Compose cold-start resources under Runtime Supervisor ownership

crates/lazaret-host-coordinator/src/bootstrap.rs

@@ -0,0 +1,483 @@
+use std::fmt;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use lazaret_core::configuration::ResolvedDomainConfiguration;
+use lazaret_core::identity::{BundleDigest, ContainerId, MountIdentity};
+use lazaret_core::network::{CanonicalDerivedNetworkPolicy, ReachabilityExpectation};
+use lazaret_core::records::Session;
+use lazaret_core::records::{Generation, GenerationIdentity, PreparedArtifact};
+use lazaret_linux::process::current_process_identity;
+use lazaret_runtime_supervisor::network::NetworkProbeRequest;
+use lazaret_runtime_supervisor::sandbox::SandboxBundle;
+use lazaret_runtime_supervisor::supervision::RuntimeSupervisorLaunch;
+use lazaret_state_anchor::StateAnchorLayout;
+use lazaret_storage_catalog::Catalog;
+
+use crate::control::ControlSessionAdmission;
+use crate::control::{PublishedControlFace, SessionRegistry, StagedControlFace, TrustAnchor};
+use crate::domain_store::HostCoordinatorDomainStore;
+use crate::lifecycle::{
+    HostCoordinatorLifecycle, LifecycleAction, LifecycleAdapter, SessionAdmissionIdentity,
+    SessionLifecycleAdapter,
+};
+use crate::network::NetworkCoordinator;
+use crate::runtime::{AwaitingAttachedRunsc, CommittedSandboxRuntime, SandboxCoordinator};
+use crate::state_anchor::{StateAnchorCoordinator, StateAnchorReleaseAuthority};
+
+#[derive(Debug)]
+pub struct ProductionLifecycleError(String);
+
+impl fmt::Display for ProductionLifecycleError {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter.write_str(&self.0)
+    }
+}
+
+impl std::error::Error for ProductionLifecycleError {}
+
+fn failure(error: impl fmt::Display) -> ProductionLifecycleError {
+    ProductionLifecycleError(error.to_string())
+}
+
+pub struct ProductionBootstrapPlan<'a> {
+    pub store: Arc<HostCoordinatorDomainStore>,
+    pub catalog: &'a Catalog,
+    pub configuration: &'a ResolvedDomainConfiguration,
+    pub generation_identity: GenerationIdentity,
+    pub prepared_artifact: &'a PreparedArtifact,
+    pub state_anchor_layout: &'a StateAnchorLayout<'a>,
+    pub runtime_launch: RuntimeSupervisorLaunch,
+    pub sandbox_bundle: &'a SandboxBundle,
+    pub network_policy: &'a CanonicalDerivedNetworkPolicy,
+    pub network_probe: NetworkProbeRequest,
+    pub reachability: &'a ReachabilityExpectation,
+    pub container_id: ContainerId,
+    pub runtime_root: PathBuf,
+    pub bundle_digest: BundleDigest,
+    pub invoking_uid: u32,
+}
+
+pub struct ProductionLifecycleAdapter<'a> {
+    plan: ProductionBootstrapPlan<'a>,
+    registry: Arc<SessionRegistry>,
+    trust_anchor: Option<TrustAnchor>,
+    staged_control: Option<StagedControlFace>,
+    published_control: Option<PublishedControlFace>,
+    awaiting_runtime: Option<AwaitingAttachedRunsc>,
+    runtime: Option<CommittedSandboxRuntime>,
+    state_anchor_mount: Option<MountIdentity>,
+    state_anchor_release: Option<StateAnchorReleaseAuthority>,
+}
+
+pub type ProductionHostCoordinatorLifecycle<'store, 'resources> = HostCoordinatorLifecycle<
+    'store,
+    ProductionLifecycleAdapter<'resources>,
+    ProductionSessionRegistry,
+>;
+
+/// Composes the cold-start lifecycle with the exact shared production session registry.
+///
+/// # Errors
+/// Refuses a plan whose Domain Store is not the caller-retained lock-owning store.
+pub fn production_lifecycle<'store, 'resources>(
+    store: &'store Arc<HostCoordinatorDomainStore>,
+    plan: ProductionBootstrapPlan<'resources>,
+) -> Result<ProductionHostCoordinatorLifecycle<'store, 'resources>, ProductionLifecycleError> {
+    if !Arc::ptr_eq(store, &plan.store) {
+        return Err(ProductionLifecycleError(
+            "production bootstrap Domain Store authority differs".into(),
+        ));
+    }
+    let (adapter, registry) = ProductionLifecycleAdapter::new(plan);
+    Ok(HostCoordinatorLifecycle::new(store, adapter, registry))
+}
+
+#[derive(Clone)]
+pub struct ProductionSessionRegistry {
+    registry: Arc<SessionRegistry>,
+}
+
+impl ProductionSessionRegistry {
+    #[must_use]
+    pub fn registry(&self) -> Arc<SessionRegistry> {
+        Arc::clone(&self.registry)
+    }
+}
+
+impl SessionLifecycleAdapter for ProductionSessionRegistry {
+    type Error = ProductionLifecycleError;
+    type Admission = ControlSessionAdmission;
+
+    fn admission_identity(&self, admission: &Self::Admission) -> SessionAdmissionIdentity {
+        <Arc<SessionRegistry> as SessionLifecycleAdapter>::admission_identity(
+            &self.registry,
+            admission,
+        )
+    }
+
+    fn register_session(
+        &self,
+        store: &HostCoordinatorDomainStore,
+        entry_action: LifecycleAction,
+        admission: Self::Admission,
+    ) -> Result<Session, Self::Error> {
+        <Arc<SessionRegistry> as SessionLifecycleAdapter>::register_session(
+            &self.registry,
+            store,
+            entry_action,
+            admission,
+        )
+        .map_err(failure)
+    }
+
+    fn disconnect_session(
+        &self,
+        store: &HostCoordinatorDomainStore,
+        session: &Session,
+    ) -> Result<(), Self::Error> {
+        <Arc<SessionRegistry> as SessionLifecycleAdapter>::disconnect_session(
+            &self.registry,
+            store,
+            session,
+        )
+        .map_err(failure)
+    }
+}
+
+impl<'a> ProductionLifecycleAdapter<'a> {
+    #[must_use]
+    pub fn new(plan: ProductionBootstrapPlan<'a>) -> (Self, ProductionSessionRegistry) {
+        let registry = Arc::new(SessionRegistry::new(
+            Arc::clone(&plan.store),
+            plan.generation_identity.clone(),
+        ));
+        (
+            Self {
+                plan,
+                registry: Arc::clone(&registry),
+                trust_anchor: None,
+                staged_control: None,
+                published_control: None,
+                awaiting_runtime: None,
+                runtime: None,
+                state_anchor_mount: None,
+                state_anchor_release: None,
+            },
+            ProductionSessionRegistry { registry },
+        )
+    }
+
+    #[must_use]
+    pub fn published_control(&self) -> Option<&PublishedControlFace> {
+        self.published_control.as_ref()
+    }
+
+    fn require_generation(&self, generation: &Generation) -> Result<(), ProductionLifecycleError> {
+        if generation.domain_id == self.plan.generation_identity.domain_id
+            && generation.generation_id == self.plan.generation_identity.generation_id
+            && generation.configuration_fingerprint
+                == self.plan.generation_identity.configuration_fingerprint
+            && generation.coordinator_id == self.plan.generation_identity.coordinator_id
+            && generation.sandbox_id == self.plan.generation_identity.sandbox_id
+        {
+            Ok(())
+        } else {
+            Err(ProductionLifecycleError(
+                "production bootstrap generation identity changed".into(),
+            ))
+        }
+    }
+
+    fn store(&self) -> &HostCoordinatorDomainStore {
+        &self.plan.store
+    }
+
+    fn sandbox(&self) -> SandboxCoordinator<'_> {
+        SandboxCoordinator::new(self.store())
+    }
+
+    fn network(&self) -> NetworkCoordinator<'_> {
+        NetworkCoordinator::new(self.store())
+    }
+
+    fn state_anchor(&self) -> StateAnchorCoordinator<'_> {
+        StateAnchorCoordinator::new(self.store(), self.plan.catalog)
+    }
+
+    fn create_trust_anchor(&mut self) -> Result<(), ProductionLifecycleError> {
+        if self.trust_anchor.is_some() {
+            return Err(ProductionLifecycleError(
+                "Trust Anchor already exists".into(),
+            ));
+        }
+        self.trust_anchor =
+            Some(TrustAnchor::create(self.plan.generation_identity.clone()).map_err(failure)?);
+        Ok(())
+    }
+
+    fn start_runtime_supervisor(&mut self) -> Result<(), ProductionLifecycleError> {
+        let sandbox = self.sandbox();
+        sandbox
+            .begin_intent(
+                self.plan.generation_identity.generation_id,
+                self.plan.container_id,
+                &self.plan.runtime_root,
+                self.plan.bundle_digest,
+                current_process_identity().map_err(failure)?,
+            )
+            .map_err(failure)?;
+        let launch = &self.plan.runtime_launch;
+        self.awaiting_runtime = Some(
+            AwaitingAttachedRunsc::launch_networked(&sandbox, &self.network(), launch)
+                .map_err(failure)?,
+        );
+        Ok(())
+    }
+
+    fn lease_prepared_lower(&self) -> Result<(), ProductionLifecycleError> {
+        self.state_anchor()
+            .begin_acquisition(
+                self.plan.generation_identity.generation_id,
+                self.plan.prepared_artifact,
+                self.plan.state_anchor_layout,
+            )
+            .map(|_| ())
+            .map_err(failure)
+    }
+
+    fn mount_state_anchor(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let coordinator = StateAnchorCoordinator::new(&store, self.plan.catalog);
+        let mount = self
+            .awaiting_runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("Runtime Supervisor is absent".into()))?
+            .mount_state_anchor(
+                &coordinator,
+                self.plan.prepared_artifact.prepared_artifact_id,
+            )
+            .map_err(failure)?;
+        self.state_anchor_mount = Some(mount);
+        Ok(())
+    }
+
+    fn start_connectivity_helper(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let network = NetworkCoordinator::new(&store);
+        self.awaiting_runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("Runtime Supervisor is absent".into()))?
+            .receive_helper_staged(&network)
+            .map(|_| ())
+            .map_err(failure)
+    }
+
+    fn start_sidecar(&mut self) -> Result<(), ProductionLifecycleError> {
+        self.awaiting_runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("Runtime Supervisor is absent".into()))?
+            .start_sidecar()
+            .map_err(failure)
+    }
+
+    fn mark_network_ready(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let network = NetworkCoordinator::new(&store);
+        let mut probe = self.plan.network_probe.clone();
+        self.awaiting_runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("Runtime Supervisor is absent".into()))?
+            .receive_network_readiness(
+                &network,
+                self.plan.network_policy,
+                |namespace| {
+                    probe.enforcement_namespace = namespace;
+                    probe
+                },
+                self.plan.reachability,
+            )
+            .map(|_| ())
+            .map_err(failure)
+    }
+
+    fn start_sandbox(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let sandbox = SandboxCoordinator::new(&store);
+        let mut pending = self
+            .awaiting_runtime
+            .take()
+            .ok_or_else(|| ProductionLifecycleError("Runtime Supervisor is absent".into()))?;
+        let staged = pending.receive_staged(&sandbox).map_err(failure)?;
+        let handoff = self
+            .plan
+            .sandbox_bundle
+            .capture_handoff_authority(self.plan.configuration)
+            .map_err(failure)?;
+        let (runtime, _) = pending
+            .acknowledge(staged, &handoff)
+            .map_err(failure)?
+            .commit_networked(&sandbox)
+            .map_err(failure)?;
+        self.runtime = Some(runtime);
+        Ok(())
+    }
+
+    fn stage_publication(&mut self) -> Result<(), ProductionLifecycleError> {
+        let anchor = self
+            .trust_anchor
+            .take()
+            .ok_or_else(|| ProductionLifecycleError("Trust Anchor is absent".into()))?;
+        self.staged_control = Some(
+            StagedControlFace::stage(
+                path_from_host(
+                    self.plan
+                        .configuration
+                        .runtime_domain_root()
+                        .canonical_label()
+                        .as_bytes(),
+                ),
+                self.plan.invoking_uid,
+                anchor,
+                Arc::clone(&self.registry),
+            )
+            .map_err(failure)?,
+        );
+        Ok(())
+    }
+
+    fn expose_publication(&mut self) -> Result<(), ProductionLifecycleError> {
+        self.published_control = Some(
+            self.staged_control
+                .take()
+                .ok_or_else(|| ProductionLifecycleError("staged control face is absent".into()))?
+                .expose()
+                .map_err(failure)?,
+        );
+        Ok(())
+    }
+
+    fn stop_sandbox(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let sandbox = SandboxCoordinator::new(&store);
+        self.runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("committed runtime is absent".into()))?
+            .stop_sandbox(&sandbox)
+            .map_err(failure)
+    }
+
+    fn stop_sidecar(&mut self) -> Result<(), ProductionLifecycleError> {
+        let store = Arc::clone(&self.plan.store);
+        let network = NetworkCoordinator::new(&store);
+        network
+            .begin_release(self.plan.generation_identity.generation_id)
+            .map_err(failure)?;
+        self.runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("committed runtime is absent".into()))?
+            .stop_sidecar()
+            .map_err(failure)?;
+        network
+            .complete_clean_after_absence(self.plan.generation_identity.generation_id)
+            .map_err(failure)
+    }
+
+    fn unmount_state_anchor(&mut self) -> Result<(), ProductionLifecycleError> {
+        let coordinator = self.state_anchor();
+        self.state_anchor_release = Some(
+            coordinator
+                .begin_release(
+                    self.plan.generation_identity.generation_id,
+                    self.plan.state_anchor_layout,
+                )
+                .map_err(failure)?,
+        );
+        self.runtime
+            .as_mut()
+            .ok_or_else(|| ProductionLifecycleError("committed runtime is absent".into()))?
+            .unmount_state_anchor()
+            .map_err(failure)?;
+        self.state_anchor_mount = None;
+        Ok(())
+    }
+
+    fn release_prepared_lower(&mut self) -> Result<(), ProductionLifecycleError> {
+        let release = self
+            .state_anchor_release
+            .take()
+            .ok_or_else(|| ProductionLifecycleError("State Anchor release is absent".into()))?;
+        self.state_anchor()
+            .complete_release_after_unmount(&release, self.plan.state_anchor_layout)
+            .map_err(failure)
+    }
+
+    fn graceful_exit(&mut self) -> Result<(), ProductionLifecycleError> {
+        self.runtime
+            .take()
+            .ok_or_else(|| ProductionLifecycleError("committed runtime is absent".into()))?
+            .graceful_exit()
+            .map_err(failure)
+    }
+}
+
+impl LifecycleAdapter for ProductionLifecycleAdapter<'_> {
+    type Error = ProductionLifecycleError;
+
+    fn execute(
+        &mut self,
+        action: LifecycleAction,
+        generation: &Generation,
+    ) -> Result<(), Self::Error> {
+        self.require_generation(generation)?;
+        match action {
+            LifecycleAction::CreateTrustAnchor => self.create_trust_anchor(),
+            LifecycleAction::StartRuntimeSupervisor => self.start_runtime_supervisor(),
+            LifecycleAction::LeasePreparedLower => self.lease_prepared_lower(),
+            LifecycleAction::MountStateAnchor => self.mount_state_anchor(),
+            LifecycleAction::StartSidecar => self.start_sidecar(),
+            LifecycleAction::StartConnectivityHelper => self.start_connectivity_helper(),
+            LifecycleAction::MarkNetworkReady => self.mark_network_ready(),
+            LifecycleAction::StartSandbox => self.start_sandbox(),
+            LifecycleAction::PublishDomain => self.stage_publication(),
+            LifecycleAction::StopSandbox => self.stop_sandbox(),
+            LifecycleAction::StopSidecar => self.stop_sidecar(),
+            LifecycleAction::UnmountStateAnchor => self.unmount_state_anchor(),
+            LifecycleAction::ReleasePreparedLowerLease => self.release_prepared_lower(),
+            LifecycleAction::GracefulSupervisorExit => self.graceful_exit(),
+            LifecycleAction::FinishTeardown => {
+                self.published_control.take();
+                Ok(())
+            }
+            LifecycleAction::StartNewDomain
+            | LifecycleAction::StartReadySession
+            | LifecycleAction::WaitForDomain
+            | LifecycleAction::RestartWaitingDomain
+            | LifecycleAction::ArmRuntimeCausalEdge
+            | LifecycleAction::AdvanceWaitingSession
+            | LifecycleAction::RejectConfigurationMismatch
+            | LifecycleAction::RegisterSession
+            | LifecycleAction::DisconnectSession
+            | LifecycleAction::BeginTeardown => Ok(()),
+        }
+    }
+
+    fn after_commit(
+        &mut self,
+        action: LifecycleAction,
+        generation: &Generation,
+    ) -> Result<(), Self::Error> {
+        self.require_generation(generation)?;
+        if action == LifecycleAction::PublishDomain {
+            self.expose_publication()
+        } else {
+            Ok(())
+        }
+    }
+}
+
+fn path_from_host(bytes: &[u8]) -> &Path {
+    use std::os::unix::ffi::OsStrExt;
+
+    Path::new(std::ffi::OsStr::from_bytes(bytes))
+}

