing_initializations(runtime, &mut commands)
            .map_err(|error| map_backing_initialization_error(runtime, batch_identity, error))?;
        let invocation = BatchedOperationInvocation::from_resolved(
            runtime,
            resolved,
            provider.dispatch(),
            batch_identity,
            completion.invocation(),
            active_bindings,
        )
        .map_err(OperationDispatchError::Contract)?;
        let plan_node = resolved
            .execution_plan()
            .payload()
            .nodes()
            .iter()
            .find(|node| node.id() == node_identity.node_id())
            .ok_or_else(|| {
                OperationDispatchError::Contract(invalid_operation(
                    "operation workspace node is absent from the immutable plan",
                ))
            })?;
        if let Some(requirement) = plan_node.provider_resources().scratch() {
            let scratch_view = invocation
                .participants()
                .first()
                .and_then(OperationInvocation::scratch_view)
                .ok_or_else(|| {
                    OperationDispatchError::Contract(invalid_operation(
                        "operation scratch requirement has no invocation view",
                    ))
                })?;
            encode_provider_workspace_initialization::<R, DefinitelyNotSubmittedRetryAuthority<R>>(
                runtime,
                0,
                node_identity,
                requirement,
                invocation.work_shape().resource_work(),
                scratch_view,
                SubmissionScratchInitialization::ProviderContract,
                &mut commands,
            )?;
        }
        let expected_phase = invocation.operation().profile_phase;
        let operation = match provider.provider().encode_selected(invocation) {
            Ok(operation) => operation,
            Err(failure)
                if batch_identity.contains_identity(failure.identity())
                    && failure.phase() == expected_phase =>
            {
                return Err(OperationDispatchError::Provider(failure));
            }
            Err(_) => {
                return Err(OperationDispatchError::Contract(invalid_operation(
                    "operation provider returned a failure for a different execution identity or profile phase",
                )));
            }
        };
        if !lane.current_descriptor_matches_snapshot() {
            return Err(OperationDispatchError::Contract(invalid_operation(
                "operation encode completion runtime drifted",
            )));
        }
        commands.push_operation(0, operation);
        let timing_mode = commands.timing_mode();
        let mut lane_reservation = lane
            .reserve_enqueue()
            .map_err(OperationDispatchError::Contract)?;
        completion.mark_submission_started();
        match lane_reservation.submit(commands) {
            LaneSubmitOutcome::DefinitelyNotSubmitted(error) => {
                drop(lane_reservation);
                let retry = completion
                    .definitely_not_submitted()
                    .map_err(OperationDispatchError::Contract)?;
                let failures = batch_identity
                    .participants()
                    .iter()
                    .map(|participant| {
                        classify_device_error(runtime, participant.identity().clone(), &error)
                    })
                    .collect::<Result<Vec<_>, _>>()
                    .map_err(OperationDispatchError::Contract)?;
                Err(OperationDispatchError::DefinitelyNotSubmitted { failures, retry })
            }
            LaneSubmitOutcome::PossiblySubmittedPanic => {
                drop(lane_reservation);
                let recovery = completion.submission_indeterminate();
                Err(OperationDispatchError::SubmissionIndeterminate { recovery })
            }
            LaneSubmitOutcome::Submitted(fence) => {
                drop(lane_reservation);
                let completion = match completion.arm(fence, timing_mode) {
                    Ok(completion) => completion,
                    Err((error, completion)) => {
                        return Err(OperationDispatchError::PostSubmitContract {
                            error,
                            completion,
                        });
                    }
                };
                if !lane.current_descriptor_matches_snapshot() {
                    lane.fail_closed();
                    return Err(OperationDispatchError::PostSubmitContract {
                        error: invalid_operation("operation submit completion runtime drifted"),
                        completion,
                    });
                }
                Ok(completion)
            }
        }
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_wave<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        Self::encode_and_submit_wave_with_inputs(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            &[],
            wave,
            lane,
            reaper,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_wave_with_inputs<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            SubmissionExecutionPolicy::adaptive(),
            None,
            None,
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )
        .map(|profiled| profiled.into_parts().0)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_wave_with_inputs_and_policy<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        execution_policy: SubmissionExecutionPolicy,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            execution_policy,
            None,
            None,
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )
        .map(|profiled| profiled.into_parts().0)
    }

    /// Dispatches one prepared wave while attributing host time to exact typed
    /// ownership boundaries. The diagnostic sink receives no access to the
    /// command, submission, completion, or failure value.
    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_wave_with_inputs_and_timing<'binding, R, I, S>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        execution_policy: SubmissionExecutionPolicy,
        timing_sink: &S,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<ProfiledSubmissionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
        S: SubmissionWaveDispatchTimingSink,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            execution_policy,
            None,
            None,
            timing_sink,
            wave,
            lane,
            reaper,
        )
    }

    /// Executes one complete plan-derived determinism restore through eager
    /// provider commands. The returned profiled handle retains physical path
    /// attribution for the hardware artifact.
    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_determinism_eager_wave<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        restore: &SubmissionWaveDeterminismRestore,
        scratch_fill: u8,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<SubmissionWaveDeterminismHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        let readback_plan = SubmissionWaveDeterminismReadbackPlan::from_restore(
            resolved,
            batch_identity,
            &wave,
            restore,
        )
        .map_err(SubmissionWaveDispatchError::Contract)?;
        let restore_fingerprint = restore
            .logical_fingerprint()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let initialization_identity = restore
            .initialization_identity()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let profiled = Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            &[],
            SubmissionExecutionPolicy::determinism_eager(scratch_fill),
            Some(restore),
            None,
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )?;
        Ok(SubmissionWaveDeterminismHandle::from_profiled_eager(
            profiled,
            readback_plan,
            restore_fingerprint,
            initialization_identity,
        ))
    }

    /// Executes the same complete restore through one sealed resident replay
    /// program. No adaptive capture or eager fallback is permitted.
    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_determinism_replayed_wave<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        restore: &SubmissionWaveDeterminismRestore,
        scratch_fill: u8,
        reusable_program: &DeviceReusableExecutionProgram,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<SubmissionWaveDeterminismHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        let readback_plan = SubmissionWaveDeterminismReadbackPlan::from_restore(
            resolved,
            batch_identity,
            &wave,
            restore,
        )
        .map_err(SubmissionWaveDispatchError::Contract)?;
        let restore_fingerprint = restore
            .logical_fingerprint()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let initialization_identity = restore
            .initialization_identity()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let replay_authority =
            Self::reusable_execution_wave_authority(providers, resolved, &wave, lane)
                .map_err(SubmissionWaveDispatchError::Contract)?
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "determinism replay wave has no reusable execution authority",
                    ))
                })?;
        if &replay_authority.program_id != reusable_program.program_id() {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "determinism replay program differs from the live wave topology",
            )));
        }
        let mut declared_eager_boundary_node_ids = replay_authority
            .eager_boundary_node_indices
            .iter()
            .map(|node_index| {
                usize::try_from(*node_index)
                    .ok()
                    .and_then(|node_index| batch_identity.node_id_at(node_index))
                    .cloned()
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "determinism eager boundary is absent from the batch identity",
                        ))
                    })
            })
            .collect::<Result<Vec<_>, _>>()?;
        declared_eager_boundary_node_ids.sort();
        let profiled = Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            &[],
            SubmissionExecutionPolicy::determinism_replayed(scratch_fill),
            Some(restore),
            Some(reusable_program),
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )?;
        SubmissionWaveDeterminismHandle::from_profiled_replayed(
            profiled,
            readback_plan,
            restore_fingerprint,
            initialization_identity,
            replay_authority.program_id,
            declared_eager_boundary_node_ids,
        )
        .map_err(SubmissionWaveDispatchError::Contract)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_reusable_wave_with_inputs<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        reusable_program: &DeviceReusableExecutionProgram,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            SubmissionExecutionPolicy::adaptive(),
            None,
            Some(reusable_program),
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )
        .map(|profiled| profiled.into_parts().0)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_reusable_wave_with_inputs_and_policy<'binding, R, I>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        reusable_program: &DeviceReusableExecutionProgram,
        execution_policy: SubmissionExecutionPolicy,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<CompletionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            execution_policy,
            None,
            Some(reusable_program),
            &DisabledSubmissionWaveDispatchTimingSink,
            wave,
            lane,
            reaper,
        )
        .map(|profiled| profiled.into_parts().0)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn encode_and_submit_reusable_wave_with_inputs_and_timing<'binding, R, I, S>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        reusable_program: &DeviceReusableExecutionProgram,
        execution_policy: SubmissionExecutionPolicy,
        timing_sink: &S,
        wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<ProfiledSubmissionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
        S: SubmissionWaveDispatchTimingSink,
    {
        Self::encode_and_submit_wave_with_inputs_timed(
            providers,
            resolved,
            batch_identity,
            active_bindings,
            timing_mode,
            input_uploads,
            execution_policy,
            None,
            Some(reusable_program),
            timing_sink,
            wave,
            lane,
            reaper,
        )
    }