crates/lazaret-host-coordinator/src/control.rs

@@ -49,6 +49,8 @@ pub const MAX_ENROLLMENT_BYTES: usize = 8_192;
 const MAX_CONTROL_BYTES: usize = 65_536;
 const CONTROL_SOCKET_NAME: &str = "control.sock";
 const PUBLICATION_IDENTITY_NAME: &str = "protocol-identity.postcard";
+const STAGED_CONTROL_SOCKET_NAME: &str = ".control.sock.staged";
+const STAGED_PUBLICATION_IDENTITY_NAME: &str = ".protocol-identity.staged";
 
 #[derive(Debug)]
 pub enum ControlError {
@@ -207,6 +209,11 @@ pub struct TrustAnchor {
 }
 
 impl TrustAnchor {
+    #[must_use]
+    pub const fn generation_identity(&self) -> &GenerationIdentity {
+        &self.generation_identity
+    }
+
     /// Creates one ephemeral generation Trust Anchor with role-separated Ed25519 keys.
     ///
     /// # Errors
@@ -412,7 +419,8 @@ pub struct SessionRegistry {
 }
 
 impl SessionRegistry {
-    fn new(
+    #[must_use]
+    pub fn new(
         store: Arc<HostCoordinatorDomainStore>,
         generation_identity: GenerationIdentity,
     ) -> Self {
@@ -676,6 +684,8 @@ struct ControlPaths {
     root_identity: FileIdentity,
     socket: PathBuf,
     identity: PathBuf,
+    staged_socket: PathBuf,
+    staged_identity: PathBuf,
 }
 
 impl ControlPaths {
@@ -708,6 +718,8 @@ impl ControlPaths {
             root_identity,
             socket: root.join(CONTROL_SOCKET_NAME),
             identity: root.join(PUBLICATION_IDENTITY_NAME),
+            staged_socket: root.join(STAGED_CONTROL_SOCKET_NAME),
+            staged_identity: root.join(STAGED_PUBLICATION_IDENTITY_NAME),
         })
     }
 
@@ -725,15 +737,15 @@ impl ControlPaths {
         Ok(())
     }
 
-    fn create_identity(
+    fn create_named_identity(
         &self,
+        name: &str,
         bytes: &[u8],
         uid: u32,
     ) -> Result<(File, FileIdentity), ControlError> {
-        self.validate_root(uid)?;
         let fd = openat(
             self.root_file.as_fd(),
-            PUBLICATION_IDENTITY_NAME,
+            name,
             OFlags::CREATE | OFlags::EXCL | OFlags::RDWR | OFlags::CLOEXEC | OFlags::NOFOLLOW,
             Mode::RUSR | Mode::WUSR,
         )
@@ -749,10 +761,10 @@ impl ControlPaths {
             self.root_file
                 .sync_all()
                 .map_err(|_| ControlError::InvalidPublication("directory sync failed"))?;
-            self.validate_identity(&file, identity, uid, bytes.len())
+            self.validate_identity_at(&self.root.join(name), &file, identity, uid, bytes.len())
         })();
         if let Err(error) = result {
-            remove_if_identity_matches(&self.identity, identity);
+            remove_if_identity_matches(&self.root.join(name), identity);
             return Err(error);
         }
         Ok((file, identity))
@@ -764,10 +776,21 @@ impl ControlPaths {
         expected: FileIdentity,
         uid: u32,
         expected_length: usize,
+    ) -> Result<(), ControlError> {
+        self.validate_identity_at(&self.identity, file, expected, uid, expected_length)
+    }
+
+    fn validate_identity_at(
+        &self,
+        path: &Path,
+        file: &File,
+        expected: FileIdentity,
+        uid: u32,
+        expected_length: usize,
     ) -> Result<(), ControlError> {
         self.validate_root(uid)?;
         let descriptor = file.metadata()?;
-        let current = std::fs::symlink_metadata(&self.identity)?;
+        let current = std::fs::symlink_metadata(path)?;
         if FileIdentity::from_metadata(&descriptor) != expected
             || FileIdentity::from_metadata(&current) != expected
             || !descriptor.file_type().is_file()
@@ -801,6 +824,150 @@ pub struct PublishedControlFace {
     registry: Arc<SessionRegistry>,
 }
 
+pub struct StagedControlFace {
+    listener: Option<std::os::unix::net::UnixListener>,
+    paths: ControlPaths,
+    staged_socket_identity: FileIdentity,
+    identity_file: File,
+    identity_file_identity: FileIdentity,
+    publication_identity: ProtocolPublicationIdentity,
+    publication_bytes: Vec<u8>,
+    uid: u32,
+    anchor: Option<TrustAnchor>,
+    server_config: Arc<ServerConfig>,
+    registry: Arc<SessionRegistry>,
+}
+
+impl StagedControlFace {
+    /// Stages an undiscoverable private control face before durable publication.
+    ///
+    /// # Errors
+    /// Refuses existing artifacts, changed paths, ownership/mode errors, or TLS failure.
+    pub fn stage(
+        runtime_domain_root: &Path,
+        uid: u32,
+        anchor: TrustAnchor,
+        registry: Arc<SessionRegistry>,
+    ) -> Result<Self, ControlError> {
+        if registry.generation_identity != anchor.generation_identity {
+            return Err(ControlError::StaleGeneration);
+        }
+        let paths = ControlPaths::open(runtime_domain_root, uid)?;
+        if [
+            &paths.socket,
+            &paths.identity,
+            &paths.staged_socket,
+            &paths.staged_identity,
+        ]
+        .into_iter()
+        .any(|path| path.try_exists().unwrap_or(true))
+        {
+            return Err(ControlError::InvalidPublication(
+                "publication path already exists",
+            ));
+        }
+        let publication_identity = anchor.publication_identity();
+        let publication_bytes = encode(&publication_identity)?;
+        let server_config = Arc::new(anchor.server_config()?);
+        let (identity_file, identity_file_identity) = paths.create_named_identity(
+            STAGED_PUBLICATION_IDENTITY_NAME,
+            &publication_bytes,
+            uid,
+        )?;
+        let listener = std::os::unix::net::UnixListener::bind(&paths.staged_socket)
+            .map_err(|_| ControlError::InvalidPublication("staged control socket bind failed"))?;
+        std::fs::set_permissions(&paths.staged_socket, std::fs::Permissions::from_mode(0o600))?;
+        let socket_metadata = std::fs::symlink_metadata(&paths.staged_socket)?;
+        let staged_socket_identity = FileIdentity::from_metadata(&socket_metadata);
+        if !socket_metadata.file_type().is_socket()
+            || socket_metadata.uid() != uid
+            || socket_metadata.mode() & 0o7777 != 0o600
+        {
+            return Err(ControlError::InvalidPublication(
+                "staged control socket ownership or mode is invalid",
+            ));
+        }
+        paths.root_file.sync_all()?;
+        Ok(Self {
+            listener: Some(listener),
+            paths,
+            staged_socket_identity,
+            identity_file,
+            identity_file_identity,
+            publication_identity,
+            publication_bytes,
+            uid,
+            anchor: Some(anchor),
+            server_config,
+            registry,
+        })
+    }
+
+    /// Exposes the staged face after the exact generation is durably Published.
+    ///
+    /// # Errors
+    /// Refuses stale durable state, changed staged artifacts, or exposure failure.
+    pub fn expose(mut self) -> Result<PublishedControlFace, ControlError> {
+        self.registry.require_serviceable_generation()?;
+        self.paths.validate_root(self.uid)?;
+        let staged_socket = std::fs::symlink_metadata(&self.paths.staged_socket)?;
+        if FileIdentity::from_metadata(&staged_socket) != self.staged_socket_identity {
+            return Err(ControlError::InvalidPublication(
+                "staged control socket changed",
+            ));
+        }
+        self.paths.validate_identity_at(
+            &self.paths.staged_identity,
+            &self.identity_file,
+            self.identity_file_identity,
+            self.uid,
+            self.publication_bytes.len(),
+        )?;
+        std::fs::rename(&self.paths.staged_socket, &self.paths.socket)?;
+        if let Err(error) = std::fs::rename(&self.paths.staged_identity, &self.paths.identity) {
+            remove_if_identity_matches(&self.paths.socket, self.staged_socket_identity);
+            return Err(error.into());
+        }
+        self.paths.root_file.sync_all()?;
+        let socket_identity =
+            FileIdentity::from_metadata(&std::fs::symlink_metadata(&self.paths.socket)?);
+        self.paths.validate_identity(
+            &self.identity_file,
+            self.identity_file_identity,
+            self.uid,
+            self.publication_bytes.len(),
+        )?;
+        let paths = ControlPaths::open(&self.paths.root, self.uid)?;
+        Ok(PublishedControlFace {
+            listener: self
+                .listener
+                .take()
+                .ok_or(ControlError::InvalidPublication(
+                    "staged listener is absent",
+                ))?,
+            paths,
+            socket_identity,
+            identity_file: self.identity_file.try_clone()?,
+            identity_file_identity: self.identity_file_identity,
+            publication_identity: self.publication_identity.clone(),
+            publication_bytes: self.publication_bytes.clone(),
+            uid: self.uid,
+            anchor: self.anchor.take().ok_or(ControlError::InvalidPublication(
+                "staged Trust Anchor is absent",
+            ))?,
+            server_config: Arc::clone(&self.server_config),
+            registry: Arc::clone(&self.registry),
+        })
+    }
+}
+
+impl Drop for StagedControlFace {
+    fn drop(&mut self) {
+        remove_if_identity_matches(&self.paths.staged_socket, self.staged_socket_identity);
+        remove_if_identity_matches(&self.paths.staged_identity, self.identity_file_identity);
+    }
+}
+
 impl PublishedControlFace {
     /// Publishes one generation control face in the fixed host-private runtime path.
     ///
@@ -813,83 +980,11 @@ impl PublishedControlFace {
         anchor: TrustAnchor,
         store: Arc<HostCoordinatorDomainStore>,
     ) -> Result<Self, ControlError> {
-        let paths = ControlPaths::open(runtime_domain_root, uid)?;
-        if paths.socket.try_exists()? || paths.identity.try_exists()? {
-            return Err(ControlError::InvalidPublication(
-                "publication path already exists",
-            ));
-        }
         let registry = Arc::new(SessionRegistry::new(
             store,
             anchor.generation_identity.clone(),
         ));
-        registry.require_generation_authority()?;
-        let publication_identity = anchor.publication_identity();
-        let publication_bytes = encode(&publication_identity)?;
-        let server_config = Arc::new(anchor.server_config()?);
-        let (identity_file, identity_file_identity) =
-            paths.create_identity(&publication_bytes, uid)?;
-        let listener = match std::os::unix::net::UnixListener::bind(&paths.socket) {
-            Ok(listener) => listener,
-            Err(error) => {
-                let _ = std::fs::remove_file(&paths.identity);
-                return Err(ControlError::InvalidPublication(
-                    if error.kind() == io::ErrorKind::PermissionDenied {
-                        "control socket bind was denied"
-                    } else {
-                        "control socket bind failed"
-                    },
-                ));
-            }
-        };
-        let finalize = (|| {
-            std::fs::set_permissions(&paths.socket, std::fs::Permissions::from_mode(0o600))
-                .map_err(|_| ControlError::InvalidPublication("control socket chmod failed"))?;
-            let socket_metadata = std::fs::symlink_metadata(&paths.socket)?;
-            if !socket_metadata.file_type().is_socket()
-                || socket_metadata.uid() != uid
-                || socket_metadata.mode() & 0o7777 != 0o600
-            {
-                return Err(ControlError::InvalidPublication(
-                    "control socket ownership or mode is invalid",
-                ));
-            }
-            paths.validate_root(uid)?;
-            Ok(FileIdentity::from_metadata(&socket_metadata))
-        })();
-        let socket_identity = match finalize {
-            Ok(identity) => identity,
-            Err(error) => {
-                drop(listener);
-                let _ = std::fs::remove_file(&paths.socket);
-                let _ = std::fs::remove_file(&paths.identity);
-                return Err(error);
-            }
-        };
-        if let Err(error) = paths.validate_identity(
-            &identity_file,
-            identity_file_identity,
-            uid,
-            publication_bytes.len(),
-        ) {
-            drop(listener);
-            let _ = std::fs::remove_file(&paths.socket);
-            remove_if_identity_matches(&paths.identity, identity_file_identity);
-            return Err(error);
-        }
-        Ok(Self {
-            listener,
-            paths,
-            socket_identity,
-            identity_file,
-            identity_file_identity,
-            publication_identity,
-            publication_bytes,
-            uid,
-            anchor,
-            server_config,
-            registry,
-        })
+        StagedControlFace::stage(runtime_domain_root, uid, anchor, registry)?.expose()
     }
 
     #[must_use]

crates/lazaret-host-coordinator/src/lib.rs

@@ -1,5 +1,6 @@
 #![doc = "Host-side coordination authority for one Lazaret Domain."]
 
+pub mod bootstrap;
 pub mod control;
 pub mod domain_store;
 pub mod escalation;

crates/lazaret-host-coordinator/src/lifecycle.rs

@@ -45,6 +45,18 @@ pub trait LifecycleAdapter {
         action: LifecycleAction,
         generation: &Generation,
     ) -> Result<(), Self::Error>;
+
+    /// Completes a physical boundary that must follow its durable lifecycle commit.
+    ///
+    /// # Errors
+    /// Returns the adapter's post-commit publication or cleanup failure.
+    fn after_commit(
+        &mut self,
+        _action: LifecycleAction,
+        _generation: &Generation,
+    ) -> Result<(), Self::Error> {
+        Ok(())
+    }
 }
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
@@ -209,6 +221,16 @@ where
         self.registered.len()
     }
 
+    #[must_use]
+    pub const fn adapter(&self) -> &A {
+        &self.adapter
+    }
+
+    #[must_use]
+    pub const fn adapter_mut(&mut self) -> &mut A {
+        &mut self.adapter
+    }
+
     /// # Errors
     /// Refuses a non-absent Domain, invalid initial records, or adapter/store failure.
     pub fn start_new_domain(
@@ -251,6 +273,28 @@ where
         Ok(())
     }
 
+    /// Drives one cold-start leader through the complete normal publication boundary.
+    ///
+    /// # Errors
+    /// Returns the first invalid transition, durable update, or production adapter failure.
+    pub fn bootstrap_new_domain(
+        &mut self,
+        generation: Generation,
+        session: Session,
+    ) -> Result<(), LifecycleError<A::Error>> {
+        self.start_new_domain(generation, session)?;
+        self.create_trust_anchor()?;
+        self.start_runtime_supervisor()?;
+        self.lease_prepared_lower()?;
+        self.mount_state_anchor()?;
+        self.start_sidecar()?;
+        self.start_connectivity_helper()?;
+        self.mark_network_ready()?;
+        self.arm_runtime_causal_edge()?;
+        self.start_sandbox()?;
+        self.publish_domain()
+    }
+
     /// # Errors
     /// Refuses an unpublished, stale, mismatched, duplicate, or unauthenticated session.
     pub fn admit_session(
@@ -529,7 +573,16 @@ where
         generation.durable_decision = DurableDecision::Published;
         self.store
             .apply(&[DomainStoreMutation::PutGeneration(generation.clone())])?;
-        self.generation = Some(generation);
+        self.generation = Some(generation.clone());
+        if let Err(error) = self
+            .adapter
+            .after_commit(LifecycleAction::PublishDomain, &generation)
+        {
+            self.phase = DomainPhase::Failed;
+            self.bootstrap_owner = None;
+            self.published = false;
+            return Err(LifecycleError::Adapter(error));
+        }
         self.phase = DomainPhase::Ready;
         self.bootstrap_owner = None;
         self.published = true;

crates/lazaret-host-coordinator/src/network.rs

@@ -368,6 +368,92 @@ impl<'store> NetworkCoordinator<'store> {
         Ok(evidence)
     }
 