    #[allow(clippy::too_many_arguments)]
    fn encode_and_submit_wave_with_inputs_timed<'binding, R, I, S>(
        providers: &[BoundOperationProvider<'_, R>],
        resolved: &dyn ExecutablePlanView,
        batch_identity: &BatchOperationIdentity,
        active_bindings: I,
        timing_mode: DeviceTimingMode,
        input_uploads: &[SubmissionWaveInputUpload],
        execution_policy: SubmissionExecutionPolicy,
        determinism_restore: Option<&SubmissionWaveDeterminismRestore>,
        reusable_program: Option<&DeviceReusableExecutionProgram>,
        timing_sink: &S,
        mut wave: PreparedStepSubmissionWave<R>,
        lane: &Arc<ExecutionLane<R>>,
        reaper: &Arc<CompletionReaper<R>>,
    ) -> Result<ProfiledSubmissionHandle<R>, SubmissionWaveDispatchError<R>>
    where
        R: DeviceRuntime,
        I: Clone + ExactSizeIterator<Item = &'binding TrustedActiveSequenceBinding>,
        S: SubmissionWaveDispatchTimingSink,
    {
        let contract_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::ContractValidateAndReserve,
        );
        let active_participant_count = active_bindings.len();
        if providers.is_empty()
            || providers.len() != wave.nodes().len()
            || providers.len() != batch_identity.node_count()
            || active_participant_count == 0
            || batch_identity.batch_step_id() != wave.batch_step_id()
            || batch_identity.batch_invocation_id() != wave.batch_invocation_id()
            || batch_identity.claimed_backing_fingerprint() != wave.fingerprint()
            || lane.id() != batch_identity.lane_id()
            || lane.descriptor().id != *batch_identity.device_id()
            || lane.descriptor().runtime_implementation_fingerprint
                != batch_identity.runtime_implementation_fingerprint()
        {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "wave execution lane, resources, nodes, or participants differ from batch identity",
            )));
        }
        if !matches!(
            (wave.purpose(), determinism_restore.is_some()),
            (SubmissionWavePurpose::FullPlan, false)
                | (SubmissionWavePurpose::DeterminismProbe, true)
        ) {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "submission wave purpose differs from its product or determinism dispatch path",
            )));
        }
        if let Some(restore) = determinism_restore {
            restore
                .validate_for_submission(
                    lane.runtime(),
                    providers,
                    resolved,
                    batch_identity,
                    active_bindings.clone(),
                    &wave,
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
            if !input_uploads.is_empty()
                || usize::try_from(restore.participant_count()).ok()
                    != Some(active_participant_count)
                || execution_policy.compute_path() == DeviceComputePathRequirement::Adaptive
                || execution_policy.scratch_initialization()
                    == SubmissionScratchInitialization::ProviderContract
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "determinism submission requires complete restore coverage, explicit scratch fill, and one forced compute path",
                )));
            }
        }
        for (node_index, (provider, prepared_node)) in
            providers.iter().zip(wave.nodes()).enumerate()
        {
            let node_id = batch_identity.node_id_at(node_index).ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "physical batch is missing a compiled plan node",
                ))
            })?;
            provider
                .validate_binding(resolved, node_id)
                .map_err(SubmissionWaveDispatchError::Contract)?;
            if prepared_node.node_id() != node_id
                || prepared_node.work_shape().fingerprint()
                    != batch_identity
                        .work_shape_fingerprint_at(node_index)
                        .expect("compiled physical batch node has work identity")
                || Some(provider.descriptor().provider_id())
                    != batch_identity.provider_id_at(node_index)
                || Some(provider.descriptor().operation_id())
                    != batch_identity.operation_id_at(node_index)
                || batch_identity.node_participant_count(node_index)
                    != Some(
                        usize::try_from(prepared_node.participant_count())
                            .expect("prepared wave participant count fits usize"),
                    )
                || batch_identity.node_participant_count(node_index)
                    != Some(active_participant_count)
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "wave provider or node identity differs from its prepared node",
                )));
            }
        }
        let reusable_execution_authority =
            Self::reusable_execution_wave_authority(providers, resolved, &wave, lane)
                .map_err(SubmissionWaveDispatchError::Contract)?;
        let mut effective_compute_path = execution_policy.compute_path();
        let mut declared_eager_compute_node_indices = Vec::new();
        if let Some(reusable_program) = reusable_program {
            if !timing_mode.direct_reusable_execution_allowed()
                || execution_policy.compute_path() == DeviceComputePathRequirement::EagerOnly
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission timing or compute-path policy requires eager provider encoding",
                )));
            }
            if execution_policy.compute_path() == DeviceComputePathRequirement::Adaptive
                && !reusable_program.has_resident_segments()
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "product reusable execution requires at least one resident segment",
                )));
            }
            let actual_authority = reusable_execution_authority.as_ref().ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "reusable execution program has no live program binding authority",
                ))
            })?;
            if &actual_authority.program_id != reusable_program.program_id()
                || reusable_program.segments().iter().any(|segment| {
                    segment.end_node_index() as usize > providers.len()
                        || segment.logical_command_count()
                            != segment.end_node_index() - segment.start_node_index()
                })
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "reusable execution program differs from the exact wave topology",
                )));
            }
            if execution_policy.compute_path() == DeviceComputePathRequirement::ReplayedOnly {
                validate_determinism_replay_coverage(
                    reusable_program,
                    providers.len(),
                    &actual_authority.eager_boundary_node_indices,
                    batch_identity,
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
                if !actual_authority.eager_boundary_node_indices.is_empty() {
                    effective_compute_path =
                        DeviceComputePathRequirement::ReplayedWithDeclaredEagerBoundaries;
                    declared_eager_compute_node_indices =
                        actual_authority.eager_boundary_node_indices.clone();
                }
            }
        } else if matches!(
            execution_policy.compute_path(),
            DeviceComputePathRequirement::ReplayedOnly
                | DeviceComputePathRequirement::ReplayedWithDeclaredEagerBoundaries
        ) {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "replay-required submission requires one sealed reusable execution program",
            )));
        }
        wave.begin_dispatch()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let mut completion =
            CompletionReaper::reserve_wave(reaper, wave, Arc::clone(lane), batch_identity.clone())
                .map_err(SubmissionWaveDispatchError::Contract)?;
        let runtime = lane.runtime();
        if !lane.current_descriptor_matches_snapshot() {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "wave encode runtime differs from its execution lane snapshot",
            )));
        }
        drop(contract_stage);

        let backing_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::BackingAndInputEncode,
        );
        let restore_capacity = determinism_restore.map_or(0, |restore| {
            restore
                .initializations()
                .len()
                .saturating_mul(active_participant_count)
        });
        let mut commands = DeviceCommandBatch::with_capacity_timing_and_compute_path(
            providers
                .len()
                .saturating_add(input_uploads.len())
                .saturating_add(restore_capacity),
            timing_mode,
            effective_compute_path,
        );
        if effective_compute_path
            == DeviceComputePathRequirement::ReplayedWithDeclaredEagerBoundaries
        {
            commands
                .set_declared_eager_compute_node_indices(declared_eager_compute_node_indices)
                .map_err(SubmissionWaveDispatchError::Contract)?;
        }
        if determinism_restore.is_some() {
            commands.require_logical_execution_path_attribution();
        }
        let backing_initialization_command_count = completion
            .encode_backing_initializations(runtime, &mut commands)
            .map_err(|error| map_backing_initialization_error(runtime, batch_identity, error))?;
        let workspace_initialization_command_count =
            encode_submission_wave_workspace_initializations(
                runtime,
                resolved,
                batch_identity,
                execution_policy.scratch_initialization(),
                &completion,
                &mut commands,
            )?;
        if commands.len()
            != backing_initialization_command_count
                .saturating_add(workspace_initialization_command_count)
        {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "backing initialization command accounting differs from encoded commands",
            )));
        }
        if let Some(restore) = determinism_restore {
            let restore_start = commands.len();
            let restore_command_count = encode_submission_wave_determinism_restore(
                runtime,
                resolved,
                batch_identity,
                &completion,
                restore,
                &mut commands,
            )?;
            if commands.len()
                != restore_start
                    .checked_add(restore_command_count)
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "determinism restore command accounting overflows usize",
                        ))
                    })?
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "determinism restore command accounting differs from encoded commands",
                )));
            }
        }
        encode_submission_wave_inputs(
            runtime,
            resolved,
            batch_identity,
            &completion,
            input_uploads,
            &mut commands,
        )?;
        drop(backing_stage);