+    /// Records exact helper and network release intent before RS resource stop.
+    ///
+    /// # Errors
+    /// Refuses noncommitted or incomplete role ownership.
+    pub fn begin_release(
+        &self,
+        generation_id: GenerationId,
+    ) -> Result<(), NetworkCoordinatorError> {
+        let network = self.participant(generation_id, ParticipantRole::Network)?;
+        let helper = self.participant(generation_id, ParticipantRole::Helper)?;
+        if network.state() != ParticipantState::Committed
+            || helper.state() != ParticipantState::Committed
+        {
+            return Err(NetworkCoordinatorError::InvalidState(
+                "network roles are not committed",
+            ));
+        }
+        let network_ownership =
+            network
+                .ownership()
+                .cloned()
+                .ok_or(NetworkCoordinatorError::InvalidState(
+                    "network ownership is absent",
+                ))?;
+        let helper_ownership =
+            helper
+                .ownership()
+                .cloned()
+                .ok_or(NetworkCoordinatorError::InvalidState(
+                    "helper ownership is absent",
+                ))?;
+        self.domain_store.apply(&[
+            DomainStoreMutation::PutParticipant(
+                Participant::snapshot(
+                    generation_id,
+                    ParticipantRole::Helper,
+                    ParticipantState::ReleaseIntent,
+                    Some(helper_ownership),
+                )
+                .map_err(|_| NetworkCoordinatorError::InvalidState("helper release is invalid"))?,
+            ),
+            DomainStoreMutation::PutParticipant(
+                Participant::snapshot(
+                    generation_id,
+                    ParticipantRole::Network,
+                    ParticipantState::ReleaseIntent,
+                    Some(network_ownership),
+                )
+                .map_err(|_| NetworkCoordinatorError::InvalidState("network release is invalid"))?,
+            ),
+        ])?;
+        Ok(())
+    }
+
+    /// Cleans helper and network roles only after RS reports their physical absence.
+    ///
+    /// # Errors
+    /// Refuses cleanup without exact release intent.
+    pub fn complete_clean_after_absence(
+        &self,
+        generation_id: GenerationId,
+    ) -> Result<(), NetworkCoordinatorError> {
+        for role in [ParticipantRole::Helper, ParticipantRole::Network] {
+            if self.participant(generation_id, role)?.state() != ParticipantState::ReleaseIntent {
+                return Err(NetworkCoordinatorError::InvalidState(
+                    "network release intent is absent",
+                ));
+            }
+        }
+        let clean = [ParticipantRole::Helper, ParticipantRole::Network].map(|role| {
+            Participant::snapshot(
+                generation_id,
+                role,
+                ParticipantState::Clean,
+                Some(ParticipantOwnership::empty(role)),
+            )
+            .map_err(|_| NetworkCoordinatorError::InvalidState("network Clean is invalid"))
+        });
+        let [helper, network] = clean;
+        self.domain_store.apply(&[
+            DomainStoreMutation::CleanParticipantAfterAbsence(helper?),
+            DomainStoreMutation::CleanParticipantAfterAbsence(network?),
+        ])?;
+        Ok(())
+    }
+
     fn require_staged(
         &self,
         staged: StagedProcessOwnership,

crates/lazaret-host-coordinator/src/network/tests.rs

@@ -155,6 +155,60 @@ fn independently_observed_namespace_refuses_attestation_mutation() {
     ));
 }
 
+#[test]
+fn committed_network_roles_release_together_only_after_absence() {
+    let fixture = Fixture::new();
+    let (policy, runtime) = fixture.prepare_roles();
+    let namespace = namespace_identity(std::path::Path::new("/proc/self/ns/net")).unwrap();
+    let attestation = attestation(&fixture.generation, policy.digest(), namespace, runtime);
+    let probe = NetworkReachabilityProbe {
+        domain_id: fixture.generation.domain_id,
+        generation_id: fixture.generation.generation_id,
+        enforcement_namespace: namespace,
+        denied_destinations: vec![DeniedDestinationObservation {
+            destination: IpAddress::V4([10, 1, 2, 3]),
+            unreachable: true,
+        }],
+        permitted_destinations: vec![PermittedDestinationObservation {
+            destination: IpAddress::V4([1, 1, 1, 1]),
+            reachable: true,
+        }],
+        dns_observations: vec![DnsObservation {
+            resolver: IpAddress::V4([8, 8, 8, 8]),
+            query_name: "example.com".into(),
+            resolved_addresses: vec![IpAddress::V4([93, 184, 216, 34])],
+        }],
+    };
+    let coordinator = NetworkCoordinator::new(&fixture.store);
+    coordinator
+        .commit_readiness(&policy, attestation, probe, &reachability())
+        .unwrap();
+    assert!(
+        coordinator
+            .complete_clean_after_absence(fixture.generation.generation_id)
+            .is_err()
+    );
+    coordinator
+        .begin_release(fixture.generation.generation_id)
+        .unwrap();
+    coordinator
+        .complete_clean_after_absence(fixture.generation.generation_id)
+        .unwrap();
+
+    let participants = fixture.store.snapshot().unwrap().participants;
+    for role in [ParticipantRole::Helper, ParticipantRole::Network] {
+        let participant = participants
+            .iter()
+            .find(|participant| participant.role() == role)
+            .unwrap();
+        assert_eq!(participant.state(), ParticipantState::Clean);
+        assert_eq!(
+            participant.ownership(),
+            Some(&ParticipantOwnership::empty(role))
+        );
+    }
+}
+
 fn policy() -> CanonicalDerivedNetworkPolicy {
     derive_policy(DerivedNetworkPolicyFields {
         address_families: vec![

crates/lazaret-host-coordinator/src/runtime.rs

@@ -4,8 +4,10 @@ use std::os::unix::ffi::OsStrExt;
 use std::path::Path;
 use std::time::{Duration, Instant};
 
+use crate::state_anchor::StateAnchorCoordinator;
 use lazaret_core::identity::{
-    BundleDigest, ContainerId, GenerationId, HostPath, NamespaceIdentity, PidfsProcessIdentityV1,
+    BundleDigest, ContainerId, GenerationId, HostPath, MountIdentity, NamespaceIdentity,
+    PidfsProcessIdentityV1, PreparedArtifactId,
 };
 use lazaret_core::network::{CanonicalDerivedNetworkPolicy, ReachabilityExpectation};
 use lazaret_core::records::{
@@ -375,6 +377,88 @@ impl AwaitingAttachedRunsc {
         self.supervisor.as_ref().map(RetainedProcess::identity)
     }
 
+    /// Requests RS-owned State Anchor acquisition and commits exact staged and mount evidence.
+    ///
+    /// # Errors
+    /// Refuses missing intent, event mismatch, changed ownership, or channel failure.
+    pub fn mount_state_anchor(
+        &mut self,
+        coordinator: &StateAnchorCoordinator<'_>,
+        prepared_artifact_id: PreparedArtifactId,
+    ) -> Result<MountIdentity, SandboxCoordinatorError> {
+        let control = self
+            .control
+            .as_ref()
+            .ok_or(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor control channel is absent",
+            ))?;
+        control.mount_state_anchor()?;
+        let staged = match control.receive()? {
+            SupervisorEvent::StateAnchorStaged(staged) => staged,
+            SupervisorEvent::Refused(detail) => {
+                return Err(SandboxCoordinatorError::Io(io::Error::other(detail)));
+            }
+            _ => {
+                return Err(SandboxCoordinatorError::InvalidState(
+                    "Runtime Supervisor did not stage the State Anchor daemon",
+                ));
+            }
+        };
+        coordinator
+            .stage_daemon(staged)
+            .map_err(|error| SandboxCoordinatorError::Io(io::Error::other(error)))?;
+        let acknowledgment = coordinator
+            .acknowledge_staged_daemon(staged)
+            .map_err(|error| SandboxCoordinatorError::Io(io::Error::other(error)))?;
+        control.acknowledge_state_anchor(acknowledgment)?;
+        let (ready_staged, mount) = match control.receive()? {
+            SupervisorEvent::StateAnchorReady { staged, mount } => (staged, mount),
+            SupervisorEvent::Refused(detail) => {
+                return Err(SandboxCoordinatorError::Io(io::Error::other(detail)));
+            }
+            _ => {
+                return Err(SandboxCoordinatorError::InvalidState(
+                    "Runtime Supervisor did not report State Anchor readiness",
+                ));
+            }
+        };
+        if ready_staged != staged {
+            return Err(SandboxCoordinatorError::InvalidState(
+                "State Anchor readiness changed daemon ownership",
+            ));
+        }
+        let readiness =
+            lazaret_state_anchor::VerifiedStateAnchorReadiness::from_parts(ready_staged, mount);
+        coordinator
+            .commit_daemon_ready(readiness)
+            .and_then(|()| {
+                coordinator.commit_prepared(staged.generation_id(), prepared_artifact_id, mount)
+            })
+            .map_err(|error| SandboxCoordinatorError::Io(io::Error::other(error)))?;
+        Ok(mount)
+    }
+
+    /// Releases the RS-owned sidecar install path after exact State Anchor readiness.
+    ///
+    /// # Errors
+    /// Refuses an unexpected RS response or channel failure.
+    pub fn start_sidecar(&mut self) -> Result<(), SandboxCoordinatorError> {
+        let control = self
+            .control
+            .as_ref()
+            .ok_or(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor control channel is absent",
+            ))?;
+        control.start_sidecar()?;
+        if control.receive()? == SupervisorEvent::SidecarStarted {
+            Ok(())
+        } else {
+            Err(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor did not start the sidecar path",
+            ))
+        }
+    }
+
     /// Launches the RS through the HC broker and commits its staged ownership before execution.
     ///
     /// # Errors
@@ -827,6 +911,70 @@ impl CommittedSandboxRuntime {
         }
     }
 
+    /// Stops and durably cleans only the exact attached sandbox runtime.
+    ///
+    /// # Errors
+    /// Refuses missing release intent, unexpected RS response, or exact cleanup failure.
+    pub fn stop_sandbox(
+        &mut self,
+        coordinator: &SandboxCoordinator<'_>,
+    ) -> Result<(), SandboxCoordinatorError> {
+        coordinator.begin_release(self.generation_id)?;
+        self.control.stop_sandbox()?;
+        if self.control.receive()? != SupervisorEvent::SandboxStopped {
+            return Err(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor did not stop the sandbox",
+            ));
+        }
+        coordinator.complete_clean_after_absence(self.generation_id)
+    }
+
+    /// Stops only the RS-owned network resources after sandbox absence.
+    ///
+    /// # Errors
+    /// Refuses an unexpected RS response or channel failure.
+    pub fn stop_sidecar(&mut self) -> Result<(), SandboxCoordinatorError> {
+        self.control.stop_sidecar()?;
+        if self.control.receive()? == SupervisorEvent::SidecarStopped {
+            self.network_evidence = None;
+            Ok(())
+        } else {
+            Err(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor did not stop the sidecar",
+            ))
+        }
+    }
+
+    /// Requests exact RS-owned State Anchor unmount after dependent runtime resources stop.
+    ///
+    /// # Errors
+    /// Refuses an unexpected RS response or channel failure.
+    pub fn unmount_state_anchor(&mut self) -> Result<(), SandboxCoordinatorError> {
+        self.control.unmount_state_anchor()?;
+        if self.control.receive()? == SupervisorEvent::StateAnchorUnmounted {
+            Ok(())
+        } else {
+            Err(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor did not unmount the State Anchor",
+            ))
+        }
+    }
+
+    /// Gracefully exits and reaps the resource-empty exact Runtime Supervisor.
+    ///
+    /// # Errors
+    /// Refuses an unexpected response or an RS that does not exit.
+    pub fn graceful_exit(mut self) -> Result<(), SandboxCoordinatorError> {
+        self.control.exit()?;
+        if self.control.receive()? != SupervisorEvent::Exited {
+            return Err(SandboxCoordinatorError::InvalidState(
+                "Runtime Supervisor did not acknowledge graceful exit",
+            ));
+        }
+        self.lifetime.take();
+        wait_and_reap_supervisor(&mut self.supervisor)
+    }
+
     /// Waits for an exact attached-runsc loss report and reaps the retained RS.
     ///
     /// # Errors

crates/lazaret-host-coordinator/tests/control.rs

@@ -17,7 +17,8 @@ use lazaret_core::records::{
     SessionRegistrationResponse, TlsCredential,
 };
 use lazaret_host_coordinator::control::{
-    CONTROL_ALPN, ControlError, PendingClientEnrollment, PublishedControlFace, TrustAnchor,
+    CONTROL_ALPN, ControlError, PendingClientEnrollment, PublishedControlFace, SessionRegistry,
+    StagedControlFace, TrustAnchor,
 };
 use lazaret_host_coordinator::domain_store::{DomainStoreMutation, HostCoordinatorDomainStore};
 use lazaret_linux::filesystem::try_exclusive_flock;
@@ -687,26 +688,15 @@ fn prepared_control_face_admits_only_after_generation_publication() {
     let directory = TestDirectory::new("publication-gate");
     let identity = generation_identity(7);
     let store = open_preparing_store(&directory, &identity);
-    let face = Arc::new(
-        PublishedControlFace::publish(
-            &directory.0,
-            getuid().as_raw(),
-            TrustAnchor::create(identity.clone()).unwrap(),
-            Arc::clone(&store),
-        )
-        .unwrap(),
-    );
-
-    let serving = server(&face);
-    let config = PublishedControlFace::client_config(face.publication_identity(), None).unwrap();
-    assert!(
-        PublishedControlFace::connect(face.socket_path(), config, face.publication_identity())
-            .is_err()
-    );
-    assert!(matches!(
-        serving.join().unwrap().unwrap_err(),
-        ControlError::StaleGeneration
-    ));
+    let registry = Arc::new(SessionRegistry::new(Arc::clone(&store), identity.clone()));
+    let staged = StagedControlFace::stage(
+        &directory.0,
+        getuid().as_raw(),
+        TrustAnchor::create(identity.clone()).unwrap(),
+        registry,
+    )
+    .unwrap();
+    assert!(std::os::unix::net::UnixStream::connect(directory.path("control.sock")).is_err());
 
     store
         .apply(&[DomainStoreMutation::PutGeneration(generation(
@@ -714,6 +704,7 @@ fn prepared_control_face_admits_only_after_generation_publication() {
             DurableDecision::Published,
         ))])
         .unwrap();
+    let face = Arc::new(staged.expose().unwrap());
     let client = enroll(&face, ClientId::new(bytes(90)));
     assert_eq!(client.credential.generation_identity, identity);
 }

crates/lazaret-host-coordinator/tests/lifecycle.rs

@@ -31,6 +31,32 @@ impl LifecycleAdapter for RecordingAdapter {
     }
 }
 
+struct PublicationFailureAdapter;
+
+impl LifecycleAdapter for PublicationFailureAdapter {
+    type Error = &'static str;
+
+    fn execute(
+        &mut self,
+        _action: LifecycleAction,
+        _generation: &Generation,
+    ) -> Result<(), Self::Error> {
+        Ok(())
+    }
+
+    fn after_commit(
+        &mut self,
+        action: LifecycleAction,
+        _generation: &Generation,
+    ) -> Result<(), Self::Error> {
+        if action == LifecycleAction::PublishDomain {
+            Err("exposure failed")
+        } else {
+            Ok(())
+        }
+    }
+}
+
 #[derive(Clone)]
 struct TestAdmission {
     session: Session,
@@ -272,6 +298,58 @@ fn exact_start_join_disconnect_and_teardown_actions_are_ordered() {
     );
 }
 
+#[test]
+fn cold_start_entry_reaches_publication_in_exact_modeled_order() {
+    let directory = tempfile::tempdir().unwrap();
+    let store = open_store(&directory);
+    let (domain_adapter, session_adapter) = adapters();
+    let mut lifecycle = HostCoordinatorLifecycle::new(&store, domain_adapter, session_adapter);
+
+    lifecycle
+        .bootstrap_new_domain(generation(1), session(10, 1))
+        .unwrap();
+
+    assert_eq!(lifecycle.phase(), DomainPhase::Ready);
+    assert!(lifecycle.is_published());
+    assert_eq!(
+        lifecycle.into_adapters().0.actions.lock().unwrap().clone(),
+        vec![
+            LifecycleAction::StartNewDomain,
+            LifecycleAction::CreateTrustAnchor,
+            LifecycleAction::StartRuntimeSupervisor,
+            LifecycleAction::LeasePreparedLower,
+            LifecycleAction::MountStateAnchor,
+            LifecycleAction::StartSidecar,
+            LifecycleAction::StartConnectivityHelper,
+            LifecycleAction::MarkNetworkReady,
+            LifecycleAction::ArmRuntimeCausalEdge,
+            LifecycleAction::StartSandbox,
+            LifecycleAction::PublishDomain,
+        ]
+    );
+}
+
+#[test]
+fn publication_exposure_failure_retains_recoverable_durable_decision() {
+    let directory = tempfile::tempdir().unwrap();
+    let store = open_store(&directory);
+    let (_, session_adapter) = adapters();
+    let mut lifecycle =
+        HostCoordinatorLifecycle::new(&store, PublicationFailureAdapter, session_adapter);
+
+    assert!(
+        lifecycle
+            .bootstrap_new_domain(generation(1), session(10, 1))
+            .is_err()
+    );
+    assert_eq!(lifecycle.phase(), DomainPhase::Failed);
+    assert!(!lifecycle.is_published());
+    assert_eq!(
+        store.snapshot().unwrap().generations[0].durable_decision,
+        DurableDecision::Published
+    );
+}
+
 #[test]
 fn publication_and_invalid_transitions_fail_before_adapter_execution() {
     let directory = tempfile::tempdir().unwrap();

crates/lazaret-runtime-supervisor/Cargo.toml

@@ -12,6 +12,7 @@ path = "src/main.rs"
 [dependencies]
 lazaret-core = { version = "=0.0.0", path = "../lazaret-core" }
 lazaret-linux = { version = "=0.0.0", path = "../lazaret-linux" }
+lazaret-state-anchor = { version = "=0.0.0", path = "../lazaret-state-anchor" }
 oci-spec.workspace = true
 serde.workspace = true
 rustix.workspace = true

crates/lazaret-runtime-supervisor/src/supervision.rs

@@ -21,6 +21,9 @@ use lazaret_linux::process::{
     ChildStandardIo, ParentEdge, ProcessOwnershipContext, RetainedProcess, TargetCommand,
     child_broker, current_process_identity,
 };
+use lazaret_state_anchor::{
+    RuntimeStateAnchorAuthority, RuntimeStateAnchorMount, StateAnchorLaunchAuthority,
+};
 use rustix::event::{PollFd, PollFlags, Timespec, poll};
 use rustix::fs::{MemfdFlags, memfd_create};
 use rustix::io::{FdFlags, fcntl_setfd, read, write};
@@ -58,6 +61,12 @@ impl AttachedRunscSignal {
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
 pub enum SupervisorEvent {
+    StateAnchorStaged(StagedProcessOwnership),
+    StateAnchorReady {
+        staged: StagedProcessOwnership,
+        mount: lazaret_core::identity::MountIdentity,
+    },
+    SidecarStarted,
     HelperStaged(StagedProcessOwnership),
     HelperReady(StagedProcessOwnership),
     NetworkAttestation(NetworkReadinessAttestation),
@@ -69,6 +78,10 @@ pub enum SupervisorEvent {
         sentry_pid_inventory: u32,
     },
     SignalApplied(AttachedRunscSignal),
+    SandboxStopped,
+    SidecarStopped,
+    StateAnchorUnmounted,
+    Exited,
     AttachedRunscExited,
     Stopped,
     Refused(String),
@@ -76,12 +89,19 @@ pub enum SupervisorEvent {
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
 enum CoordinatorMessage {
+    MountStateAnchor,
+    StateAnchorStagedAck(ProcessOwnershipCommitAck),
+    StartSidecar,
     HelperStagedAck(ProcessOwnershipCommitAck),
     NetworkAttestationAccepted(NetworkProbeRequest),
     NetworkReadinessCommitted,
     StagedAck(ProcessOwnershipCommitAck),
     SandboxCommitted,
     SignalAttachedRunsc(AttachedRunscSignal),
+    StopSandbox,
+    StopSidecar,
+    UnmountStateAnchor,
+    Exit,
     Stop,
 }
 
@@ -101,6 +121,30 @@ impl CoordinatorChannel {
         receive_message(&self.descriptor)
     }
 
+    /// # Errors
+    /// Returns an error when exact State Anchor acquisition cannot be requested.
+    pub fn mount_state_anchor(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::MountStateAnchor)
+    }
+
+    /// # Errors
+    /// Returns an error when the exact daemon acknowledgment cannot be delivered.
+    pub fn acknowledge_state_anchor(
+        &self,
+        acknowledgment: ProcessOwnershipCommitAck,
+    ) -> io::Result<()> {
+        send_message(
+            &self.descriptor,
+            &CoordinatorMessage::StateAnchorStagedAck(acknowledgment),
+        )
+    }
+
+    /// # Errors
+    /// Returns an error when the RS-owned sidecar install path cannot be released.
+    pub fn start_sidecar(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::StartSidecar)
+    }
+
     /// # Errors
     /// Returns an error when the exact typed acknowledgment cannot be delivered.
     pub fn acknowledge_staged(&self, acknowledgment: ProcessOwnershipCommitAck) -> io::Result<()> {
@@ -157,14 +201,41 @@ impl CoordinatorChannel {
     pub fn stop(&self) -> io::Result<()> {
         send_message(&self.descriptor, &CoordinatorMessage::Stop)
     }
+
+    /// # Errors
+    /// Returns an error when ordered sandbox stop cannot be requested.
+    pub fn stop_sandbox(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::StopSandbox)
+    }
+
+    /// # Errors
+    /// Returns an error when ordered sidecar stop cannot be requested.
+    pub fn stop_sidecar(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::StopSidecar)
+    }
+
+    /// # Errors
+    /// Returns an error when exact State Anchor unmount cannot be requested.
+    pub fn unmount_state_anchor(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::UnmountStateAnchor)
+    }
+
+    /// # Errors
+    /// Returns an error when graceful Runtime Supervisor exit cannot be requested.
+    pub fn exit(&self) -> io::Result<()> {
+        send_message(&self.descriptor, &CoordinatorMessage::Exit)
+    }
 }
 
 pub struct RuntimeSupervisorLaunch {
     target: TargetCommand,
     ownership: ProcessOwnershipContext,
     policy_file: Option<File>,
+    policy_digest: Option<DerivedPolicyDigest>,
+    configuration_fingerprint: Option<ConfigurationFingerprint>,
     network_sidecar_id: Option<NetworkSidecarId>,
     runsc_command: RunscCommand,
+    state_anchor: Option<StateAnchorLaunchAuthority>,
 }
 
 #[derive(Clone, Copy)]
@@ -213,8 +284,11 @@ impl RuntimeSupervisorLaunch {
             target: target.with_lifetime_pipe_parent_edge(),
             ownership,
             policy_file: None,
+            policy_digest: None,
+            configuration_fingerprint: None,
             network_sidecar_id: None,
             runsc_command: command.clone(),
+            state_anchor: None,
         })
     }
 
@@ -260,10 +334,68 @@ impl RuntimeSupervisorLaunch {
         launch.target = launch.target.with_lifetime_pipe_parent_edge();
         drop(target);
         launch.policy_file = Some(policy_file);
+        launch.policy_digest = Some(network.policy.digest());
+        launch.configuration_fingerprint = Some(network.configuration_fingerprint);
         launch.network_sidecar_id = Some(network.network_sidecar_id);
         Ok(launch)
     }
 