        let provider_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::ProviderNodeEncode,
        );
        let pre_provider_command_count = commands.len();
        let mut encoded_provider_command_count = 0_usize;
        let mut program_bindings = Vec::new();
        let mut program_binding_resources = BTreeSet::new();
        let mut reusable_execution_binding_nodes = Vec::new();
        let mut encoded_operations = Vec::with_capacity(providers.len());
        if let Some(reusable_program) = reusable_program {
            let mut node_index = 0_usize;
            let mut segment_index = 0_usize;
            while node_index < providers.len() {
                let segment = reusable_program
                    .segments()
                    .get(segment_index)
                    .filter(|segment| segment.start_node_index() as usize == node_index);
                if let Some(segment) = segment {
                    let mut segment_dynamic_bindings = Vec::new();
                    let mut segment_result_bindings = Vec::new();
                    for binding_node_index in reusable_program
                        .per_wave_binding_node_indices()
                        .iter()
                        .copied()
                        .filter(|binding_node_index| segment.contains_node(*binding_node_index))
                    {
                        let binding_node_index =
                            usize::try_from(binding_node_index).map_err(|_| {
                                SubmissionWaveDispatchError::Contract(invalid_operation(
                                    "reusable execution binding node exceeds usize",
                                ))
                            })?;
                        let provider = &providers[binding_node_index];
                        let node_identity = batch_identity
                            .materialize_node(binding_node_index)
                            .map_err(SubmissionWaveDispatchError::Contract)?;
                        let invocation = BatchedOperationInvocation::from_wave_node(
                            runtime,
                            resolved,
                            provider.dispatch(),
                            batch_identity,
                            node_identity,
                            completion.wave(),
                            binding_node_index,
                            active_bindings.clone(),
                        )
                        .map_err(SubmissionWaveDispatchError::Contract)?;
                        let expected_phase = invocation.operation().profile_phase;
                        let program_binding = invocation.program_binding().cloned();
                        let bindings = match provider
                            .provider()
                            .encode_reusable_execution_bindings(invocation)
                        {
                            Ok(bindings) => bindings,
                            Err(failure)
                                if node_identity.contains_identity(failure.identity())
                                    && failure.phase() == expected_phase =>
                            {
                                return Err(SubmissionWaveDispatchError::Provider(failure));
                            }
                            Err(_) => {
                                return Err(SubmissionWaveDispatchError::Contract(
                                    invalid_operation(
                                        "reusable provider returned a failure for another node identity or profile phase",
                                    ),
                                ));
                            }
                        };
                        let program_binding_count = bindings.program_binding_count();
                        validate_program_binding_patch(
                            resolved,
                            node_identity,
                            program_binding.as_ref(),
                            program_binding_count,
                            &mut program_binding_resources,
                        )
                        .map_err(SubmissionWaveDispatchError::Contract)?;
                        let (mut node_program_bindings, mut dynamic_bindings, mut result_bindings) =
                            bindings.into_parts();
                        program_bindings.append(&mut node_program_bindings);
                        segment_dynamic_bindings.append(&mut dynamic_bindings);
                        segment_result_bindings.append(&mut result_bindings);
                    }
                    let invocation = DeviceReusableExecutionInvocation::new(
                        reusable_program.program_id().clone(),
                        segment.clone(),
                        u32::try_from(active_participant_count).map_err(|_| {
                            SubmissionWaveDispatchError::Contract(invalid_operation(
                                "reusable execution participant count exceeds u32",
                            ))
                        })?,
                        completion
                            .wave()
                            .claimed_backing()
                            .work_shape()
                            .immediate_tokens(),
                    )
                    .map_err(SubmissionWaveDispatchError::Contract)?;
                    let compute = runtime
                        .encode_reusable_execution(invocation)
                        .map_err(|error| {
                            SubmissionWaveDispatchError::Contract(invalid_operation(format!(
                                "device runtime rejected a sealed reusable program: {error}"
                            )))
                        })?
                        .ok_or_else(|| {
                            SubmissionWaveDispatchError::Contract(invalid_operation(
                                "device runtime published a reusable program it cannot encode",
                            ))
                        })?;
                    let operation_command_count = segment_dynamic_bindings
                        .len()
                        .checked_add(1)
                        .and_then(|count| count.checked_add(segment_result_bindings.len()))
                        .ok_or_else(|| {
                            SubmissionWaveDispatchError::Contract(invalid_operation(
                                "reusable execution command count overflows usize",
                            ))
                        })?;
                    encoded_provider_command_count = encoded_provider_command_count
                        .checked_add(operation_command_count)
                        .ok_or_else(|| {
                            SubmissionWaveDispatchError::Contract(invalid_operation(
                                "submission wave command count overflows usize",
                            ))
                        })?;
                    encoded_operations.push((
                        segment.start_node_index(),
                        segment_dynamic_bindings,
                        compute,
                        segment_result_bindings,
                    ));
                    node_index = segment.end_node_index() as usize;
                    segment_index += 1;
                    continue;
                }
                if reusable_program
                    .segments()
                    .get(segment_index)
                    .is_some_and(|segment| (segment.start_node_index() as usize) < node_index)
                {
                    return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                        "reusable execution segment traversal is not canonical",
                    )));
                }
                let provider = &providers[node_index];
                let node_identity = batch_identity
                    .materialize_node(node_index)
                    .map_err(SubmissionWaveDispatchError::Contract)?;
                let invocation = BatchedOperationInvocation::from_wave_node(
                    runtime,
                    resolved,
                    provider.dispatch(),
                    batch_identity,
                    node_identity,
                    completion.wave(),
                    node_index,
                    active_bindings.clone(),
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
                let expected_phase = invocation.operation().profile_phase;
                let program_binding = invocation.program_binding().cloned();
                let operation = match provider.provider().encode_selected(invocation) {
                    Ok(operation) => operation,
                    Err(failure)
                        if node_identity.contains_identity(failure.identity())
                            && failure.phase() == expected_phase =>
                    {
                        return Err(SubmissionWaveDispatchError::Provider(failure));
                    }
                    Err(_) => {