+    /// Adds exact inherited State Anchor authorities without beginning physical acquisition.
+    #[must_use]
+    pub fn with_state_anchor(mut self, authority: StateAnchorLaunchAuthority) -> Self {
+        self.state_anchor = Some(authority);
+        self.rebuild_target();
+        self
+    }
+
+    fn rebuild_target(&mut self) {
+        let mut target = TargetCommand::new("/proc/self/exe")
+            .argument(PRIVATE_ROLE_ARGUMENT)
+            .argument("--domain")
+            .argument(hex(self.ownership.domain_id.as_bytes()))
+            .argument("--generation")
+            .argument(self.ownership.generation_id.get().to_string())
+            .argument("--coordinator")
+            .argument(hex(self.ownership.coordinator_id.as_bytes()));
+        if let Some(authority) = &self.state_anchor {
+            let (device, inode) = authority.expected_lower();
+            target = target
+                .argument("--state-lower-fd")
+                .argument(authority.lower_fd().to_string())
+                .argument("--state-persistent-fd")
+                .argument(authority.persistent_anchor_fd().to_string())
+                .argument("--state-runtime-fd")
+                .argument(authority.runtime_anchor_fd().to_string())
+                .argument("--state-lower-device")
+                .argument(device.to_string())
+                .argument("--state-lower-inode")
+                .argument(inode.to_string());
+        }
+        if let (Some(policy_file), Some(digest), Some(fingerprint), Some(sidecar)) = (
+            &self.policy_file,
+            self.policy_digest,
+            self.configuration_fingerprint,
+            self.network_sidecar_id,
+        ) {
+            target = target
+                .argument("--network-policy-fd")
+                .argument(policy_file.as_raw_fd().to_string())
+                .argument("--network-policy-digest")
+                .argument(hex(digest.as_bytes()))
+                .argument("--configuration-fingerprint")
+                .argument(hex(fingerprint.as_bytes()))
+                .argument("--network-sidecar")
+                .argument(hex(sidecar.as_bytes()));
+        }
+        target = target
+            .argument("--runsc")
+            .argument(self.runsc_command.program());
+        for argument in self.runsc_command.arguments() {
+            target = target.argument(argument);
+        }
+        self.target = target.with_lifetime_pipe_parent_edge();
+    }
+
     #[must_use]
     pub const fn target(&self) -> &TargetCommand {
         &self.target
@@ -328,11 +460,19 @@ pub fn dispatch_private_runtime_role(arguments: &[OsString]) -> Option<io::Resul
 
 struct RuntimeArguments {
     ownership: ProcessOwnershipContext,
+    state_anchor: Option<RuntimeStateAnchorArguments>,
     network: Option<RuntimeNetworkArguments>,
     runsc_program: OsString,
     runsc_arguments: Vec<OsString>,
 }
 
+struct RuntimeStateAnchorArguments {
+    lower_fd: i32,
+    persistent_anchor_fd: i32,
+    runtime_anchor_fd: i32,
+    expected_lower: (u64, u64),
+}
+
 struct RuntimeNetworkArguments {
     policy_fd: i32,
     policy_digest: DerivedPolicyDigest,
@@ -348,6 +488,22 @@ fn parse_runtime_arguments(arguments: &[OsString]) -> io::Result<RuntimeArgument
         .and_then(|value| value.parse().ok())
         .ok_or_else(|| io::Error::other("Runtime Supervisor generation is invalid"))?;
     let coordinator = parse_fixed_hex(named(&mut arguments, "--coordinator")?)?;
+    let state_anchor = if arguments
+        .peek()
+        .is_some_and(|value| *value == "--state-lower-fd")
+    {
+        Some(RuntimeStateAnchorArguments {
+            lower_fd: parse_number(named(&mut arguments, "--state-lower-fd")?)?,
+            persistent_anchor_fd: parse_number(named(&mut arguments, "--state-persistent-fd")?)?,
+            runtime_anchor_fd: parse_number(named(&mut arguments, "--state-runtime-fd")?)?,
+            expected_lower: (
+                parse_number(named(&mut arguments, "--state-lower-device")?)?,
+                parse_number(named(&mut arguments, "--state-lower-inode")?)?,
+            ),
+        })
+    } else {
+        None
+    };
     let network = if arguments
         .peek()
         .is_some_and(|value| *value == "--network-policy-fd")
@@ -374,12 +530,23 @@ fn parse_runtime_arguments(arguments: &[OsString]) -> io::Result<RuntimeArgument
             coordinator_id: CoordinatorId::new(coordinator),
             slot: HostProcessOwnershipSlot::SandboxAttachedRunsc,
         },
+        state_anchor,
         network,
         runsc_program: (*program).clone(),
         runsc_arguments: arguments.cloned().collect(),
     })
 }
 
+fn parse_number<T>(value: &OsStr) -> io::Result<T>
+where
+    T: std::str::FromStr,
+{
+    value
+        .to_str()
+        .and_then(|value| value.parse().ok())
+        .ok_or_else(|| io::Error::other("Runtime Supervisor numeric argument is invalid"))
+}
+
 fn parse_network_arguments<'a>(
     arguments: &mut impl Iterator<Item = &'a OsString>,
 ) -> io::Result<RuntimeNetworkArguments> {
@@ -439,9 +606,92 @@ fn run_supervisor(arguments: &RuntimeArguments) -> io::Result<()> {
     require_lifetime_held()?;
     let control = std::io::stdout();
     let parent = current_process_identity()?;
+    let mut state_anchor = prepare_state_anchor(arguments, parent, &control)?;
+    let result = run_supervisor_with_state(arguments, parent, &control, &mut state_anchor);
+    let cleanup = state_anchor.take().map(RuntimeStateAnchorMount::unmount);
+    match (result, cleanup) {
+        (Err(error), Some(Err(cleanup))) => Err(io::Error::new(
+            error.kind(),
+            format!("{error}; exact State Anchor cleanup failed: {cleanup}"),
+        )),
+        (Err(error), _) => Err(error),
+        (Ok(()), Some(Err(error))) => Err(io::Error::other(error)),
+        (Ok(()), _) => Ok(()),
+    }
+}
+
+fn prepare_state_anchor(
+    arguments: &RuntimeArguments,
+    parent: lazaret_core::identity::PidfsProcessIdentityV1,
+    control: &std::io::Stdout,
+) -> io::Result<Option<RuntimeStateAnchorMount>> {
+    let Some(state) = &arguments.state_anchor else {
+        return Ok(None);
+    };
+    let command: CoordinatorMessage = receive_message(control)?;
+    if command != CoordinatorMessage::MountStateAnchor {
+        return Err(io::Error::other(
+            "Runtime Supervisor received no State Anchor mount command",
+        ));
+    }
+    let authority = RuntimeStateAnchorAuthority::from_inherited(
+        state.lower_fd,
+        state.persistent_anchor_fd,
+        state.runtime_anchor_fd,
+        state.expected_lower,
+    )
+    .map_err(io::Error::other)?;
+    let ownership = ProcessOwnershipContext {
+        slot: HostProcessOwnershipSlot::StateAnchorDaemon,
+        ..arguments.ownership
+    };
+    authority
+        .acquire(
+            parent,
+            ownership,
+            |staged| send_message(control, &SupervisorEvent::StateAnchorStaged(staged)),
+            |staged| match receive_message(control)? {
+                CoordinatorMessage::StateAnchorStagedAck(acknowledgment)
+                    if acknowledgment.acknowledges(&staged) =>
+                {
+                    Ok(acknowledgment)
+                }
+                _ => Err(io::Error::other(
+                    "Runtime Supervisor received a mismatched State Anchor acknowledgment",
+                )),
+            },
+            |readiness| {
+                send_message(
+                    control,
+                    &SupervisorEvent::StateAnchorReady {
+                        staged: readiness.staged(),
+                        mount: readiness.mount(),
+                    },
+                )
+            },
+        )
+        .map(Some)
+        .map_err(io::Error::other)
+}
+
+fn run_supervisor_with_state(
+    arguments: &RuntimeArguments,
+    parent: lazaret_core::identity::PidfsProcessIdentityV1,
+    control: &std::io::Stdout,
+    state_anchor: &mut Option<RuntimeStateAnchorMount>,
+) -> io::Result<()> {
+    if arguments.state_anchor.is_some() {
+        let message: CoordinatorMessage = receive_message(control)?;
+        if message != CoordinatorMessage::StartSidecar {
+            return Err(io::Error::other(
+                "Runtime Supervisor received no sidecar start command",
+            ));
+        }
+        send_message(control, &SupervisorEvent::SidecarStarted)?;
+    }
     let network = match &arguments.network {
         Some(network) => Some(prepare_runtime_network(
-            arguments, network, parent, &control,
+            arguments, network, parent, control,
         )?),
         None => None,
     };
@@ -456,9 +706,9 @@ fn run_supervisor(arguments: &RuntimeArguments) -> io::Result<()> {
         attached_runsc_standard_io(&runsc)?,
         |ownership| {
             staged.set(Some(ownership));
-            send_message(&control, &SupervisorEvent::Staged(ownership))
+            send_message(control, &SupervisorEvent::Staged(ownership))
         },
-        |ownership| match receive_message(&control)? {
+        |ownership| match receive_message(control)? {
             CoordinatorMessage::StagedAck(acknowledgment)
                 if acknowledgment.acknowledges(&ownership) =>
             {
@@ -493,7 +743,7 @@ fn run_supervisor(arguments: &RuntimeArguments) -> io::Result<()> {
         return fail_with_network_cleanup(process, network, error);
     }
     if let Err(error) = send_message(
-        &control,
+        control,
         &SupervisorEvent::Ready {
             staged,
             sentry_pid_inventory,
@@ -501,7 +751,7 @@ fn run_supervisor(arguments: &RuntimeArguments) -> io::Result<()> {
     ) {
         return fail_with_network_cleanup(process, network, error);
     }
-    let committed: CoordinatorMessage = match receive_message(&control) {
+    let committed: CoordinatorMessage = match receive_message(control) {
         Ok(committed) => committed,
         Err(error) => return fail_with_network_cleanup(process, network, error),
     };
@@ -512,7 +762,7 @@ fn run_supervisor(arguments: &RuntimeArguments) -> io::Result<()> {
             io::Error::other("sandbox commit acknowledgment was not received"),
         );
     }
-    supervise(&control, process, network)
+    supervise(control, process, network, state_anchor)
 }
 
 fn prepare_runtime_network(
@@ -645,14 +895,23 @@ fn supervise(
     control: &std::io::Stdout,
     process: RetainedProcess,
     mut network: Option<PreparedNetwork>,
+    state_anchor: &mut Option<RuntimeStateAnchorMount>,
 ) -> io::Result<()> {
+    let mut process = Some(process);
     loop {
-        let exited = match process.has_exited() {
+        let exited = match process
+            .as_ref()
+            .map_or(Ok(false), RetainedProcess::has_exited)
+        {
             Ok(exited) => exited,
-            Err(error) => return fail_with_network_cleanup(process, network, error),
+            Err(error) => return fail_active_runtime(process, network, error),
         };
         if exited {
-            child_broker().reap(process)?;
+            child_broker().reap(
+                process
+                    .take()
+                    .ok_or_else(|| io::Error::other("attached runsc handle is absent"))?,
+            )?;
             stop_network(network)?;
             send_message(control, &SupervisorEvent::AttachedRunscExited)?;
             return Ok(());
@@ -669,10 +928,12 @@ fn supervise(
         }
         let lease_lost = match lifetime_lost() {
             Ok(lost) => lost,
-            Err(error) => return fail_with_network_cleanup(process, network, error),
+            Err(error) => return fail_active_runtime(process, network, error),
         };
         if lease_lost {
-            terminate_and_reap(process)?;
+            if let Some(process) = process.take() {
+                terminate_and_reap(process)?;
+            }
             stop_network(network)?;
             return Ok(());
         }
@@ -684,11 +945,11 @@ fn supervise(
                 tv_nsec: 10_000_000,
             }),
         ) {
-            return fail_with_network_cleanup(process, network, error.into());
+            return fail_active_runtime(process, network, error.into());
         }
         let events = descriptors[0].revents();
         if events.intersects(PollFlags::HUP | PollFlags::ERR) {
-            return fail_with_network_cleanup(
+            return fail_active_runtime(
                 process,
                 network,
                 io::Error::new(
@@ -700,41 +961,110 @@ fn supervise(
         if events.intersects(PollFlags::IN) {
             let message = match receive_message(control) {
                 Ok(message) => message,
-                Err(error) => return fail_with_network_cleanup(process, network, error),
+                Err(error) => return fail_active_runtime(process, network, error),
             };
-            match message {
-                CoordinatorMessage::SignalAttachedRunsc(signal) => {
-                    if let Err(error) = process.send_signal(signal.signal()) {
-                        return fail_with_network_cleanup(process, network, error);
-                    }
-                    if let Err(error) =
-                        send_message(control, &SupervisorEvent::SignalApplied(signal))
-                    {
-                        return fail_with_network_cleanup(process, network, error);
-                    }
-                }
-                CoordinatorMessage::Stop => {
-                    terminate_and_reap(process)?;
-                    stop_network(network)?;
-                    send_message(control, &SupervisorEvent::Stopped)?;
-                    return Ok(());
-                }
-                CoordinatorMessage::HelperStagedAck(_)
-                | CoordinatorMessage::NetworkAttestationAccepted(_)
-                | CoordinatorMessage::NetworkReadinessCommitted
-                | CoordinatorMessage::StagedAck(_)
-                | CoordinatorMessage::SandboxCommitted => {
-                    return fail_with_network_cleanup(
-                        process,
-                        network,
-                        io::Error::other("unexpected duplicate coordinator message"),
-                    );
-                }
+            match handle_supervisor_message(
+                control,
+                &message,
+                &mut process,
+                &mut network,
+                state_anchor,
+            ) {
+                Ok(true) => return Ok(()),
+                Ok(false) => {}
+                Err(error) => return fail_active_runtime(process, network, error),
             }
         }
     }
 }
 
+fn handle_supervisor_message(
+    control: &std::io::Stdout,
+    message: &CoordinatorMessage,
+    process: &mut Option<RetainedProcess>,
+    network: &mut Option<PreparedNetwork>,
+    state_anchor: &mut Option<RuntimeStateAnchorMount>,
+) -> io::Result<bool> {
+    match message {
+        CoordinatorMessage::SignalAttachedRunsc(signal) => {
+            process
+                .as_ref()
+                .ok_or_else(|| io::Error::other("attached runsc is already stopped"))?
+                .send_signal(signal.signal())?;
+            send_message(control, &SupervisorEvent::SignalApplied(*signal))?;
+        }
+        CoordinatorMessage::StopSandbox => {
+            terminate_and_reap(
+                process
+                    .take()
+                    .ok_or_else(|| io::Error::other("attached runsc is already stopped"))?,
+            )?;
+            send_message(control, &SupervisorEvent::SandboxStopped)?;
+        }
+        CoordinatorMessage::StopSidecar => {
+            if process.is_some() {
+                return Err(io::Error::other("sidecar stop preceded sandbox stop"));
+            }
+            stop_network(network.take())?;
+            send_message(control, &SupervisorEvent::SidecarStopped)?;
+        }
+        CoordinatorMessage::UnmountStateAnchor => {
+            if process.is_some() || network.is_some() {
+                return Err(io::Error::other(
+                    "State Anchor unmount preceded runtime resource stop",
+                ));
+            }
+            state_anchor
+                .take()
+                .ok_or_else(|| io::Error::other("State Anchor is absent"))?
+                .unmount()
+                .map_err(io::Error::other)?;
+            send_message(control, &SupervisorEvent::StateAnchorUnmounted)?;
+        }
+        CoordinatorMessage::Exit => {
+            if process.is_some() || network.is_some() || state_anchor.is_some() {
+                return Err(io::Error::other(
+                    "Runtime Supervisor exit preceded resource cleanup",
+                ));
+            }
+            send_message(control, &SupervisorEvent::Exited)?;
+            return Ok(true);
+        }
+        CoordinatorMessage::Stop => {
+            if let Some(process) = process.take() {
+                terminate_and_reap(process)?;
+            }
+            stop_network(network.take())?;
+            send_message(control, &SupervisorEvent::Stopped)?;
+            return Ok(true);
+        }
+        CoordinatorMessage::HelperStagedAck(_)
+        | CoordinatorMessage::MountStateAnchor
+        | CoordinatorMessage::StateAnchorStagedAck(_)
+        | CoordinatorMessage::StartSidecar
+        | CoordinatorMessage::NetworkAttestationAccepted(_)
+        | CoordinatorMessage::NetworkReadinessCommitted
+        | CoordinatorMessage::StagedAck(_)
+        | CoordinatorMessage::SandboxCommitted => {
+            return Err(io::Error::other("unexpected duplicate coordinator message"));
+        }
+    }
+    Ok(false)
+}
+
+fn fail_active_runtime(
+    process: Option<RetainedProcess>,
+    network: Option<PreparedNetwork>,
+    error: io::Error,
+) -> io::Result<()> {
+    if let Some(process) = process {
+        fail_with_network_cleanup(process, network, error)
+    } else {
+        let _ = stop_network(network);
+        Err(error)
+    }
+}
+
 fn terminate_and_reap(process: RetainedProcess) -> io::Result<()> {
     let term_error = process.send_signal(Signal::TERM).err();
     let deadline = Instant::now() + TERMINATION_GRACE;
@@ -890,6 +1220,42 @@ mod tests {
         );
     }
 
+    #[test]
+    fn coordinator_channel_preserves_phased_resource_commands() {
+        let (coordinator, supervisor) = private_seqpacket_pair().unwrap();
+        let channel = CoordinatorChannel::new(coordinator);
+        channel.mount_state_anchor().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::MountStateAnchor
+        );
+        channel.start_sidecar().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::StartSidecar
+        );
+        channel.stop_sandbox().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::StopSandbox
+        );
+        channel.stop_sidecar().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::StopSidecar
+        );
+        channel.unmount_state_anchor().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::UnmountStateAnchor
+        );
+        channel.exit().unwrap();
+        assert_eq!(
+            receive_message::<CoordinatorMessage>(&supervisor).unwrap(),
+            CoordinatorMessage::Exit
+        );
+    }
+
     #[test]
     fn runtime_supervisor_and_runsc_keep_distinct_edge_and_stdio_policies() {
         let ownership = ProcessOwnershipContext {
@@ -902,12 +1268,15 @@ mod tests {
             target: TargetCommand::new("/proc/self/exe").with_lifetime_pipe_parent_edge(),
             ownership,
             policy_file: None,
+            policy_digest: None,
+            configuration_fingerprint: None,
             runsc_command: RunscCommand::new(
                 "runsc",
                 [OsString::from("run"), OsString::from("container")],
                 [0; 32],
             ),
             network_sidecar_id: None,
+            state_anchor: None,
         };
         let (_coordinator_control, supervisor_control) = private_seqpacket_pair().unwrap();
         let (supervisor_lifetime, _coordinator_lifetime) = pipe_with(PipeFlags::CLOEXEC).unwrap();
@@ -925,6 +1294,7 @@ mod tests {
                 slot: HostProcessOwnershipSlot::SandboxAttachedRunsc,
                 ..ownership
             },
+            state_anchor: None,
             network: None,
             runsc_program: "runsc".into(),
             runsc_arguments: ["-root=/runtime", "run", "container"]

crates/lazaret-state-anchor/src/lib.rs

@@ -388,7 +388,223 @@ pub struct StateAnchorAcquisition<'path> {
     pub ownership: ProcessOwnershipContext,
 }
 
+pub struct StateAnchorLaunchAuthority {
+    lower: OwnedFd,
+    persistent_anchor: OwnedFd,
+    runtime_anchor: OwnedFd,
+    expected_lower: (u64, u64),
+}
+
+impl StateAnchorLaunchAuthority {
+    /// Captures exact directory authorities for inheritance by the Runtime Supervisor.
+    ///
+    /// # Errors
+    /// Refuses a changed layout or prepared lower.
+    pub fn capture(
+        layout: &StateAnchorLayout<'_>,
+        artifact: &PreparedArtifact,
+    ) -> Result<Self, StateAnchorError> {
+        layout.revalidate()?;
+        let expected_lower = (artifact.device.get(), artifact.inode.get());
+        let lower = open_exact_directory_path(&path_from_host(&artifact.path), expected_lower)?;
+        Ok(Self {
+            lower: inherit_for_exec(lower)?,
+            persistent_anchor: inherit_for_exec(&layout.persistent_anchor)?,
+            runtime_anchor: inherit_for_exec(&layout.runtime_anchor)?,
+            expected_lower,
+        })
+    }
+
+    #[must_use]
+    pub fn lower_fd(&self) -> i32 {
+        self.lower.as_raw_fd()
+    }
+
+    #[must_use]
+    pub fn persistent_anchor_fd(&self) -> i32 {
+        self.persistent_anchor.as_raw_fd()
+    }
+
+    #[must_use]
+    pub fn runtime_anchor_fd(&self) -> i32 {
+        self.runtime_anchor.as_raw_fd()
+    }
+
+    #[must_use]
+    pub const fn expected_lower(&self) -> (u64, u64) {
+        self.expected_lower
+    }
+}
+
+pub struct RuntimeStateAnchorAuthority {
+    lower: OwnedFd,
+    persistent_anchor: OwnedFd,
+    runtime_anchor: OwnedFd,
+    merged_descriptor: OwnedFd,
+}
+
+impl RuntimeStateAnchorAuthority {
+    /// Reopens inherited authorities without falling back to ambient configured paths.
+    ///
+    /// # Errors
+    /// Refuses missing, replaced, or incorrectly typed descriptors and fixed children.
+    pub fn from_inherited(
+        lower_fd: i32,
+        persistent_anchor_fd: i32,
+        runtime_anchor_fd: i32,
+        expected_lower: (u64, u64),
+    ) -> Result<Self, StateAnchorError> {
+        let lower = reopen_inherited_directory(lower_fd)?;
+        require_directory_descriptor_identity(&lower, expected_lower)?;
+        let persistent_anchor = reopen_inherited_directory(persistent_anchor_fd)?;
+        let runtime_anchor = reopen_inherited_directory(runtime_anchor_fd)?;
+        for name in ["upper", "work"] {
+            drop(open_directory(&persistent_anchor, Path::new(name))?);
+        }
+        let merged_descriptor = open_directory(&runtime_anchor, Path::new("merged"))?;
+        Ok(Self {
+            lower,
+            persistent_anchor,
+            runtime_anchor,
+            merged_descriptor,
+        })
+    }
+
+    /// Mounts the exact inherited prepared lower and anchors under RS child ownership.
+    ///
+    /// # Errors
+    /// Refuses acquisition, acknowledgment, readiness, or authority failures.
+    pub fn acquire<Stage, Acknowledge, Ready>(
+        self,
+        expected_parent: lazaret_core::identity::PidfsProcessIdentityV1,
+        ownership: ProcessOwnershipContext,
+        stage: Stage,
+        acknowledge: Acknowledge,
+        ready: Ready,
+    ) -> Result<RuntimeStateAnchorMount, StateAnchorError>
+    where
+        Stage: FnOnce(StagedProcessOwnership) -> io::Result<()>,
+        Acknowledge: FnOnce(StagedProcessOwnership) -> io::Result<ProcessOwnershipCommitAck>,
+        Ready: FnOnce(VerifiedStateAnchorReadiness) -> io::Result<()>,
+    {
+        if ownership.slot != HostProcessOwnershipSlot::StateAnchorDaemon {
+            return Err(StateAnchorError::InvalidLayout(
+                "process slot is not State Anchor daemon",
+            ));
+        }
+        let baseline = mount_identity(&self.merged_descriptor)?;
+        let lower_launch = inherit_for_exec(&self.lower)?;
+        let persistent_launch = inherit_for_exec(&self.persistent_anchor)?;
+        let runtime_launch = inherit_for_exec(&self.runtime_anchor)?;
+        let target = TargetCommand::new("fuse-overlayfs")
+            .argument("-f")
+            .argument("-o")
+            .argument(overlay_options(
+                &descriptor_path(&lower_launch),
+                &descriptor_child_path(&persistent_launch, "upper"),
+                &descriptor_child_path(&persistent_launch, "work"),
+            ))
+            .argument(descriptor_child_path(&runtime_launch, "merged"));
+        let process =
+            child_broker().acquire(expected_parent, ownership, &target, stage, acknowledge)?;
+        let staged = StagedProcessOwnership::v1(
+            ownership.domain_id,
+            ownership.generation_id,
+            ownership.coordinator_id,
+            ownership.slot,
+            process.identity(),
+        );
+        let deadline = Instant::now() + MOUNT_TIMEOUT;
+        loop {
+            if process.has_exited()? {
+                child_broker().reap(process)?;
+                return Err(StateAnchorError::DaemonExited);
+            }
+            let observed = mount_identity(&open_mounted_directory(
+                &self.runtime_anchor,
+                Path::new("merged"),
+            )?)?;
+            if observed.mount_id != baseline.mount_id {
+                let readiness = VerifiedStateAnchorReadiness {
+                    staged,
+                    mount: observed,
+                };
+                if let Err(error) = ready(readiness) {
+                    process.send_signal(rustix::process::Signal::KILL)?;
+                    while !process.has_exited()? {
+                        thread::yield_now();
+                    }
+                    let _ = rootless_unmount_anchor(&self.runtime_anchor);
+                    child_broker().reap(process)?;
+                    return Err(error.into());
+                }
+                return Ok(RuntimeStateAnchorMount {
+                    authority: self,
+                    process,
+                    mount: observed,
+                });
+            }
+            if Instant::now() >= deadline {
+                process.send_signal(rustix::process::Signal::KILL)?;
+                while !process.has_exited()? {
+                    thread::yield_now();
+                }
+                let merged = open_mounted_directory(&self.runtime_anchor, Path::new("merged"))?;
+                let observed = mount_identity(&merged)?;
+                if observed.mount_id != baseline.mount_id {
+                    let _ = rootless_unmount_anchor(&self.runtime_anchor);
+                }
+                child_broker().reap(process)?;
+                return Err(StateAnchorError::MountTimeout);
+            }
+            thread::sleep(MOUNT_POLL_INTERVAL);
+        }
+    }
+}
+
+pub struct RuntimeStateAnchorMount {
+    authority: RuntimeStateAnchorAuthority,
+    process: RetainedProcess,
+    mount: MountIdentity,
+}
+
+impl RuntimeStateAnchorMount {
+    #[must_use]
+    pub const fn mount_identity(&self) -> MountIdentity {
+        self.mount
+    }
+
+    /// Unmounts and reaps only the retained exact RS-owned State Anchor.
+    ///
+    /// # Errors
+    /// Refuses mount replacement, unmount failure, or exact child cleanup failure.
+    pub fn unmount(self) -> Result<(), StateAnchorError> {
+        let observed = mount_identity(&open_mounted_directory(
+            &self.authority.runtime_anchor,
+            Path::new("merged"),
+        )?)?;
+        if observed != self.mount {
+            return Err(StateAnchorError::OwnershipMismatch);
+        }
+        rootless_unmount_anchor(&self.authority.runtime_anchor)?;
+        let after = mount_identity(&open_directory(
+            &self.authority.runtime_anchor,
+            Path::new("merged"),
+        )?)?;
+        if after == self.mount {
+            return Err(StateAnchorError::OwnershipMismatch);
+        }
+        child_broker().reap(self.process)?;
+        Ok(())
+    }
+}
+
 impl VerifiedStateAnchorReadiness {
+    #[must_use]
+    pub const fn from_parts(staged: StagedProcessOwnership, mount: MountIdentity) -> Self {
+        Self { staged, mount }
+    }
+
     #[must_use]
     pub const fn staged(self) -> StagedProcessOwnership {
         self.staged
@@ -757,6 +973,24 @@ fn inherit_for_exec(descriptor: impl AsFd) -> Result<OwnedFd, StateAnchorError>
     dup(descriptor).map_err(io::Error::from).map_err(Into::into)
 }
 