                        return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                            "wave provider returned a failure for another node identity or profile phase",
                        )));
                    }
                };
                let program_binding_count = operation.program_binding_count();
                validate_program_binding_patch(
                    resolved,
                    node_identity,
                    program_binding.as_ref(),
                    program_binding_count,
                    &mut program_binding_resources,
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
                let operation_command_count = operation
                    .dynamic_binding_count()
                    .checked_add(1)
                    .and_then(|count| count.checked_add(operation.result_binding_count()))
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "provider operation command count overflows usize",
                        ))
                    })?;
                encoded_provider_command_count = encoded_provider_command_count
                    .checked_add(operation_command_count)
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "submission wave command count overflows usize",
                        ))
                    })?;
                let encoded_node_index = u32::try_from(node_index).map_err(|_| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission wave node index exceeds u32",
                    ))
                })?;
                let (mut node_program_bindings, dynamic_bindings, compute, result_bindings) =
                    operation.into_parts();
                program_bindings.append(&mut node_program_bindings);
                encoded_operations.push((
                    encoded_node_index,
                    dynamic_bindings,
                    compute,
                    result_bindings,
                ));
                node_index += 1;
            }
            if segment_index != reusable_program.segments().len() {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "reusable execution program contains an unreachable segment",
                )));
            }
        } else {
            for (node_index, (provider, node_identity)) in
                providers.iter().zip(batch_identity.nodes()).enumerate()
            {
                let invocation = BatchedOperationInvocation::from_wave_node(
                    runtime,
                    resolved,
                    provider.dispatch(),
                    batch_identity,
                    node_identity,
                    completion.wave(),
                    node_index,
                    active_bindings.clone(),
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
                let expected_phase = invocation.operation().profile_phase;
                let program_binding = invocation.program_binding().cloned();
                let operation = match provider.provider().encode_selected(invocation) {
                    Ok(operation) => operation,
                    Err(failure)
                        if node_identity.contains_identity(failure.identity())
                            && failure.phase() == expected_phase =>
                    {
                        return Err(SubmissionWaveDispatchError::Provider(failure));
                    }
                    Err(_) => {
                        return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                        "wave provider returned a failure for another node identity or profile phase",
                    )));
                    }
                };
                let program_binding_count = operation.program_binding_count();
                validate_program_binding_patch(
                    resolved,
                    node_identity,
                    program_binding.as_ref(),
                    program_binding_count,
                    &mut program_binding_resources,
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
                let operation_command_count = operation
                    .dynamic_binding_count()
                    .checked_add(1)
                    .and_then(|count| count.checked_add(operation.result_binding_count()))
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "provider operation command count overflows usize",
                        ))
                    })?;
                let has_per_wave_bindings = program_binding_count > 0
                    || operation.dynamic_binding_count() > 0
                    || operation.result_binding_count() > 0;
                encoded_provider_command_count = encoded_provider_command_count
                    .checked_add(operation_command_count)
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "submission wave command count overflows usize",
                        ))
                    })?;
                let node_index = u32::try_from(node_index).map_err(|_| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission wave node index exceeds u32",
                    ))
                })?;
                if has_per_wave_bindings {
                    reusable_execution_binding_nodes.push(node_index);
                }
                let (mut node_program_bindings, dynamic_bindings, compute, result_bindings) =
                    operation.into_parts();
                program_bindings.append(&mut node_program_bindings);
                encoded_operations.push((node_index, dynamic_bindings, compute, result_bindings));
            }
        }
        if let Some(layout) = completion.wave().claimed_backing().program_binding_layout() {
            let selected_plan_nodes = completion
                .wave()
                .nodes()
                .iter()
                .map(|node| node.plan_node_index())
                .collect::<BTreeSet<_>>();
            let expected_resources = layout
                .slots()
                .iter()
                .filter(|slot| selected_plan_nodes.contains(&slot.node_index()))
                .map(|slot| slot.resource_id())
                .collect::<BTreeSet<_>>();
            if program_binding_resources.len() != expected_resources.len()
                || expected_resources
                    .iter()
                    .any(|resource| !program_binding_resources.contains(*resource))
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "provider program binding patches do not cover the selected compiled layout exactly",
                )));
            }
        }
        let uncoalesced_program_binding_count = program_bindings.len();
        let coalesced_program_bindings = runtime
            .coalesce_program_bindings(program_bindings)
            .map_err(|error| {
                SubmissionWaveDispatchError::Contract(invalid_operation(format!(
                    "device runtime could not coalesce program bindings: {error}"
                )))
            })?;
        if (uncoalesced_program_binding_count == 0) != coalesced_program_bindings.is_empty()
            || coalesced_program_bindings.len() > uncoalesced_program_binding_count
        {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "device runtime changed the program binding boundary cardinality illegally",
            )));
        }
        encoded_provider_command_count = encoded_provider_command_count
            .checked_add(coalesced_program_bindings.len())
            .ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "coalesced program binding command count overflows usize",
                ))
            })?;
        for command in coalesced_program_bindings {
            commands.push_dynamic_binding(command);
        }
        for (node_index, dynamic_bindings, compute, result_bindings) in encoded_operations {
            commands.push_operation_parts(node_index, dynamic_bindings, compute, result_bindings);
        }
        if reusable_program.is_none() {
            if let Some(authority) = reusable_execution_authority {
                commands
                    .set_reusable_execution_capture(
                        DeviceReusableExecutionCapture::new(
                            authority.program_id,
                            u32::try_from(providers.len()).map_err(|_| {
                                SubmissionWaveDispatchError::Contract(invalid_operation(
                                    "reusable execution capture node count exceeds u32",
                                ))
                            })?,
                            authority.eager_boundary_node_indices,
                            reusable_execution_binding_nodes,
                        )
                        .map_err(SubmissionWaveDispatchError::Contract)?,
                    )
                    .map_err(SubmissionWaveDispatchError::Contract)?;
            }
        }
        if commands.len()
            != pre_provider_command_count.saturating_add(encoded_provider_command_count)
            || !lane.current_descriptor_matches_snapshot()
        {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "wave encode command phases differ from provider-declared operation boundaries",
            )));
        }
        drop(provider_stage);

        let lane_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::LaneReserveSubmitAndArm,
        );
        let lane_reserve_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::LaneReserve,
        );
        let mut lane_reservation = lane
            .reserve_enqueue()
            .map_err(SubmissionWaveDispatchError::Contract)?;
        drop(lane_reserve_stage);

        let device_submit_stage = SubmissionWaveDispatchStageTimer::start(
            timing_sink,
            SubmissionWaveDispatchStage::DeviceRuntimeSubmit,
        );
        completion.mark_submission_started();
        let submit_outcome = lane_reservation.submit_with_timing(commands, timing_sink);
        drop(lane_reservation);
        drop(device_submit_stage);

        let outcome = match submit_outcome {
            LaneSubmitOutcome::DefinitelyNotSubmitted(error) => {
                let retry = completion
                    .definitely_not_submitted_wave()
                    .map_err(SubmissionWaveDispatchError::Contract)?;
                let failures = batch_identity
                    .participants()
                    .iter()
                    .map(|participant| {
                        classify_device_error(runtime, participant.identity().clone(), &error)
                    })
                    .collect::<Result<Vec<_>, _>>()
                    .map_err(SubmissionWaveDispatchError::Contract)?;
                Err(SubmissionWaveDispatchError::DefinitelyNotSubmitted { failures, retry })
            }
            LaneSubmitOutcome::PossiblySubmittedPanic => {
                let recovery = completion.submission_indeterminate();
                Err(SubmissionWaveDispatchError::SubmissionIndeterminate { recovery })
            }
            LaneSubmitOutcome::Submitted(fence) => {
                let device_attribution = runtime.submission_attribution(&fence);
                let completion_arm_stage = SubmissionWaveDispatchStageTimer::start(
                    timing_sink,
                    SubmissionWaveDispatchStage::CompletionArm,
                );
                let completion = match completion.arm(fence, timing_mode) {
                    Ok(completion) => completion,
                    Err((error, completion)) => {
                        return Err(SubmissionWaveDispatchError::PostSubmitContract {
                            error,
                            completion,
                        });
                    }
                };
                if !lane.current_descriptor_matches_snapshot() {
                    lane.fail_closed();
                    return Err(SubmissionWaveDispatchError::PostSubmitContract {
                        error: invalid_operation("wave submit completion runtime drifted"),
                        completion,
                    });
                }
                let attribution = match device_attribution
                    .map(|device| {
                        BoundDeviceSubmissionAttribution::new(
                            batch_identity.clone(),
                            completion.receipt().fingerprint().to_owned(),
                            device,
                        )
                    })
                    .transpose()
                {
                    Ok(attribution) => attribution,
                    Err(error) => {
                        return Err(SubmissionWaveDispatchError::PostSubmitContract {
                            error,
                            completion,
                        });
                    }
                };
                drop(completion_arm_stage);
                Ok(ProfiledSubmissionHandle::new(completion, attribution))
            }
        };
        drop(lane_stage);
        outcome
    }
}

fn map_backing_initialization_error<R, Retry>(
    runtime: &R,
    batch_identity: &BatchOperationIdentity,
    error: BackingInitializationEncodeError<R::Error>,
) -> OperationDispatchError<R, Retry>
where
    R: DeviceRuntime,
    Retry: DispatchRetryAuthority,
{
    match error {
        BackingInitializationEncodeError::Contract(error) => {
            OperationDispatchError::Contract(error)
        }
        BackingInitializationEncodeError::Runtime { participant, error } => {
            let identity = batch_identity
                .nodes()
                .iter()
                .flat_map(BatchOperationNodeIdentity::participants)
                .find(|candidate| {
                    candidate.node_key().sequence_authority() == participant.sequence_authority()
                        && candidate.node_key().request_authority()
                            == participant.request_authority()
                })
                .map(BatchOperationParticipantIdentity::identity)
                .cloned();
            let Some(identity) = identity else {
                return OperationDispatchError::Contract(invalid_operation(
                    "backing initialization failure has no matching batch participant",
                ));
            };
            match classify_device_error(runtime, identity, &error) {
                Ok(failure) => OperationDispatchError::Initialization(failure),
                Err(error) => OperationDispatchError::Contract(error),
            }
        }
    }
}