+fn reopen_inherited_directory(descriptor: i32) -> Result<OwnedFd, StateAnchorError> {
+    if descriptor < 0 {
+        return Err(StateAnchorError::InvalidLayout(
+            "inherited directory descriptor is invalid",
+        ));
+    }
+    let file = File::open(format!("/proc/self/fd/{descriptor}"))?;
+    let descriptor: OwnedFd = file.into();
+    let metadata = rustix::fs::fstat(&descriptor).map_err(io::Error::from)?;
+    if FileType::from_raw_mode(metadata.st_mode).is_dir() {
+        Ok(descriptor)
+    } else {
+        Err(StateAnchorError::InvalidLayout(
+            "inherited authority is not a directory",
+        ))
+    }
+}
+
 fn overlay_options(lower: &Path, upper: &Path, work: &Path) -> OsString {
     let mut bytes = b"lowerdir=".to_vec();
     append_fuse_option_path(&mut bytes, lower);
@@ -781,7 +1015,12 @@ fn append_fuse_option_path(target: &mut Vec<u8>, path: &Path) {
 
 #[allow(clippy::disallowed_methods, clippy::disallowed_types)]
 fn rootless_unmount(layout: &StateAnchorLayout<'_>) -> Result<(), StateAnchorError> {
-    let inherited_anchor = inherit_for_exec(&layout.runtime_anchor)?;
+    rootless_unmount_anchor(&layout.runtime_anchor)
+}
+
+#[allow(clippy::disallowed_methods, clippy::disallowed_types)]
+fn rootless_unmount_anchor(runtime_anchor: impl AsFd) -> Result<(), StateAnchorError> {
+    let inherited_anchor = inherit_for_exec(runtime_anchor)?;
     let path = descriptor_child_path(&inherited_anchor, "merged");
     let status = std::process::Command::new("fusermount3")
         .arg("-u")

crates/lazaret-state-anchor/tests/authority.rs

@@ -2,13 +2,22 @@ use std::collections::BTreeMap;
 use std::ffi::OsString;
 use std::fs;
 use std::os::unix::ffi::OsStringExt;
+use std::os::unix::fs::MetadataExt;
 use std::path::PathBuf;
 use std::sync::atomic::{AtomicU64, Ordering};
 
 use lazaret_core::configuration::{
     ConfigurationLoadRequest, SandboxLayoutV1, load_trusted_configuration,
 };
-use lazaret_state_anchor::{ProtectedRuntimeFiles, StateAnchorLayout};
+use lazaret_core::identity::{
+    DeviceId, HostPath, IdMapDigest, InodeId, PolicyId, PreparedArtifactId, SelectedManifestDigest,
+    SourceIdentity,
+};
+use lazaret_core::records::{ArtifactState, PreparedArtifact};
+use lazaret_state_anchor::{
+    ProtectedRuntimeFiles, RuntimeStateAnchorAuthority, StateAnchorLaunchAuthority,
+    StateAnchorLayout,
+};
 
 static SEQUENCE: AtomicU64 = AtomicU64::new(0);
 
@@ -48,6 +57,45 @@ fn replaced_runtime_root_is_refused_before_runtime_file_write() {
     assert!(!replacement_files.join("hosts").exists());
 }
 
+#[test]
+fn inherited_runtime_authority_accepts_only_the_exact_prepared_lower() {
+    let tree = TestTree::new();
+    let configuration = configuration(&tree);
+    let layout = StateAnchorLayout::create(&configuration).unwrap();
+    let lower = tree.path.join("prepared-lower");
+    fs::create_dir(&lower).unwrap();
+    let metadata = lower.metadata().unwrap();
+    let artifact = PreparedArtifact {
+        prepared_artifact_id: PreparedArtifactId::new([8; 32]),
+        source_identity: SourceIdentity::new(b"fixture".to_vec()),
+        selected_manifest_digest: SelectedManifestDigest::new([9; 32]),
+        policy_id: PolicyId::new([10; 32]),
+        id_map_digest: IdMapDigest::new([11; 32]),
+        path: HostPath::new(lower.as_os_str().as_encoded_bytes().to_vec()),
+        device: DeviceId::new(metadata.dev()),
+        inode: InodeId::new(metadata.ino()),
+        state: ArtifactState::Available,
+    };
+    let launch = StateAnchorLaunchAuthority::capture(&layout, &artifact).unwrap();
+    let expected = launch.expected_lower();
+    RuntimeStateAnchorAuthority::from_inherited(
+        launch.lower_fd(),
+        launch.persistent_anchor_fd(),
+        launch.runtime_anchor_fd(),
+        expected,
+    )
+    .unwrap();
+    assert!(
+        RuntimeStateAnchorAuthority::from_inherited(
+            launch.lower_fd(),
+            launch.persistent_anchor_fd(),
+            launch.runtime_anchor_fd(),
+            (expected.0, expected.1 + 1),
+        )
+        .is_err()
+    );
+}
+
 struct TestTree {
     path: PathBuf,
 }

crates/lazaret-state-anchor/tests/state-anchor-lifecycle.rs

@@ -19,8 +19,9 @@ use lazaret_linux::process::{
     ProcessOwnershipContext, current_process_identity, dispatch_private_process_role,
 };
 use lazaret_state_anchor::{
-    FuseOverlayfsMount, ProtectedRuntimeFiles, StateAnchorAcquisition, StateAnchorLayout,
-    StateAnchorSandboxView, unmount_after_coordinator_loss,
+    FuseOverlayfsMount, ProtectedRuntimeFiles, RuntimeStateAnchorAuthority, StateAnchorAcquisition,
+    StateAnchorLaunchAuthority, StateAnchorLayout, StateAnchorSandboxView,
+    unmount_after_coordinator_loss,
 };
 use rustix::process::Signal;
 
@@ -77,6 +78,45 @@ fn run() -> io::Result<()> {
     };
     let layout = StateAnchorLayout::create(&configuration).map_err(io::Error::other)?;
 
+    let launch =
+        StateAnchorLaunchAuthority::capture(&layout, &artifact).map_err(io::Error::other)?;
+    let runtime_authority = RuntimeStateAnchorAuthority::from_inherited(
+        launch.lower_fd(),
+        launch.persistent_anchor_fd(),
+        launch.runtime_anchor_fd(),
+        launch.expected_lower(),
+    )
+    .map_err(io::Error::other)?;
+    let expected = std::cell::Cell::new(None);
+    let runtime_mount = runtime_authority
+        .acquire(
+            current_process_identity()?,
+            ProcessOwnershipContext {
+                domain_id: DomainId::new([1; 32]),
+                generation_id: GenerationId::new(1),
+                coordinator_id: CoordinatorId::new([3; 32]),
+                slot: HostProcessOwnershipSlot::StateAnchorDaemon,
+            },
+            |staged| {
+                expected.set(Some(staged));
+                Ok(())
+            },
+            |staged| Ok(ProcessOwnershipCommitAck::v1(staged)),
+            |readiness| {
+                if expected.get() != Some(readiness.staged())
+                    || lazaret_linux::filesystem::current_mount_identity(layout.merged())?
+                        != readiness.mount()
+                {
+                    return Err(io::Error::other(
+                        "inherited State Anchor readiness changed authority",
+                    ));
+                }
+                Ok(())
+            },
+        )
+        .map_err(io::Error::other)?;
+    runtime_mount.unmount().map_err(io::Error::other)?;
+
     let first = acquire(&layout, &artifact)?;
     fs::write(
         layout.merged().join("usr/bin/installed-tool"),