fn encode_submission_wave_determinism_restore<R>(
    runtime: &R,
    resolved: &dyn ExecutablePlanView,
    batch_identity: &BatchOperationIdentity,
    completion: &CompletionReservation<R>,
    restore: &SubmissionWaveDeterminismRestore,
    commands: &mut DeviceCommandBatch<R::Command>,
) -> Result<usize, SubmissionWaveDispatchError<R>>
where
    R: DeviceRuntime,
{
    restore
        .validate_for(resolved)
        .map_err(SubmissionWaveDispatchError::Contract)?;
    let participant_count = usize::try_from(restore.participant_count()).map_err(|_| {
        SubmissionWaveDispatchError::Contract(invalid_operation(
            "determinism restore participant count exceeds host address space",
        ))
    })?;
    let mut command_count = 0_usize;
    for (initialization_index, initialization) in restore.initializations().iter().enumerate() {
        let location = initialization.location();
        if !initialization
            .consumer_node_ids()
            .iter()
            .any(|node_id| node_id == location.node_id())
            || initialization.consumer_node_ids().iter().any(|node_id| {
                batch_identity
                    .node_index(node_id)
                    .and_then(|index| batch_identity.node_participant_count(index))
                    != Some(participant_count)
            })
        {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "determinism restore consumer coverage differs from its submitted wave",
            )));
        }
        let node_index = batch_identity
            .node_index(location.node_id())
            .ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "determinism restore anchor node is absent from its submitted wave",
                ))
            })?;
        let node_index_u32 = u32::try_from(node_index).map_err(|_| {
            SubmissionWaveDispatchError::Contract(invalid_operation(
                "determinism restore node index exceeds u32",
            ))
        })?;
        let node_identity = batch_identity
            .materialize_node(node_index)
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let prepared_node = completion.wave().nodes().get(node_index).ok_or_else(|| {
            SubmissionWaveDispatchError::Contract(invalid_operation(
                "determinism restore anchor node is absent from its prepared wave",
            ))
        })?;
        for participant_index in 0..participant_count {
            let participant_index_u32 = u32::try_from(participant_index).map_err(|_| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "determinism restore participant index exceeds u32",
                ))
            })?;
            let participant = node_identity
                .participants()
                .get(participant_index)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "determinism restore participant is absent from its anchor node",
                    ))
                })?;
            let participant_work = prepared_node
                .work_shape()
                .participant_work()
                .get(participant_index)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "determinism restore participant lacks exact token work",
                    ))
                })?;
            let logical_work = DeviceCommandLogicalWork::for_participant_range(
                DeviceBatchingForm::Scalar,
                participant_index_u32,
                1,
                participant_work.token_span().immediate_tokens(),
            )
            .map_err(SubmissionWaveDispatchError::Contract)?;
            let bytes = restore
                .participant_payloads(participant_index_u32)
                .and_then(|payloads| payloads.get(initialization_index))
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "determinism restore payload matrix is incomplete",
                    ))
                })?;
            let range = restore
                .initialization_range(participant_index_u32, initialization_index)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "determinism restore range matrix is incomplete",
                    ))
                })?;
            let backing = completion
                .backing_view(
                    location.node_id(),
                    participant_index_u32,
                    location.resource_id(),
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
            let encoded = encode_submission_wave_backing_upload(
                runtime,
                participant.identity(),
                &backing,
                location.usage(),
                location.element_type(),
                range.logical_offset_bytes(),
                bytes,
                "determinism restore",
                |command| {
                    commands.push_node_initialization(node_index_u32, logical_work, command);
                },
            )?;
            command_count = command_count.checked_add(encoded).ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "determinism restore command count overflows usize",
                ))
            })?;
        }
    }
    Ok(command_count)
}

fn encode_submission_wave_inputs<R>(
    runtime: &R,
    resolved: &dyn ExecutablePlanView,
    batch_identity: &BatchOperationIdentity,
    completion: &CompletionReservation<R>,
    uploads: &[SubmissionWaveInputUpload],
    commands: &mut DeviceCommandBatch<R::Command>,
) -> Result<(), SubmissionWaveDispatchError<R>>
where
    R: DeviceRuntime,
{
    struct ValidatedInputUpload<'a> {
        upload: &'a SubmissionWaveInputUpload,
        participant: &'a BatchOperationParticipantIdentity,
        destination: std::ops::Range<u64>,
    }

    let mut upload_cursor = 0_usize;
    while upload_cursor < uploads.len() {
        let first_upload = &uploads[upload_cursor];
        let mut run_end = upload_cursor + 1;
        while uploads.get(run_end).is_some_and(|upload| {
            upload.node_id() == first_upload.node_id()
                && upload.input_ordinal() == first_upload.input_ordinal()
        }) {
            run_end += 1;
        }
        let node = resolved
            .execution_plan()
            .payload()
            .nodes()
            .iter()
            .find(|node| node.id() == first_upload.node_id())
            .ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload references an unknown plan node",
                ))
            })?;
        let identity_node_index = batch_identity
            .node_index(first_upload.node_id())
            .ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload has no physical node identity",
                ))
            })?;
        let node_identity = batch_identity
            .materialize_node(identity_node_index)
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let value = node
            .values()
            .iter()
            .find(|value| {
                value.role() == ResolvedValueRole::Input
                    && value.ordinal() == first_upload.input_ordinal()
            })
            .ok_or_else(|| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload references an unknown node input",
                ))
            })?;
        let [component] = value.storage().components() else {
            return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                "submission input upload requires one activation storage component",
            )));
        };
        let descriptor = completion
            .wave()
            .step_resources()
            .dynamic_descriptor(component.resource_id())
            .map_err(SubmissionWaveDispatchError::Contract)?;
        let participant_packed = descriptor.lifetime() == AllocationLifetime::Step
            && descriptor.kind() == &AllocationKind::Value;
        let prepared_work_shape = if participant_packed {
            Some(
                completion
                    .wave()
                    .nodes()
                    .iter()
                    .find(|wave_node| wave_node.node_id() == first_upload.node_id())
                    .ok_or_else(|| {
                        SubmissionWaveDispatchError::Contract(invalid_operation(
                            "submission input upload has no prepared wave node",
                        ))
                    })?
                    .work_shape(),
            )
        } else {
            None
        };

        let mut validated = Vec::with_capacity(run_end - upload_cursor);
        for upload in &uploads[upload_cursor..run_end] {
            let participant_index = usize::try_from(upload.participant_index()).map_err(|_| {
                SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload participant index exceeds host address space",
                ))
            })?;
            let participant = node_identity
                .participants()
                .get(participant_index)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission input upload participant is absent from its plan node",
                    ))
                })?;
            let byte_len = upload.source_layout().byte_len().map_err(|error| {
                SubmissionWaveDispatchError::Contract(invalid_operation(error.to_string()))
            })?;
            let value_end = upload
                .logical_offset_bytes()
                .checked_add(byte_len)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission input upload value range overflows",
                    ))
                })?;
            if value.usage() != BufferUsage::Activations
                || !matches!(value.access(), TensorAccess::Read | TensorAccess::ReadWrite)
                || value.tensor().element_type() != upload.source_layout().element_type()
                || component.element_type() != upload.source_layout().element_type()
                || value_end > component.length_bytes()
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload differs from its resolved activation binding",
                )));
            }
            let semantic_destination_start = component
                .offset_bytes()
                .checked_add(upload.logical_offset_bytes())
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission input upload destination overflows",
                    ))
                })?;
            let semantic_destination_end = semantic_destination_start
                .checked_add(byte_len)
                .ok_or_else(|| {
                    SubmissionWaveDispatchError::Contract(invalid_operation(
                        "submission input upload destination range overflows",
                    ))
                })?;
            let destination = if let Some(work_shape) = prepared_work_shape {
                translate_step_participant_upload_range(
                    descriptor.demand(),
                    work_shape,
                    participant_index,
                    semantic_destination_start..semantic_destination_end,
                )
                .map_err(SubmissionWaveDispatchError::Contract)?
            } else {
                semantic_destination_start..semantic_destination_end
            };
            if destination.end.checked_sub(destination.start) != Some(byte_len) {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload translated range differs from its source layout",
                )));
            }
            let backing = completion
                .backing_view(
                    upload.node_id(),
                    upload.participant_index(),
                    component.resource_id(),
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
            if backing.usage() != BufferUsage::Activations
                || backing.element_type() != upload.source_layout().element_type()
                || destination.end > backing.size_bytes()
            {
                return Err(SubmissionWaveDispatchError::Contract(invalid_operation(
                    "submission input upload backing differs from its resolved activation",
                )));
            }
            validated.push(ValidatedInputUpload {
                upload,
                participant,
                destination,
            });
        }

        let mut validated_cursor = 0_usize;
        while validated_cursor < validated.len() {
            let mut contiguous_end = validated_cursor + 1;
            if participant_packed {
                while contiguous_end < validated.len()
                    && validated[contiguous_end - 1].destination.end
                        == validated[contiguous_end].destination.start
                {
                    contiguous_end += 1;
                }
            }
            let first = &validated[validated_cursor];
            let backing = completion
                .backing_view(
                    first.upload.node_id(),
                    first.upload.participant_index(),
                    component.resource_id(),
                )
                .map_err(SubmissionWaveDispatchError::Contract)?;
            if contiguous_end == validated_cursor + 1 {
                encode_submission_wave_backing_upload(
                    runtime,
                    first.participant.identity(),
                    &backing,
                    BufferUsage::Activations,
                    first.upload.source_layout().element_type(),
                    first.destination.start,
                    first.upload.bytes(),
                    "submission input upload",
                    |command| commands.push_dynamic_binding(command),
                )?;
            } else {
                let aggregate_byte_len = validated[validated_cursor..contiguous_end]
                    .iter()
                    .try_fold(0_usize, |total, input| {
                        total
                            .checked_add(input.upload.bytes().len())
                            .ok_or_else(|| {
                                SubmissionWaveDispatchError::Contract(invalid_operation(
                                    "submission input upload aggregate byte count overflows usize",
                                ))
                            })
                    })?;
                let mut aggregate_bytes = Vec::with_capacity(aggregate_byte_len);
                for input in &validated[validated_cursor..contiguous_end] {
                    aggregate_bytes.extend_from_slice(input.upload.bytes());
                }
                encode_submission_wave_backing_upload(
                    runtime,
                    first.participant.identity(),
                    &backing,
                    BufferUsage::Activations,
                    first.upload.source_layout().element_type(),
                    first.destination.start,
                    &aggregate_bytes,
                    "submission input upload aggregate",
                    |command| commands.push_dynamic_binding(command),
                )?;
            }
            validated_cursor = contiguous_end;
        }
        upload_cursor = run_end;
    }
    Ok(())
}